Here's what's on our plate today:

• 🧪 The fit you can't rent: why every major AI lab now draws its own chips.

• 🍪 Anthropic's IPO math, Stripe's $7B gateway buy, investors hunting the next AI winners.

• 💡 Prompt of the Day: work out whether your company should design its own hardware.

• 📊 Poll: why did every AI lab end up designing its own chip?

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

Renting the machine was fine until the machine became the product.

• The list is now complete: Anthropic confirmed on August 5 that it is building a team to design its own chips, making it the last major AI company to stop leaving the hardware to someone else. Google, Amazon, Meta, Microsoft, and OpenAI all got there first.

• Not a supplier fight: Anthropic already runs on over 1M of Amazon's chips, with up to 1M more promised by Google. It has plenty of hardware. What it lacked was a design of its own.

• Why the timing: Deloitte expects running finished models to eat two-thirds of AI computing this year. A model built once and run billions of times a day rewards a chip made for that one job.

• The record is mixed: Tesla shut down its chip team. Microsoft's first attempt never ran at scale. Half a billion dollars buys an attempt, not an outcome.

• What's at stake: if these bets land, competing means owning a matched model and machine. Renting gets you the power, never the fit.

Why AI labs stopped renting their machines & started designing them

Few technology success stories come close to Apple's growth. Once the underappreciated choice of enthusiasts, especially when it came to computers, the company changed what a personal machine could be once it started building its own chips. The change did not happen overnight because, for 15 years, the MacBook maker had bought the processors for its computers from the industry standard, Intel. Those were capable chips, built for the whole industry at once, which meant Apple could only go as far as a part designed for everybody else allowed. That is what ended in June 2020, when the company announced that the Mac would move to chips of its own design within two years.

The reason Apple gave was that its own chips could do more work while drawing far less power. The proof arrived five months later in the first MacBook Air built on the M1 chip, which, as MacRumors noted, shipped without a fan. The chip ran cool enough that the laptop no longer needed one, and a machine designed as a single piece could finally do something no bought part had allowed.

Artificial intelligence spent its first boom years on the arrangement Apple walked away from. The companies building AI models wrote the software and rented the computers that ran it, by the hour, from whoever had room. The machines underneath were treated as a commodity that any supplier could hand over. Under that split, the model was the product, the machinery was somebody else's business, and for a few years it looked like the natural shape of a young industry.

That split is now coming apart, and how it settles will decide three things at once. It will decide what these companies are really selling, the model on its own or the whole machine that answers on its behalf. It will decide where the money in AI ends up, because whoever designs the machine controls how fast and how cheap every answer is. And it will decide what a newcomer can still buy, since computing power rents by the hour while a good fit between a model and its machine has never been for sale. The question under all three is whether an AI company can finish its product on machines it did not design, and last week the final holdout gave its answer.

On August 5, 2026, Anthropic confirmed to Reuters that it is building an in-house team to design its own chips for its Claude models, a step the news agency tied to a shortage of the chips needed to build more advanced systems. The company said it is hiring engineers who work on both hardware and software, so the chips and the models that run on them get built together. TechCrunch reported a job listing describing the group as a 'custom silicon team’, and noted that Anthropic's deals with Amazon, Google, NVIDIA, and AMD all remain in place. Industry sources in the same Reuters report put the cost of a new advanced AI chip near $500M, and Anthropic has named neither a date nor a factory.

The last name on the list

Anthropic's announcement finishes a pattern that has been forming for years. A Tom's Hardware survey of the field counts every big AI builder as a chip designer already. Google is seven generations deep into chips it designs with Broadcom, Amazon has shipped over a million of its own Trainium chips, and Meta and Microsoft each run machines built around their own designs. OpenAI joined in June with a chip called Jalapeño, which the company says was designed for the way its models run, with the first units due by the end of 2026. Across those names, not one company now writes models and leaves the machines to somebody else, and Anthropic held out longer than any of them.

What Anthropic already had

The obvious reading is that Anthropic wanted out from under a supplier, and its own history rules that out. The company has said Claude already runs on over 1M Trainium chips, which Amazon designs for its own cloud, and Google Cloud has promised it up to 1M TPUs, the chips Google designs for its own cloud. Microsoft announced late last year that Anthropic would also spend $30B on Azure capacity, which left the lab running on very nearly every kind of machine on the market.

A company with that much hardware available still decided that renting somebody else's chips leaves its product unfinished. Its job postings ask for engineers who have carried a chip from a blank page to production, which is how a company talks when it intends to hold the pen itself.

Why now

The timing comes down to a change in what all this computing power is used for. Deloitte predicts that running finished models will take two-thirds of all AI computing this year, up from about a third in 2023. A model is built once and then run billions of times a day, so every answer a person receives is generated fresh in a building full of computers somewhere. Once the answering itself is the product, the machine stops being a background detail. How quickly the reply arrives, what it costs to produce, and how much power it burns are exactly what the customer notices, which is the lesson Apple learned in laptops, where buyers judged the computer on heat and battery life.

A chip built to handle a bit of everything is the right choice, while nobody knows what next year's models will need. It becomes the wrong choice once the work settles into a single job repeated billions of times, a job the chip was built to do among a thousand others.

Buyers have already started moving, and TrendForce projects that chips built for a single job will grow almost three times as fast as the general-purpose kind this year, reaching nearly 28% of AI computers sold. Built for a single job, in this business, means built around a particular model, which is why the companies writing the models have ended up drawing the chips.

Designing the chip, not making it

None of this turns the AI companies into manufacturers, and none of them is attempting to do so. OpenAI drew Jalapeño, while Broadcom describes its own job as building and producing the chip, and Anthropic has named no factory at all beyond early talks with Samsung, as TechCrunch reported. What these companies want is the design, the part that decides how their product behaves, while the factories stay with the people who already own them.

What could complicate the picture?

The bet has not paid off yet, and the record so far runs both ways. That same Tom's Hardware survey records Tesla shutting down its Dojo chip team after years of work, and notes that Microsoft's first Maia chip never ran production AI at scale, so half a billion dollars guarantees very little. Apple reached the M1 carrying a decade of iPhone chips behind it, while the AI companies are starting these teams from job listings. The general-purpose machines are not fading either, with NVIDIA's filings showing data center sales up 92% to a record last quarter, carried partly by training work, where a flexible chip still does the job best.

What comes next

If the bets do land, the industry that follows will be built around pairings of model and machine, with each company sitting inside one made for it. A newcomer could still rent excellent computing by the hour, and would find that the fit between the two was never on the price list. The name to watch is xAI, the one big AI company without an announced chip program, and three years of this pattern suggest it will arrive there too.

Apple's proof was something anybody could feel in their hands, a laptop that went quiet because the chip and the machine had finally been designed as one thing. Nothing like it exists yet for a building full of computers answering millions of people at once, and whether the AI industry ever produces one will decide if designing the machine truly finishes the product, or simply buys the same performance twice.

Prompt of the Day

Bite-Sized Brains

• Anthropic's IPO math: Anthropic's listing is being priced off a $190–200B revenue forecast for 2028, a number sitting a long way above the $47B run rate the company reported in May.

• Stripe buys the gateway: Stripe is paying more than $7B for OpenRouter, the single interface to 400-plus models that raised at a $1.3B valuation in May, roughly five times over in three months.

• Investors hunt the next winners: Big investors are moving past the companies building the data centers and into the ones expected to profit from them, a sign that anxiety over AI capital spending is easing.

Tuesday Poll

The Toolkit

• Deepgram: Speech-to-text API built for scale, handling real-time transcription and voice intelligence for production apps.

• Descript: AI-powered audio and video editor that lets you edit recordings by editing the transcript like a doc.

• Drift.ai: AI add-in for Excel that makes financial models context-aware and queryable, built for investment banking, VC, and corporate finance teams.

Rate This Edition