OpenAI spent the second week of July losing its revenue crown to Anthropic, getting sued by Apple over stolen chip designs, and watching a flagship launch delete customers’ files. The market reading is harsher than any single headline: the company is being pinned between a rival taking the profitable top of the market and a Chinese lab destroying the price floor underneath it.
The Week, Priced Out
Start with the scoreboard, because that is what an IPO book gets built on. Anthropic crossed roughly $30 billion in annualized revenue, ahead of the $24 billion to $25 billion pace OpenAI has described, having gone from about $1 billion fifteen months earlier. That is not a rounding error between rivals. That is the lead changing hands in the category OpenAI defined.
Then the legal exposure. Apple sued OpenAI on July 10 over what it alleges was a two-year scheme to take confidential chip designs, and CNBC reported that Apple described the conduct as occurring “at every level”. The specific ask matters more than the rhetoric: Apple is seeking a preliminary injunction that could stop OpenAI’s device program before a single unit ships. A hardware program is one of the few stories that justifies a valuation beyond model licensing. An injunction would put that story on hold with a court’s calendar, not OpenAI’s.
Add the product failures. The July 9 rollout of ChatGPT Work and GPT-5.6 Sol burned through usage limits faster than expected on high reasoning modes, broke existing multi-agent workflows, and produced at least two independent reports of the model deleting user data nobody told it to touch. For a consumer product that is embarrassing. For the enterprise buyer OpenAI needs in order to catch Anthropic, an agent with unrequested delete permissions is a procurement blocker.
One thing on that week’s tally deserves to be pulled out of the pile. Fidji Simo, the company’s second-ranking executive, stepped back from her full-time role on July 9 because of a severe exacerbation of a chronic illness, and she is staying on as an advisor. That is a health matter, not a vote of no confidence, and reading it as a defection would be both wrong and cheap. The business fact that remains is narrower: the role that consolidated product, business, and operations under one leader is now split across three executives, and that reorganization is happening while the company runs its most complicated year.
Anthropic Took the Half That Pays
The revenue flip is not a marketing story, it is a customer-mix story, and the mix is where the margin lives.
Roughly 85% of Anthropic’s revenue comes from enterprise and developer customers, and the company counts more than 1,000 accounts spending over $1 million a year, double the 500 it reported at its Series G in February. OpenAI’s mix runs the other direction, with the large majority tied to ChatGPT consumer subscriptions, and the overwhelming majority of those consumer users paying nothing at all.
Those are two different businesses wearing the same category label. Enterprise seats are contracted, expansion-priced, and sticky, and they carry the gross margin that services an infrastructure bill. A free consumer tier is a customer acquisition cost that shows up as compute. OpenAI is carrying frontier-scale capital expenditure on a base that is mostly not paying, which is how a company posting about $2 billion a month in revenue is still projected to lose around $14 billion this year.
DeepSeek Removes the Floor
Now the other jaw. DeepSeek is raising fresh capital at roughly $74 billion ahead of an onshore listing, and TechCrunch reported in July that a filing could come before year-end. We covered the structural argument on our sister site: that listing is better read as a state-integration event than a fundraise, given that Tencent and Beijing’s National Artificial Intelligence Industry Investment Fund are already on the cap table and export controls have closed the alternatives.
For OpenAI the relevant number is not DeepSeek’s valuation but its price card. DeepSeek’s cheaper tier runs around $0.14 per million input tokens and $0.28 per million output, roughly two orders of magnitude below flagship Western pricing on comparable work. Over the past year the curves diverged, with American flagship input prices climbing while DeepSeek cut and extended context.
And it is landing where it counts. On Vercel’s enterprise gateway in June, DeepSeek handled nearly 23% of tokens processed, against Anthropic’s 32%. A Chinese lab holding close to a quarter of a serious Western developer channel is a distribution fact, not a curiosity.
Here is what makes it structurally nasty rather than merely competitive. DeepSeek’s cost of capital is not set by a market that demands returns. If inference becomes cheap and undifferentiated, Beijing gets a strategic win whether or not DeepSeek ever earns its money back. You do not have to beat the frontier if you can make holding the frontier unprofitable.
The Contradiction the IPO Forces Into Daylight
This is the why, and it is not “OpenAI had bad luck in one week.”
OpenAI’s valuation requires it to be two companies at once: the premium frontier lab that commands pricing power, and the default volume platform that everything gets built on. Those were compatible when there was no serious peer, because being first meant being both. They are not compatible now. Anthropic has taken the premium enterprise position on customer mix, and DeepSeek has made the volume position close to unpriceable. What is left in the middle is a consumer franchise with enormous reach, weak monetization, and the capital requirements of both businesses it is no longer winning.
An IPO is what turns that from a strategy debate into a disclosure problem. Private companies can hold two contradictory narratives indefinitely. A prospectus makes you show the customer mix, the gross margin per segment, the compute commitments, and the actual losses, in one document, to investors who will price the contradiction rather than politely absorb it. That is the real reason the Apple injunction stings, and the reason OpenAI’s regulatory exposure across dozens of state attorneys general has become a financing question rather than a legal one. Prediction markets reportedly cut the odds of a 2026 listing from about 22% to 18.5% after the Apple filing, which is a small move in probability and a large one in what it signals.
What Would Change the Read
None of this makes OpenAI weak in the way a bad week makes a company look weak. It still has the most recognized brand in consumer AI, roughly $2 billion a month coming in, and reach no rival matches.
The question for the next two quarters is whether it can convert reach into contracted enterprise revenue fast enough to matter before the listing window, because that is the only lever that resolves the squeeze from either direction. Watch the enterprise account disclosures rather than the model benchmarks. Benchmarks are what the frontier argument runs on, and the frontier argument is precisely the one that both Anthropic’s customer mix and DeepSeek’s price card have stopped rewarding.