August 13, 2026
Two stories dominated AI today, and they are really the same story wearing different clothes. Anthropic told investors it expects to turn its first operating profit. OpenAI edged closer to opening its books to the public. Both are attempts to answer the question that has hung over this industry since the first billion-dollar training run: can any of this actually make money?
The answers so far are interesting, incomplete, and — in at least one case — contested.
Anthropic’s $559 million question
The headline is genuinely striking. According to financial projections reviewed by The Wall Street Journal and shared with investors as part of an ongoing funding round, Anthropic expects Q2 2026 revenue of $10.9 billion, up roughly 130% from Q1’s $4.8 billion, and an operating profit of about $559 million. That would be the company’s first profitable quarter.
The mechanism is cost discipline rather than pure growth. Anthropic’s compute spending fell from about 71 cents per dollar of revenue in Q1 to a projected 56 cents in Q2. When your single largest input cost drops fifteen points as a share of revenue while revenue itself more than doubles, the arithmetic takes care of the rest.
It is worth pausing on how fast this moved. As recently as summer 2025, Anthropic’s own guidance to investors suggested no full-year profitability until 2028 at the earliest. Pulling that in by two years is not a rounding error; it is a different company.
Now the caveats, which matter more than usual here.
These are projections, not results. Anthropic is a private company. It does not file with the SEC, it is not bound by public-company reporting standards, and the Journal itself noted it is unclear what accounting methods produced these numbers. What we have are figures a company circulated to prospective investors while raising money. That is not a category of document known for pessimism.
The definition of “operating profit” is doing work. The $559 million figure includes model training costs but excludes stock-based compensation. Excluding SBC is common in private-company reporting and is also one of the most reliable ways to make an AI company look better than it is, given how much of the compensation at these firms is equity.
The quarter may have been unusually kind. Anthropic has reportedly cautioned that Q2 benefited from favorable compute pricing that will not persist, and has not committed to profitability in Q3 or Q4, citing planned infrastructure spending.
That last point is where the sharpest criticism lands. Ed Zitron, writing at Where’s Your Ed At, argues the profitable quarter was effectively manufactured. His case: Anthropic pays SpaceX in the neighborhood of $1.25 billion a month for compute, but with reduced fees during the ramp-up period — a discount that happened to fall in May and June, precisely the months underpinning the profitable quarter. Strip out the temporary rate and the picture changes. He also points to inference costs running about 23% above expectations, and to the fact that compute costs scale more or less linearly with usage, meaning there is no obvious point at which the economics get structurally better rather than incrementally cheaper.
Zitron additionally flags revenue figures that are hard to reconcile: CFO Krishna Rao stated under oath in March that revenues exceeded “$5 billion to date,” while the company was simultaneously describing $19 billion in annual recurring revenue. ARR and recognized revenue are different measures and the gap is not automatically damning — but “annualize the best month and call it revenue” is a very familiar move, and it is fair to ask which number is load-bearing.
My read: the underlying trend is real and the specific quarter is soft. Compute costs per unit of output genuinely have been falling across the industry, and 1,000-plus enterprise accounts spending seven figures a year is a real business that did not exist two years ago. But a one-quarter, non-GAAP, projected operating profit that excludes stock comp and rides a temporary vendor discount is not the same as a company that has crossed into sustainable profitability. It is evidence that the destination exists, not proof of arrival.
OpenAI’s paperwork problem
The other half of today’s news is OpenAI’s march toward the public markets, which is generating more speculation than fact.
Here is what appears solid, per TECHi’s running tracker of the filing: OpenAI submitted a confidential draft S-1 to the SEC on June 8, 2026. Its most recent private valuation, from a March 2026 financing, was $852 billion post-money. Reported revenue is running around $2 billion a month. Leaked 2025 financials show $13.07 billion in revenue against a $20.92 billion operating loss.
Here is what is not solid: essentially everything about timing and valuation. Some reports have the public prospectus hitting EDGAR in mid-to-late August, roughly fifteen days ahead of a roadshow, with a September listing. Others say no date has been set and 2027 remains live. Valuation targets in circulation range from the $852 billion private mark to a trillion-dollar aspiration. When the spread on a rumored valuation is $150 billion, the rumor is not carrying much information.
What will actually matter is the disclosure, not the date. A public S-1 would force OpenAI to answer questions it has spent years declining to answer:
- The Microsoft relationship. Revenue-share terms run through 2030 and remain undisclosed. For a company whose largest distribution partner is also its largest shareholder-adjacent entity, this is not a footnote.
- Compute obligations. Long-term cloud and infrastructure commitments are the single biggest swing factor in whether OpenAI’s losses converge or diverge. Right now they are invisible from the outside.
- The capital structure. Amazon has roughly $35 billion remaining in committed capital and SoftBank $64.6 billion total, with conversion mechanics tied to listing timing. Untangling who owns what after an IPO is genuinely difficult.
- Governance. The OpenAI Foundation retains board appointment rights and other special powers. Public shareholders would be buying into a structure where they do not control the company — an arrangement investors tolerate at Meta and Alphabet, but those companies were profitable.
At $852 billion against 2025 revenue, you are paying roughly 65 times sales for a business that lost $21 billion. That is a bet on a specific future, not a valuation of a present.
The rest of the day
Model releases. SpaceXAI shipped Grok 4.6, which it says matches GPT-5.6 Sol on benchmarks, priced at $2 per million input tokens and $6 per million output. Aimed squarely at agentic and coding workloads — the segment where price per token actually moves purchasing decisions.
Consolidation. Anthropic is reportedly in talks to acquire Israeli startup Decart AI for around $6 billion, for real-time generative video and GPU optimization work. PYMNTS frames the deal as a cost play, which — given everything above about compute economics — is the tell. If your profitability depends on driving cost per token down, buying the team that does GPU optimization is a more direct route than waiting for Nvidia.
Valuations, still climbing. Cognition is in talks at $40 billion-plus on roughly $1 billion in annualized revenue from its Devin coding agent — a company that raised at $10 billion less than a year ago. Swedish legal AI firm Legora is seeking funding above $10 billion, up from $5.6 billion four months ago, on about $150 million ARR. OpenAI-backed Thrive Holdings raised $2 billion at a $12 billion valuation to push AI into traditional service industries.
Infrastructure. Vantage Data Centers is exploring an IPO at around a $100 billion valuation, or an outright sale. In India, L&T won a $1.57 billion contract to deploy roughly 10,000 Nvidia B300 GPUs in Chennai for Together AI — the country’s largest single-cluster installation. Lenovo posted $26.9 billion in quarterly revenue, up 43%, with AI products and services up 60% and now about 35% of the total. The picks-and-shovels trade remains undefeated.
Governance. Several outlets recirculated Demis Hassabis’s call for an independent AI oversight body today, some describing it as an IAEA-style regulator. That is not quite what he said. In remarks on July 14, the DeepMind chief proposed a US-led, industry-funded watchdog that would evaluate frontier models before deployment — explicitly modeled on FINRA, the private body that polices Wall Street under SEC oversight, not on the IAEA. The distinction matters: FINRA is self-regulation with a government backstop, which is a considerably lighter proposal than an international inspections regime.
What today actually told us
Strip away the numbers and the pattern is clear enough. The AI industry is transitioning from a phase where capital was raised on narrative to one where it must be raised on disclosure. Anthropic’s projections and OpenAI’s S-1 are both, in different ways, the industry submitting to arithmetic.
That transition is healthy and it is going to be uncomfortable. Private projections get to exclude stock comp and lean on favorable vendor timing. A public prospectus does not. Whatever OpenAI’s S-1 eventually says about compute obligations and Microsoft revenue share will tell us more about this industry’s economics than any number of leaked investor decks.
The interesting question is not whether AI companies can produce a profitable quarter. Today suggests at least one can. The question is what happens when the discounts expire, the infrastructure bills come due, and someone has to sign the filing.
Sources
- Anthropic On Track for First Operating Profit as Revenue Surges — PYMNTS, on the WSJ-reviewed projections
- Anthropic projects first operating profit of $559M in Q2 after 130% revenue jump — Crypto Briefing
- Anthropic’s “Profitability” Swindle — Ed Zitron, Where’s Your Ed At
- OpenAI IPO: S-1 Status, Valuation, Timeline, and Risks — TECHi
- Anthropic Said in Talks to Buy AI Startup Decart for $6 Billion — Bloomberg
- Anthropic Pursues $6 Billion Decart Deal to Cut AI Costs — PYMNTS
- AI coding startup Cognition reportedly already in talks to raise at $40B valuation — TechCrunch
- Vantage Data Centers explores IPO at US$100 billion valuation or sale — Reuters via BNN Bloomberg
- Demis Hassabis calls for U.S.-led AI standards body like FINRA — Quartz
- Top Tech News Today, August 13, 2026 — Tech Startups, for the Grok 4.6, Legora, Thrive, Lenovo and L&T items
Figures in this post come from reporting on private company projections and leaked financials, not audited public filings. Nothing here is investment advice.