
SIGNAL / NOISE
The Pipe Doesn't Get Paid
Eighty-nine cents of every enterprise AI dollar now goes to a closed frontier model, up from 81 a year ago. David Sacks dropped that number on All-In this week and the room heard a coronation. Open source owns the downloads, the GitHub stars, the token counts — and it just fell from 19% of actual enterprise spend to 11%, while the closed labs climbed to 89. Follow the money and the frontier looks untouchable.
Read it that way and you'll misprice the whole board. 89% isn't the frontier winning the argument. It's the sound of ten thousand companies that want out and can't find the door. Sacks said it himself: the spirit is willing, the flesh is weak. Everybody would love to route the cheap work to a cheap model and save the expensive one for the jobs that earn it. The two companies he could name who actually pulled it off were Coinbase and DoorDash, because they built the plumbing themselves. Everyone else is still talking to the head, because the head at least answers.
Chamath threw the counterpunch from the same couch: those open-source jobs are "dark tokens." They run on your own hardware and bill as compute, not as lab revenue, so the 89% can't even see them. The scoreboard everyone's crowning the winner with is blind in one eye.
We've seen this movie. The phone carriers spent a decade and a fortune laying the fastest networks on earth, and Apple strolled in and kept the customer. AT&T became the dumb pipe, enormous and essential and paid like a water utility, while the company that owned the thing you actually touch every morning kept the margin. That's the cliff the frontier labs are standing on. Whoever owns the tool the enterprise opens daily — the harness that decides which model does what — owns the account. The model underneath is bandwidth.
And the industry knows it. The New Stack's headline this week was blunt: the harness is the product. OpenAI shipped ChatGPT Work days ago, straight into Slack and Salesforce. Every lab is racing to sell you the console. Here's the trap, and it's a beauty: a real router steers you off the labs' own fat-margin tokens toward the cheap ones. Build it well and you're voting for your own Thanksgiving. So watch one thing — whether OpenAI's harness actually routes you down, or just orchestrates you into staying premium. The one player who can route cheap without bleeding sells seats, not tokens. That's Cursor, and Cursor now belongs to SpaceX.
Stop staring at the head. Build the console, or be the pipe someone else plugs into.
At COAI today: the full Signal/Noise — why the "just route to open source" answer is right on paper and impossible in the building, and why the labs can't ship the fix without eating themselves — is live at getcoai.com.
Which workloads justify frontier tokens and which get routed down — and whether the tool you're being sold actually routes them, or just bills you a premium and calls it a harness. Let’s talk.
ONE — A NUMBER THAT SUMMARIZES THE DAY
89%. That's the share of enterprise AI dollars now going to closed frontier models, up from 81% a year ago, even as open source runs away with the downloads and token counts. David Sacks put it plainly on All-In: the money says the opposite of the hype. But 89% isn't the frontier winning. It's the harness nobody's shipped yet — the man behind the curtain who'd let you route the cheap work to cheap models. Until he shows up, you keep paying the head.
THREE — ACTIONS TO TAKE TODAY
Build or buy the routing layer while switching is still cheap. Coinbase and DoorDash beat the frontier bill by building middleware that sends easy work to cheap models and saves the frontier for jobs that earn it. Nobody sells that off the shelf yet. Pull your ten biggest AI workloads today and mark each one "frontier-worthy" or "route-down," so you own the switch before the invoice forces it.
Ask your harness vendor one question: does it route me down? OpenAI shipped ChatGPT Work this week and every lab is racing to sell you the orchestration layer. A real router sends the easy 80% to a Haiku-class model and pockets you the savings. A fake one orchestrates beautifully and keeps every task on premium. Make them show you the routing logic before you sign anything.
Position above the model, not on it. A frontier lab that routes you to cheaper models is cannibalizing its own token revenue, so it won't do it eagerly. The player selling seats instead of tokens — Cursor, now inside SpaceX — has the opposite incentive and owns the usage loop that compounds. If you're allocating, the value is climbing from the model to the console. Buy the console.
FIVE — STORIES TO KEEP YOU INFORMED
Monday, July 13
Sacks follows the money, and the money says closed is running away with it. Open source fell from 19% to 11% of enterprise AI spend while closed climbed to 89%. Chamath's counter on the same show: those are "dark tokens" the revenue line can't even see. (Full analysis above.)
OpenAI clocks into your office — ChatGPT Work ships. OpenAI's Workspace Agents plug GPT-5.6 straight into Slack and Salesforce. The chatbot just became a harness. The only question that matters is whether it routes you down to cheaper models or quietly keeps you maxed. (Full analysis above.)
Apple sues OpenAI over hardware secrets. Apple claims OpenAI lifted trade secrets to build AI hardware. Two of the biggest names in the business now fighting in court over the device layer — the exact place "local open models on your laptop" would someday live. The real fight is where AI runs.
Washington may kill open models before the market decides. Nathan Lambert warns a White House order could bar open-weight models above a capability line within six months, aimed at Chinese models, and, he argues, regulatory capture wearing a safety badge. The open-versus-closed question may never get settled on price at all.
Grok 4.5 undercuts the frontier on price. SpaceX and Cursor's Grok 4.5 lands at $2/$6 per million tokens against Opus 4.8's $5/$25, burning roughly 60% fewer output tokens on the same work. Token efficiency is a procurement line now, and Cursor owns the harness the cheap routing runs through.
MARK TO MARKET
Where the cycle caught up to us this week.
The harness was the product before it was a headline. The chatbot was never the moat — the orchestration you build around the model is, and the lab shipped its harness from day one while you hadn't shipped yours (us, Your Company Needs a Harness, Not an Upgraded Chatbot, May 28) → "the harness is the product" (The New Stack, Jul 2026).
The tape doesn't lie. We just read it early.
— Harry and Anthony
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)
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
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
The AI race is shifting from bigger models to cheaper, smarter systems — CNBC, Jul 10, 2026
Apple sues OpenAI over AI-hardware trade secrets — as covered across The Neuron, Techpresso, TAAFT, Jul 11–12, 2026