George Hotz published an essay on June 21 titled “the doom justifies the valuation”, arguing that frontier AI companies cannot justify current prices with actual technical progress, so the doom narrative does the work instead. Anthropic is named specifically. The observation is more useful as an accounting argument than as a prediction: what terminal-value assumption actually closes the gap between current AI cash flows and current AI prices, and who among the parties capable of answering that question is financially positioned to do so honestly?
What the Numbers Actually Require
No consumer SaaS model closes the gap to $500 billion on negative margins. OpenAI’s October 2025 share sale valued the company at $500 billion. That same year, OpenAI reported $13.1 billion in revenue and a net income of negative $9 billion: a price-to-revenue multiple above 38x from a business losing approximately 70 cents for every dollar of revenue it earns.
Mature enterprise software companies trade at 5-15x revenue. Hyper-growth SaaS with visible margin expansion might reach 20-30x. At 38x with deeply negative margins, terminal value is carrying nearly the entire weight of the valuation in any plausible DCF model. That terminal value must justify not just today’s $9B annual loss but every accumulated loss on the path to profitability, plus a return premium for the risk of holding equity in a company that has not yet demonstrated it can generate positive net income.
ChatGPT’s current product lineup sells access to GPT-5.5, positioned as “our most capable and efficient frontier model for professional work”, through a tiered subscription from free to Pro. That is a recognizable software business model. It produces recognizable software margins: negative, at current scale, with no visible near-term path to the kind of margin profile that would justify an earnings multiple even vaguely consistent with $500 billion.
Enterprise contracts and API revenue add to the numerator. They do not change the structural problem. The gap between plausible near-term cash flows and the current price is too wide for incremental improvements to bridge. The only assumption that closes it is a qualitative step-change in the economic value of AI output: a system capable of substituting for human labor broadly enough that the current loss period becomes a rounding error in retrospect. Call it AGI, or AGI-adjacent, or whatever label survives the marketing review. The math requires whatever that is.
Why OpenAI’s Financial Structure Complicates the Picture
The circular financial relationship between OpenAI and Microsoft makes the standalone $500B figure harder to evaluate than it appears. Microsoft holds approximately 27% of OpenAI Group PBC, a stake valued at roughly $135 billion. Under the same arrangement, OpenAI committed to purchasing $250 billion of Azure services over the deal’s term. A substantial fraction of OpenAI’s compute costs flows to Microsoft; a substantial fraction of Microsoft’s AI revenue flows back from OpenAI’s user base. The largest investor and the largest cloud vendor are the same company, which makes standalone enterprise value analysis genuinely difficult.
Bloomberg reported in May 2026 that OpenAI was preparing a confidential IPO filing targeting a fall 2026 debut. An S-1 forces a specificity that private funding rounds do not require. Risk factors, revenue concentration, and the assumptions behind projected terminal value get disclosed in a document that lawyers and accountants sign their names to. The current private-round framing, an agreed-upon figure among consenting parties with shared interests, gets replaced by a public document that analysts, short sellers, and institutional allocators who passed on the private round can read and contest.
What goes in the terminal-value section? Consumer subscription growth could plausibly take OpenAI to $25-30 billion in annual revenue within five years with improving margins. That does not close a $500B valuation at standard discount rates without applying a multiple that implies either dominant market share won against well-resourced competitors running the same race, or a step-change in what the models can do. The former is a competitive prediction. The latter is the AGI assumption under a different label.
How the Safety Narrative and the DCF Story Converged
The safety narrative and the DCF story converge on the same terminal-value assumption: AGI is real, near-term, and of sufficient magnitude to dwarf current revenue in the terminal period. OpenAI’s founding charter defined the company’s mission as ensuring AGI “benefits all of humanity”. The framing was safety-first: the company exists to navigate the development of powerful AI responsibly and broadly. It became, concurrently, the investor pitch.
If AGI is the mission, AGI is what the company is building toward. If the company is building toward AGI, the safety spending is not a cost center but a prerequisite for the revenue event. The same terminal-value claim that makes the existential risk argument serious makes the valuation defensible. Safety researchers and late-stage investors are making the same underlying claim about magnitude and proximity; they describe the conclusion in different vocabularies for different audiences, but they require the same prior.
Hotz’s sharpest observation is about what happens in the absence of that prior. He contrasts Anthropic’s doom-oriented communications with what he describes as honest technical posts from companies like GLM and comma.ai: incremental improvement claims, documented benchmarks, no recursive-self-improvement narratives. His characterization of the resulting Bay Area culture as a “cult” and “mind virus” is more polemical than analytically precise, but the structural observation holds: if AGI is neither real nor near-term, neither the safety spending nor the valuation makes sense. Both arguments fail identically, from the same failed premise.
OpenAI’s founding charter is relevant here not as irony but as a mechanism. The company’s stated mission is not to build good products or generate returns; it is to ensure the benefits of AGI are distributed broadly. That statement commits the company’s purpose to an assumption about AGI’s arrival. The investor pitch followed from the mission statement rather than being retrofitted onto it, which is why the two narratives are so structurally unified: they were always describing the same thing.
Why No Incumbent Will Revise the Timeline Publicly
The correction problem is structurally asymmetric: the first mover takes the full loss, and no one else is required to follow.
If any major frontier lab published a material AGI-timeline revision, say, from “5-10 years” to “20-30 years”, the valuation consequences would be immediate. Discount rates alone would cut the present value of a 25-year terminal event by 80% or more relative to a 5-year model. A company currently priced at $500B on a 5-year AGI assumption would be repriced, on the longer timeline, toward a multiple consistent with a consumer software business posting negative margins. The repricing would not be gradual.
No board, no management team, and no current shareholder has a financial incentive to trigger that repricing. The rational strategy for every incumbent is to maintain current framing, because deviation punishes the deviating party without requiring any competitor to match. The first lab to say “our timelines were too aggressive” gets repriced immediately; others face reputational pressure but can wait out the market reaction before committing to any similar revision.
Hotz frames the eventual unwinding as consequence-free, like previous failed predictions that quietly expired without accountability. That may describe the private-market endgame. The IPO changes the mechanics. Institutional allocators who bypassed the $500B private round can run their own models on a public S-1. Short sellers can structure positions. Analysts can publish price targets with explicit AGI-timeline assumptions laid out for clients to evaluate. The belief that private markets agreed not to examine becomes publicly testable for the first time.
What a Correction Actually Looks Like
Corrections of this type rarely come from honest acknowledgment. They come from events that force updates despite the incentive structure.
The most plausible external mechanisms: a benchmark plateau that persists across multiple model generations; a high-profile enterprise product failure that makes current limitations viscerally apparent to buyers; an IPO priced below the October 2025 private-round valuation. Each works differently. Benchmark plateaus are slow and deniable until they aren’t. Enterprise failures are visible but affect one company’s narrative rather than the structural assumption underneath all of them. A discounted IPO is the most immediately propagating: it sets a reference valuation that flows through every AI company that used OpenAI as a comparable in its most recent fundraise.
The IPO mechanism is structurally different from the others because it is already scheduled. An offering priced at, say, $300 billion instead of $500 billion would not declare AGI false. It would declare that public-market investors applying standard discount rates cannot sustain the same terminal-value bet that late-stage private investors made in 2025. That is a localized statement about pricing discipline that becomes a sector-wide statement about comparables.
Hotz’s essay frames the current situation as a belief held in common by parties who all benefit from holding it. That is the stable configuration: it requires no coordination, no dishonesty, and no one to be wrong about AGI’s eventual arrival. It requires only that the parties capable of examining the assumption are all better off not examining it. The S-1 process is one of the few events that forces the question into the open regardless of whether the incumbents want it there.
Frequently Asked Questions
Does Anthropic’s public benefit corporation structure separate it from the same valuation logic?
No. The PBC designation affects governance and liability exposure, not the DCF math. Anthropic raised at a valuation estimated above $60 billion through 2025 with comparable negative margins and a safety-first AGI framing. Investors in a PBC still require a terminal-value assumption that justifies entry price; the governance wrapper does not change what that assumption must be. Hotz names Anthropic explicitly in the essay, treating the PBC structure as no logical firewall against the same argument.
What would the S-1 risk-factor section need to say to satisfy securities law on the AGI assumption?
Under SEC guidance, a material business assumption must appear as a risk factor if its failure would impair the stated business model. If OpenAI’s terminal value depends on AGI-scale capability gains, the S-1 would likely require language along the lines of: ‘our long-term revenue model assumes continued advancement in model capability that may not occur on the timelines we project.’ That sentence, once signed by attorneys and auditors, converts the AGI timeline from a marketing claim into a liability-bearing public statement that analysts and short-sellers can stress-test against reported model benchmarks each quarter.
How does the $250 billion Azure commitment constrain OpenAI’s path to positive margins?
The $250 billion commitment flows from OpenAI to Microsoft, locking a large share of OpenAI’s future compute costs to its largest equity holder. This creates a structural ceiling on margin improvement that standard enterprise SaaS companies do not face: OpenAI cannot renegotiate compute pricing downward without straining the relationship with the party whose $135 billion stake underpins the current valuation. Margin expansion stories typically require the ability to play vendors against each other; that lever is largely unavailable here.
Which AI companies face direct valuation pressure if OpenAI prices its IPO below the October 2025 private round?
Any company that used the October 2025 OpenAI round as a comparable in its own fundraising faces the same pressure. Anthropic, xAI, and Mistral all raised at valuations implying the same frontier-AI terminal-value framework. A public OpenAI re-price does not force automatic private markdowns, but it shifts the burden of proof: fund LPs reviewing portfolio marks and any company planning a 2026-2027 raise would need to explain why their AGI-timeline assumption survives a public market that declined to sustain OpenAI’s.
Has any technology company previously justified a comparable multiple on a future capability claim, and what does the precedent suggest?
Amazon traded at extreme price-to-revenue multiples in the late 1990s on the argument that logistics infrastructure would generate disproportionate long-run returns, and it took over a decade of margin expansion to vindicate that bet at the operating level. The key structural difference is that Amazon’s moat was physical infrastructure requiring years of capital to replicate. An AGI capability leap, if it arrives, would be software-native and far more rapidly copyable by any lab running equivalent training infrastructure, which makes the winner-take-all assumption embedded in a 38x multiple harder to defend on competitive grounds.
