Skate to Where the Puck Is Going

THE NUMBER: $84. That’s what Harvey, the legal-AI company, paid an open-source model to run a hundred tasks on its own internal benchmark — and beat Claude Opus, which cost $954 for the same hundred tasks and lost on the scoreboard too. Eleven times cheaper and better. Sit with that ratio, because it’s the whole issue. When the expensive thing loses to the cheap thing on the expensive thing’s home court, the expensive thing was never the moat. It was just the bill.

On Friday we told you capital was the moat — that even Google got caught short and had to rent compute from Elon Musk. That was the story read from the balance sheet. Today the market asked a harder question, the one underneath: when intelligence becomes a commodity, what business are these companies even in?

Here’s the day in one frame. OpenAI confidentially filed for its IPO this afternoon. The same afternoon, it announced it’s turning ChatGPT into an app store — a super-app bundling agents, coding tools, and third-party services. Read those two press releases side by side and they contradict each other. One says the model is our crown jewel, come buy a piece of it. One says the model is our crown jewel, come buy a piece of it. The other says we’re not really a model company, we’re becoming a distribution business. Both can’t be the headline. And the app store is the tell. When your core product is commoditizing underneath you, you climb to where the margin still lives. That’s not a pivot born of ambition. It’s a pivot born of math.

Wayne Gretzky had the line everyone quotes — skate to where the puck is going to be, not where it’s been. What people forget is that it wasn’t really Wayne’s line. It was his father Walter’s, drilled into him on a backyard rink in Brantford, Ontario, before the kid could spell dynasty. The most famous insight in the history of sports strategy was a dad coaching his son. Keep that in your pocket. We’ll come back to it, because the puck in this story is the margin, and right now every player on the ice is skating toward the same corner.

💲 The Substitution Wave Isn’t Three Trends. It’s One Trapdoor.

Tomasz Tunguz published a piece this morning called “The Substitution Wave in AI,” and it’s the best framing of the day — which is exactly why it’s worth arguing with. Tunguz says three forces are reshaping the AI cost structure. One, foundation labs are moving up the stack into applications. Two, frontier model prices keep climbing for the smartest models. Three, open-source models have crossed the good-enough threshold for most use cases.

All true. But he lays them out like three weather systems that happened to blow in together. They’re not parallel. They’re a chain, and the order of the links is the entire argument. Open source got good enough first. Once that happened, paying frontier rates for the last five percent of IQ stopped penciling out on the eighty percent of work that never needed it. And once the model layer has no margin left in it, there’s only one direction a rational company can go. Up. Cause, cause, effect. Tunguz’s three bullets are really one trapdoor, and the labs are the ones standing on it.

The receipts came in all week, and they’re brutal because they’re not from skeptics — they’re from the labs’ best customers. Lindy switched a hundred percent of its traffic to DeepSeek v4, churned off Anthropic entirely, said it saved millions and saw performance go up on its core use cases. Harvey’s $84-versus-$954 benchmark is up top. Coinbase has been quietly routing prompts to cheaper models and keeping costs flat while token usage grows exponentially — Brian Armstrong’s read is that within twelve to eighteen months, eighty percent of workloads run on models that are ninety-nine percent cheaper, and only twenty percent stay on the latest-gen frontier where IQ-maxing actually matters. Cursor didn’t even bother routing. It post-trained an open model into its own production system, Composer, and called it ten times more efficient than comparable models.

Tunguz reached for the right history, too: the minimill. Nucor’s electric-arc minimills started small, capital-light, and close to demand. The integrated steel giants — Bethlehem, US Steel — laughed at them, because minimills could only make cheap rebar at first. Then the minimills crept up the quality ladder, one product category at a time, and within a generation they’d hollowed out the giants from below. That’s Clayton Christensen’s disruption curve drawn in molten steel, and it’s the exact shape of what open models are doing to frontier inference. The frontier labs are the integrated mills. They make the best steel in the world. Nobody’s arguing otherwise. The minimills just don’t need to be the best. They need to be good enough and a tenth the price, and they get a little better every quarter.

What this tells you: the closed models are getting more expensive at the very top while the open models get cheaper at parity. Those two slopes are scissoring. The only question a buyer has to answer is which slope they want running underneath their unit economics — and most of them already answered it this week.

🏒 Your Margin Is My Opportunity

So where does the margin go when it leaves the model? Up the stack, into the application. And here’s the part that should make a frontier lab’s stomach drop: the application layer is occupied. By their own customers.

Jeff Bezos said the cleanest sentence in modern business thirty years ago — your margin is my opportunity. He built Amazon on it. He’d watch a category fatten up on comfortable margins and walk straight into it with a thinner price and a longer time horizon. What’s happening now is the AI version, and the labs are playing both roles in the same week. The substitution wave is the market treating the model layer as Bezos treated retail margin — squeezing it to nothing. And the labs’ response is to go become Bezos themselves, hunting the fat margin one floor up, in the apps.

OpenAI’s super-app is that move, named out loud. It used to sell picks and shovels — the API — to the people building the apps. Now it wants to be the app. Which means it’s about to compete with Cursor, with the GPT-wrapper economy, with every company that built a business on top of its tokens. The supplier is becoming the competitor. That’s the puck moving, and OpenAI is finally skating after it.

But — and this is the question we kept circling on the desk this morning — what if the corner they’re all skating toward is just as cold as the one they’re leaving? Because if the model isn’t a moat, the app might not be either.

🍎 Apple Already Owns the Rink

To see the end state of this whole migration, you didn’t have to imagine it today. Apple showed it.

At WWDC, Apple shipped Siri as a standalone app and did the most quietly devastating thing in the entire AI cycle: it made the frontier models compete to be a setting. Bring your own model — Claude, ChatGPT, or Gemini — pick one in a dropdown, swap it tomorrow when one gets cheaper or smarter. ChatGPT’s exclusivity, gone. Apple doesn’t care which lab wins the model war, and that indifference is the flex. It owns the iPhone, which is the one harness every model on earth has to run through to reach a billion-plus people. And it owns the silicon — Apple’s unified-memory chips are, right now, the best-positioned consumer hardware on the planet for a world where you switch inference between a frontier model in the cloud and an open model on the device.

Think about how prescient that silicon bet looks today. Apple spent a decade and a fortune building its own chips while everyone asked why it didn’t just buy Intel’s like everybody else. The answer arrived this week: when inference becomes something you route — hard problems to the cloud, the easy eighty percent to a local model — the device that runs the local model best wins the daily-driver position. The iPhone is the ultimate harness. The Mac with unified memory is, as Tunguz put it, a minimill on your desk. The only credible challenge on the horizon is the new wave of NVIDIA-and-Microsoft N1X Windows machines built for the same job, and they’re a lap behind.

Tim Cook handed off his final keynote on the way out the door — John Ternus takes the chair in September. Cook’s parting gift to his successor is a moat Apple has been pouring concrete on for fifteen years: it doesn’t need to be ready for an IPO, doesn’t need to win the model war, doesn’t need to skate anywhere. It built the rink, and it charges admission no matter who scores.

📉 The OpenAI Spiral, and the One Person Looking at the Books

Now the protagonist, because every good story has one and today it’s Sam Altman.

OpenAI is the purest victim of its own thesis. Roughly 800 million weekly users, somewhere around 35 million of them paying. In any other business that ratio is a crisis, not a triumph — it means every new user is a cost, not a customer. More people show up, the inference bill climbs, and the conversion to paid stays stubborn. That’s not a flywheel. That’s a treadmill that bills you by the step. Frontier prices are rising, open models are eating the easy majority of the workload, and OpenAI is walking toward the IPO window third — behind a $1.75 trillion SpaceX that prices Wednesday and trades Thursday, and behind Anthropic, which is targeting June 16. By the time OpenAI gets to the public’s wallet, two of the biggest AI offerings in history will already have drained it.

So the super-app isn’t a luxury. It’s the lifeboat. Convert that billion-user distribution into a platform — take a cut of the agents and the third-party services and the coding tools — before the S-1 forces Wall Street to price the model itself as a commodity. Move up the stack to where the margin is, like everybody else, but do it from behind and with the heaviest cost base in the field.

And here’s the human anchor, the detail that turns a press release into a story. OpenAI’s own CFO, Sarah Friar, didn’t want to do this yet. She targeted 2027. Altman wanted Q4 2026. Her reason was the most sober number in the building: more than $600 billion in compute commitments over the next five years, stacked against revenue growth that’s actually slowing. Today’s filing, with its careful hedge — we filed, but it may be a while, some things are easier as a private company — is the sound of those two splitting the difference in public. The brake and the accelerator, pressed at once.

Friar is the one person in this whole story looking at the books instead of the horizon. Everybody else is skating to where the puck is going. She’s the only one in the room asking whether the puck is actually going there, or whether the whole league just assumed it was and took off at the same time.

Contrast that with the foil. Anthropic already climbed. It crossed roughly $44 billion in run-rate revenue and is on track for its first operating profit — around $559 million in Q2 — because it went up the enterprise stack early with Claude Code and the developer business. It’s posting recursive-self-improvement science to 18 million views eleven days before its offering while fielding viral “running out of compute” reports, which is its own kind of pre-IPO tightrope. But Anthropic is not the company in the spiral. OpenAI is. Same window, opposite altitude.

Lombards, Toll Booths, Tenants

Here’s the lens that sorts the whole IPO wave, and it’s worth more than any valuation model you’ll read this week. When intelligence becomes a commodity, stop asking who has the best model. Ask what each company actually owns. Three answers, three very different businesses wearing the same “AI” costume.

SpaceX owns the power. It’s the landlord. Anthropic pays it roughly $1.25 billion a month to rent Colossus; Google pays $920 million a month — about $11 billion a year — for compute capacity at the xAI data centers SpaceX now controls. That’s rent, the fattest and most durable margin in the entire stack, and it’s why the market hands SpaceX a $1.75 trillion price that looks insane on a revenue multiple until you realize a landlord’s multiple and a tenant’s multiple are supposed to be different. Harry’s point on the desk this morning was exactly right: SpaceX at $1.75 trillion is baking in a lot of growth, but it’s more profitable growth, because the Anthropic and Google deals prove SpaceX collects the rent instead of paying it.

Apple owns the toll booth — the device, the harness, the place the user already is. Different mechanism, same durability. It taxes the traffic regardless of which model is in the car.

OpenAI owns the eyeballs — real distribution, a billion of them, and the thinnest margin of the three. Which is precisely why it files last and hedges hardest.

Now drop Cursor into that frame, because it’s the cleanest test. Cursor hit $4 billion in ARR this week, up from $2 billion in February — the fastest ARR ramp in software history, seventy-five percent enterprise. Musk holds a reported $60 billion option to fold it into his empire, roughly fifteen times run-rate revenue. Harry flagged that fifteen-times number as looking cheap next to SpaceX’s nosebleed multiple, and he’s right on the arithmetic — but the gap isn’t about growth assumptions. It’s about which side of the lease you’re on. SpaceX collects compute rent. Cursor pays it — that’s the whole reason Cursor post-trained its own Composer model, to claw its inference costs down. A tenant with a big COGS line doesn’t earn a landlord’s multiple, and shouldn’t. Fifteen times for a tenant and ninety-plus for a landlord can both be fair on the same afternoon. That’s the lesson.

And then the cut that goes one level deeper than anyone wants to admit. If the app layer is where the margin lives, is it defensible? Look at Kirkland & Ellis, reportedly building its own in-house Harvey from its own matter files. Stop on that. The proprietary data — a century of deal documents and litigation files — was never Harvey’s. It belongs to the law firm. Which means the enterprise can pull the app in-house and disintermediate the vendor the same way the vendor disintermediated the model. Harvey’s best customers are its successors-in-waiting. The moat was never the app, and it was never the model. It was the data — and the data usually belongs to whoever generated it, which is often neither the lab nor the app company. Everybody’s eating everybody. The only thing that doesn’t get eaten is whatever sits on a pile of data nobody else can copy.

🐻 The Other Hand I Have to Show You

I’ve picked my hand here — the move up the stack is rational, forced, and already underway. But I’d be selling you the same hype I distrust if I didn’t show you the counterweight, because the app-layer math isn’t proven either.

Jensen Huang spent the week calling AI stocks “very cheap right now” and likening the coming IPOs to Amazon, Google, and Meta at their debuts. Maybe. He also sells every shovel in the gold rush, so consider the source. On the other side, Gary Marcus published a piece this week titled, with characteristic restraint, “An entire industry is being propped up by math that is insane,” and he’s not entirely wrong on the unit economics — one analysis pegged the labs as spending more than $1,000 to earn every $100 of revenue. Only twenty-six percent of companies even fully track what their AI costs. So here’s the honest tension: the labs are fleeing the model layer because its margins are brutal, but nobody has yet proven the app layer they’re fleeing to is any kinder. They may be skating from one cold corner to another, just faster than the puck. The bull case and the bear case are reading the same spreadsheet. The difference is the time horizon.

What This Means For You

Stop buying “AI” as one trade. A landlord, a toll booth, and a tenant are three different risk profiles wearing the same logo. SpaceX collects rent, Apple taxes the traffic, OpenAI carries the cost. Sort your exposure by what each one owns, not by whose model topped a benchmark last week. The benchmark is the most perishable asset in this whole business.

Find your eighty percent and route it cheap — today, not next quarter. Lindy moved all of its traffic to an open model and got better results for a fraction of the cost. You don’t need to be that bold. Take one high-volume, low-stakes workflow, run it on an open or mid-tier model beside your frontier default, and compare. If you can’t tell the outputs apart, the difference between $84 and $954 is sitting in your budget waiting to be claimed.

Know what you’d own if the vendor vanished. Kirkland & Ellis is proof the data is the asset and the app is rented. Write down the three AI tools you lean on hardest and name what’s actually yours when they disappear — the prompts, the workflow, the data, or nothing. The blanks are your risk register. The data you generate and control is the one moat in this entire issue that doesn’t get eaten.

The bubble and the buildout are still the same event read from two chairs, same as Friday. But the question moved. Friday it was who can write the check. Today it’s who owns something the commodity can’t reach. The model was the thing everyone thought they were buying. Turns out it was just the cost of admission to a rink somebody else built.

Three Questions We Think You Should Be Asking Yourself

When your core product commoditizes, do you know which floor the margin moved to? OpenAI does — that’s what the super-app is. The harder question is whether you’ll see your own version of this coming, or whether you’ll defend the model layer long after the margin has left it for the floor upstairs.

Are you a landlord, a toll booth, or a tenant? Be honest about which one your business actually is in the AI stack. Tenants pay rent and earn tenants’ multiples. If you’re a tenant, your survival job is to either own your data so thoroughly the landlord can’t squeeze you, or climb a floor before the rent does.

What do you own that the commodity can’t reach? Not the model — that’s renting by the token now. Not the app, necessarily — your customer might rebuild it from their own data. The durable answer is almost always distribution or proprietary data. If you can’t name yours, that’s not a strategy gap. That’s the strategy meeting.

Skate to where the puck is going to be, not where it has been.”
— Wayne Gretzky, quoting his father, Walter

— Harry and Anthony

Sources

• The Substitution Wave in AI — Tomasz Tunguz (Lindy, Harvey, Cursor, Coinbase data and quotes)
• The Minimill of AI — Tomasz Tunguz (Nucor minimill framing, June 5)
• OpenAI confidentially files for IPO — CNBC
• OpenAI Turns ChatGPT Into a Platform Play — PYMNTS
• ‘We expect it to leak, so we’re just announcing it’ — Gizmodo (valuation ~$850B, eyeing $1T)
• CFO Sarah Friar wanted 2027; flagged $600B compute risk vs slowing revenue — Gizmodo / The Information
• OpenAI files for US IPO after Anthropic; SpaceX prices June 11 — INDmoney
• Apple launches standalone Siri AI at WWDC, opens to Claude/ChatGPT/Gemini — Apple / WWDC 2026
• Anthropic: “When AI builds itself” — Anthropic on X (RSI post, ~18M views)
• Silicon Valley’s new buyout playbook is hitting Wall Street — CNBC
• An entire industry is being propped up by math that is insane — Gary Marcus
• Brian Armstrong: 80% of workloads on 99% cheaper models in 12-18 months — Coinbase CEO on X
• Google to pay SpaceX $920M/month for xAI compute — CNBC
• SpaceX’s $60B option to acquire Cursor; Cursor $4B ARR — The Next Web