TLDR

The AI cost debate has a villain, and it is the wrong one. Yes, the tools now cost more than the people they replaced. Uber torched its entire 2026 AI coding budget in 4 months. Nvidia admits its compute bill beats its payroll. None of that is the real story, because cost was never the metric that mattered. Productivity was. And the industry already knows it.

This is not an anti-AI piece. The technology is real, the gains in raw output are real, and the shift the market just made, from paying per seat to paying per outcome, is the correct shift. Paying only when the software works is exactly how this should be priced.

The market has not priced this in. The “outcome” is defined, measured, and marketed by the company collecting the check. Every headline ROI figure in the market was commissioned by the vendor being graded. And on the rare occasions someone with no stake in the answer has actually measured the work, a randomized trial, 2 years of telemetry across 22,000 developers, a 300-deployment academic study, the result is worse than either side of the cost fight assumed. Output goes up. Delivered value does not.

The subsidies that hid all of this are unwinding in real time, and the repricing has started. When the bill gets honest, the question stops being what AI costs and becomes whether anyone ever checked that it worked.

Subscribe now

A vendor just bet its own revenue on the answer

On June 25, 2026, Salesforce relaunched the pricing for Agentforce, its flagship AI agent. The new model is pay-per-resolution. You are charged roughly $2 only when the agent resolves a customer issue on its own, start to finish. If the customer asks for a human, or walks away unhappy, there is no charge. The cost is tied to the outcome, not the activity.

Read that as what it is: a concession. For 2 years the pitch was that you should buy intelligence and trust that value would follow. This is the vendor quietly admitting that customers stopped believing it, and demanding proof before they pay.

It is also the third pricing model for the same product in 18 months. It launched at $2 per conversation. Customers could not tell what counted as a conversation, so it moved to $0.10 per action. Then per-user licenses at $125 a month. Now, per outcome.

That zigzag is the whole market compressed into one company. The unit of billing has been walking steadily away from “how much AI you consume” and toward “what the AI actually accomplished.” Which is the right direction. It is also an admission that the old way of measuring was selling something nobody could verify.

The cost panic is real. But it is also last quarter’s fight.

The numbers behind the panic are not in dispute. Uber’s CTO disclosed the company burned through its entire 2026 AI coding budget in 4 months. By March 2026, 84% of Uber’s engineers had adopted Claude Code, and roughly 70% of committed code originated with AI. The president conceded that token usage did not seem to correlate with useful features shipped.

It is not an outlier. Nvidia’s VP of applied deep learning said the cost of compute for his team now exceeds what the company spends on the employees using it.

The company that sells the shovels admits the digging costs more than the diggers.

And OpenAI spends close to $2 for every $1 it earns on inference, and loses money on its $200-a-month subscriptions by its own CEO’s admission.

So the tools cost more than the humans. Everyone agrees. That is precisely the problem with the debate: it has already been won, loudly, by both sides, and it is a fight over the wrong number.

Cost tells you what you spent. It tells you nothing about what you got. The only comparison that decides whether any of this was worth it is productivity, and productivity is the thing almost nobody is measuring honestly.

“Costs are falling” is a mirage too

The reflexive rebuttal is that prices are collapsing, so the cost problem solves itself. Per-token prices have genuinely dropped, and Gartner projects that running the largest models could be nearly 90% cheaper by 2030.

The catch is that consumption is scaling faster than the price is falling. Cheaper tokens do not produce a smaller bill when you burn far more of them, and every incentive in the building pushes people to burn more of them.

Amazon built an internal leaderboard to rank AI usage across engineering teams. It came down after staff started burning tokens on meaningless tasks purely to climb it. When you reward consumption, consumption becomes the output.

Falling unit prices do not save you from a culture that treats spending as a scoreboard. They just lower the price of each point.

The newest evidence for the rebuttal landed on July 16, 2026, when Moonshot AI released Kimi K3 in open weights, claiming frontier-competitive performance at a fraction of the price. Rival AI stocks fell the following day. Fortune called it a second DeepSeek shock.

Take that claim at face value, and the arithmetic still does not move. The bill is unit price times consumption, and cheap capability is the strongest invitation to consume more of what this industry has produced.

The industry found the right metric. Then it graded its own homework.

Now go back to that launch and read the coverage.

Bloomberg reported that Kimi K3 beat every rival except Claude Fable 5 and GPT-5.6 on overall capability, citing Moonshot itself. Quartz placed it below GPT-5.6 Sol overall, ahead of the prior generation on coding and agent evaluations. Fortune reported Moonshot claiming K3 substantially outperformed GPT-5.6 Sol, the model the other outlets had it trailing.

Three financial outlets, 24 hours, one vendor’s self-reported benchmark run, and no agreement on what the vendor had even claimed. Markets repriced anyway.

Nothing independent had been checked, because there was nothing independent to check.

This is not a China story. It is a measurement story, and it runs through every lab. On July 8, 2026 and July 9, 2026, Grok 4.5 and GPT-5.6 shipped a day apart, and neither led with “smartest.” Both led with efficiency: tokens per task, performance per dollar, cost per result.

The whole industry repriced its pitch around productivity. That is the strongest confirmation yet that productivity, not raw capability, is now the metric. It is also exactly where the grading problem begins.

Every one of those efficiency figures is the vendor’s own, on benchmarks the vendor picked. When Databricks built its own coding-agent benchmark in July 2026, it refused the public ones outright because the tasks are public and the solutions leak into training data over time. It also refused to grade with an LLM judge, on the grounds that a judge rewards sounding right over being right.

Then it went further. Mid-run, Databricks had to seal off its own git history, after finding that agents with shell access were walking through past commits to dig up the merged solution. That is what it costs to get an honest number, and almost nobody pays it.

OpenAI’s own footnote concedes its cost and latency estimates come from offline simulation and could vary substantially in practice. This applies to every lab, Anthropic included.

There is no neutral referee timing the race.

The self-grading does not stop at benchmarks. Once you accept that outcomes are what matters, the question is who decides what counts as an outcome. Today, it is the vendor being paid for it.

Take that pay-per-resolution model. A “resolution” is defined by the seller. Salesforce’s own CRMArena-Pro benchmark found that an out-of-the-box agent hit an accuracy of roughly 35% before heavy customization.

Yet the customer-facing resolution numbers that it and its peers advertise run anywhere from 25% to 95%, and no two of them mean the same thing. Some count ticket deflection. Some count customer-confirmed resolution. Some count raw response accuracy.

When one word stretches from 25 to 95 depending on who is defining it, the word is not a measurement. It is marketing wearing a percentage sign. And the customer, who is now paying per resolution, is paying against a number that the vendor gets to draw.

Every ROI figure you have seen was ordered by the company being graded

Widen the lens from customer service to the whole enterprise-AI ROI industry and the pattern holds. The wall of returns at the top of this piece, from 293% to 465%, comes from Forrester’s “Total Economic Impact” studies. Every one of them was commissioned by the vendor whose product is being evaluated.

This is not an accusation of fraud. Forrester’s methodology is disclosed, the studies are labeled, and the underlying shape they describe is often real. It is a point about who is holding the ruler. A vendor does not commission and publish a study that makes its own product look like a waste of money. These figures are the ceiling of what the seller wants you to believe, not an independent read on what buyers actually got.

The question is not whether the returns are real. It is who is holding the ruler.

And there is at least one number floating around the discourse that traces to nobody at all. A widely reshared claim of “540% ROI across 287 audited enterprise deployments” appears on a single vendor content page with no findable underlying report. It is the kind of stat that spreads precisely because no one checks. This piece does not use it.

When someone with no stake actually measured

Now put the self-graded numbers next to the independent ones. The gap is the entire story.

METR, an AI research nonprofit, ran a randomized controlled trial with experienced open-source developers on real tasks in codebases they knew well. Economists predicted AI would make them 39% faster. Machine-learning experts said 38%. After using the tools, the developers themselves estimated they had been sped up 20%. Measured, they were 19% slower. The people doing the work could not tell that the tool was costing them time.

That is one study on a hard setting, so hold it loosely on its own. It does not stand alone.

Faros AI analyzed 2 years of engineering telemetry across 22,000 developers and more than 4,000 teams, comparing each team’s lowest-AI quarters against its highest. Output rose: epics completed per developer up 66%, tasks up 34%.

Then the invoice arrived downstream. Bugs per developer are up 54%. Incidents per pull request are up 243%. Median review time is up more than 5x. Code churn, work thrown away shortly after it was written, is up 861%.

More code got written. More of it also got reverted, reviewed, and repaired. That is not a productivity gain. That is motion. And when MIT’s NANDA initiative reviewed more than 300 real enterprise GenAI deployments, it found that roughly 5% produced measurable impact on the bottom line. The other 95% never crossed from pilot to profit.

The same MIT research found the other half of the story, and it is the part that isolates the problem. While 95% of sanctioned deployments stalled, employees were not waiting. Roughly 90% reported using personal AI tools for work, often daily, even inside companies whose official pilots had failed. And that unsanctioned “shadow AI” frequently delivered more real value than the six-figure deployments sitting next to it.

Sit with what that means. The technology works best exactly where nobody is selling you an outcome, a person picking a tool for a task they understand. It stalls where someone is, a vendor booking revenue against a number they defined. The failure is not the AI. It is the sold outcome wrapped around it.

Self-graded, the story is 300% to 465% returns and 95% resolution. Independently measured, it is 19% slower, 861% more churn, and 5% with real P&L impact. Both cannot be true. Only one of them was checked by someone who did not get paid for the answer.

The obvious objection, taken head on

The strongest rebuttal is that the independent studies are already stale. METR tested early-2025 tools, and the models have improved sharply since. Fair. If the slowdown were a one-time artifact of early-2025 tooling, the effect should be shrinking as the tools get better.

It is not shrinking. It is widening. Faros’s own bug-rate finding went from a 9% increase in its 2025 study to a 54% increase in its 2026 study, on newer tools. GitClear, tracking more than 211 million changed lines, found code churn roughly doubled from its pre-AI baseline as adoption climbed, an independent second source for the same rework problem.

The tools got better and the repair bill got bigger, because better tools generate more code faster, and the bottleneck was never generation. It was review, testing, and integration, the parts a human still has to own.

Faster generation into an unchanged review pipeline does not speed delivery. It floods it. The counter-argument predicts the gap should be closing. The data says it is opening.

Where this leaves the money

This is the connection to a thesis this newsletter has made before: after the AI repricing, value does not accrue to whoever runs the biggest model. It accrues to whoever owns the workflow, the domain logic, and the governance, the layer where “did it work” can actually be measured, audited, and stood behind.

This piece is about that measurement gap, the layer where AI vendors grade their own outcomes. It is not about application-layer software serving real customers at real enterprise scale, which runs on completely different unit economics.

As a solution consultant at Adobe I have watched augmentation deployments work when they are scoped correctly. Forrester’s own Total Economic Impact studies point at the same shape once you look past the headline number: triple-digit returns when the use case is tight and a human stays in the loop, and near-zero when a company tries to pull the human out of the workflow entirely.

That is the tell. The deployments that work are the ones where someone can prove they worked. The ones drowning in churn and stalled pilots are the ones sold on a number the buyer could never verify.

The durable position in AI is not the most intelligence. It is the most accountability.

Databricks is what that looks like in practice, and it is the reason this piece is not a case against the technology. It built its benchmark out of its own engineers’ merged pull requests, graded against the tests those engineers wrote, on code no model had trained on. The agents genuinely helped. The company got real efficiency out of them.

What it also got was a number nobody sold it. Token price turned out to be a poor predictor of task cost: one model roughly 1.7x cheaper per token finished the work more expensively, at $2.09 per task against $1.94, while scoring 6 points lower, because it read more and burned 1.9x the tokens getting there. Running the same model through a different harness swung cost per task by more than 2x at identical quality.

None of that is visible from a price list, and none of it appears in a vendor deck. It is only visible if you measure your own work yourself.

Every company with a history of merged pull requests already owns an evaluation set, written by its own engineers, that no vendor can game.

That is the option almost nobody exercises, and it is free.

The parallel is Cisco in 1999. The internet was real. Cisco’s routers were real. The company still lost most of its market value, because a real technology and a sustainable business are two different bets, and the market had confused them.

AI is a real technology. Whether it pays for itself before the subsidies run out is a separate question, and right now the industry is answering it with numbers it wrote itself.

Nobody is checking if AI actually works. They are checking whether the company selling it says it worked.

The tools cost more than the people. The vendors moved the goalposts to outcomes. Then they appointed themselves referee.

AI’s productivity is mispriced.

TLDR

Last week, Coinbase moved its engineers off the American AI frontier and onto two open-weight models from China. Its AI bill fell close to 50%. Usage kept climbing.

That is the whole thesis in one move. The US still leads in raw capability. It is losing on the thing that actually decides revenue: what companies run in production, and what it costs them per token.

This is not a collapse. The frontier still matters, and America still owns it. But the market is pricing that lead as a moat when the usage data says it is a margin, and a margin is a very different thing to own.

The leading indicator is already public. On OpenRouter, where developers route real workloads, the US share of tokens fell from ~70% to ~30% in 12 months. Both US frontier labs filed to go public in June, at the exact moment their pricing power is being questioned at both ends. Here’s the full case.

Subscribe now

---

The switch that should bother the bulls

Brian Armstrong runs Coinbase, one of the largest public crypto companies in the world. Last week he changed one setting.

He moved his engineers’ default model off the US frontier and onto GLM 5.2 from Zhipu and Kimi from Moonshot. Both are open-weight. Both are Chinese.

He did not frame it as geopolitics. He framed it as arithmetic. 91% of his engineers had never once hit their old usage limits.

So instead of rationing access, he switched the default to models that cost a fraction of the American ones. He kept the expensive US frontier on hand for the genuinely hard jobs. The bill dropped by ~50% while usage rose.

He was not first. Months earlier, Flo Crivello took his AI agent company Lindy entirely off Claude and onto DeepSeek. He did not talk about which model was smarter either. He talked about survival.

One company changing one default is not a story. It is an anecdote, and anecdotes are cheap.

But Coinbase is an early instrument reading, the first gauge to move in public before the rest of the panel catches up. What it is registering is the most important thing in AI right now that the market has not priced.

Two scoreboards

There are two ways to keep score in AI, and they have quietly come apart.

The first scoreboard measures capability. Who has the smartest model, the largest training run, the most chips, the top benchmark? On this board, the US is comfortably ahead.

The second scoreboard measures what people actually run and what it costs per token. This is the board that decides who collects the revenue.

The cleanest public read on it is OpenRouter, where developers route real workloads across hundreds of models with no contract and no lock-in. Nobody routes an app to a model out of loyalty. They route it to whatever does the job at a price they can live with.

On that second board, the picture has inverted fast. A year ago, around June 2025, US models were roughly 70% of the tokens flowing through OpenRouter.

By June 2026, that share was about 30%. Chinese models crossed US ones for the first time in February, and by spring, they were the clear majority.

The market is pricing American AI off the first scoreboard. The cash is being decided on the second.

Act one: the bottom fell out on price

To see why whole companies are voting the way they are, start with the model that set off the latest stampede.

In mid-June, the Chinese lab Zhipu released GLM 5.2, an open-weight model anyone can download and run on their own hardware. Its scorecard reset the conversation.

On the coding benchmarks that decide real engineering work, GLM 5.2 landed within 1 point of Claude Opus 4.8, the model at the top of the capability rankings. It beat OpenAI’s GPT-5.5 outright on the same tests.

The honest version matters more than the hype. Claude still holds the top score on most of these benchmarks, and China has not suddenly built something smarter than the US frontier.

What it has built is something close enough that the gap rounds away on most real work. Then it released that for nothing.

Now set the price beside the score. GLM 5.2 runs at about $4.40 per million output tokens.

Claude Opus 4.8 charges $25. GPT-5.5 charges $30. That is 5 to 6 times cheaper for work that scores within a point, and the floor is lower still, with DeepSeek’s open model under $1.

This stopped being a story about a cheap tier losing to a premium one on quality. It became a story about the premium tier’s quality lead shrinking to a rounding error while its price stayed 5 to 6 times higher.

And the squeeze lands hardest where the volume is going. Coding grew from ~10% of OpenRouter usage at the start of last year to more than 50% today.

Coding is both the most price-sensitive workload in the building and the one where the open Chinese models are strongest.

The old reassurance was that cheap models were toys and serious money would always pay up. GLM 5.2 retired that argument.

When an open model matches the closed one within a point and costs a sixth as much to run, the burden of proof flips. Every team with a finance function eventually has to explain why it continues to pay the premium.

The counterargument that died in June

The bull’s best response is that enterprises do not actually behave this way. They standardize on a trusted vendor, sign a contract, and stop optimizing.

That counterargument died in June. UBS, surveying companies that actually track their AI spend, found that around 60% have begun imposing hard limits on token usage.

These are not penny-pinching startups. These are the buyers the premium case depends on, and they are watching the meter.

The spend behind that caution is exploding. By CloudZero’s count, the share of organizations planning to spend over $100K a month on AI doubled from 20% in 2024 to 45% in 2025.

When the bill gets that big, a 5x price gap stops being a line item and becomes a board-level decision. That is exactly the moment a 50% cut from switching defaults becomes irresistible.

They are raising prices into the war

Here is the part that should worry anyone underwriting US pricing power. The American labs are raising prices, not cutting them, right as the cheaper rivals arrive.

By JPMorgan’s reading, OpenAI roughly doubled token prices between two recent generations. Premium-tier increases ran 3x to 9x, and a handful of users reported moves as steep as 100x.

Take the steepest number with a grain of salt, since it is user-reported rather than a clean measurement. Take the direction seriously.

You raise prices into a war like that only if you believe your product has no substitute. The OpenRouter data says the substitute has already shipped.

Two weeks in June

Then the US government did something genuinely new, and it cut the wrong way.

On June 12, the Commerce Department recalled Anthropic’s two most capable models, Mythos 5 and Fable 5, under an export-control directive. They were barred from all foreign nationals, including Anthropic’s own non-citizen staff.

Anthropic pulled both worldwide. The NSA and the UK’s AI Safety Institute, one of the most important model-testing bodies in the world, lost access in the process.

The government did not restrict these models because they were weak. It restricted them because they were too good at finding software vulnerabilities, a capability that cuts both ways.

The episode bred its own scandal when a general told the Senate that in an authorized exercise, Mythos broke into nearly all of an agency’s classified systems in hours.

That line raced around the internet as “AI hacked the NSA.” It had not. It was a controlled red-team test, the agency pointing the model at its own networks, and Anthropic’s position is that the underlying research was for defense.

Now watch the backfire. Within days of the ban, Zhipu shipped an open-weight model that security researchers found could beat Anthropic’s publicly available coding models at exactly that kind of vulnerability detection.

The careful version is still the damning one. The capability the US went to extraordinary lengths to contain is now free, open, and downloadable by anyone with a graphics card.

You cannot recall a model that has already been downloaded 100,000 times. An export-gated frontier model cannot win a global market, because most of the world is the set of customers it is now forbidden to serve.

And when the best American model is unavailable, the global default does not become the second-best American model. It becomes the Chinese open weight with no kill switch and no government letter that can arrive on a Friday and shut it off.

The mispricing, named

Step back, and the shape is clear. By JPMorgan’s own count, the AI trade has driven somewhere between 65% and 80% of the S&P 500’s returns, earnings, and capital spending since ChatGPT launched.

The entire index is leaning on one bet, and that bet assumes US capability is a durable moat that converts into a durable margin. The two scoreboards say the moat is real, but the conversion is broken.

The numbers underneath do not help. JPMorgan’s midyear outlook raised total AI capital spending to $5.5T through 2030, most of it debt-financed.

To earn even a 10% return on that buildout, the industry needs roughly $650B of revenue every year, in perpetuity. The bank illustrated that as $35 a month from every iPhone user on the planet, or $180 from every Netflix subscriber.

Today, the industry spends roughly $8 to $10 of capital for every $1 of revenue it earns. It is funding hardware that wears out in about a year with debt that assumes a useful life of 7 to 15.

There is even a circularity propping up reported profit. Last quarter the catch-all “other income” line made up 60% of Google’s profit, 51% of Amazon’s, and 27% of NVIDIA’s.

A meaningful slice of that is gains these companies book from marking up their own stakes in OpenAI and Anthropic. The hyperscalers fund the labs, the labs spend it on the hyperscalers’ chips, and everyone revalues everyone else upward.

It is not fraud. It is a loop, and loops look like growth right up until they do not.

All of this sat in plain sight while valuations stayed anchored to the first scoreboard. Then, in the last week of June, the second scoreboard reached the tape.

Korean markets, home to the memory chipmakers at the center of the buildout, tripped their circuit breakers twice in one morning. Samsung and SK Hynix shed roughly 10% of their value, and the selling spread to the Nasdaq.

The market started to price the problem. It just did it for the wrong reason, panicking about debt and valuation, when the deeper mechanism is quieter: demand is migrating to a substitute that needs far less capital.

And the two labs at the center of the trade both filed to go public in June. Anthropic first, at a private valuation near $965B, with an October listing in view. OpenAI days later, around $$852B.

Combined, that is roughly $1.8T racing to sell shares into the public market at the exact moment their pricing power is being questioned at both ends.

Where I could be wrong

A thesis you cannot argue against is a mood, not a thesis. So here is the honest other side.

The migration could be slower than it looks. Some of the falling prices reflect US labs getting more efficient at inference, not customers fleeing to China, and the evidence on open-model adoption is still early.

UBS frames its own finding as growth moderating, not demand reversing. Spending that was tracking toward 150 is landing nearer to 120 to 130. That is a haircut, not a collapse.

The frontier genuinely matters and is not going away. The hardest work, in cybersecurity defense, in scientific discovery, in long-horizon agents, still wants the best model regardless of price, and that tier is American.

This is bifurcation, not extinction. A vast commodity layer that China wins on cost, and a smaller premium layer where US labs keep real pricing power.

This piece is about that foundation-model layer and the capital cycle funding it. It is not about application-layer software serving real customers at real enterprise scale, which runs on completely different unit economics.

As a solution consultant at Adobe, I have watched augmentation deployments work when scoped correctly, and Forrester’s Total Economic Impact studies show the same shape: triple-digit ROI when the use case is tight, near-zero when companies try to remove humans from the workflow entirely.

The bull case is not stupid. It is measuring the wrong board, and underwriting the whole index on the assumption that the premium tier is most of the prize, when the usage data says it is the minority.

What to watch

If this read is right, it shows up in specific, checkable places over the next few quarters.

Watch whether hyperscaler inference pricing starts to compress and whether margin guidance softens. Watch whether the premium tier narrows from a general-purpose default to a niche reserved for high-stakes work.

Watch the two IPOs, and whether the bankers can hold their valuations once the prospectuses force the unit economics into daylight.

And watch the open-weight share on the second scoreboard. That single line is the leading indicator for whether the revenue JPMorgan needs ever actually arrives.

America still builds the best AI in the world. That part is true and worth saying clearly.

But the market is pricing that lead as a moat, when the evidence says it is a margin, contested from below on price and constrained from above by the government. One of those justifies the valuations. The other does not.

The best models in the world, priced like a monopoly, were sold into a market that had already found a cheaper substitute.

A margin wearing the price tag of a moat.

The American AI trade is mispriced.

TL;DR

- The AI cost panic of 2026 is real. Uber burned its annual AI budget in four months, Microsoft cancelled Claude Code for whole divisions, and the labs ended subsidized pricing. Tokenmaxxing, the practice of gaming raw token usage, is over. Fortune wrote the obituary in May.

- The viral fixes miss the point. The “caveman” prompt that promised 75% savings delivers 14 to 21, sometimes costs more, and where it does save, the real lever was a hidden reasoning setting, not the broken grammar.

- The waste was never in the prompt. It lives in the system around the model: retrieved context, tool schemas, reasoning traces, and memory. Token efficiency is an architecture problem, not a model-pricing one.

- The whole field already knows this. The discipline moved from prompt engineering to context engineering to harness engineering, and the operative metric in 2026 is cache-hit rate, not prompt quality.

- More tokens is not just costlier, it is worse. Every frontier model tested degrades as context grows. Curated context is cheaper and sharper.

- That is the moat. Value accrues to whoever owns the architecture that turns tokens into work. The labs commoditize raw capability; the layer that owns the system of work keeps the margin.

Subscribe now

---

The most popular fix was to make the AI talk like a caveman

For a few weeks this spring, the most popular answer to the AI cost crisis was to make the model speak like a caveman.

A prompt skill called Caveman went up on GitHub and collected four thousand stars in days. The pitch was simple and seductive. Strip the articles, the pleasantries, the connective grammar, and instruct the model to answer in dense fragments. No “sure, happy to help.” No “you can use a JOIN operation to combine data from two tables based on a common key.” Just: use JOIN. The headline claim was a seventy-five percent cut in tokens.

It does not deliver that, and the way it fails is the whole story. When developers actually benchmarked the skill on Claude, the real savings landed between fourteen and twenty-one percent, and a stripped-down six-line version outperformed the full skill on every run.

On coding tasks it frequently backfired, returning more tokens than a plain instruction to be concise, because the model started adding telegraphic annotations and quietly expanding the scope of the work it was asked to do. And the single case where someone measured a genuine eleven-fold reduction had nothing to do with caveman grammar at all.

A classifier buried in the coding harness had silently switched the model into extended-thinking mode and burned roughly two thousand reasoning tokens before writing one visible word. Pin the reasoning effort low and the identical answer cost two hundred tokens instead of two thousand.

That is the AI cost panic in a single anecdote. Everyone went hunting for waste in the words they could see, and the waste was somewhere else entirely.

[The caveman prompt, promised vs delivered. Claimed 75% token savings against a measured 14-21% on typical tasks, and a 43-56% increase on coding. Sources: developer benchmarks on Claude.]

The panic is real, and tokenmaxxing is the casualty

None of this means the bills are imaginary. They are enormous, and they arrived faster than anyone budgeted. Uber capped employees at fifteen hundred dollars apiece after burning its entire 2026 AI budget in four months. Microsoft cancelled Claude Code subscriptions across several product divisions. Ramp’s own data shows average monthly token consumption up thirteenfold since the start of 2025, even as the dollars spent rose only fourfold, the gap being the falling price of a token. And the subsidy that hid all of it is ending: Anthropic began charging Claude Enterprise the full cost of compute in April, and GitHub moved Copilot to usage-based billing on June 1, citing inference costs it could no longer absorb.

The cultural artifact of that era was the leaderboard. Amazon, Meta, and others ran internal dashboards that ranked employees by token consumption, which produced precisely the behavior you would expect, engineers throwing trivial work at expensive models all day to climb the rankings. The practice earned a name, tokenmaxxing, and for about a year it stood in as a proxy for how seriously a company was taking AI. It was always a bad proxy. In May, Fortune declared it dead, calling it a flawed way to measure return on AI. Ramp’s CEO, whose company just raised seven hundred and fifty million dollars to help finance teams track exactly this spend, called it the twilight of tokenmaxxing.

What replaces it is a different question. Not how many tokens a system consumes, but how much useful work it produces per token. Glean’s CEO Arvind Jain gave the metric a name, token yield, and because Glean sells enterprise search built around precisely this idea, treat the framing as both sharp and self-interested. The metric is bigger than any one vendor, and the rest of this piece is the case for why.

The tokens are not where you are looking

Start with where the spend actually goes, because it is not the prompt.

In any real agentic system, the user’s question is a sliver of the tokens in play. The bulk is everything the system wraps around it: the system instructions, the schemas describing every tool the agent can call, the documents retrieved to ground the answer, the model’s own reasoning traces, and the memory of prior turns. Agentic loops then compound the problem, because each step re-sends the accumulated context, and the cost of a long session grows quadratically with its length. A single misconfigured run can produce a bill of hundreds of dollars over a weekend, and almost none of it is the words the user typed.

[Where the tokens actually go. The real token load of a single “small” agentic request: system instructions, tool schemas, retrieved context, reasoning traces, and memory dwarf the user’s prompt.]

Which leads to the conclusion the caveman crowd skipped past. Token efficiency is not a model-pricing problem you solve by shopping for a cheaper API or trimming your prose. It is an architecture problem, and it lives in the design of the system, not the phrasing of the request. Jain put it plainly in the post that named the metric: the waste is in the system design, not the prompt.

[Why a long agentic session blows up. Each step re-sends the whole accumulated context, so the cumulative cost grows with the square of the session length, unless it is curated. Illustrative.]

The field already moved; the income statements just haven’t

This is not a contrarian take. It is the consensus that the entire field has reached over the past year, and the only thing lagging is the accounting.

The discipline even has a name and a dated lineage. In June 2025, Shopify CEO Tobi Lütke described the real skill as context engineering, the art of supplying all the context a task needs to be plausibly solvable by the model. Andrej Karpathy endorsed the term a week later, calling it the delicate art and science of filling the context window with just the right information for the next step. The institutions ratified it from there.

Anthropic formalized context engineering in September 2025 as the strategies for curating the optimal set of tokens during inference. Gartner published research under the blunt headline that context engineering is in and prompt engineering is out. By 2026, the frontier had moved one level deeper again, to what practitioners now call harness engineering, the design of the whole agent environment, with Martin Fowler’s team, Ethan Mollick, and others writing the early canon.

Three eras in three years: prompt engineering, then context engineering, then harness engineering. Each shift happened because the previous lever stopped paying off. And the tell of where the value sits now is the metric people optimize. In 2026, it is no longer prompt quality. It is the cache-hit rate. Seen against that backdrop, the caveman prompt was the last gasp of prompt engineering, a way to fight a war the serious builders had already left.

More tokens is not just costlier. It is worse.

Here is the part that turns a cost argument into something harder to wave away. Adding tokens does not just raise the bill. It degrades the answer.

The AI lab Chroma published a study in July 2025 testing eighteen frontier models, including the latest from Anthropic, OpenAI, and Google, on how performance changes as input grows. Every one of them got worse as the context lengthened, on simple tasks, well before the window was anywhere near full. The phenomenon has a name, context rot, and three drivers: models attend poorly to the middle of long inputs, costing thirty percent or more in accuracy in Stanford’s earlier work; attention dilutes as the token count climbs; and irrelevant but similar content pulls the model off course. For coding agents, the primary failure mode is no longer raw capability. It is context rot, the noise the agent piles up as it searches and backtracks.

So the tokenmaxxer pays twice. A bigger bill, and a dumber model. The corollary is the most important sentence in this piece for anyone building or buying: curated, minimal, high-yield context is the rare thing that is cheaper and better at once. Efficiency and quality stop being a tradeoff and start being the same project.

What good architecture actually looks like

If yield is an architecture problem, the obvious question is what the architecture is. Jain’s four levers are a useful frame: context quality, model routing, continual learning, and harness design. The reason to trust the levers rather than the vendor who named them is that everyone else arrived at the same four from different directions.

NVIDIA Research argues in its widely read paper that small language models are the future of agentic AI, that a seven-billion-parameter model is ten to thirty times cheaper to run than a frontier model and should handle the bulk of agentic work, with larger models reserved for the hard parts. Gartner says the same in different words: route routine, high-frequency tasks to small or domain-specific models, and gate expensive frontier inference for genuine reasoning. Andreessen Horowitz, in a much-cited May 2026 piece by partner Joe Schmidt, lists the defenses that let an application company survive the labs, and they reduce to the same set, a data flywheel, routing across vendors, cost optimization across model tiers, and governance.

When a vendor, a chipmaker, a research firm, and a venture fund independently describe the same machine, it has stopped being marketing. And the waste is not hypothetical. Jain, whose company sells exactly this, so weigh it accordingly, estimates that roughly ninety-five percent of enterprise AI usage still runs on the most expensive frontier models, even for work a cheaper model could handle. Cognition’s Scott Wu, whose firm builds the coding agent Devin, puts the cost efficiency available on routine work at five to ten times. The default, in other words, is to overpay for almost everything.

And routing buys more than savings. When a government can switch off a single lab’s flagship overnight, as the US did to Anthropic’s two most capable models in June under an export-control order, spreading work across providers stops being only a cost strategy and becomes a resilience one. Concentration is a bill that arrives all at once.

[The four levers of execution efficiency. Context quality, model routing, continual learning, and harness design, shown as the stack that determines token yield, with the converging sources that name each.]

It helps to see the machine assembled, because the contrast with the naive system is the entire argument.

The naive system, the one most enterprises are running right now, points one frontier model at every request, shovels the whole document corpus into context, routes nothing, caches nothing with any discipline, and remembers nothing about what worked last time. High token count, low yield, and thanks to context rot, worse answers. The high-yield system has five working parts:

- A context layer that retrieves the minimum viable context for a query rather than the whole corpus. This is the part that fights context rot directly.

- A router that classifies the task and sends it to the right model tier, gating frontier reasoning instead of defaulting to it.

- A harness that controls reasoning effort, reuses the cache, and evicts dead context from the working set. This is the part the caveman crowd never saw.

- A learning loop in which every run leaves behind evals, labels, and corrections that sharpen retrieval and routing over time.

- A metering layer that attributes token cost down to the workflow, so yield is something you can actually see.

I have been building one of these, a knowledge-graph platform I call Project Nexus, so let me be concrete about the context layer, which is where most of the waste and most of the moat live. The design choices map straight onto the levers.

It is ontology-first, meaning a typed schema constrains what extraction can produce, which is the only thing that gives the resulting graph reliable structure instead of a pile of loosely related text. Retrieval is hybrid: vector similarity for semantic matches, keyword search for exact phrases, and graph traversal for the relationship-shaped questions that vector search cannot answer, with the rankings fused cheaply at the end. That is what lets the system pull the few connected facts a question actually needs rather than the nearest large blob of text.

Every node carries a confidence score and a validity window, and the score decays over time on a curve, climbing back up when a fresh source corroborates it, so stale facts fade out of retrieval on their own without anyone pruning them by hand. That last piece is the continual-learning lever made literal. And every claim in an answer cites the specific source it came from, which is the governance lever, the difference between an answer you can audit and one you have to trust.

The honest version, because a reference architecture sold as a free lunch is just a brochure: I built it to attack the context-quality and learning levers specifically, the system meters the real token cost of every query against what naive retrieval would have spent on the same question, and in my own corpus, that gap is large. The number I care about more is that the gap is visible at all, because you cannot manage a yield you do not measure. The point is not that a knowledge graph is the answer. The point is that the work is in the architecture, and a graph is one way to do it.

It also comes with real tradeoffs, and pretending otherwise is how you get a brochure. Routing adds orchestration overhead. Knowledge graphs are costly to build, add traversal latency, and break in their own ways, the same entity splitting into three, queries pointed at the wrong shape of the problem, because models reason over text rather than graph topology. Even caching has a ceiling, since cached context still occupies the window and still costs attention on every output token. Proper architecture means managing those tradeoffs, not escaping them.

[Anatomy of a high-yield system. The naive architecture (one frontier model, full corpus dumped in, no routing, no caching, no loop) versus the high-yield architecture (curated context layer, router, disciplined harness, learning loop, metering), with the same task producing far more useful work per token on the right.]

For an investor, the architecture is also a lens. Pair it with the three tests Schmidt offers for telling a durable company from a thin wrapper, and you have a diagnostic you can run from the outside. How many steps does the work take, and how complex are the tools required to support it? Is the company a system that the customer runs their work through, or a tool sitting on top of a system they already have? And is its performance judged against benchmark scores or against the customer’s own profit and loss? A company that owns the system of work passes all three. A wrapper on someone else’s model passes none.

Where the value accrues

Which is the investment thesis, finally stated plainly? If execution efficiency is the lever, value accrues to whoever owns the architecture that produces it, and that is a very different map of winners than the one the market is currently pricing.

Schmidt’s framing is the cleanest version. The labs own what a16z calls the Yellow Brick Road, the horizontal, low-step work that improves with raw model capability, code generation, writing, image creation, the things a generic coworker does well. Everything else lives in the rest of Oz, the vertical, multi-step, regulated work where the value comes from the scaffolding around the model and not the model itself. Both can win.

But the application companies that wander onto the road, plugging off-the-shelf connectors into a frontier model and shipping an orchestration layer, are walking into the labs’ own product. The ones that survive own the data capture, the workflow, and the governance for a specific domain, and they compound an advantage with every deployment. The model underneath is fungible. The system of work is not.

Even the labs concede the point: OpenAI and Anthropic have both stood up forward-deployed enterprise units to configure their models for specific industries, which is not what you do if you believe the next model release handles everything.

If that still sounds like a venture pitch, consider who else is now making it. In mid-June, Satya Nadella, whose own company is busy building models to lessen its dependence on OpenAI, told businesses to stop reaching for the most powerful model for work that does not need it, and argued the prize is a frontier ecosystem, not a frontier model.

His framing is almost word-for-word Schmidt’s. Companies should own the learning loop that compounds their institutional knowledge, treat the underlying general-purpose model as something to swap out as the technology turns over, and recognize that the durable intellectual property is the proprietary system built on top, not the model underneath. When the chief executive of the most valuable software company on earth tells you the model is the fungible part, the contrarian read has become the consensus.

The macro floor under all of this comes from IBM’s Arvind Krishna, who did the arithmetic on the buildout. Roughly a hundred gigawatts of committed AI data-center capacity at sixty to eighty billion dollars of silicon per gigawatt implies six to eight trillion dollars of infrastructure, which on a five-to-seven-year payback would require one to two trillion dollars of new annual revenue to justify. “That much incremental revenue I don’t believe is there,” he said, and he expects the frontier models themselves to commoditize, with maybe two or three builders surviving. Stack his view on Schmidt’s, and the picture resolves. The builders of raw capability commoditize toward cost. The owners of the durable layer keep the margin. Execution efficiency is what defines that durable layer.

[Krishna’s $8 trillion gap. Committed AI infrastructure spend (~$6-8T) against the $1-2T in new annual revenue it would need to generate to justify the payback.]

[Where the value accrues. The labs eat the horizontal “Yellow Brick Road” layer; durable value collects in the infrastructure beneath and the vertical, embedded, governance-heavy application layer above.]

And the market is already pricing the shift. In early June, CNBC’s Deirdre Bosa reported that model routing has become a direct problem for OpenAI and Anthropic, both of which built their businesses, and the IPO expectations around them, on the assumption of enormous demand at premium prices. If easy, high-volume work moves to cheaper models, including open-source ones, the labs keep only the hard jobs, and pricing power shifts toward whoever decides which model runs when.

The labs see it coming. Anthropic has moved from flat-rate enterprise pricing toward per-token billing and cut off some heavy third-party token consumers, and both ship tiered model families, which is itself an admission that not every task needs the top of the line. The fair read, and the one SemiAnalysis’s Dylan Patel offers, is that routing does not sink the frontier labs but bends their pricing from charging more toward being used more efficiently. Which is the same conclusion from the other end of the telescope: the model is becoming a commodity input, and the margin migrates to whoever architects around it.

The capital is already moving there. In June, OpenRouter, which routes across more than four hundred models and sells companies a way out of single-vendor lock-in, raised a hundred and thirteen million dollars at a 1.3 billion dollar valuation, led by Alphabet’s growth fund with the venture arms of NVIDIA, Snowflake, and Databricks alongside. Its chief executive reduced the thesis to a sentence: running inference at scale is a multi-model problem, and the era of picking a single model is over. The Linux Foundation, for its part, has launched a Tokenomics Foundation to standardize how all of this gets measured in the first place.

The finance function is wiring the meters to match. The FinOps Foundation reports that ninety-eight percent of practices now manage AI spend, up from sixty-three percent a year earlier, and the question has shifted from how much we spend to cost-per-workflow and return-on-workflow, with model rightsizing now squarely inside the mandate. The market is learning, through its accounting departments, to price yield.

This is also where I should disclose my own position, because it cuts against type. I work as a solution consultant at Adobe, which Morningstar downgraded this spring from a wide to a narrow economic moat, alongside Salesforce, Workday, and a cluster of other application-layer software names, on exactly the thesis that the labs will eat the application layer. I do not think that thesis is wrong so much as incomplete.

Morningstar itself noted it has watched several software-is-dead episodes over twenty-six years and does not believe software moats evaporated overnight. The surviving application-layer moats are the ones built on the system of work, the embedded context, the governance, and the execution efficiency that a horizontal model cannot replicate from the outside. Whether any given incumbent has built that is the question worth asking, and it is the question this framework is for.

The objection that does not survive contact

The strongest case against everything above is that none of it lasts, because models simply get cheap enough that efficiency stops mattering. Tokens have fallen in price before and will again, so why architect around a cost that is trending to zero.

The numbers say the opposite. Gartner projects that by 2030 the cost of running inference on a trillion-parameter model will fall more than ninety percent. In the same breath it warns that agentic workloads consume five to thirty times more tokens per task than a chatbot does, so as consumption outruns price, total inference bills rise rather than fall. Goldman Sachs models a twenty-four-fold increase in token consumption by 2030. And Gartner’s sharpest point is the one that should end the debate: cheap tokens do not fix architectural inefficiency, they hide it, until agentic scale drags it back into the light.

The cheaper a token gets, the more of them a careless system will burn, which makes yield matter more, not less. Falling prices are not a reprieve from this discipline. They are the reason it becomes unavoidable. The labs are behaving as if they know it. OpenAI, under the same routing pressure, is reportedly weighing token price cuts and has built out a full ladder, from a ten-cent nano model to a thirty-dollar flagship, so that customers who are going to route do it inside the OpenAI stack rather than across to open-source rivals. That is not the posture of a company that expects to keep charging a premium for every token.

[Cheaper tokens, bigger bills. Gartner’s projected ~90% drop in per-token cost by 2030 against the 5-30x rise in tokens consumed per agentic task, and the net effect on total spend.]

The point the caveman missed

The caveman prompt was not stupid. It was looking in the only place a prompt can look, at the words on the way out, and trying to make them shorter. The whole industry made the same mistake at a larger scale, mistaking a usage number on a leaderboard for a measure of value, and calling the resulting waste a strategy.

The real lever was always one level down, in the system around the model, where the context gets curated before it is sent, the work gets routed to the cheapest model that can do it, the cache gets reused instead of rebuilt, and the system quietly learns what worked the last ten thousand times.

The proof arrived this month, on a schedule no one would have scripted. On a Friday in June, the US government switched off Anthropic’s two most capable models worldwide with an export-control order. By the next day, OpenRouter had shipped Fusion, a tool that fans a question out to a panel of ordinary models, has a judge model weigh the answers, and synthesizes one result. On a hard research benchmark, a panel of three budget models, none of them frontier, landed within a single point of the banned Fable 5 at roughly half the cost.

It is no free lunch. A panel call runs several times the price of a single one, and it is no substitute for a frontier model on the hardest long-horizon work. But the headline is hard to miss. The most capable model on the planet went dark on Friday, and by Saturday a stack of commodity models had matched it, because the answer was assembled in the architecture rather than bought in the model. As one writer put it, that is not magic, it is architecture.

[A cheap panel matched the model that the government banned. OpenRouter’s Fusion on the DRACO benchmark, days after Fable 5 was pulled. Source: OpenRouter.]

Tokenmaxxing measured the wrong thing, and it is dead. Token yield is the thing, and the companies learning to produce more of it per dollar are not running a cost-savings program. They are building the one moat the labs cannot reach across, because the model is fungible, and the system of work is not. The cost panic of 2026 will be remembered as the moment the market finally learned to tell the two apart.

---

Disclosure: The author works as a solution consultant at Adobe, referenced above. This piece concerns the architecture of execution efficiency across the AI application and infrastructure layers and is not investment advice. Project Nexus is a personal project described here to illustrate an architectural argument, not as a commercial offering, and the cost figures cited from it are self-measured. Do your own work before making any decision.

TL;DR

- SpaceX prices Thursday, June 11, and starts trading Friday, June 12, as SPCX: $135 a share, $75 billion raised, a $1.75 trillion valuation. The largest IPO in history, at 94 times trailing revenue.

- The filing says what the photos don’t. SpaceX registered as a data processing company. Starlink’s $4.4 billion segment profit funds a $657 million launch loss and a $6.4 billion AI loss. 61 percent of capex goes to AI.

- Nasdaq rewrote its index rules effective May 1: entry 15 trading days after an IPO, no minimum float, and a 3x weight multiplier for low-float names. S&P reviewed the same idea and refused. Only one of them can be right.

- Roughly $5 to 6 billion of forced index buying arrives around July 7. Insider lock-ups release weeks later, in tranches that grow the float and raise the index weight, which forces more buying. The plumbing fits together without a gap.

- The S-1 books the Anthropic compute deal as $45 billion through 2029. Musk publicly calls it a 180-day arrangement. Both reference the same contract.

- The mispricing is not the rocket. It is the rules rewritten to deliver the rocket into your index fund.

Subscribe now

---

On June 4, S&P Dow Jones Indices rejected its own idea. The proposal would have done two favors for giant IPOs: cut the waiting period before index entry from twelve months to six, and waive the rule that a company must be profitable. S&P asked market participants what they thought. Then it walked the whole thing back. The rejection ran one sentence, with a knife inside it: exceptions should not be granted, the committee wrote, solely based on market capitalization.

Five weeks earlier, Nasdaq went the other way. Effective May 1, its revised Nasdaq-100 methodology allows any newly listed company ranked in the top 40 by market capitalization to enter the index after 15 trading days. The old minimum free float requirement of 10 percent was eliminated outright. The seasoning period that existed for price discovery is gone for companies of sufficient size.

SpaceX prices its IPO on Thursday, June 11, and trades Friday, June 12, under the ticker SPCX, at $135 per share, 555.6 million shares, a $75 billion raise, a $1.75 trillion valuation. The largest IPO in history, listed on the exchange that changed its rules in time to receive it.

Most coverage this week is arguing about whether SpaceX is worth $1.75 trillion. Morningstar says $780 billion. Damodaran says $1.3 trillion. The S-1 says whatever a 94 times multiple on trailing revenue says. That debate is loud, public, and already in the price of the argument.

The quieter question is structural. When two committees apply different rules to the same security, one of them is wrong, and the capital that tracks the wrong one pays for it mechanically, on a schedule, without anyone making a decision. That is the setup. Here is the filing underneath it.

The plumbing, in order. Sources: Nasdaq-100 methodology FAQ, SEC EDGAR, S&P DJI, CNBC.

One profitable business, wearing two unprofitable ones

Start with the industry code. SpaceX filed its S-1 under SIC 7370: Services, Computer Programming, Data Processing. Not aerospace. Not defense. The company that launches more than 80 percent of global payload mass classified itself, for regulatory purposes, as a data processing company. The first 18 pages of the prospectus are rocket photographs. The industry code is the honest part.

The segment disclosures, public for the first time, explain why. SpaceX is three businesses with opposite financial signatures, recast after the February 2026 all-stock acquisition of xAI, which had itself absorbed X in March 2025.

Connectivity is the engine. Starlink generated $11.4 billion of revenue in 2025, up 49.8 percent, with $4.4 billion of segment operating income at roughly 63 percent adjusted EBITDA margins. Subscribers doubled from 4.4 million to 8.9 million during 2025 and crossed 10.3 million by early 2026. It is the only segment that makes money. It produced effectively all of the consolidated EBITDA.

Space is a platform investment. The launch business posted $4.1 billion of 2025 revenue and a $657 million operating loss, with roughly $3 billion of Starship R&D consuming the cost base. Falcon 9 itself is profitable. The segment is not.

AI is the sink. The xAI segment lost $6.36 billion from operations in 2025 and absorbed $12.7 billion of capital expenditure, 61 percent of the company’s $20.7 billion total capex. In the first quarter of 2026 the pattern held: Connectivity earned $1.19 billion in operating profit, Space lost $619 million, AI lost $2.47 billion. Consolidated Q1 net loss: $4.3 billion. Q1 capex alone ran $10.1 billion, of which $7.7 billion was AI. At that pace, SpaceX will spend more on capex in 2026 than it generated in revenue in 2025.

Segment operating income, FY2025. Data: SpaceX Form S-1.

Morningstar’s framing is the correct one: Starlink’s success is effectively subsidizing xAI’s expenditures. Investors who want the satellite business, and there are good reasons to want it, are required to swallow the AI segment to get it. The AWS-inside-Amazon comparison circulating this week flatters the structure. Amazon’s retail business was not burning six billion dollars a year when AWS emerged.

Capital expenditure by destination. Data: SpaceX Form S-1 and Q1 2026 disclosures.

The related-party disclosures complete the picture. SpaceX spent $131 million on Cybertrucks in 2025 and $697 million on Tesla battery storage systems across 2024 and 2025, at retail. Musk holds 93.6 percent of the Class B shares, which carry ten votes each, giving him 85 percent of voting power on 42 percent of the equity. SpaceX is opting out of Nasdaq’s independent board requirement as a controlled company. The compensation plan ties new Musk equity awards to milestones up to and including a Mars colony.

The valuation debate takes all of this as input and produces a number. Fine. The number is not the story.

Three answers to one question. Sources: Morningstar; Aswath Damodaran via WSJ; SpaceX S-1.

The plumbing

Here is what changed at Nasdaq, specifically, and when.

In March 2026, with SpaceX, Anthropic, and OpenAI all expected to list within the year, Nasdaq adopted a new fast entry rule for the Nasdaq-100, effective May 1. Three revisions matter.

First, speed. A newly listed company ranked in the top 40 by market capitalization becomes index-eligible 15 trading days after its IPO. The prior path ran through annual reconstitution in December, with a three to twelve month seasoning period that existed, in the words of one Wall Street critic, so the market could establish real price discovery before passive investors were forced in.

Second, float. The old methodology excluded securities with less than 10 percent free float. The new methodology has no minimum at all. Instead, a security with float below 20 percent gets its index market capitalization set at three times its float value, capped at 100 percent. Read that mechanism against SpaceX’s offering: roughly 4 to 5 percent of shares floated at IPO, insiders controlling the rest. The multiplier inflates SpaceX’s effective benchmark weight well beyond what its tradable share base supports.

Third, scope. Eligibility and ranking now consider the entire market capitalization, listed and unlisted shares together. A company can be weighted against a share count the public cannot trade.

FTSE Russell adjusted as well; SpaceX is already eligible for Russell inclusion. S&P alone declined, which means S&P 500 trackers sit out for at least a year while Nasdaq-100 and Russell trackers are required to buy.

The size of the requirement is calculable. Roughly $527 billion sat in 63 ETFs tracking the Nasdaq-100 as of December, per ETFGI, before counting structured products and proprietary indices built on top of it. At an initial weight near one percent, inclusion generates approximately $5 to 6 billion of mechanical demand, per the math laid out by heise’s analysis of the methodology FAQ. Those funds must sell Apple, Microsoft, and Nvidia, pro rata, to fund the purchase. No analyst forms a view. The index decides, around July 7, fifteen trading days after Friday.

A reasonable person asks why an exchange would do this. The commercial answer is that exchanges compete for listings, the decade’s largest listings are richly valued private companies with small floats and no GAAP profits, and an index seat is a product feature. The timing has prompted commentators to ask whether the rule change was a marketing play for the exchange business. Nasdaq has not characterized it that way. The sequence of dates is in the public record either way: rule adopted March, effective May 1, SpaceX chose Nasdaq, lists June 12.

The release valves

The standard IPO lock-up bars insider sales for 180 days. SpaceX replaced it with a staircase.

Per the S-1, after the company reports earnings for the quarter through June, its first as a public company, insiders may sell up to 20 percent of their eligible locked shares. Further tranches release on additional triggers. CNBC’s read of the structure is the polite one: phased selling prevents a single cliff where everyone exits at once. The less polite read follows immediately after in the same article. The staircase grows the float, and a growing float raises SpaceX’s index weight under the new low-float weighting rules, which compels tracking funds to buy more, on a schedule, into the insider selling.

Hold the two mechanisms next to each other. Forced passive demand arrives 15 trading days after the IPO. Insider supply unlocks weeks later, in tranches, each tranche raising the index weight and summoning the next round of forced demand. Supply has been engineered to arrive precisely when the buyers who cannot say no show up. Whether by design or coincidence, the plumbing fits together without a gap.

Who is selling into that demand is also in the record. The xAI merger in February converted xAI equity, which had absorbed X’s equity in 2025, which had absorbed the Twitter acquisition’s losses before that, into SpaceX shares. The early backers of a $44 billion social media purchase that lost most of its value now hold stock in a $1.75 trillion company, with a staircase out, and a buyer of last resort that is legally required to show up. That sentence describes the mechanics. It does not require a villain. It only requires the rules as written.

The consensus exhibit, and the caveat the consensus skips

The most circulated chart of the week comes from Truist’s Keith Lerner, who compiled the first-year record of 30 major tech IPOs. The medians: up 3 percent after one week, down 9 percent after twelve months, with only 43 percent of names positive a year out. The column that matters is the last one. Median first-year maximum drawdown: negative 54 percent. Every one of the 30 names drew down. The best case was Okta at negative 20. Robinhood reached negative 90. Meta, the closest analog as a giant consumer-facing IPO with a control structure, fell 54 percent inside year one before becoming one of the great compounders of the era. The exhibit follows, attributed to Truist. We did not compute it; we verified its headline figures against independent coverage.

Year 1 maximum drawdown, 30 major tech IPOs. Data: Truist Advisory Services (Keith Lerner), June 2026, via @saxena_puru

Now the caveat, because this table is on every finance site this week and the herd is reading it one way. Jay Ritter’s IPO research, the academic standard for this question, finds that long-run IPO underperformance concentrates in the smallest companies. Large IPOs, on average, neither outperform nor underperform after their first day. SpaceX is the largest IPO ever. The honest statistical prior for a megacap listing is not doom. It is roughly market performance, with a violent path.

So the drawdown table does not predict that SpaceX falls. It predicts that whoever owns it in year one should expect a moment when it is down a lot, and the relevant question becomes who is holding when that moment arrives. The answer, uniquely in this structure, includes everyone with a Nasdaq-100 tracker, added at day 15, weighted at three times float, without consent. The historical drawdown statistics were generated by IPOs whose passive inclusion came after seasoning. This one front-loads the passive base into the most statistically violent stretch. That is the new variable. It has no historical table.

The Anthropic discrepancy

One contract inside the AI segment deserves its own ledger entry, because the filing and the chief executive describe it differently.

The S-1 discloses a compute agreement under which Anthropic pays $1.25 billion per month for access to the Colossus GPU clusters in Memphis, over 220,000 Nvidia GPUs and more than 300 megawatts of power. Coverage of the filing, including WIRED’s, frames the contract as running through May 2029, roughly $45 billion in total, and reports that Anthropic confirmed the figures. That framing makes the deal the anchor of the AI segment’s revenue story and a load-bearing wall of the $1.75 trillion valuation.

On May 28, Musk publicly described the arrangement as a six-month lease. Both sides hold a 90-day cancellation right.

Both descriptions reference the same contract. They are not the same number. A $45 billion commitment through 2029 and a 180-day arrangement with a quarterly exit are different assets, and the gap between them is most of the AI segment’s claim to durable revenue. The segment lost $2.47 billion in Q1 on roughly $800 million of revenue, which means the Anthropic rate sits below SpaceX’s current cost to operate the cluster. The price reflects strategy, not unit economics. Whichever description of the term is operative, the buyer of SPCX on Friday is underwriting it.

The strongest version of the bull rebuttal should be stated plainly, because it is real: subsequent filings disclosed additional AI infrastructure agreements, including with Google, that bring reported contracted AI revenue to roughly $26 billion annually across the two largest customers, per Barchart’s read of the amended filing. If those contracts hold at term, the AI capex stops looking speculative and starts looking like productive assets. If. The cancellation clauses are where the thesis lives.

What the committee said

The legal and regulatory overhang gets one paragraph, because it is disclosed and priced rather than hidden. The S-1 reserves more than $500 million for litigation, driven in part by lawsuits over non-consensual sexualized images generated by Grok, including, per research by the Center for Countering Digital Hate cited in the complaints, an estimated 23,338 images of minors over an eleven-day stretch. Plaintiffs include three Tennessee minors, the City of Baltimore, and a UK member of parliament. Two countries temporarily banned the product. Buyers of SPCX are buying that docket too. It is in the risk factors, all 38 pages of them, next to the company’s own statement that it has a history of net losses and may not achieve profitability.

None of that stopped the order book. Demand reportedly reached $150 billion, roughly twice the offering, with an unusual 30 percent of the float targeted for retail, against a typical 5 to 10. The deal will price. The stock will trade. The index will do what its rules now say it must.

Which returns us to the one institution that looked at the largest IPO in history and applied its existing standards. S&P’s committee wrote that exceptions should not be granted solely based on market capitalization, and then granted none. In a week when an exchange rewrote its index to receive a single company, the most contrarian act in American finance was a committee doing nothing.

A closing disclosure, since this publication runs on them. I have never paid for Starlink. I used it once, on a Qatar Airways flight from Singapore to New York. No login, no payment screen, fourteen hours of flawless connectivity over two oceans. The product is not the question. The product was never the question. The price of the wrapper around it, and the rules rewritten to deliver that wrapper into your index fund fifteen days after it starts trading, is.

Position disclosure: no position in SPCX, and no intention to initiate one.

In May 2026, Ethereum did more than it has ever done. 64% of crypto’s transactions, half of all stablecoins, $628 billion settled in a single month. And ETH the token, a $244 billion asset trading near $2,025, earned almost nothing for it. Base fees collapsed after the Dencun upgrade, the blob fix shipped in December and changed nothing, and the network now burns about 70 ETH a day against 2,820 issued. ETH is inflating. The “ultrasound money” thesis is dead, and even the $250,000 bull deck quietly admits the issuance.

This is not an anti-Ethereum piece, and I hold ETH. Ethereum the network is winning, decisively, on every axis that matters for the next decade. The problem is that everything that makes it win is engineered to give value away, and the market has spent two years pricing the token as if the network were dying. The truth is stranger than that.

The seven loudest voices on ETH are all fighting about one question: whether the token captures the network through fees. They run the full range, from David Hoffman, who just sold every coin he owned, to Standard Chartered, which is calling for $40,000. That argument can be laid to rest. It doesn't.

ETH is not a fee asset and never really was. It is the reserve collateral of the most neutral settlement network on Earth, and you price that the way you price gold: on demand to hold it, not on the toll it collects.

The repricing is not $250,000, and it is not zero. It is a slow migration of monetary and collateral demand toward the asset the whole system is denominated in, plus a free option on institutions, AI agents, and tokenization, eventually routing value through it. It can stay mispriced for years. It has. Here’s the full case.

Subscribe now

---

The Week Two Smart People Looked At The Same Chart And Saw Opposite Things

In the last week of May, David Hoffman sold all of his ETH.

This matters because Hoffman is not a tourist. He built Bankless, a media empire, and a decade of his identity on Ethereum. So when he writes that the “ETH is Money” thesis played out rather than failed, that the token got the price it deserves and won’t be rerated up or down, the people who bought ETH because of him listen. His verdict was generous and brutal at once: massively bullish on Ethereum the network, finished with ETH the asset. Ethereum, he wrote, is a giver, not a taker.

The same week, Standard Chartered’s Geoffrey Kendrick told clients that ETH would reach $4,000 this year and $40,000 by 2030. His comparison was Amazon during the dot-com crash, the stock that fell to six dollars while the business got better every quarter.

And here is the most amusing detail. JPMorgan, in the same month, did both at once. On May 19, its research desk told clients ETH was unlikely to reverse years of underperformance against Bitcoin without real improvement in network activity and use, and that upcoming upgrades might not fix it. Then JPMorgan put its own product, a tokenized deposit token called JPMD, live on Base, an Ethereum layer-2. Bullish the network. Skeptical of the token. One institution, living the entire contradiction in a single month.

That contradiction is the whole story. And the data resolves it cleanly, just not in either camp’s favor.

[Ethereum’s share of crypto activity vs share of market cap. Bars at 64% transactions, 54% stablecoins, 53% tokenized assets, 52% DeFi TVL, all in white. Market cap share alone at 9.77% in red, far below the pack. Caption: “Ethereum does the work. ETH doesn’t capture the value.” Source: Mougayar monthly dashboard / rwa.xyz / DefiLlama / CoinMarketCap.]

The Burn Is Dead, And The Bulls’ Own Math Admits It

Start with the number that ends the loudest bull argument.

Every bull case eventually rests on the burn: usage rises, ETH gets destroyed, supply shrinks, price climbs. So look at what actually burns. Over roughly sixty days after December’s Fusaka upgrade, Ethereum burned 4,251 ETH through ordinary base fees and 3.16 ETH through blobs, the layer-2 data mechanism that was supposed to be the future of the network’s revenue. Blobs are seven hundredths of one percent of the burn. That is not a rounding error you grow out of. That is the mechanism not existing.

Zoom out, and it gets worse. The network issues about 2,820 ETH a day to validators and burns about 70. ETH is inflating at roughly 0.83% a year. Not ultrasound, not deflationary. Inflationary, today, at an all-time high usage.

The remarkable part is that the bulls know this. Etherealize, the firm behind the $250,000 target, states on its own website that ETH supply has been “flat-to-deflationary since the Merge.” Its own report, three pages from that line, concedes net issuance “has hovered around 0.8%.” The chain says 0.83% inflation. Three numbers, one of them contradicting itself inside one firm’s own materials. When the people calling for $250,000 cannot keep their scarcity story straight across two documents, the scarcity story is over.

Didn’t December’s upgrade fix this? Fusaka shipped a floor under blob fees so they could never collapse to zero again. It worked. Blob burn rose something like eightfold. But eight times almost nothing is almost nothing, and Ethereum’s monthly revenue chart tells the truth: a peak above $1.5 billion a month in 2021, and a flat line scraping the bottom through 2025 and into 2026.

[Blob burn donut: 99.9% EIP-1559 base fee / 0.1% EIP-4844 blobs, over ~60 days post-Fusaka. Source: Dune (@xeift).]

[Ethereum monthly chain revenue, 2021 to 2026, showing the $1.5B+ peak collapsing to a flat line. ETH-denominated and USD both noted. Source: DefiLlama.]

This Is Not A Bug. It Is The Architecture Working As Designed.

Hoffman’s deepest point is that none of this is an accident.

Ethereum chose the rollup-centric roadmap on purpose. Layer-2 networks do the cheap, fast transactions and pay Ethereum a thin slice for security. How thin? Base, Coinbase’s layer-2, earned more than $94 million in profit and paid Ethereum about $4.9 million for the privilege. Across 2025, every layer-2 combined paid Ethereum roughly $10 million for security, under a tenth of what they kept. The value moved up to the rollups and out to the apps, exactly as the design intended.

And the revenue that does reach the base layer increasingly skips ETH holders entirely. Break down what’s left, as the analyst The DeFi Report does on-chain, and the surviving income is mostly priority fees and MEV tips, which flow to validators. The base fee, the only part that gets burned and the only part that reaches every holder through scarcity, is the shrinking slice of a shrinking pie.

However, this is not an Ethereum disease. Solana, the chain the bears hold up as the winner, runs about $10 million a day in ecosystem fees and routes under 10% to the protocol. SOL inflates around 4%. It has no burn at all. Every general-purpose smart-contract chain gives its value to apps and validators and hands the token holder a thinning toll. The value that the entire sector fights about is mostly absent everywhere. The bears who sold ETH for SOL traded one version of the problem for a worse one.

And the last hope, that blob demand grows until fees finally spike, has the arrow pointing backwards. Blob demand peaked in mid-2025 and has fallen since, even as Ethereum raised the blob target. Rising capacity, falling use. The market is emptying, not filling.

The Institutions Came. They Came To The Layer Above ETH.

The bull story says Wall Street is coming and ETH rides the wave. Wall Street is coming. Watch where it actually lands.

JPMorgan’s JPMD, a tokenized claim on real bank deposits, went live not on Ethereum but on Base, the layer-2. It is the first time a major commercial bank put a deposit product on a public blockchain, and the fees accrue to JPMorgan, to Base, and to Coinbase. Ethereum the base layer earns the same near-zero blob fee as everyone else. The bank validated the ecosystem and routed the token nothing. It is Hoffman’s thesis with a Fortune 50 logo stamped on it.

The same pattern runs through stablecoins, the asset Ethereum supposedly owns. Ethereum holds roughly $163 billion of them, about 54% of global supply, per rwa.xyz. But PayPal’s PYUSD now does more transaction volume on Solana than on Ethereum, and has since the middle of 2025. Ethereum holds the balances. The activity rotates elsewhere.

Eighteen months ago, Ethereum’s fee dominance looked permanent. By total fees, the gross amount users pay, Solana, Tron, and Ethereum ran neck and neck in February, around $27, $24, and $23 million for the month. But fees are the top line. What matters for the token is revenue, the slice each chain keeps, and Ethereum is worse than at parity. Over the last thirty days, it kept about $5.2 million, a sixth of Tron's $32 million. And here is the line that should end the argument: Ethereum barely out-earned Base, its own layer-2, which kept $4.4 million. The parent and its own child, nearly tied, both a rounding error next to the volume Ethereum settles. The crown is gone, and the toll barely reaches the token.

But look closer at William Mougayar’s monthly dashboard, where every line traces to a primary source, and a pattern appears that both camps miss. Ethereum’s share is highest exactly where institutions live: 64% of transactions, 54% of stablecoins, 53% of tokenized assets. It is lowest where retail speculates: 26% of NFT volume, 21% of DEX volume. Ethereum is quietly becoming the settlement and collateral layer for serious money. That is the signal the bears overlook and the bulls misprice.

[Stablecoin supply by chain: Ethereum ~$163B / Tron ~$87B / Solana / BSC, with Ethereum’s share line flat-to-eroding over the year. Source: DefiLlama / rwa.xyz.]

The Strongest Bull Case Ever Written, And Why It Breaks Itself

If you want the bull case at full strength, read Etherealize’s “Ethereum and the Era of Productive Money.” It is the most sophisticated argument for ETH in existence, which is exactly why it deserves to be met head-on rather than waved away.

One disclosure first, because it matters: Etherealize describes itself as a product, business-development, and marketing arm for the Ethereum ecosystem, and it represents Ethereum in Washington. This is a registered advocate publishing a $250,000 target for the asset it exists to promote. That does not make it wrong. It makes it the most polished piece of bull marketing the cycle has produced, and we should treat it that way.

The case runs like this. Warren Buffett dismissed gold because it just sits there: own an ounce for eternity, and you still have an ounce. Bitcoin has the same flaw. ETH does not, because you stake it and earn a yield generated by the protocol itself, with no borrower and no counterparty. So ETH becomes the first money in history that compounds in your own hands. Add the rest of the monetary checklist, and if ETH captures even a slice of the roughly $30 trillion premium sitting in gold and Bitcoin, the arithmetic spits out $250,000 a coin.

It is genuinely elegant. It breaks on two facts.

First, the productivity is mostly dilution. We already saw that staking yield comes overwhelmingly from new issuance, not from fees, because the fees collapsed. Paying stakers with freshly printed ETH while everyone else holds still is not a dividend funded by profit. It is the printer moving coins from one pocket to another. “You will own more ETH” is true and nearly empty, because every other holder’s pile inflated right alongside yours. Real productivity is fees, and the fees are gone.

Second, and this is the heart of the whole debate: a monetary premium is the price the market pays for an asset to be boring.

Gold and Bitcoin earn their premium by being inert. Fixed supply, no governance, no surprises, nothing ever changes. That stillness is the entire product. Now look at what Etherealize celebrates about ETH: a quantum-resistance overhaul, a privacy roadmap, zkEVMs, a gas limit tripling, and a hard fork every six months. For a technology platform, that velocity is a triumph. For a monetary premium, it is a disqualification, because money is trusted precisely for what it refuses to do. Etherealize even concedes that ETH is weakest on established history, and then asks you to value it as established money.

You cannot be the most dynamic computer in the world and the most trusted dead asset in the world at the same time. Those are two different products, sold to two different buyers, for opposite reasons. Every bull conflates them. Hoffman, staring at the same chain, simply believes the platform and gives up on the premium. They are each half right, and the half they share is the wrong half.

There is a sharper version of this, and quantum computing just made it concrete. When the cryptographic ground itself can crack, inertness stops being a virtue and becomes a liability. The same refusal to change that makes Bitcoin trustworthy as money is exactly why it is paralyzed on the quantum threat, while Ethereum’s relentless dynamism, the thing that disqualifies it from a pure monetary premium, is what lets it adapt and survive. In a world where the substrate shifts, the adaptive chain is the one that stays money at all. More on that below.

What’s Actually Mispriced, And It Isn’t On Anyone’s Chart

So if it isn’t fees, and it isn’t a $250,000 monetary premium, what is the trade?

Follow the capex. Two things are tightening that have nothing to do with the burn.

The first is float. ETH is being locked and absorbed faster than it is being created. Around 30% of all ETH is staked and illiquid. One corporate treasury, Bitmine, holds about 5.4 million ETH, roughly 4.5% of the entire supply, a figure Messari confirmed crossing five million this spring. This is the demand the fee bears cannot see, because it has nothing to do with revenue and everything to do with ETH being the reserve asset of the system. The honest counterweight: spot ETH ETFs bled about $216 million over ten days in late May, and ETH against Bitcoin just printed a year-to-date low near 0.027. Supply is being absorbed by treasuries and staking and released by ETFs at the same time. Which side wins is the real open question, and it is a far more useful one than “how much does the network burn.”

The second is the development moat, and this is where the bears are most wrong. Ethereum is not standing still while it loses fees. It is building what the next decade needs, in the open, faster than anyone else. The clearest proof landed in March, when Google’s Quantum AI team published a paper showing the cryptography behind Bitcoin and Ethereum could be broken with roughly twenty times fewer qubits than anyone thought. One of its co-authors was Justin Drake, an Ethereum Foundation researcher, alongside Stanford’s Dan Boneh.

Ethereum’s own people helped write the paper that exposes the threat, and Ethereum already has a concrete four-part plan to rip out and replace its vulnerable signatures, on a 2029 migration target shared with Google and Cloudflare, while the Bitcoin community is still arguing over whether the threat is even real and its ossification makes the fix unlikely. No cryptographically relevant quantum computer exists today, and the timeline is genuinely debated; Drake’s own odds are aggressive and personal. But the gap in readiness is not debated.

In January, Ethereum also shipped ERC-8004, an identity and reputation standard for AI agents, authored jointly by MetaMask, the Ethereum Foundation, Google, and Coinbase, paired with the x402 payment rail. Its privacy stack, Railgun, has shielded $4.5 billion and is being pulled directly into wallets by a Foundation initiative. The upgrades ship twice a year, on schedule.

Here is the correction the bulls need, though, because it is the same trap in reverse. None of that development reaches ETH through fees. We buried that idea three sections ago. It reaches the token, if it reaches it at all, by making ETH the unavoidable reserve and collateral of the venue where all of this gets built, and by keeping alive a cheap option that one of these threads, institutional settlement, autonomous agents, tokenized assets, eventually routes real value through the coin. That option is out of the money right now. It is also nearly free at today’s price.

[ETH/BTC ratio, Aug 2025 peak 0.0432 falling to ~0.027 now, with StanChart’s 0.08 target and the implied “monetary premium” prices drawn as dotted lines far above. Visualizes the gap between the models and the market. Source: market data / StanChart note.]

[The float war: stacked view of ETH absorbed (staking + treasuries) vs ETH released (ETF outflows + net issuance). Build once the net-float figure is locked. Source: strategicethreserve.xyz / Farside / ultrasound.money.]

The Same Mistake As The AI Trade, Pointed The Other Way

Mispriced has one recurring claim: markets get the technology right and the economics wrong.

In the last piece, on big tech’s AI profits, the error ran in one direction. The market priced the technology as real, which it is, and ignored that a chunk of the economics underneath was circular and partly imaginary. The gap was bearish. The narrative was richer than the cash.

ETH is the mirror image. The market is pricing the economics as broken, which they are, fees collapsed, burn dead, token inflating, and ignoring that the technology and the capex are compounding a position no competitor can touch. The gap is bullish, but only if you stop using the income statement to value an asset that was never an income statement in the first place.

Same mistake, opposite sign. With AI, people paid for economics that were not there. With ETH, people are refusing to pay for an asset because they are measuring the wrong economics. In both cases the capex told the truth before the price did. That is the whole job: find the place where what is being built and what is being paid for have come apart, and stand in the gap.

Price The Collateral, Not The Clearinghouse

There is a clean historical warning here, and it is not Amazon.

It is Cisco. Cisco was real infrastructure, the actual plumbing of the internet, and the dot-com market priced it like a religion, then left it roughly flat for two decades while the internet it powered remade the world. Real technology and a tradeable asset turned out to be two different bets. The bulls quoting Amazon should sit with Cisco for a minute.

But ETH inverts even that. The market is not overpricing the infrastructure. It is underpricing it, by measuring the wrong thing. Tom Dunleavy, in the noisiest bull thread of the spring, got exactly one thing right and the rest wrong: ETH is the collateral, not the clearinghouse. You do not value the world’s settlement layer by the toll it collects. You value the asset the system is forced to hold. But you do not arrive at a number by decree, the way his model and Etherealize’s both try to, and you do not earn the premium of dead money while rebuilding your own cryptography every single year.

So price the collateral, on demand, not on a formula.

What that gives you is not $250,000 and not zero. It is a slow repricing of the reserve asset of the chain that actually won, plus a free option on the institutions and the agents now arriving. Sustained growth, not a 100x. And it can stay mispriced for years, because nothing forces the gap shut on a schedule. I hold ETH on exactly that thesis, with my eyes open to exactly that risk.

The network already won. The token is the part still waiting. And that wait, not a six-figure fantasy and not a funeral, is where the mispricing actually lives.

---

Sources and full reference list: https://telegra.ph/Ethereum-Already-Won-Your-ETH-Didnt-References-06-03

---

Written by Udipta Basumatari. Nine years in crypto. Former CMO of two crypto-native startups, took the BWLD token to a $75 million market cap at launch, and, working solo and anonymously, launched a prediction-market token that reached a $3 million cap. Angel investor across multiple blockchain organizations and a long-time advocate for decentralized data ownership. I hold ETH long-term, which is rather the point: this is a bull reading his own bag honestly.

Mispriced asks one question: what is the market getting wrong, and what does the capex tell us that the narrative doesn’t? Subscribe for the next one.

TLDR:

Three of the most powerful people in AI (Sam Altman, Jensen Huang, and Dario Amodei) walked back the jobs apocalypse narrative in the span of two weeks, right as OpenAI and Anthropic prepare to file for blockbuster IPOs. That timing is not a coincidence. The labor narrative is softening because the underlying financial structure is becoming harder to defend.

Here is what that structure actually looks like. Alphabet booked $28.7 billion in profit last quarter, which was a paper markup on its Anthropic stake, not a customer dollar. Amazon did roughly the same. The combined $1.6 trillion cloud backlog held by the four hyperscalers is concentrated in two AI labs that lose money on every token they sell. Those same labs spend most of their revenue right back on cloud compute rented from the companies that funded them in the first place.

It is a circle. Money goes from hyperscaler to lab, the lab spends it on the hyperscaler’s cloud, and both sides book the same dollars as growth. We have seen this movie before. In 2001 it was called Global Crossing and Qwest, and it ended in bankruptcy and fraud charges. The only difference today is that the AI version is legal.

Here’s the full case.

The tell

In the span of two weeks in late May 2026, three of the most powerful figures in artificial intelligence quietly walked back the story they had been telling for two years.

Sam Altman, speaking at a Commonwealth Bank of Australia conference in Sydney on May 26, said he had been wrong about AI eliminating jobs. His words: “I’m delighted to be wrong about this. I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened.” He added that the jobs picture was “likely to be very different from what we thought.”

The same week, Nvidia CEO Jensen Huang called the AI layoff explanation “lazy” and “irresponsible,” asking the obvious question nobody pushing the narrative wanted to hear: “AI has just arrived. How is it possible they’re already losing jobs?” Anthropic’s Dario Amodei, who a year earlier had warned AI could wipe out half of entry-level white-collar jobs, was softening his language in the same window.

Here is the context that makes the timing matter. OpenAI is reportedly preparing a confidential filing for a US IPO in the coming weeks, potentially targeting a $1 trillion valuation and raising at least $60 billion. Anthropic is on a similar track. When the people selling a story start dismantling it right before they cash out, the right question is not whether they had a change of heart. The right question is what they can see on the balance sheet that the public cannot.

The answer is that the revenue underneath the AI boom is not what it appears to be. Some of it is a paper entry. Some of it is the same dollar counted twice. And a dangerous amount of it depends on two companies that have never turned a profit.

The paper-gain machine

Start with the most striking number of the last earnings season.

In the first quarter of 2026, Alphabet reported $62.6 billion in net income, up 81% year over year. A blowout quarter by any measure. Except that $28.7 billion of that profit was not a profit in any sense a normal person would recognize. It was a mark-to-market gain on Alphabet’s equity stake in Anthropic.

Here is how that works. When a company holds a stake in a private business and that business raises money at a higher valuation, accounting rules let the stakeholder mark up the value of its holding and book the increase as income. No cash changes hands. No customer pays anything. The startup simply raised a new round at a higher price, and the investor records the unrealized gain as if it were earnings.

Amazon ran the same play in the same quarter. Its net income surged 77% to $30.3 billion, with roughly half of that attributable to the markup on its own Anthropic stake. Two of the largest companies on earth posted record profits, and a very large share of those profits came from marking up the value of a private AI lab that has never made money.

The tax expert Robert Willens put his finger on the strange part. These companies, he noted, are able to influence the value of an asset they own and then mark that asset to market through transactions with the very same entity. Amazon invests in Anthropic, does business with Anthropic, and books gains when Anthropic’s valuation rises. The investor, the customer, and the supplier are arranged in a circle, and the circle generates reported profit.

The mechanism is not theoretical. Anthropic’s Series H funding round closed at a $965 billion post-money valuation. Every step up in that number flows onto the balance sheets of its investors as reported profit, without a single customer dollar changing hands. When the valuation rises, the stakeholders book the gain. When it eventually stops rising, the same mechanism runs in reverse.

This is not illegal. It is not even unusual under current accounting standards. But it means that a meaningful slice of Big Tech’s headline AI-era profitability is an accounting entry tied to private valuations that the same companies help set.

The circle

The paper gains are the visible symptom. The vendor financing loop underneath is the engine.

The clearest documented example involves Amazon and Anthropic. Amazon has invested billions into Anthropic across multiple rounds, taking a stake reported at between 15 and 19 percent. In late 2025 it committed another tranche, and simultaneously Anthropic committed to roughly $100 billion in spending on Amazon Web Services over the following decade. Amazon gives Anthropic money. Anthropic gives the money back to Amazon as cloud spending. Amazon books that spending as AWS revenue.

The numbers on the Anthropic side are stark. According to an investigation by Ed Zitron drawing on AWS billing data, Anthropic spent $2.66 billion on Amazon Web Services in the first nine months of 2025 against $2.55 billion in revenue over the same period. That is 104% of revenue going straight back to its largest investor in the form of cloud bills. In January 2025 the ratio was even more extreme, with a reported $185.5 million AWS bill against $72.9 million in revenue, or more than 300%.

The same structure appears across the industry. Microsoft put $13 billion into OpenAI, much of it in the form of Azure cloud credits, which OpenAI then spent running its models on Azure, which Microsoft recorded as cloud revenue. Amazon struck a parallel arrangement with OpenAI. Nvidia invests in the labs that then commit to buying Nvidia chips.

In each case the same dollar is doing double duty. It counts as an investment on one ledger and as revenue on another. The growth story that justifies hundreds of billions in capital expenditure is, in part, a function of money moving in a circle rather than new money entering the system from paying customers

Two customers holding up a trillion-dollar backlog

When hyperscalers report their growth, the headline figure analysts watch is remaining performance obligations, or RPO. It represents contracted future revenue, the backlog of deals already signed. The combined RPO across Microsoft, Oracle, Amazon, and Google now stands at roughly $1.6 trillion. That number is the foundation of the entire AI infrastructure bull case.

Look at what it actually rests on.

Oracle reported $552.6 billion in remaining performance obligations as of the end of February 2026. More than $300 billion of that, roughly 54%, comes from a single customer: OpenAI. Microsoft’s commercial RPO of around $627 billion is, by analyst estimates, close to half tied to OpenAI as well. A backlog that Wall Street treats as evidence of broad, durable demand is, in reality, concentrated in two AI labs that have never been profitable and that fund their compute purchases through capital raises rather than operating cash flow.

The market has started to notice. Oracle raised $18 billion in debt in September 2025 and was reportedly seeking tens of billions more to finance the data center buildout it promised OpenAI. Oracle’s credit default swap spreads, the cost of insuring against an Oracle default, have widened to their highest levels since 2009. Bond investors are pricing in the risk that a backlog dependent on one unprofitable customer might not convert into cash.

The numbers do not work even when you cheat

Here is the part that should end the debate, and it comes from an investment bank, not a short seller.

Joachim Klement, a strategist at the brokerage Panmure Liberum, modeled the implied return on the AI capital expenditure of the five largest hyperscalers from 2025 to 2030. He did it under the most generous possible assumptions: zero operating costs, top-end revenue gains, pure projected revenue measured against capital spending. No GPU depreciation, no power bills, no salaries. Just the best case.

Even then, four of the five posts negative implied returns. Microsoft comes in at -9.2%. Alphabet at -15.7%. Meta at -28.8%. Oracle, the most OpenAI-dependent of the group, lands worst at -35.6%. Only Amazon cleared positive, at a slim +7.2%, helped by its custom silicon. And remember, those are the numbers before a single real cost is subtracted. The actual returns are worse.

Klement’s broader math is just as stark. The five hyperscalers would need to find somewhere between $2 trillion and $5 trillion in additional annual revenue to justify their planned data center spending, which is projected to exceed $600 billion in 2026 alone. To get there at current capex-to-sales ratios, they would have to roughly quadruple their revenues. His conclusion, in his own words, is that “if the hyperscalers continue on the current trajectory, the AI boom will become one of the largest destructions of shareholder value in history.” He also calculates that the buildout is already 60% larger than the dotcom bubble, with technology capital expenditure now accounting for nearly all of US GDP growth.

This is not a fringe view from a permabear. It is an investment bank running the arithmetic and finding that the best case loses money.

We have seen this before

Strip away the word “AI” and the structure is familiar. It is the telecom bubble of 2001.

In the late 1990s, telecom companies needed to show booming demand for fiber-optic capacity to justify their valuations and their capital spending. Some of them manufactured that demand by swapping capacity with each other. Global Crossing and Qwest Communications famously exchanged essentially identical fiber-optic network capacity, each booking the other’s purchase as revenue. The trades created the appearance of a thriving market, while no real end demand existed underneath.

When the music stopped, Qwest was forced to erase $1.4 billion in improperly recognized income. Global Crossing filed for bankruptcy in early 2002. Executives at both companies faced prosecution. The swaps had been securities fraud.

The AI version is mechanically the same. Capital flows in a circle between a small number of large players, each booking the flow as growth, creating the appearance of demand that outruns the real paying market underneath. The one meaningful difference is that the 2001 swaps were illegal, while the 2026 vendor financing loop is fully permitted under current accounting rules. That should be reassuring and is not. The fraud in 2001 was the deviation from the rules. In 2026, the rules themselves accommodate the circularity.

Where it breaks: the real world

A circular economy of subsidized compute can sustain itself only as long as the money keeps circling. The moment the technology leaves the subsidy loop and meets a real enterprise customer with a hard budget, the unit economics show their true shape. And that is exactly what has started to happen.

Uber rolled out Claude Code to 5,000 engineers in December 2025. By April 2026, the company had exhausted its entire annual AI budget, four months into the year, against a $3.4 billion R&D base. Heavy users were costing up to $2,000 per person per month. The CTO’s line was telling: “I’m back to the drawing board, because the budget I thought I would need is blown away already.”

The admission then climbed the org chart. In late May, Uber’s president and COO Andrew Macdonald described the moment he learned the company had blown its annual AI budget in four months as “head-exploding.” His assessment of the spend was blunt: “If you’re not actually able to draw a direct line to how many useful features and functionality you’re shipping to your users, that trade becomes harder to justify.” When the COO of a company where 95% of engineers use AI monthly and 70% of committed code is AI-generated cannot draw a line from the spend to the product, the ROI is not lagging. It is missing.

Then came the data point that should worry every CFO in tech. In May 2026, Microsoft began canceling thousands of internal Claude Code licenses across its Experiences and Devices division, covering Windows, Microsoft 365, Outlook, Teams, and Surface, with a June 30 deadline. The official reason was tool consolidation. The real one was the cost. The company that put $13 billion into OpenAI, which runs the Azure infrastructure powering much of the model layer, that has effectively unlimited cloud capacity, looked at the bill for a competitor’s coding tool and decided it was not worth paying. If Microsoft cannot justify the cost, the question of who can becomes very short.

The pattern is now widespread enough that Axios called it “Corporate America’s AI reckoning.” In that reporting, an AI consultant said one of their clients spent half a billion dollars in a single month after forgetting to set usage limits on Claude licenses for employees. The CEO of CloudBees offered the darkest interpretation of the layoff wave: workforce cuts, he suggested, may simply be “the only lever they can pull” to offset their AI bills. Read that against my last piece. The layoffs were never really about AI replacing workers. They were partly about paying for the AI. One CTO told Axios that employees were using frontier models to check the weather, which gets expensive fast when enterprise plans are not actually all-you-can-eat.

The broader spend data confirms the trajectory. Procurement firm Ramp reports that the average business now spends roughly 13 times more on AI tokens than it did in January 2025. Tokens remain under 2% of total spend even for the heaviest users, but a 13x annual growth rate does not stay small for long. GitHub, owned by Microsoft and running on Azure, is abandoning flat-rate pricing for Copilot in June 2026 precisely because the flat-rate model was bleeding money on heavy agentic users. And Nvidia’s own Bryan Catanzaro admitted in April that “the cost of compute is far beyond the costs of the employees,” which is the GPU maker conceding that the thing meant to replace expensive humans currently costs more than the humans.

The pitch was that tokens would get cheaper and agents would replace workers. Outside the subsidy loop, tokens are expensive and getting more so on a per-workload basis, and the labs are quietly walking back the replacement story.

The best argument against all of this

The strongest counter to this thesis is not a denial of the numbers. It is the Solow Paradox.

In 1987, the economist Robert Solow quipped that “you can see the computer age everywhere but in the productivity statistics.” Massive corporate IT investment through the 1970s and 1980s had failed to show up in measured productivity. Then, beginning around 1996, the productivity payoff from computing and the internet finally arrived, years after the capital had been spent. The lag was real, and betting against the technology during the lag would have been a mistake.

In May 2026, the Federal Reserve Bank of San Francisco argued that AI may be on the same curve. Labor productivity is rising even as total factor productivity stays flat, exactly the divergence seen in the mid-1990s before the internet’s gains materialized in the aggregate data. Maybe, the argument goes, the AI ROI is real and simply lagging, and the capital expenditure will be vindicated in a few years.

It is a serious argument, and it deserves a serious answer. Here is mine. The 1990s internet buildout was funded by real demand and real customer cash, not by a circle of hyperscalers marking up their own investments and booking each other’s spending as revenue. The Solow lag explains why productivity gains might show up late. It says nothing about whether reported revenue is circular or whether reported profit is a paper markup. Those are accounting questions, not timing questions. The Atlanta Fed found that executives perceive larger AI productivity gains than researchers can actually measure in company revenue, attributing the gap to “delayed output realizations.” Maybe. Or maybe it is the same wishful thinking that led 75% of executives in a separate survey to admit their AI strategy was “more for show” than real guidance. The lag may well be real. But a lag in productivity does not turn a circular dollar into a customer dollar.

Why nobody is pricing it in

If the structure is this visible in public filings, why has the market not corrected?

Because the loop is self-reinforcing on the way up. Paper gains inflate reported profits. Inflated profits push up market capitalizations. Higher market caps mean index funds and passive retirement accounts, which buy in proportion to market weight, mechanically purchase more of the inflated stocks. That buying pushes prices higher still. At no point in this cycle does anyone need to verify that the AI revenue has been converted into actual cash profit. The machine runs on reported figures, and the reported figures are partly manufactured by the circularity.

The most mispriced risk in global finance right now is not a single company or a single number. It is the widening gap between the optimism on earnings calls and the invoices landing on corporate desks. One side is a story. The other side is a budget. They cannot both be right indefinitely.

What breaks first

A few leading indicators are worth watching because they will move before the headline numbers do.

Oracle’s credit spreads are the clearest early warning, already at their widest since 2009. If they keep widening, the bond market is telling you it does not believe the OpenAI backlog converts to cash. Watch for more hyperscalers quietly canceling AI tools internally, the way Microsoft did, because internal cancellation is the most honest signal a company can send about unit economics. Watch the IPO pricing for OpenAI and Anthropic, because the bankers setting those prices have seen the real numbers and the valuations they land on will reveal what the smart money actually believes. And watch whether the labs hit their internal revenue targets, because the entire backlog assumes they grow into commitments that currently exceed their revenue many times over.

When the markup machine slows, the sequence is predictable. Capital expenditure slows. The labs miss revenue targets. The paper gains reverse into paper losses. And the circular flow that looked like growth on the way up looks like contagion on the way down.

Disclosure: I work at Adobe as a solution consultant. This piece is about the financial structure of the foundation layer, the labs, and the hyperscalers building the AI infrastructure. It is not about application-layer software serving real customers who pay real money for real outcomes, which operates on entirely different unit economics. The bubble I am describing is in the foundation, not in the products built on top of it.

The connection

My last piece argued that the AI layoff narrative was breaking and that a selective rehire wave had already begun. This is the other half of that story. The labor market repricing was the symptom. The capital market mirage is the cause.

When the loop unwinds, the layoffs reverse, but the rehire is selective, exactly as the labor data already shows. The companies that survive the correction will hire the people who can actually wield these tools productively, and they will let everyone else compete for a shrinking pool of unautomated work. The capital story and the labor story are the same story told at two different layers of the stack.

The technology is real. The productivity gains may even be real, eventually. But a real technology can sit on top of a fake balance sheet for a surprisingly long time, right up until it cannot.

In 2000, the internet was real, too. Cisco was the company selling the picks and shovels for it, the most valuable company in the world at its peak. The internet went on to remake the entire economy, exactly as promised. Cisco’s stock took twenty-six years to climb back to where it traded in March 2000. Being right about the technology and being right about the returns are two completely different bets.

Where do you sit when the correction comes? I built ismyjobscrewed.com to help you find out, and to map the upskilling path that puts you in the cohort that gets hired rather than the one that gets cut.

Mispriced covers tech at the intersection of AI, data, and crypto. Following the capex, finding what the market got wrong. Subscribe for the next one.

Sources and full reference list: https://telegra.ph/Half-Of-Big-Techs-AI-Profit-Is-Imaginary-References-05-30

The agentic AI replacement narrative that justified 18 months of tech layoffs is cracking. Token prices are going UP across all three frontier labs. Microsoft just canceled Claude Code internally because the bill was too high. GitHub is killing flat-rate pricing in June. Uber blew its entire 2026 AI budget by April. 95% of enterprise AI pilots are still failing despite spending going up 65%. The data center buildout looks exactly like the 1999 fiber bubble.

This isn’t a case against agentic AI as a category. The deployments that work, augmentation with humans in the loop, scoped to specific workflows, are quietly delivering real ROI. The category that’s breaking is the full-replacement fantasy.

Meanwhile, the companies running the working playbook are hiring hard. Amazon: 11,000 engineers. Salesforce: 1,000 new grads. IBM is tripling entry-level hiring. ClickUp just announced $1M salary bands for AI-native operators. Gartner forecasts 50% of AI-attributed layoffs will be quietly reversed by 2027. Forrester says the same. But the rehire is selective, and the new roles pay less unless you can wield agents.

This is the rehire wave nobody is pricing in. It started in April. It is likely to accelerate over the next 18 months. After that, the cohort closes, and a new equilibrium sets in for the next decade of knowledge work.

Here’s the full case.

Meta laid off 8,000 people on Wednesday. The stock barely moved.

Two months ago, Block laid off 4,000, and the stock jumped 20% after hours. Same playbook. Same AI narrative. Same press release language. Completely different market reaction.

Something broke in between. Let me show you why.

The pump is dead

The AI layoff stock pump worked for 18 months, and then it died sometime in Q2 2026. Wall Street stopped rewarding the playbook, which means CEOs just lost their main reason to keep cutting. The timeline tells the story better than any analysis can.

Block, March 2026. Jack Dorsey announced 4,000 layoffs, 40% of the company, and framed the entire cut as a pivot to AI. The stock jumped 20% after hours, rose another 17% during Friday morning trading, and the framing seemed to be working perfectly. Classic pump.

But the cracks were already visible if you knew where to look. When Wired pressed Dorsey directly on whether this was AI washing, he gave a wishy-washy answer that satisfied nobody. Then, in a follow-up post on X, he quietly admitted the real reason: “over-hired during covid because I incorrectly built 2 separate company structures (Square & Cash App) rather than 1, which we corrected mid 2024.” Aaron Zamost, Block’s head of communications from 2015 to 2020, went further and wrote a New York Times op-ed calling the cuts to the policy team and DEI roles “standard prioritization and cost management, not an A.I.-driven reinvention.”

The market didn’t care in March. By May, it absolutely did.

GitLab, May 11. CEO Bill Staples announced layoffs and an “agentic era” restructuring using almost the same script as Block. The stock dropped 9% the next day and closed at $23.08, down from $52 a year ago, leaving the company worth less than a third of its 2021 peak. Wall Street wasn’t buying the agentic era pitch any more than it was buying the broader narrative.

Meta, May 20. Mark Zuckerberg cut 8,000 people, 10% of the workforce, in a move that would have moved the stock 5% in either direction a year ago. The stock stayed flat in midday trading. Meta is now down 8.7% year to date, while Alphabet is up 24% and Amazon is up 12%, and Senator Bernie Sanders publicly criticized the cuts within hours of the announcement.

Sam Altman, of all people, has been the loudest voice calling this AI washing. His framing is that companies are using AI as a rationale for layoffs they were going to make anyway, regardless of any actual productivity story. The market took longer than I expected to catch on to the pattern, but once it did, the trick stopped working almost overnight.

The pump worked for as long as it did because the AI narrative gave CFOs perfect moral cover for cost-cutting. Boards rewarded the framing because the margin expansion story was easy to model and even easier to communicate to investors. Four quarters later, the margin expansion never showed up in earnings, and once GitLab dropped 9% on an “agentic era” announcement, every other CEO got the memo. The trick is done.

This matters because layoffs aren’t free. They cost severance, morale damage, customer trust, knowledge loss, and the recruitment expense of rehiring later when you realize you cut too deep. CEOs were willing to absorb all of those costs as long as the stock pumped on the announcement, but without the pump, the math simply doesn’t work anymore.

The token economics are broken

The pitch was simple -> Tokens get cheaper -> Agents replace humans -> Margins expand.

Reality flipped on all three, and the bills coming due are big enough that even Microsoft just blinked.

Frontier model prices are going up, not down.

Gemini 3.5 Flash launched on May 19 at $1.50 input and $9 output per million tokens, which is 3x the price of Gemini 3 Flash. Artificial Analysis ran the full Intelligence Index benchmark and found the new Flash was 5x more expensive to run than the previous Flash, and 75% more expensive than Gemini 3.1 Pro. That’s for a “Flash” tier model, the one specifically positioned as the cheap workhorse for high-volume agentic workloads.

The same pattern shows up across the labs. GPT-5.5 launched at 2x the price of GPT-5.4. Claude Opus 4.7 is 1.46x the price of Opus 4.6 once you account for the new tokenizer. Simon Willison summed up the trend perfectly in his post on the Gemini 3.5 launch: “All three of the major AI labs are starting to probe the price tolerance of their API customers.” This is the exact opposite of the pricing trajectory the layoffs were sold on.

The AI Tax is real, and CFOs are now tracking it.

Procurement intelligence firm Tropic published data showing AI-driven software prices have jumped 20% to 37% across vendor categories on enterprise renewals, with vendors restructuring their SKUs to force migrations into more expensive AI-bundled plans. They call it the “AI tax,” and even buyers who negotiate hard still see final pricing land 12% above pre-AI baselines.

Fortune ran a piece on April 28 quoting Nvidia exec Pat Lee directly: “the cost of compute is far beyond the costs of the employees.” That’s the maker of the GPUs telling enterprise buyers that AI is currently more expensive than the humans it was supposed to replace, which should give you a sense of how upside-down the unit economics have become.

Anthropic got caught nerfing.

On April 23, Anthropic published a public postmortem confirming what developers had been complaining about for six weeks. Claude Code, Claude Agent SDK, and Claude Cowork had all been degraded between March 4 and April 20 due to three separate product changes. Two of the three changes were deliberate decisions rather than bugs, specifically a reasoning effort downgrade and a verbosity cap on the system prompt.

Worth saying clearly: this wasn’t model regression at the weights level. The underlying model was fine. Anthropic ran the math, decided to reduce output quality to manage inference costs, and got caught when developers started filing GitHub issues comparing outputs side by side. They eventually issued the postmortem and reset usage limits, though the phantom-token bug from the same incident was still present in v2.1.137 as of May 13.

I’m not telling you this to bash Anthropic. I’m telling you this because it reveals that the supply side is more constrained than the marketing suggests. If frontier labs are willing to nerf their flagship products to manage demand, the “infinite cheap intelligence” thesis is fiction.

GitHub just killed flat-rate pricing.

On April 27, GitHub announced that Copilot is moving to usage-based billing effective June 1, 2026. Premium request units are gone, replaced with GitHub AI Credits priced by token consumption. The base subscription prices stay the same, but the value you get inside those prices drops sharply for anyone doing agentic work.

GitHub Chief Product Officer Mario Rodriguez justified the shift as essential for a “sustainable business model,” which is corporate-speak for the fact that the flat-rate experiment was bleeding money on heavy agentic users. This is the canary in the coal mine that everyone building on top of LLMs should be paying attention to. GitHub is owned by Microsoft and runs on Azure with effectively unlimited compute, and if GitHub can’t sustain flat-rate AI pricing under those conditions, nobody can. Every other developer tool is going to follow within 12 months.

One developer in the GitHub community thread did the math on his actual workflow and posted the result. He was paying €39.99/month for Pro+ in 2025 with 1,500 premium requests. Under the new model, his projected cost goes from €67 in April to €966 in June. That’s a 14x increase for the same usage pattern, and that’s the bill enterprise CFOs are about to be staring at across every developer in their org.

Uber blew the entire 2026 AI budget by April.

CTO Praveen Neppalli Naga revealed on April 15 that Uber had exhausted its full-year AI budget just four months in. The company’s R&D budget is $3.4 billion. They rolled out Claude Code to 5,000 engineers in December 2025, usage nearly doubled by February, and heavy users were costing up to $2,000 per person per month. The direct quote from the CTO is the one I keep coming back to: “I’m back to the drawing board, because the budget I thought I would need is blown away already.”

Now scale this up. Uber is one company with a $3.4B R&D budget. If they can’t sustain 5,000 engineers on Claude Code for a full year, what does that say about every other Fortune 500 trying to pivot to agents? Pegasystems already put a token usage calculator on their homepage, token budgeting is now standard enterprise procurement, and CFOs across the industry are reviewing line items called “Anthropic credits” for the first time in their careers.

Microsoft pulled the plug on Claude Code internally.

This is the data point that should make every CFO in tech sit up and pay attention.

The Verge reported on May 14 that Microsoft is canceling thousands of internal Claude Code licenses across its Experiences and Devices division, which covers Windows, Microsoft 365, Outlook, Teams, and Surface. The deadline is June 30, 2026, and engineers are being moved to GitHub Copilot CLI. The official framing is “tool consolidation,” but the underlying reality is more revealing than that framing suggests.

Microsoft invited Claude Code in last December. Thousands of developers picked it up over the following six months, and sources told The Verge that it became “very popular” inside the company. Then the company that put $13 billion into OpenAI and runs the Azure infrastructure powering most of Anthropic’s compute looked at the bill, looked at the cannibalization of their own Copilot product, and decided it wasn’t worth paying.

That is the single most important data point in this entire essay, and not because Microsoft is special. The point is that if Microsoft can’t justify the bill even with effectively infinite cloud capacity and a fiscal partnership at the model layer, nobody else in tech can either.

Personal data point worth adding here. I’m on an Anthropic Enterprise plan, and I burned 80% of my monthly credits in half a month doing genuinely productive work. The math that the layoffs were predicated on assumed I’d cost a fraction of a junior developer. The reality is that I’m spending agent budget that already exceeds the salary of a new grad in Bangalore, and I’m one person.

The CXO economics simply don’t work the way they were sold. Tokens are not the cheap labor substitute that the full-replacement narrative promised, at least not at the frontier tier where autonomous replacement work has to happen. Augmentation economics are different and they’re working.

The data center bubble

China built 500+ AI data centers in 2023 and 2024. Up to 80% sit unused.

The United States is on the same trajectory at 5x to 10x the capex.

This is the part of the story that doesn’t get enough airtime. Even if the agentic thesis were correct on the demand side, the physical supply layer is over-committed in ways that look exactly like every prior infrastructure bubble.

The China precedent. Local governments, state-backed firms, and real estate speculators launched over 500 AI data centers in 2023 and 2024. 150+ are operational. But reports indicate up to 80% of the newly built compute resources remain unused, per MIT Technology Review. Most were built in remote areas for cheap power, too far from AI hubs where low-latency inference actually matters. The driver was political incentives and hype. The outcome is empty buildings.

The US version is bigger. Alphabet, Amazon, Meta, and Microsoft are expected to spend more than $650 billion combined in 2026 alone on AI capacity expansion. McKinsey projects $7 trillion in global data center construction through 2030. The WTO reports that imports account for 70-90% of the value of US AI investment. That’s a level of import intensity with few parallels in modern American industrial history. And the supply chain runs through China for transformers, switchgear, and batteries.

Half of the planned US builds for 2026 are already delayed or canceled per Bloomberg. TD Cowen reported that Microsoft has already cancelled leases with at least two data center operators, totaling a few hundred megawatts.

The bubble warning came from inside the house. Alibaba Chairman Joe Tsai, who has every incentive to talk his own AI book up, instead publicly called the US data center spend a bubble at the HSBC Hong Kong summit. His exact frame: “I’m still astounded by the type of numbers that’s being thrown around in the U.S. about investing into AI. People are talking, literally talking about $500 billion, several hundred billion dollars. I don’t think that’s entirely necessary.”

His specific red flag was developers building “on spec” without binding agreements from the big AI buyers. Same pattern that broke China’s buildout.

The grid is the hard ceiling. Morgan Stanley forecasts that US data centers face a 44 gigawatt electricity shortfall in the next three years. China added 543 GW of power capacity in 2024 alone, more than the total power capacity the US has added in its entire history. Goldman Sachs: “AI’s insatiable power demand is outpacing the grid’s decade-long development cycles, creating a critical bottleneck.”

This is the structural problem. Token prices stay high because the supply layer can’t scale. The supply layer can’t scale because you can’t recommission a nuclear reactor in 18 months, and you can’t build a gigawatt solar farm in two years. Stifel Nicolaus is already forecasting an S&P 500 correction tied to this dynamic. They describe the data center capex boom as “a one-off build-out of infrastructure, while consumer spending is clearly on the wane.”

The demand side is constrained by token economics. The supply side is constrained by capex risk and grid physics. If either of these cracks, the entire full-replacement thesis cracks with it. If both crack, the layoffs reverse faster than anyone is pricing in.

The ROI math isn’t there

Companies spent 65% more on AI in 2026 than in 2025. Average enterprise AI spend went from $7 million to $11.6 million. The success rate moved from 5% to 29%, which sounds like progress until you do the arithmetic and realize that’s the most expensive 24 points of progress in enterprise software history. And the curve is flattening fast.

The data comes from Writer’s 2026 AI Adoption in the Enterprise survey, released April 7, covering 2,400 leaders, including 1,200 C-suite executives. The numbers are brutal across every dimension. 97% of executives deployed AI agents in the past year, but only 29% report significant ROI from generative AI, and only 23% see significant returns from agents specifically. 48% describe AI adoption at their company as a “massive disappointment,” 75% admit their AI strategy is “more for show” than actual guidance, and 38% of CEOs report high or crippling stress around AI strategy. 64% fear losing their job if they fail to lead the transition.

The abandonment curve is even worse than the ROI numbers suggest. 42% of companies abandoned most AI initiatives in 2025, up from 17% the year before, which is a 147% year-over-year increase. The average organization scrapped 46% of POCs before they ever hit production, which means roughly half of all AI experiments inside the enterprise are dying before they ever generate revenue.

Independent corroboration is now stacked deep across multiple research firms. RAND Corporation found that AI projects fail at twice the rate of non-AI IT projects. Carnegie Mellon’s TheAgentCompany benchmark showed that the best AI agents fail nearly 70% of real-world office tasks. Gartner projects 40% of agentic AI projects will be cancelled by 2027 due to governance failures. The pattern is consistent regardless of who is measuring or how.

Now look at where the 5-29% who actually succeed are winning. Back-office automation, specialized vendor tools, tightly scoped use cases. Not flashy chatbots, not agentic replacement of knowledge workers. The MIT NANDA research found that purchased AI solutions succeed 67% of the time versus 22% for internal builds, which is a brutal indictment of every company that tried to build agentic platforms in-house instead of buying focused tools from specialized vendors.

The pattern across the winners is consistent: scoped use cases, humans reviewing the output, agents automating the repeatable middle of the pipeline. As a solution consultant at Adobe I’ve watched this play out across enterprise deployments in APAC, and Forrester’s Total Economic Impact studies show the same shape: triple-digit ROI when the deployment is scoped right, near-zero when companies try to remove humans from the workflow entirely.

The companies that bet hard on full agentic replacement of human workers are sitting in the 71% that already abandoned the initiative or are about to. The companies that succeed are doing humans plus agents, not agents alone. This is the entire upskilling thesis in one paragraph.

The rehire wave

The companies that fired the loudest in 2025 are quietly hiring in 2026. Watch what they do, not what they say.

The headline data point comes from Gartner. Published February 3, 2026. Direct forecast: by 2027, 50% of companies that attributed headcount reductions to AI will rehire staff to perform similar functions, but under different job titles. Forrester reached the same conclusion in their late 2025 future of work outlook with one critical addendum: “We expect half of AI-attributed layoffs to be quietly reversed, with jobs returning offshore or at lower wages.”

That second part matters. The rehire is real. But the new roles pay less unless you’re in the cohort that can actually wield agents. Which is exactly the labor market split this essay is about.

55% of employers already regret their AI-driven layoffs per HR Director coverage. The macro data is even more striking. Citadel Securities just published Indeed posting data showing software engineering postings crashed from 73 in January 2024 to 61 by mid-2025, then ripped vertical to over 70 by early 2026. Software engineering is leading the broader labor recovery, which is the exact opposite of what the AI-replacement narrative would predict. The BLS independently projects 15% software developer growth from 2024 to 2034, against 3% for all occupations.

The economic logic is Jevons paradox. When a resource becomes cheaper or more efficient to produce, total demand for it goes up, not down. The 19th-century example is coal: when steam engines got more efficient, coal consumption exploded rather than collapsed. Software is now playing out the same way. AI didn’t replace engineers; it expanded the surface area of what’s worth building. The companies that internalized this fast are the ones currently hiring.

The macro recovery is real. The company-level data points back it up.

Amazon. AWS CEO Matt Garman announced at the 2026 What’s Next event that Amazon will hire 11,000 software engineers and interns this year, which comes after the company cut 30,000 people in late 2025 and early 2026. The exact words are worth reading in full because the framing matters: “We are hiring just as many software developers as we ever had inside of Amazon. And in fact, I see the demand for that really accelerating. I think as we think about where AI really takes off and where agents can help free us from the monotony of the repeatable parts of our task, it’s really about leaning in.” That’s a 37% rehire ratio at the same company, just for a different skill profile.

Salesforce. Marc Benioff announced on X on April 25, “We’re hiring 1,000 new grads & interns right now to ride the AI exponential,” and the timing is significant because it came months after laying off roughly 1,000 in February 2026 and another 1,000 in February 2025. Salesforce’s Agentforce ARR hit $800 million on the back of this hiring push, up 169% year over year, with 29,000 deals closed since launch. Full replacement ratio, different roles, and they’re now building the AI products that are doing the replacement at other companies.

IBM. Tripling entry-level hiring as of February 2026, with CHRO Nickle LaMoreaux explicitly making the contrarian argument: “The companies three to five years from now that are going to be the most successful are those companies that doubled down on entry-level hiring in this environment.” This is the strongest contrarian signal in the entire essay because the HR head of a Fortune 50 company is explicitly arguing the opposite of the AI replacement narrative that the rest of the industry has been selling.

Meta. Cut 8,000 people, but is reassigning roughly 7,000 of them into newly formed AI-focused teams, which works out to an 87.5% internal reassignment ratio. Zuckerberg explicitly told staff that AI was not driving the layoffs, and the cuts went to areas of the business that were over-resourced while the AI teams kept growing.

The macro signal. The National Association of Colleges and Employers projects employers will increase hiring by 5.6% for the class of 2026, and many of the industries assumed most vulnerable to AI automation (information services, engineering, professional services) are exactly the ones planning to boost hiring most aggressively.

The pattern is clear once you stop reading press releases and start counting jobs. Roughly half the AI-attributed cuts will be reversed by 2027 per Gartner, and the specific company examples sit in a wide range from Amazon at 37% to Salesforce at full replacement, plus significant internal reassignment at companies like Meta. But the new roles are AI-native, and per Forrester, they pay less unless you can prove agent-native productivity. The old roles are not coming back at the old wages.

What the new playbook actually looks like

ClickUp CEO Zeb Evans published a post today, May 22, that reads like the operating manual for the new equilibrium. He cut 22% of the headcount. Then he said the quiet part out loud about what the remaining workforce will look like.

The framing is what he calls the “100x org.” Quote: “The 100x org is actually heavily dependent on people - infinitely more than today. This is only possible with 10x people who have embraced and adopted new ways of working.”

He coined a phrase I want to steal: “the great reckoning of AI coding.” His take is that companies pushing every engineer to use infinite tokens are “celebrating 500% more pull requests” while “customer outcomes don’t match the volume of code being generated.” More code is just another bottleneck. The skill is judgment, not output.

But here is the line that should reset every compensation conversation in tech: “We’re introducing $1 million cash/year salary bands with a path available to nearly everyone in the company if they produce 100x impact by creating or managing AI systems.”

A million dollars cash for AI-native operators. Severance for everyone else.

This is the labor market bifurcation happening in real time, with a number attached. The Writer survey predicted it. ClickUp just put it in writing.

Worth being honest about the tension. ClickUp ALSO cut 22%. Some readers will see this as just another AI layoff dressed up nicely. The honest read is that both things are true. The dying playbook (cut and hope the stock pumps) and the emerging playbook (cut hard, pay survivors like founders) look identical from the outside if you only count the layoffs. The difference shows up in what happens to the people who stay.

If your company is doing the first version, the cuts continue, and morale collapses. If it’s doing the second, the top decile gets a path to seven figures, and the middle has to choose which side of the line they want to be on.

Evans wrote one more line worth holding onto: “Nearly every company will make changes like these. The ones that do it proactively will define what comes next.”

He’s right. And the workers who become AI-native proactively will get the seven-figure bands. Everyone else is competing for what’s left.

The upskilling window

The job market reversal won’t restore the old jobs at the old wages. It creates new ones at slightly lower volume, with sharply bifurcated comp bands, and the window to become AI-native and survive the reset is roughly 18 months. After that, the cohort closes and the new equilibrium sets in for the next decade.

The Writer survey makes the labor market split brutally clear across every dimension. 92% of the C-suite is actively cultivating “AI elite” employees as a deliberate retention strategy. 60% of companies plan layoffs specifically for non-adopters, 77% of executives say employees who refuse to become AI-proficient won’t be considered for promotions, and 90% say AI super-users will require completely rethinking performance evaluation frameworks.

Forrester’s 2026 future of work outlook found that only 16% of workers had a high “AIQ” (their measure of AI readiness) in 2025, projected to reach just 25% by the end of 2026. Which means the other 75% is the addressable market for the rehire wave that’s about to start. Get into the top quartile, and you get the new roles at the new comp bands; stay in the bottom three quarters, and you compete against a much larger labor pool for what’s left.

The productivity gap between cohorts is already showing up in the data. AI super-users save 9 hours per week compared to laggards, which is 4.5x the productivity differential. They are 3x more likely to have received both a promotion and a raise in the past year. This is the gap that’s about to widen into a chasm as the rehire wave selects on these exact attributes.

So what does AI-native actually mean in 2026, as a practical matter? Prompt engineering is table stakes at this point, not a differentiator. The real skill is agent orchestration, which covers knowing how to design multi-agent workflows with proper handoffs, building eval suites that actually catch failure modes, architecting RAG systems over private data, and managing token budgets at the workflow level. It also covers knowing when to use Frontier Intelligence versus when Flash-Lite will do the job for a fraction of the cost. And the most underrated skill of all: knowing when not to use AI in the first place, because forcing it into the wrong workflows is how the 71% of failed implementations got there.

The companies hiring right now are hiring exactly this profile. Amazon’s job descriptions explicitly require “experience with AI tools for development productivity.” Salesforce’s new grad program is building Agentforce. The training programs at IBM and AWS are heavily agent-orchestration focused. If you can ship an agentic workflow today, you’re in the rehire cohort. If you can only use ChatGPT to write emails, you’re not.

The new equilibrium

The layoff trade is dead, the rehire trade is on, but it’s only on for one specific kind of worker. The next 18 months is a one-time window, and the cohort that becomes AI-native between now and the end of 2027 will own the new equilibrium of knowledge work for the next decade. Everyone else competes for a smaller pool of jobs that weren’t automated, at lower wages, against a much bigger labor pool that’s also been displaced.

The companies that figured this out are already hiring. Amazon, Salesforce, IBM, and ClickUp are running the new playbook openly while most of the industry is still running the dying one. The companies still trying to pump their stock with AI-flavored layoffs are about to learn what GitLab and Meta just learned, which is that Wall Street stopped buying the narrative sometime in Q2 2026 and isn’t coming back.

So where do you sit on the line?

I built ismyjobscrewed.com to answer exactly that question. Score your role against actual displacement data, see your risk percentage, and get a tailored upskilling path so you’re in the AI-native cohort when the rehire wave hits your industry. Takes 90 seconds, costs nothing, and might save your career trajectory in the bifurcation that’s already underway.

A note on certainty. This is forecasting, not prophecy. The future surprises everyone, and macro, algorithmic, or geopolitical shocks could change the timing of all of this. But too many independent data points across token economics, ROI surveys, capex warnings, public CEO statements, and hiring reversals are triangulating in the same direction for the structural pattern to be wrong.

References: https://telegra.ph/Tech-Hiring-Is-Coming-Back-But-Most-Wont-Make-The-Cut-References-05-23