The Man Behind the Curtain

THE NUMBER: 89% — the share of enterprise AI spending now going to closed frontier models, up from 81% a year ago, per the a16z survey of 100 verified Global 2000 CIOs that David Sacks laid out on All-In this week. Open source fell from 19% to 11% of the money, even as its share of raw tokens climbed. Read one way, it’s the headline every AI bull wants: the frontier is winning, the DeepSeek scare was wrong, closed models took the enterprise. Read the right way, it’s the opposite — it isn’t the frontier winning at all. It’s the curtain, still drawn. It’s the sound of ten thousand companies who want to leave and can’t work the mechanism that would let them. Hold that number. The whole issue hangs off which way you read it.

There’s a moment near the end of The Wizard of Oz that everybody remembers wrong. They remember the Emerald City, the ruby slippers, the flying monkeys. They forget the machinery.

Dorothy and her companions walk the length of a great hall to beg an audience with the Wizard, and what greets them is theater: a colossal head wreathed in green smoke, flame erupting on cue, a voice that shakes the walls. I am Oz, the Great and Powerful. They came for power, and power is exactly what it looks like — enormous, frightening, other. Then a small dog named Toto, who has no interest in theater, trots over and tugs a curtain aside. Behind it: an ordinary man at a console, pulling levers, leaning into a microphone, frantically running the show. The booming head goes slack. Pay no attention to that man behind the curtain, the voice pleads, a half-second too late. Because now they’ve seen it. The head was never the power. The head was the output. The power was the man who knew how to work the levers.

We have spent two years, as an industry, staring at the head.

The head is the frontier model — the great and terrible floating face of this whole era. It’s the thing Washington wants to license and leash. It’s the thing that scored 16.1% on the Remote Labor Index and the thing Commerce pulled off the market in June with a single letter, the thing every keynote points at and every doomer loses sleep over. It is genuinely, legitimately impressive. And for a while it was also genuinely scarce — for a window there, only a handful of labs could conjure a head worth fearing at all. That window has closed, and almost nobody has noticed, because the head is still booming. This week the number that pulls the curtain back finally showed up. And in the great tradition of this letter, it showed up wearing a mask of its own opposite.

🎭 The Number That Isn’t What It Says

David Sacks went looking on All-In for the metric that actually matters and found the one everybody’s been misreading. Ignore the GitHub stars, the Hugging Face downloads, the 700 million DeepSeek pulls. Follow the checks. And the checks say open source lost ground in the enterprise this year: 19% of spend down to 11%, closed models up to 89 and climbing. His phrase for why was the best thing said on the show: “the spirit is willing, but the flesh is weak.” Enterprises would love to diversify off the expensive closed models. Token bills are doubling. Nobody wants to hand a frontier lab that might one day compete with them a clean look at their own alpha. They want to leave. Most of them can’t.

Here is the tell, and it’s the whole issue. When Sacks names the companies that did pull it off — Coinbase, DoorDash — he’s naming two of the most sophisticated engineering shops on the planet. They escaped by building the man behind the curtain themselves: a routing layer, middleware that sends the easy eighty percent of the work to a cheap model and saves the frontier head for the jobs that genuinely need it. That’s not a model. That’s a console. And the reason 89 cents of every enterprise dollar still flows to the frontier isn’t that the frontier won the argument on merit. It’s that the average company can’t build the console, so it keeps talking to the head — because the head, at least, answers.

🧵 The Flesh Is Weak

Every writer on the bleeding edge prescribes the same rational answer, and on paper it’s unimpeachable: route the grunt work to cheap open models, reserve the frontier for the heavy lifting. The problem is that the answer is inoperable in the actual building, for three reasons that lock together like a Chinese finger trap.

The first is the bubble. The people who write about this live on the absolute frontier, and they make standing up your own model sound like a Saturday project. It isn’t. Hitting an open model through a hosted endpoint is easy — but running one yourself, on hardware you control, in a way that would satisfy your security team, is a real engineering effort most companies can’t staff.

The second is the harness. There is still no off-the-shelf way to route a task from one model to another and collect the finished work in one place. Coinbase and DoorDash built their own because there was nothing to buy. The tool everyone needs to execute the rational answer does not yet exist as a product you can purchase on a Tuesday.

The third is trust, and it’s rightly earned: nobody wants a Chinese model with a clean line of sight to their internal data. Here’s where the trap closes. The easy way to use an open model — a hosted API — sends your data right back out the door you were trying to shut, which reintroduces the exact risk you were fleeing. The safe way — weights on your own air-gapped box, where an inert file can’t phone home — is the hard one from reason one that nobody can staff. Easy doesn’t solve trust. Safe isn’t easy. So the median enterprise does the only rational thing left: it uses Anthropic or OpenAI or Google, because everything lives in one place, the vendor is unlikely to steal your data even if it might one day compete with you, and the interface — while nobody’s idea of a real harness — is good enough. The frontier bundle wins on convenience and trust. Not capability. Convenience.

🌑 Dark Tokens

Then Chamath, from the same couch, put a crack in the whole scoreboard. Those open-source jobs, he pointed out, are dark tokens. When you self-host an open model you pay for compute — Nvidia’s line, the neoclouds’ line — and nothing shows up as lab revenue. So the 89% that looks like closed models winning is structurally blind to the open usage it claims is losing. The metric everyone’s crowning the champion with is blind in one eye.

And he went further, into the question nobody on an earnings call wants asked. His own CTO reported token costs doubling every 45 days for a productivity lift of “maybe 5% max.” He ran the numbers on AI-attributable EPS growth across the S&P 493, stripped out Nvidia selling the shovels, and landed on a real return somewhere between zero and two percent. Set that against Brad Gerstner’s honest bull case — “we’re so early nobody cares” — and you see the fault line. Both men are right, which is exactly what makes it dangerous. A market can run for years on so early nobody cares, right up until one bad quarter makes everyone care at once. Revenue measures what enterprises can’t yet escape. Tokens measure what they’re escaping to. We wrote in Show Me the Money that the unit of account was shifting from effort to accepted outcomes; this is the same tape, read at the ledger level. Everyone is scoring the prison. Nobody’s counting the tunneling.

🎚️ The Man Who Works the Levers

So the value already moved. It moved off the head and onto the console, and the industry said so out loud this week: The New Stack’s headline was flatly “the harness is the product.” Days earlier, OpenAI shipped ChatGPT Work — Workspace Agents that plug GPT-5.6 straight into Slack and Salesforce. We called this shape in The Age of the Router back in March and again in Harness Not Chatbot in May: the chatbot was never the product, the orchestration around it is. What’s new is that the race to own the console is now the main event, and it splits into two species.

A captive harness — ChatGPT Work, Claude Cowork — routes down its own ladder. Haiku for the easy stuff, Opus for the middle, Fable for the true frontier. It defends the lab from disintermediation and it hands the enterprise the cost curve it wanted from open source, without the setup, without the trust problem, without ever leaving the building. It neutralizes all three of your frictions at once by internalizing the routing. That is the endgame, and it’s a good one — if you’re the lab that builds it.

A neutral harness — OpenRouter, OpenClaw, Ollama, Cursor — routes across everyone. And a neutral router does something the labs can’t survive: it turns every model, frontier included, into an interchangeable part. It makes the enterprise’s dream real. Nikesh Arora, quoted by Sacks, named that dream precisely — “model fungibility,” the ability to hot-swap the cheapest model that clears the bar. And then Arora named the one thing standing in the way: “what do you do about memory? What do you do about context? What do you do about history?” Nobody’s figured out how to abstract that away from the model yet. Whoever does — whoever builds the neutral console that makes your context portable — commoditizes the frontier overnight. The enterprise’s dream is the lab’s nightmare, and they are the same sentence.

🦃 Turkeys and Thanksgiving

Which sets up the trap, and it’s the purest Innovator’s Dilemma I’ve seen in this cycle. Think about what a great router does to Anthropic’s own P&L. It deliberately steers customers off high-margin Fable and Opus tokens and onto cheap Haiku. It compresses the exact ARR that’s currently skyrocketing — the ARR built because CTOs incented token-maxing and nobody’s routing efficiently yet. A frontier lab building the honest, cost-cutting harness is turkeys voting for Thanksgiving.

But here’s the next step, and it’s the guillotine: if you don’t build the tool everyone opens every morning, someone else does — and then you’re disintermediated anyway. You become the dumb pipe. This is the movie we’ve seen before. The phone carriers spent a decade and a fortune laying the fastest networks on earth, and Apple walked in and kept the customer. AT&T became enormous, essential, and paid like a water utility, while the company that owned the thing you touch every morning kept the margin. That is the cliff every frontier lab is standing on. Fat-margin-but-doomed, or thin-margin-but-durable. There is no third door. Christensen’s only known survival move is exactly the painful one: build the thing that eats you, in-house, before the entrant does.

So the tell to watch — and it’s checkable, which is why it belongs in the issue — is whether OpenAI’s shiny new ChatGPT Work actually routes you down. A real router sends the easy 80% to a Haiku-class model and pockets you the savings. A harness in costume orchestrates beautifully and quietly keeps every task on premium, because the P&L can’t stomach the alternative. If it routes cheap, OpenAI has accepted the Thanksgiving vote and is playing to own the account. If it keeps you maxed, it’s defending margin, and it’s disintermediable by the first neutral router that doesn’t have to.

And that is why the most interesting player on the board sells neither tokens nor a head. Cursor sells seats. A seat-priced company can route to the cheapest model that clears the bar happily, because every cheap token fattens its margin instead of thinning it. Anthropic routing cheap is self-harm; Cursor routing cheap is self-interest. Now add that Cursor is itself a harness, owns the usage data, and — since SpaceX bought it — owns a model trained on that data. Model, harness, data loop, and the right incentive, vertically integrated, iterating on the tightest cycle in the industry. Chamath spent the episode betting that intelligence goes recursive and the frontier pulls away. He’s right that the flywheel is real. He’s just pointing it one layer too low. The recursion doesn’t live in the head anymore. It lives in the console: own the router, see the most usage, train the next model and the next router on it, route better, own more usage. The smartest weights don’t win. The daily tool does.

👠 The Ruby Slippers

Dorothy had the power to go home the entire time. She just couldn’t work the slippers until Glinda showed her how. So can you — but wanting it isn’t working it, and that gap is the whole ballgame.

If you run an enterprise, stop asking which model is cheapest and start asking who owns the harness you work every day, and whether owning it yourself changes the math. The build-vs-wait decision on the routing layer is the 2026 call, and the window to build it cheap is now, before the bill forces the question for you. Map your ten biggest workloads to frontier-worthy or route-down this week, the way DoorDash did, so you own the switch instead of renting it.

If you allocate capital, the value is climbing from the model to the console, and the fork is trillion-dollar: do the incumbents capture the harness (captive) or does an independent layer (neutral)? Watch whether the labs’ own harnesses route down. A lab that won’t cannibalize its token line is telling you it plans to be the pipe.

And if you run a lab — pay attention to the man behind the curtain. Better yet, be him. Stop spending the quarter cloning Figma and chasing the consumer head-fake. The prize is the enterprise console, and the clock on building it before a neutral router does is already running.

🐕 Toto Always Pulls the Curtain

Open source is having its Linux moment. It will win the server room — the sovereign stacks, the infrastructure, the tip-of-the-spear 1% who can run it — and it will lose the desktop, the median enterprise, until the console gets good enough that it doesn’t. 89% isn’t the head winning. It’s the curtain, still drawn, for now. But there’s always a Toto, and the curtain never stays shut. This week it was a number.

Stop staring at the head. Build the console, or be the pipe someone else plugs into.

---

Sources:

• All-In E280 — David Sacks, Chamath Palihapitiya, Brad Gerstner (recorded ~Jul 10–11, 2026); the 19%→11% / 81%→89% figures cite the a16z 2026 enterprise CIO survey (100 Global 2000 executives); Sacks “spirit is willing, flesh is weak,” Coinbase/DoorDash routing, Nikesh Arora “model fungibility / memory, context, history”; Chamath “dark tokens,” token costs “doubling every 45 days,” S&P 493 ROI 0–2%
• The Open-Source AI Boom May Be a Mirage — 24/7 Wall St., Jul 12, 2026
• America is Building the Best AI. China is Building the Default. — Shelly Palmer, Jul 12, 2026
• The harness is the product; they disagree on price — The New Stack, Jul 2026
• OpenAI Debuts ChatGPT Work with GPT-5.6 — Forbes, Jul 9, 2026 · OpenAI unveils Workspace Agents — VentureBeat
• The AI race is shifting from bigger models to cheaper, smarter systems — CNBC, Jul 10, 2026 (Perplexity/Srinivas “the model alone is no longer the product,” Fenton on open-weight token share, Ollama in 85% of the Fortune 500)
• 6 months to live for open models — Nathan Lambert, Interconnects, Jul 12, 2026
• SpaceXAI’s Grok 4.5 Undercuts Anthropic and OpenAI on Coding Agent Pricing — DevOps.com, Jul 10, 2026
• CO/AI prior issues referenced: The Age of the Router (Mar 30), Harness Not Chatbot (May 29), Show Me the Money (Jul 1), Ghost in the Machine (Jul 9)