The standard account of mid-2026 is that the AI industry is having its best year and its worst year at the same time, and that one of those readings must be wrong. The bulls point to Anthropic filing to go public at a valuation larger than the GDP of the Netherlands. The bears point to a tracker that has logged 150,000 technology layoffs since January and a Johns Hopkins economist saying the productivity boom never showed up. Each camp treats the other as a failure of nerve or a failure of arithmetic. The standard account leaves out the possibility that both ledgers are accurate, and that the gap between them is the actual story.
The boom is real, and it is enormous
Start with the bull case, because it is built on hard filings rather than vibes. On June 1, 2026, Anthropic submitted a confidential S-1 to the Securities and Exchange Commission, the first formal step toward a public offering. The filing followed a $65 billion funding round that set the company's valuation at roughly $965 billion. To put that number in proportion: Anthropic did not exist as a company eight years ago, and it is now valued within sight of the trillion-dollar club that took Apple and Microsoft four decades each to reach.
The valuation is not pure speculation either. Anthropic told investors its revenue run-rate reached about $47 billion in May 2026, up from roughly $10 billion a year earlier, and projected the annualised figure would cross $50 billion within weeks. Revenue growing more than fourfold in twelve months is not the signature of a company selling a story. It is the signature of a company that cannot provision capacity fast enough.
Which is where the compute numbers come in, and they are the part that should make even a sceptic sit up. To feed demand for its Gemini Enterprise platform, Google agreed to pay SpaceX $920 million every month, from October 2026 through June 2029, for access to roughly 110,000 Nvidia GPUs housed at xAI's Colossus data centre outside Memphis. Read that again. One company is renting another company's graphics chips, which sit inside a third company's building, for nearly a billion dollars a month, and Google is described in the reporting as the largest single owner of AI compute in the world. It is buying this as bridge capacity, a stopgap, because its own build-out cannot keep up.
If you stopped here, you would conclude that the only honest position is to be long on everything. The capability story supports the capital story. Frontier models that could barely close a third of the tasks on the SWE-bench Verified coding benchmark a couple of years ago now resolve nearly all of them. Agentic systems have moved from chat windows into actual workflows: filing legal documents, reconciling payments, writing and shipping production code. The boom is not a mirage. That is the first thing to get right, because the second thing only matters once you have admitted the first.
The backlash is also real, and it is also enormous
Now turn the page. By early summer, independent trackers had counted more than 150,000 job cuts across the technology sector in 2026, with 81,747 of them landing in the first quarter alone, the heaviest three months the industry had seen in years. Amazon accounted for roughly 16,000 corporate roles by itself, even as it reported its fastest cloud growth in over three years. Salesforce cut staff for the third time in nine months while its Agentforce AI product crossed a billion dollars in revenue. The pattern is consistent enough to name: the companies shedding workers are frequently the same companies posting AI-driven gains.
The official explanation is efficiency. The contested explanation is that some of this is what critics call "AI washing": dressing up ordinary cost-cutting, or a correction after pandemic-era over-hiring, in the more flattering language of automation, because "we are becoming an AI-first company" reads better to shareholders than "we hired too many people." Both things are almost certainly happening at once, and the honest position is that nobody can yet cleanly separate the genuine displacement from the convenient cover story.
What is not contested is the mood. A widely-cited survey reported that a striking share of white-collar workers were refusing or quietly resisting employer mandates to adopt AI tools, a phenomenon that picked up the unlovely acronym FOBO, the fear of becoming obsolete. This is the part the capital ledger cannot see. You can lead a knowledge worker to a copilot. You cannot make them trust it, and a tool that half your staff is working around rather than working with does not produce the productivity miracle the valuations are pricing in.
The bear case has a sharp empirical edge, and it belongs to economist Steve Hanke of Johns Hopkins University. His argument is simple: if a general-purpose technology were truly transforming the economy, it would show up in the aggregate productivity statistics, the way electricity and the internet eventually did. So far, at the level of the whole economy, it mostly has not. Spending is vertical. Measured output per worker is not following at the same angle.
There are good reasons a real revolution might lag in the statistics for years (it took decades for electrification to register). But "it will show up later" is a forecast, not a measurement, and the bears are within their rights to keep pointing at the gap until the data closes it.
"Forget the AI bubble. If AI delivered on its promises, productivity would be way up. It hasn't."— Steve Hanke, economist, Johns Hopkins University, 2026
Two ledgers, one machine
Here is the move both camps refuse to make: lay the two ledgers side by side and ask whether they are describing different worlds or the same world from two desks. Set out plainly, the symmetry is hard to miss.
| Dimension | The Capital Ledger | The Labour Ledger |
|---|---|---|
| Headline number | $965B Anthropic valuation; near-trillion-dollar IPO | 150,000+ tech jobs cut in 2026; 81,747 in Q1 |
| What it measures | Expected future value of AI capability | Present cost of getting there, paid by workers |
| Direction | Vertical and accelerating | Vertical and accelerating |
| Driving force | Agentic systems doing real work | Agentic systems doing real work |
| Unresolved by data | Whether revenue justifies the valuation | Whether layoffs are automation or cover story |
Figure 1 — The same engine, booked twice. Capability is the entry on both sides.
Look at the two middle rows. The driving force is identical. The thing inflating Anthropic's valuation and the thing emptying corporate floors at Amazon and Salesforce is one and the same: AI systems that have crossed the threshold from demonstration to deployment. A bubble, in the classic sense, is a price detached from any underlying reality. That is not quite what is happening here. The reality is real. What is uncertain is who captures its value and on what timeline, and that uncertainty is precisely what lets two honest analysts read the same engine and write down opposite conclusions.
The investor sees capability and books it as future revenue. The laid-off analyst sees the same capability and books it as a vanished salary. Neither is hallucinating. They are standing on opposite sides of the same transaction. The boom and the backlash are not competing theories about AI. They are the credit and debit columns of a single ledger, and a single ledger is supposed to have both.
This reframing dissolves the favourite question of the season, "is it a bubble?", into a better one. The valuations could still be wrong. Markets routinely overshoot even when the underlying technology is genuine, and a correction in AI equities is entirely possible without the technology itself failing. The internet was real in 1999 and the dot-com crash was also real in 2000. Both can be true. The thing to watch is not whether the line goes down. It is whether the productivity finally arrives to settle the account.
The number that settles the argument
If the boom and the backlash are two readings of one engine, then there is a single measurement that adjudicates between the bulls and the bears, and it is the one Hanke keeps pointing at: economy-wide productivity. Not the revenue of AI companies, which can grow for years on capital expenditure and enterprise FOMO. Not the layoff count, which can be inflated by cover stories. The clean test is whether the average worker, across the whole economy, produces measurably more per hour because these tools exist. When that number moves decisively, the capital ledger is vindicated. If it stays flat for long enough, the labour ledger was the leading indicator and a reckoning is coming for the valuations.
As of June 2026, that number has not yet delivered its verdict, which is exactly why the argument is so loud. Loud arguments tend to cluster around questions the data has not yet closed. The honest forecaster says the productivity case is plausible and unproven, that the lag could be benign (the way it was for electricity) or terminal (the way it was for a hundred forgotten technologies that genuinely worked but never paid for themselves at scale), and that anyone claiming certainty in either direction is selling something.
There is a human cost folded into that uncertainty, and the two-ledger frame should not be allowed to launder it into an accounting abstraction. The person whose entry-level analyst job disappeared in the first quarter is not a debit entry waiting to be offset by some future productivity credit. The displacement is happening now, in real households, ahead of whatever benefits eventually arrive, and "the aggregate will work out" is cold comfort to someone living through the transition rather than reading about it.
So hold both ledgers open. The boom is genuine. The pain is genuine. The bubble question is the wrong question, because it assumes one of the two columns has to be fake. Watch productivity. It is the line where the two ledgers finally meet, and until it moves, the only intellectually honest position is the uncomfortable one: nobody filing an S-1 and nobody filing for unemployment this summer yet knows which of them was right.
Buy me a coffee