For two years the AI business ran on private valuations and founder assertions. This week the actual numbers start walking into daylight — a draft S-1, a revenue comparison nobody could run before, and a funding round that tells you where the smart money thinks the margins live. Yesterday the fight was over ideas and charters. Today it's over the income statement. Here's July 12.
OpenAI is lining up an IPO, which means the economics finally go on the record
OpenAI is reportedly preparing a confidential IPO filing with Goldman Sachs and Morgan Stanley, targeting a public debut as soon as September 2026 at a private-market valuation around $730 billion (Build Fast with AI). Going public means the company would have to show audited financials for the first time — real numbers, not decks (unrot).
The operator's take: an S-1 is the most honest document a hype cycle ever produces, because the lawyers make you write down what's actually true. Whatever the valuation, the useful signal for you is that the unit economics of frontier AI — cost of compute, gross margin per token, how much of "revenue" is discounted enterprise pilots — are about to become legible. If you're building a budget around a vendor whose finances you've never seen, the day that filing drops is the day you get to price your dependency on reality instead of narrative. Read it before your renewal.
The revenue crown may already have changed hands
Per reporting from Fortune, Anthropic has pulled ahead of OpenAI on revenue — roughly $47 billion annualized for Anthropic against a projected $25 to $33 billion for OpenAI in 2026 — with coding tools doing the heavy lifting (unrot). The standout line: Claude Code reportedly grew from about $1 billion to over $2.5 billion in annualized revenue in roughly two months earlier this year (unrot).
The operator's take: the loudest model and the most-monetized model are not always the same thing, and that gap is the whole ballgame for a buyer. Coding assistance is where AI stopped being a demo and became a line item people renew, because the value is measurable in shipped work. When you're choosing a vendor to standardize on, weight "is this thing actually generating durable revenue from customers like me" over leaderboard position. Revenue is retention with a dollar sign in front of it.
A billion dollars just bet against the GPU-cluster default
The inference layer is getting funded like it's the next fight. SambaNova closed a $1 billion Series F at an $11 billion post-money valuation led by General Atlantic, and it follows Together AI's $800 million Series C earlier this month (Build Fast with AI). The thesis behind the checks: inference-focused challengers can peel high-volume workloads away from both Nvidia GPUs and rented mega-clusters (Build Fast with AI).
The operator's take: for most companies, training is somebody else's cost and inference is your cost — the recurring bill that scales with every user you add. A billion dollars chasing cheaper, faster inference is a competitive tailwind for the line item you actually pay. You don't need to move workloads today, but you should stop treating "we run it on the default cluster" as a permanent decision. Keep your inference layer swappable, because the market is about to give you options that make your current bill look lazy.
Also on my radar
- The API is quietly going multipolar. Chinese models now reportedly account for 30 to 46 percent of US enterprise API traffic, up from about 4.5 percent a year ago (unrot, citing CNBC). If your data-governance policy still assumes every model call lands on a US provider, that assumption is out of date — audit where your prompts actually go.
- A VC gets a seat near monetary policy. A newly announced Federal Reserve body studying AI's effect on jobs and productivity is reportedly co-led by a16z's Marc Andreessen, and the appointment is already contested given his AI investments (Build Fast with AI). Watch who writes the rules you'll be operating under.
- Next model wave has a date. Google's Gemini 3.5 Pro is reportedly slated for July 17 (unrot). Launch-week benchmark claims deserve skepticism, but a real release is a real reason to re-run your own evals.
The throughline: after years of arguing about capability, the AI story is becoming a balance-sheet story — who can file, who can bill, and who can serve a request cheaply. Benchmarks tell you what a model can do; revenue and unit economics tell you whether the company behind it will still be standing when your contract renews. As a buyer, that second question is the one that pays your salary.
That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.