Consider a single week in the summer of 2026, and two men employed by the two companies at the center of the AI economy, saying opposite things.
On July 25, on the Relentless podcast, Sam Altman says: "We are now, like, in the singularity." He adds, "I've been waiting for this my whole life, and I think it's going to be incredible, hugely positive, awesome for the world."
Five days later, The Decoder reports on a position paper by Tom Zahavy, a research scientist at Google DeepMind. Its title is "LLMs can't jump." Its argument is that the systems Altman sells are "structurally incapable" of the one cognitive move—inventing an explanation with no prior term to select from—that any real singularity would require.
Most readers file this as a scientific disagreement in progress: the optimist against the skeptic, the CEO against the researcher, and time will tell who's right.
That reading is a mistake. The two statements are not competing descriptions of reality. They are competing financial instruments, and they clear in different markets.
The two statements are not competing descriptions of reality. They are competing financial instruments, and they clear in different markets.
Last week, this column traced the Hassabis/Doudna timeline and what I called algorithmic denialism—the habit of asserting capability faster than it can be checked. This week the question sharpens: does any checkable milestone or hype accountability actually emerge from the gap between what the labs claim and what their own scientists concede? The answer, once you follow the money, is that the gap is not an accident. It is the product.
Start with the arithmetic, because the rhetoric only makes sense downstream of it.
Ed Zitron, reviewing OpenAI's audited financials for the Financial Times, reports the company lost $20.9 billion in 2025. Not slowing losses—"the situation is still deteriorating."
Now look at where the money circulates. Per UBS analyst Stephen Ju, roughly 27% of Google Cloud's revenue this year reportedly comes from just two customers: OpenAI and Anthropic. Next year (2027) that share is projected to exceed 48%—more than $76 billion, per the same analysis.
And the chain doubles back on itself. Google sells its own TPU chips to those firms via Broadcom, then leases the same chips back to them as cloud service. It profits twice on one transaction.
Zitron has a name for this: circular financing. The giants build capacity for two or three customers, fund those customers, and then book the customers' spending as demand. The demand and the supply are the same dollar wearing two coats.
Microsoft runs the same play. Zitron found that the 69% growth of its Intelligent Cloud division in 2025 came from OpenAI alone; strip that out and the segment grew 8%, "barely outpacing inflation."
Here is why the rhetoric has to escalate. A structure like this cannot survive an honest conversation about this quarter's revenue, because this quarter's revenue is the problem, not the solution.
So the story must move to a horizon where no invoice can reach it.
Read the July 25 quote again with that in mind. "We are now, like, in the singularity" is not a claim about a product you can buy. It is a claim about a destination you must keep funding to reach.
Joseph Alalou charted the pattern and found it clean: at OpenAI, every escalation in existential rhetoric corresponds to a fundraise. The language is doing financial work.
The evidence is in what the round survived. OpenAI's own investor deck, per The Information, pitches something pedestrian—replace Salesforce, Workday, Adobe, Slack, Atlassian. A SaaS displacement story, not a superintelligence thesis.
And the product stumbled. Weekly active users passed 910 million by February 2026, short of the billion the company wanted by the end of 2025. GPT-5's launch was, by Alalou's account, more debacle than triumph.
The $110 billion round closed anyway.
It closed because the bet was never on this quarter's model. The bet is on superintelligence by 2028. As Alalou puts it: you cannot falsify a claim about 2028 with data from 2025. That is not a flaw in the argument. It is the design.
This is the load-bearing sentence of the whole edifice. The implausibility is the feature. A modest, checkable claim—"we will displace Workday by Q3"—can be measured and missed. A grand, unfalsifiable one—"the singularity is here"—cannot be indicted by any single earnings report.
The worse the near-term numbers, in other words, the more the timeline has to stretch, and the more the language has to inflate to cover the distance. Doom and utopia are not descriptions. They are capital.
Now watch the same logic move from the fundraise into the state.
In late July, OpenAI floated giving the United States government a stake in the company—reportedly modeled on the sovereign-shareholder approach used with Intel, where the government took a 9.9% stake (OpenAI's own proposed stake was reportedly 5 percent). At OpenAI's reported $852 billion valuation, that share would reportedly be worth roughly $42.6 billion.
It arrived as generosity: an Alaska Permanent Fund for the AI age, ordinary Americans sharing the gains. Mike Brock reads it more coldly, and correctly.
Note the timing. The offer surfaced less than a week after OpenAI delayed the broad release of GPT-5.6 at the government's request, and weeks after a coalition of 42 state attorneys general opened an investigation into the company.
A company under this much pressure does not volunteer to put its regulator on its cap table out of civic feeling. It does so because a regulator who owns a stake in you has a fiduciary reason to keep you solvent.
Once the state is a shareholder, any restriction that dents your revenue dents a public asset, and any competitor who displaces you threatens an investment "held on behalf of every American." Brock's phrase is exact: this is not a conventional bailout. It is "the architecture of a future bailout assembled in advance."
The singularity claim and the equity offer are the same instrument aimed at two audiences. To investors: the upside is infinite, so no price is irrational. To the government: the downside is systemic, so you cannot afford to let us fall. Both messages require the stakes to be existential. Neither survives a sober accounting.
Which brings us back to the researcher, and to the one honest register in this whole story.
Zahavy's paper is careful in a way the market cannot afford. He borrows Peirce's three modes of reasoning—deduction, induction, abduction—and grants that models already do the first two well; AlphaProof and Gemini reportedly clear gold-medal Olympiad math.
His claim is narrow. What models cannot do, he argues, is manipulative abduction: coining an explanation with no prior term to select from. Einstein inventing general relativity when Newton fit the data and the one anomaly, Mercury's drift, had already been "explained" by a hypothetical planet named Vulcan.
A system trained to minimize error against reality would have had no error to chase, and would have kept patching Newton. The leap came from an embodied thought experiment—a man in a falling elevator—that no text-trained model can run.
That is a serious argument. But notice what Zahavy does not do.
He does not say it can never be done. His own hedge, in the reviewer's reading and in his own text: a training-signal problem is not the same claim as an architectural impossibility. The paper argues the former while sometimes reading like the latter.
His proposed fix—world models grounded in physical simulation, systems like Genie that let an agent cut a cable and see what falls—is a research direction, not a verdict. And DeepMind, his employer, sells those too.
This is the tell. The skeptic and the prophet are not as far apart as their headlines. Altman says the singularity is here; Zahavy says today's models can't jump but tomorrow's architectures might. Between them sits an enormous zone of genuine uncertainty—and that zone is precisely what the market cannot price.
Because sober uncertainty doesn't raise a round. "It might work, on an unknown timeline, if we solve a hard training problem" is the true state of the field, and it is worth zero to a fundraise that needs 2028 to be a certainty and this quarter to be irrelevant.
So the honest sentence—the researcher's hedge—gets buried under the confident one. Not because it's wrong. Because it doesn't clear.
Return, finally, to the accountability question this column has been tracking. Did a checkable milestone emerge this week?
One did, quietly, and it cuts against the hype. The billion-user target for end of 2025 was missed—910 million in February 2026—and the miss changed nothing about the valuation, because the valuation was never indexed to it.
That is the mechanism laid bare. When the near-term goal is real enough to be measured, it is missed and ignored. When the far-term claim is grand enough to matter, it is engineered to be unmeasurable.
The gap between Altman and Zahavy was never a debate waiting to be settled. It is a spread—priced, quoted, and traded—and both ends of it are worth more to their employers the longer it stays open.
Follow the arithmetic, and the singularity stops looking like a prophecy. It looks like an instrument for keeping $20.9 billion in annual losses out of the same sentence as a trillion-dollar valuation.
That is the work the word is doing. It was never a forecast. It was always a hedge.