What's on the Table
$700,000 a day. As of October 27, 2026, that early-2024 figure — what OpenAI was reported to be spending just to keep ChatGPT answering questions — is still the most quoted number in the AI cost debate, and it is still the one most people round off wrong. Annualize it and you get roughly $256 million a year to operate a single product. The same body of reporting puts GPT-4's training bill at over $100 million. Training, in other words — the part everyone calls expensive — costs less than half of what one year of serving the finished model costs.
According to Google News, which surfaced The Washington Post's reporting on AI firms hunting for capital without an obvious source for it, the industry's funding question has moved from how much to from whom. Bloomberg has separately worked through burn rates and runway estimates for the major labs, while The Information has reported on funding negotiations and down-round anxiety among second-tier companies. Those three vantage points do not fully agree, and the disagreement is the most useful data in the story.
Chart: GPT-4's reported training cost of over $100 million, set against the two competing daily operating estimates for ChatGPT annualized — $700,000/day and the higher "over $1 million/day" figure some outlets cite when R&D is folded in.
Note the width of that gap. The spread between the two operating estimates is roughly $110 million a year — more than the entire reported cost of training GPT-4. When the error bar on a company's run-rate exceeds its flagship capital project, outside investors are not underwriting a business model. They are underwriting a guess.
Who's Actually Writing the Checks — and What the Filings Don't Say
The non-obvious part of the funding story is not that tech giants are backstopping the labs. It is how little of the total AI capital stack those famous checks represent.
Start with what is verifiable rather than reported. Microsoft's SEC filings disclose a $13 billion investment commitment to OpenAI through 2024 (SEC EDGAR), and Amazon announced a $4 billion investment in Anthropic in September 2023 via its own press release. Google's commitments to Anthropic and DeepMind run into the billions but remain undisclosed. Collectively, Microsoft, Google and Amazon have put over $25 billion into AI startups — enough to draw antitrust and competition scrutiny.
Do the arithmetic the headlines skip. The two disclosed commitments, $13 billion and $4 billion, total $17 billion — roughly two-thirds of that $25 billion-plus figure. The remaining third is essentially unaudited from the outside. And set the whole $25 billion against the projection that global AI infrastructure spending exceeds $200 billion by 2025: the money flowing into the labs is on the order of one-eighth of the money being spent on the physical layer those labs run on. The capital intensity is not concentrated in the model builders. It is in the concrete, the power, and the silicon.
There is a second thing the filings don't settle. A disclosure quantifies the size of a commitment; it does not tell a reader how much arrives as deployable cash versus other consideration. That distinction decides whether a lab's runway is genuinely independent or structurally tethered to one vendor — and it is exactly the question analysts keep flagging when they warn that massive upfront costs paired with experimental revenue is not sustainable without continued mega-funding from the giants.
The Two-Tier Market: Who Wins Under Which Condition
Put the tiers side by side and the picture sharpens considerably.
Tier one is the frontier lab with a hyperscaler behind it. OpenAI was discussed at an $80–90 billion valuation in late 2023 funding talks despite no clear path to profitability; Anthropic raised over $7 billion across multiple rounds in 2023–2024 from Google, Salesforce and others. These companies win under one specific condition: that their strategic backers continue to value the relationship — distribution, cloud consumption, competitive positioning — more than they value near-term returns on the stake itself. That is a patronage condition, not a market condition.
Tier two is everyone else. Venture capital for AI remains strong in aggregate but is increasingly concentrated, with mega-rounds landing on a handful of names while smaller AI startups struggle to close Series A and B rounds. VCs describe it as winner-take-most; the practical result for founders is a funding desert. Tier two wins only under the opposite condition: that it stops competing on frontier scale entirely. That is precisely what has been happening — several mid-tier AI companies have pivoted to enterprise-focused offerings or been acquired rather than fight for the consumer foundation-model slot.
So the real split is not "big versus small." It is subsidized compute versus purchased compute. A tier-one lab consumes infrastructure partly supplied by its own investor. A tier-two company buys it at list price out of a venture round. On identical technology with identical engineers, the second company's unit economics are worse by construction. The moat compresses when model quality converges — and it widens again the moment compute is priced differently for different buyers.
Where the Bubble Comparison Breaks Down
The easy read is "this is the dot-com bubble with bigger numbers." A careful skeptic should push back on that twice.
First, the revenue is real, even if it lags. The major labs are monetizing through API access, enterprise licenses and consumer subscriptions — three genuinely distinct motions, two of which (enterprise and API) have conventional, forecastable contract structures. The problem described across the reporting is not an absence of revenue; it is that monetization has trailed far behind investment cost for most players. That is a timing problem, and timing problems resolve differently than demand problems.
Second, the better historical analogy is railroads or electrification, not pets.com. In both cases, the infrastructure got built, many of the original operators went bankrupt, and the economy kept the rails and the grid. The capital was not wasted; it was simply not recovered by the people who supplied it. If that is the pattern here, the correct question for a reader's investment portfolio is not "is AI real?" but "which layer of this stack captures the margin after the buildout?" Those are very different bets, and the equity volatility around AI names — a pattern NewsLens Finance traced through the Nasdaq's 2.6% AI sell-off — mostly reflects the market arguing with itself about exactly that question.
Where the bubble framing does hold: capital requirements here exceed anything in software history, which shortens the time a mispriced company can survive on narrative alone.
What to Watch Next
Three indicators matter more than valuation headlines over the next 6–18 months. First, whether disclosed investment commitments from the hyperscalers get renewed rather than merely extended — renewal implies the strategic logic still holds. Second, whether the operating-cost estimates converge; the current $700K-versus-$1M-per-day spread is a transparency failure, and labs that volunteer clean inference economics will price better than those that don't. Third, the Series A and B window for tier-two AI companies, which is the single cleanest read on whether the funding desert is widening.
For anyone treating this as input to financial planning rather than entertainment, the practical move is to separate the three layers — infrastructure, models, applications — and evaluate each on its own cash conversion, because the stock market today increasingly prices them as one trade when they are not one business.
Bottom line: our read is that the training-cost obsession has misdirected the conversation. Inference is the recurring line item, it scales with success rather than against it, and at $700,000 a day it already annualizes to more than twice what the flagship model cost to build. On balance, the more likely outcome is not a sudden collapse but a quiet consolidation — tier-two labs absorbed or repositioned into enterprise niches, frontier capability concentrated among three or four patronage-backed names, and the infrastructure layer keeping most of the margin for several years regardless of which model wins.
Frequently Asked Questions
How do AI companies actually make money right now?
Primarily three ways: API access sold per token to developers, enterprise licenses sold as contracts to large organizations, and consumer subscriptions. As of October 27, 2026, the consistent theme across reporting is that these revenue streams exist and are growing, but monetization has lagged far behind investment costs for most players in the sector.
Why are AI companies losing money if demand is so high?
Because demand is the cost. Unlike traditional software, where serving an additional user is nearly free, each query consumes compute. Reported operating costs of roughly $700,000 per day for ChatGPT as of early 2024 scale with usage, which means a popular product can deepen losses rather than relieve them.
How much does it cost to train a frontier AI model today?
Training costs have escalated from hundreds of millions to billions of dollars per model, with GPT-4 estimated at over $100 million. The figure is a moving target and rarely disclosed by the labs themselves, which is why outside estimates vary so widely.
Will AI companies become profitable, and on what timeline?
No source in the public record provides a credible dated answer, and any claim of one should be treated skeptically. Industry analysts have warned that the current structure — enormous upfront costs against experimental revenue — is not sustainable without continued mega-funding from tech giants. That is a statement about dependency, not a profitability forecast.
Who is funding AI research if venture capital is tightening?
Increasingly, corporate strategic investors rather than traditional VCs. Microsoft disclosed a $13 billion commitment to OpenAI through 2024, Amazon announced $4 billion into Anthropic in September 2023, and Google has invested undisclosed billions in Anthropic and DeepMind. Collectively the three have put over $25 billion into AI startups — which is also why the arrangements are attracting antitrust attention.
Disclaimer: This article is editorial commentary for informational purposes only and does not constitute financial, investment, or legal advice. It reflects analysis of publicly reported information, not independent product testing. Research based on publicly available sources current as of October 27, 2026.