Nobody Ever Got Rich Selling Electricity

We Sell Electricity

JPMorgan put a number on the AI buildout this week: $5.5 trillion through 2030. Read past the headline to the financing line and the real story shows up — $4.1 trillion of it is debt, at loan-to-cost ratios north of 85%, some deals over 90%. That's not a venture round. Venture is equity with a little debt sprinkled on top. Eighty cents of borrowed money on the dollar is a leveraged buyout. And nobody does an LBO on a science project. You lever a business when the cash flows are boring and bankable — cigarettes and cookies, RJR Nabisco, the kind of cash flow you can set a covenant to. They've stapled buyout debt to a moonshot.

Here's why that lands now. The same week, OpenAI agreed to ship GPT-5.6 customer by customer, only to buyers the federal government signs off on. Anthropic took Alibaba to Congress, accusing it of draining Claude through 25,000 fake accounts. Both moves are defensive. When your moat is being the smartest model, and a distillation six weeks behind you runs 50× cheaper, "smartest" stopped paying the bills. So you wall the frontier and call it security.

Strip it all down and every player is selling the same thing: electricity. The frontier lab, the fast-follower, last year's model — all of it is compute turned into tokens. A commodity. Some generators sit on cheap power and custom silicon; some buy gas on the spot market. Some lines run lossless, some lossy. None of it changes what the product is. And here's what the trillion-dollar marks forget: nobody ever got rich generating electricity. The utilities were the lowest-return, most capital-starved, most heavily regulated businesses of the last century. The money went to everyone who plugged in — the factories, the chip fabs, the whole digital economy that ran on top of cheap power.

There's exactly one way an electric company carries 80% debt safely: a regulator guarantees a return on the rate base. The labs took the leverage without the guarantee. The only way to earn it after the fact is to get Washington to legislate one — and gating the model, banning the open-weight competitor, reclassifying distillation as theft is how you buy that. The trillion-dollar equity isn't a bet on intelligence. It's a bet that someone makes competing electricity illegal.

Don't make that bet. Own the load — the customer, the proprietary data, the routing table that picks the model — and rent whichever electricity is cheapest and good enough this quarter. Let the labs and their bankers fight over who eats the depreciation.

At COAI today: the full Signal/Noise — the LBO math, the Medallion cap, and why even Citadel never cleared a trillion — is live at getcoai.com.

Outsider Labs Advisory →

$4.1 trillion. That's the borrowed slice of JPMorgan's $5.5 trillion AI buildout — roughly 75 cents of debt on every dollar, at loan-to-cost ratios north of 85%. Venture rounds don't look like that. Leveraged buyouts do. The difference is an LBO works on predictable cash flow, and you don't lever a moonshot. The AI boom stopped being a bubble question this week and quietly became a credit question. When the meter and the loan covenant collide, which model gets shut off first?

Rent the plant; don't build it. A distillation six weeks behind the frontier already runs about 50× cheaper. Any architecture that marries you to one model vendor on a multi-year commit is buying yesterday's electricity at a premium. Today: rip out the hard dependency. Make your model layer swappable so you can move to whatever's cheapest-and-good-enough next quarter.

Put a scoreboard on every workload. Intelligence only turns into margin where there's ground truth — code that compiles, an invoice that reconciles, a lead that converts. Today: list your AI use cases and mark which ones have a verifiable result. Fund those. The ones with no scoreboard are chat, and chat is the commodity you buy at spot, not the thing you build.

Own the load, not the plant. What survives the price war isn't the model. It's the customer relationship, the proprietary data, and the routing table that decides which model runs. Today: name the one asset only you have, and put this week's effort there. Let the labs and their lenders argue over the depreciation schedule.

Thursday, June 25

OpenAI gates GPT-5.6 behind the government. Altman told staff the model ships "customer by customer," each approved by federal officials, and called it "not our preferred long-term model." Voluntary review stopped being voluntary. The frontier is now a members-only club with Washington at the door. (Full analysis above.)

Anthropic drags Alibaba to Congress. The claim: 25,000 fraudulent accounts, 28.8 million Claude exchanges harvested in roughly six weeks to distill its best capabilities. If burner accounts can drain your model that fast, the model was never the moat. The remedy Anthropic wants is political, not technical. (Full analysis above.)

JPMorgan: $5.5 trillion, and 75% is borrowed. The bank calls the capex cycle "profitable — for now." The for-now is doing the work. $4.1 trillion of debt at 85%-plus loan-to-cost, against assets that go obsolete on a three-year clock. That's not infrastructure. It's a margin call waiting for a trigger. (Full analysis above.)

OpenAI builds its own chip. Jalapeño, with Broadcom — a custom inference accelerator aimed squarely at Nvidia and at OpenAI's own cost curve. A "model company" turning into a chip company is the tell that the model alone no longer clears the cost of capital. Drive your token cost down, hold your prices up.

Claude moves into your Slack. Anthropic swapped its Slack app for a persistent agent that lives in the channel. The pitch is delegation. The model is that your team's accumulated context now sits inside someone else's agent — and context is the switching cost. Convenience today, lock-in at renewal.

— Harry and Anthony

Sources: