The people warning that frontier AI economics do not add up used to be short sellers and provocateurs. Now they are operators, auditors, and the labs’ own regulatory filings.
Inside a single week, Palantir CEO Alex Karp called the token business model insane on live television, leaked audited financials put OpenAI’s 2025 operating loss near $21 billion, and Oracle warned the SEC that its AI datacenter buildout could unravel on customer nonpayment.
Skeptic count matters because venture capital has priced these labs at unprecedented multiples. OpenAI closed a $122 billion round in March 2026 at an $852 billion valuation, according to leaked figures verified by the Financial Times. Anthropic has raised at comparable altitude. The bet underwriting those marks is that durable value in AI accrues to the model layer. Ed Zitron has argued the sector is closer to a modest market dressed up as a trillion-dollar one. The growing pile of credible critics is a direct challenge to that thesis.
Starting with the audited numbers; documents obtained by Ed Zitron and confirmed by the Financial Times show OpenAI posting a $20.9 billion operating loss on $13.07 billion of revenue in 2025, against total costs near $34 billion. The reported net loss widened to $38.5 billion after a one-time charge tied to the company’s for-profit conversion. Those figures landed days after OpenAI made a confidential filing toward an IPO, which means public-market investors will soon price a company that spent roughly $1.60 for every dollar it earned.
Karp’s critique runs on a parallel track. On CNBC’s Squawk Box on July 1, the Palantir chief said frontier models had been completely oversold and that enterprise customers were paying for tokens that generate no value while handing over their alpha, the competitive edge a business holds in its market, to third-party providers. He called the prospect of routing national-security software through Silicon Valley’s default models effing insane. Karp is talking his book, since Palantir sells the sovereignty layer he is pitching, yet enterprises have already begun to rein in token spending after a year of buying as much AI as possible.
The intellectual-property worry is not hypothetical. When Anthropic launched Claude Design in April, a tool that competes with Figma, Figma shares fell 7% the same day, three days after Anthropic’s chief product officer resigned from Figma’s board. Founder Dylan Field later said the lab had not been consistently candid in its communications. Activist investor Findell Capital has since mounted an activist push urging Figma to reassess the partnership, turning a product dispute into a governance question about feeding a rival deep product insight.
Then comes systemic risk; Oracle signed a $300 billion deal to supply OpenAI’s compute and, in its latest annual filing, flagged exposure to customer nonpayment risk and non-renewal. Zitron’s analysis estimates Oracle is building roughly 7.1 gigawatts of Stargate capacity for one dominant tenant, OpenAI, against a construction bill he pegs near $340 billion, money OpenAI cannot yet generate from operations. The flagship Stargate venture, unveiled at the White House in January 2025, had hired no staff and stalled over financing more than a year later.
The capital intensity now reaches the strongest balance sheets. Alphabet moved to raise $80 billion through stock sales in June to fund its AI buildout, on top of tens of billions in recent bond issuance. Combined hyperscaler capex is on track to exceed $700 billion this year, per Wall Street estimates. When the most profitable hyperscaler taps both equity and debt to keep pace, the claim that AI capex is outrunning cash flow becomes harder to wave away.
Bulls hold real counterpoints. Anthropic reported an annual revenue run rate near $47 billion and an operating profit in the second quarter of 2026, though Zitron called engineered the way that profit was reached. Nvidia booked $81.6 billion in a single quarter, proof the compute layer is minting genuine cash. OpenAI’s expense ratio improved from the prior year. None of that settles the question the critics keep circling, which is; whether the model layer is a durable business or a subsidized commodity heading into a price war over token pricing between OpenAI and Anthropic.
What makes this a hot topic among VC is that there is meaningful change is in who is doing the criticizing.
A short seller is easy to dismiss. An operator, an auditor, and a company’s own risk disclosures are significantly more serious acusations. The near-term test is OpenAI’s IPO, where audited figures replace narrative and the market sets a price on losses that can no longer hide behind legacy monopolies. For founders building on frontier APIs, Karp’s warning reads as a concrete instruction to control their data, their weights, and their alpha before their edge becomes someone else’s training set. The bubble does not have to pop for the thesis to break. It needs only enough credible people, holding receipts, to keep showing up.
