Skate to Where the Puck Is Going

The Puck Left the Model

Monday, Lindy quietly fired Anthropic. Switched all of its traffic to an open model, DeepSeek v4, saved millions of dollars, and — the part that should keep a frontier lab up at night — got better answers. Harvey, the legal-AI darling, ran the same test: an open model beat Claude Opus on Harvey's own benchmark and cost $84 instead of $954 for the same hundred tasks. Cursor stopped renting altogether and trained its own. Tomasz Tunguz wrote it up this morning and called it the Substitution Wave.

Tunguz lists three forces driving it: labs are moving up the stack into apps, frontier prices keep climbing for the smartest models, and open source crossed the "good enough" line. All true. But he lays them side by side like three weather systems that happened to arrive together. They're not parallel. They're a chain. Open source got good enough, so paying frontier rates for the last 5% of IQ stopped penciling out on the 80% of work that never needed it — and once the model layer has no margin left in it, the only direction is up. Cause, cause, effect. That's not three trends. That's one trapdoor.

Now look at who fell through it. OpenAI filed for its IPO today and announced it's turning ChatGPT into an app store on the same afternoon. Read those two together. A company telling Wall Street the model is the crown jewel is also telling Wall Street it's becoming a distribution business. Both can't be the headline. The app store is the tell: when your product commoditizes underneath you, you climb to where the margin still lives. Bezos said it cleaner thirty years ago — your margin is my opportunity. OpenAI just became its own Bezos, hunting the application layer it used to sell picks and shovels to.

Apple spent the day showing the finished version of the same idea. Siri now lets you bring your own model — Claude, ChatGPT, Gemini, pick one, swap it tomorrow. Apple doesn't care who wins the model war, because it owns the iPhone, the one harness every model has to run on, and the silicon built to run them on the device. That isn't skating to the puck. That's owning the rink.

Then there's the one person looking at the books instead of the horizon. OpenAI's own CFO, Sarah Friar, wanted to wait until 2027 — $600 billion in compute promises stacked against revenue that's slowing. She's the only one in the room asking whether the puck is actually going where everybody's skating, or whether they all just assumed.

At COAI today: the full Signal/Noise — Friar's math, the landlord-vs-tenant read on SpaceX, and why even Harvey's customers are its successors-in-waiting — is live at getcoai.com.

Outsider Labs Advisory →

$84. That's what Harvey paid an open-source model to beat Claude Opus on a hundred legal tasks. Opus cost $954 for the same work — and lost. Multiply that across every company now routing the easy 80% of its AI to models that run at a penny on the dollar, and you've got today's IPO rush in a single line: the model isn't the moat anymore, it's the thing everyone's racing to stop selling. The puck already moved. The labs are just now skating after it.

Route one workflow to a cheaper model before your next invoice clears. Lindy moved 100% of its traffic to an open model and watched cost drop while quality went up. You don't have to be that bold. Pick one high-volume, low-stakes job today, run it on an open or mid-tier model next to your frontier default, and compare. If you can't tell the outputs apart, you just found your margin.

Write down what you'd own if your AI vendor vanished tomorrow. Kirkland & Ellis is reportedly building its own Harvey from its own files — proof the data, not the app, is the asset. List the three AI tools you lean on hardest and name what's actually yours when they're gone: the prompts, the data, the workflow, or nothing. The blanks are your risk register.

Sort the IPO wave by what each company truly owns. SpaceX owns the compute and collects the rent. Apple owns the device every model runs on. OpenAI owns a billion users and the thinnest margin of the three, which is why it files last and hedges hardest. Before you treat "AI IPO" as one trade, decide whether you're buying a landlord, a toll booth, or a tenant.

Monday, June 8

OpenAI files to go public, then files down its own crown. (Full analysis above.) Confidential S-1 today, eyeing a $1 trillion valuation, plus a ChatGPT super-app reveal. Friar wanted 2027, Altman wanted Q4 — "it may be a while" is the two of them shaking hands in public.

Apple makes the world's best models fight to be a dropdown. (Full analysis above.) Siri goes standalone at WWDC — bring your own Claude, ChatGPT, or Gemini. Tim Cook's final keynote before he hands off in September. The good on-device Siri needs 12GB of RAM, so only the iPhone 17 Pro and Air get it.

Anthropic says its AI is now improving its own successor. Eleven days before its IPO, it posted that Claude is accelerating AI research toward recursive self-improvement — 18 million views — while "running out of compute and energy" reports spread the same week. The flex and the constraint, one breath.

The AI rollup hits Wall Street. VCs stopped selling software to legacy firms and started buying the firms outright to rebuild them with AI inside. General Catalyst calls it "service as software." The same puck thesis, in a private-equity suit.

Cursor hits $4 billion ARR, fastest in software history. From $2B in February, 75% enterprise, now running a model it trained itself. Musk's $60B option to buy it is 15x revenue — cheap next to SpaceX until you remember Cursor pays for compute and SpaceX collects it.

— Harry and Anthony

Sources: