Apple Makes the Customer Buy the Compute

Anthropic is asking public markets to pay for its compute. Apple just asked you to.

The two answers landed within days of each other, and they answer the same question. Anthropic filed its S-1 right after a punishing week for the AI trade. Bloomberg reports that when Wall Street looked closely, it generally liked what it saw, and the FT has Asia following Wall Street higher on the filing. Ben Thompson’s read at Stratechery is the sharpest one: the core business is very profitable, but competition may eat all of the profits, and investors are set up for endless dilution. Then Apple announced a 14-inch MacBook Pro with M6. AI runs through the release, from the chip to macOS 27.

Start with Thompson’s headline question: who is paying for the compute? For a frontier lab, the answer has layers. Shareholders fund the data centers. Customers pay per token. Dilution lives in the gap between the two. My read of why the dilution never ends: each new model generation resets the capex bill, and the lab has to raise again to stay at the frontier. An IPO doesn’t end that treadmill. It puts the treadmill on public display and lets anyone with a brokerage account step onto it.

Apple’s answer is simpler. The customer pays up front, at Apple’s hardware margin. The M6 has a 12-core GPU with Neural Accelerators built in and what Apple calls a “Dual 16-core Neural Engine.” Apple claims up to 25 percent faster AI workloads than M5 and up to 13x faster than M1. The release lists “running LLMs locally” as a headline use case. Every one of those transistors is inference capacity Apple doesn’t have to finance, house, cool or depreciate. The buyer handles all of that, and the buyer’s wall socket pays the electricity bill. Apple books a profit on the capex instead of carrying it.

Everyone talks about on-device AI as a privacy feature. I’d argue it matters more as a capital structure. One company sells shares to fund GPUs it rents out by the token. The other sells GPUs, at a markup, to the people who would otherwise be renting tokens. Both are in the inference business. Only one gets paid to build out capacity.

Here’s the strongest case against me. Local models are toys next to Claude. Nobody runs a serious agentic coding session on a laptop’s neural engine. Demand at the frontier is growing faster than any consumer chip can absorb, so the labs keep pricing power exactly where the hard, valuable work happens. That’s all true, at the top of the curve.

But Thompson’s worry is competition, and the competition that matters isn’t only OpenAI and Google. Some of it is free. Once a task is good enough on a machine the user already owns, the marginal price of doing it there is zero. Summarizing an email, rewriting a paragraph, sorting a photo library: none of these needs a data center once the laptop can handle it. That zero sets a price ceiling, and the ceiling climbs every time Apple ships a chip. Apple claims up to 13x faster LLM prompt processing from M1 to M6. My bet is that the line keeps going and swallows the bottom of every lab’s price list a little more each year.

The labs escape the ceiling by pushing the frontier higher. That takes more compute, which takes more capex, which takes more dilution. The endless dilution Thompson warns about and the ceiling Apple keeps raising are one mechanism seen from two sides. Technology compounds, but it matters whose money it compounds on. For Apple, chip improvements compound on the customer’s money. For a lab, model improvements compound on shareholders’ money, and each cycle costs more than the last.

That doesn’t make the S-1 a bad document or the IPO a bad idea. The market liked the growth, and the market is usually right about the next few quarters. The inflection point is somewhere else: the moment the volume work moves to hardware customers have already paid for, leaving the labs to fight over the frontier, where every step forward is the most expensive one yet. When you read the filing, today’s margin matters less than how fast the free tier on the desk eats the cheap end of the menu.

The cheapest data center in the world is the one your customer already bought.