Cheap AI models threaten OpenAI, Anthropic trillion-dollar IPOs

The most valuable AI companies in the world are about to ask public-market investors to buy a thesis that their own biggest customers have already started to reject: that premium AI models command premium prices, and that those premiums will keep growing.

OpenAI is expected to confidentially file for an initial public offering as early as Friday, laying the groundwork for a listing that could value the company near $1 trillion. Anthropic, its closest rival, is in the middle of a funding round targeting a $600 billion to $900 billion valuation — up from $380 billion in February — while projecting $10.9 billion in second-quarter revenue and its first operating profit of $559 million. Cerebras, the AI chipmaker, debuted May 14 and surged 70 per cent on its first day, handing the company a market capitalisation of roughly $95 billion. And SpaceX, whose IPO prospectus landed this week, revealed that Anthropic alone is paying it $1.25 billion a month — $15 billion a year — for compute access through May 2029.

It’s very hard to care about anything other than the $3 trillion potential IPOs that, in theory, are going to happen in the next year.
— Sam Lessin, Partner at Slow Ventures

The pipeline is easily the most concentrated run of mega-cap tech listings in history.

But the assumptions that underpin those valuations are colliding with a force that every infrastructure boom eventually meets: commoditisation.

The price floor is falling

DeepSeek, the Chinese AI lab, charges $1.74 per million input tokens for its V4-Pro model. Anthropic’s Claude costs roughly $25 per million. OpenAI charges $30. The spread is not an edge case — it is 15 to 17 times cheaper, and enterprise chief financial officers are doing the arithmetic.

The gap widens on real workloads. An evaluation that costs $4,811 using Claude runs on the cheapest Chinese alternative for $544, according to benchmark data from Artificial Analysis, cited in a CNBC analysis published May 20. Nine times less.

Over the past twelve months, Chinese AI models have gone from roughly 1 per cent of usage on OpenRouter, a routing platform that connects developers to dozens of models, to more than 60 per cent in May 2026. OpenAI’s share of the global AI market, meanwhile, has shed 15 percentage points — from 55 per cent to roughly 40 per cent — in the same period.

That is the path to an AI industry that looks a lot more like a commodity business than a premium software monopoly.

The adviser model

Ali Ghodsi, the chief executive of data platform company Databricks, has a name for what is happening inside enterprise procurement departments. He calls it the “adviser model” — route 80 per cent of AI workloads to the cheapest capable open-source model, and reserve the expensive frontier models for the hardest five or ten per cent of tasks.

You can curb costs really well this way.
— Ali Ghodsi, CEO of Databricks

The numbers bear him out. Sixty-nine per cent of enterprise AI workloads are already routed to mid-tier models rather than frontier ones. A separate analysis by Beri of 2.4 billion API calls across 8,000 enterprises found that multi-model routing drove a 67 per cent year-over-year drop in per-request costs. CloudZero, a cost-management platform, reported that 45 per cent of companies are now spending more than $100,000 a month on AI, up from 20 per cent a year earlier — and those companies are the ones who will push hardest for cheaper alternatives.

Sundar Pichai, Google’s chief executive, acknowledged the dynamic from the other side of the table this month, telling investors that “many companies are already blowing through their annual token budgets, and it’s only May.” Google has responded by positioning its cheaper Flash models as the solution — implicitly conceding the price-sensitivity problem.

The adviser model is, in economic terms, the structural compression of premium AI revenue. Every dollar an enterprise saves by routing work to DeepSeek or Qwen or Mistral is a dollar that OpenAI and Anthropic built into their revenue projections.

The trust moat — and its limits

The defence from the labs themselves has two pillars. First, that regulated industries — banks, hospitals, defence contractors — cannot and will not use Chinese models regardless of price, creating a defensible premium segment. Second, that the product stickiness of platforms such as ChatGPT, which claims over 800 million users, creates a switching-cost moat that pure pricing cannot breach.

Cohere, a Toronto-based AI company that sells exclusively into regulated-enterprise settings, offers the strongest evidence for pillar one. Chief executive Aidan Gomez told CNBC that his revenue has grown sixfold, and that “you can’t use a Chinese model for any of these regulated customers.” Cohere’s trajectory suggests a real premium segment exists.

But the question is whether that segment — financial services, healthcare, government — is large enough to sustain valuations approaching or exceeding $1 trillion. And Anthropic itself has clouded the argument. A policy paper the company published this spring acknowledged that Beijing is “winning on adoption on cost” and that the capability gap between U.S. and Chinese frontier models is “only several months.” If the labs’ own analysis says Chinese models are close enough in capability and winning on cost, the trust moat looks thinner than the IPO pitch decks suggest.

OpenAI, for its part, has said pricing pressure “isn’t on the top ten list of concerns.” That level of confidence reads differently when the company is projecting $14 billion in losses for 2026 — a number that implies the current revenue trajectory, even at premium pricing, is nowhere near breakeven.

The margin math

Here are the numbers that make the IPO narratives hard to close.

OpenAI is burning through cash at a rate that would require capturing essentially the entire global enterprise AI software budget for several years to reach breakeven at current premium margins. DeepSeek’s training costs are estimated at $5 million to $10 million per model — against $500 million to $1 billion for each Western frontier model — meaning the cost structure of the competition is fundamentally different.

Anthropic, for all its revenue growth, is locked into a compute agreement with SpaceX that costs $15 billion a year through May 2029. That is a fixed cost roughly equal to one and a half times its projected Q2 annualised revenue. No frontier AI lab — not OpenAI, not Anthropic, not xAI — is profitable.

The $582 billion global AI capital-expenditure cycle is building infrastructure for a product whose unit price is collapsing. Every prior infrastructure boom — railways, fibre optics, cloud computing — followed the same arc: capacity glut, price compression, valuation reset. The cycle is playing out at speed in AI because the open-weight model ecosystem ships new competitive releases every four to six weeks.

If you compare xAI to a traditional SaaS company, the financials look reckless.
— Harrison Rolfes, Analyst at PitchBook

The same sentence, applied to the $1 trillion OpenAI narrative, is the question the S-1 will force public-market investors to answer.

What the S-1 tests

When OpenAI’s confidential filing becomes public — which, under SEC rules, happens at least 15 days before the roadshow begins — investors will see for the first time the company’s unit economics, customer concentration, and gross margins under the pricing pressure that has already reshaped the market.

The crux is straightforward. If DeepSeek V4-Pro scores 80.6 per cent on the SWE-bench Verified coding benchmark — within 0.2 percentage points of Western frontier models — at one-seventh the price, the premium that any lab can charge is bounded by the capability gap. That gap is measured in months, not years. And the market share numbers — 60 per cent Chinese on OpenRouter, 15 points of share lost by OpenAI — suggest enterprise buyers are already making their choice.

The three-way collision of OpenAI, Anthropic and SpaceX racing to public markets simultaneously — the densest mega-cap tech IPO pipeline ever — means that every assumption in every S-1 will be scrutinised against the others. If OpenAI needs monopoly-grade margins to justify $1 trillion, and the market structure already looks like a commodity, the pricing is the question the S-1 will answer.

Deirdre Bosa and her CNBC colleagues contributed reporting to the analysis this piece draws on.