Anthropic Custom Chips Are a Margin Play, Not an Nvidia Divorce

Did you know AI companies are no longer satisfied buying chips? They want to design them too.

That’s right! On August 5, Anthropic confirmed it is building an in-house silicon team to develop custom chips for Claude. The company says it will keep using hardware from Nvidia, AMD, AWS, and Google. But now it wants to co-design the hardware alongside the models that run on it.

You see, Nvidia GPUs became the gold standard for AI. Yet when you operate Claude at enormous scale, even tiny efficiency gains translate into large savings. So why not design the hardware and the software together?

The Arithmetic Behind Anthropic Custom Chips

Here’s what makes this affordable to attempt. Anthropic’s run rate revenue passed $30 billion, up from roughly $9 billion at the end of 2025. More than 1,000 business customers each spend over $1 million a year. The company reported over $11.5 billion in second quarter revenue and turned adjusted operating income positive for the first time.

Tripling revenue in a year is what makes a chip programme worth contemplating at all.

Because the costs are brutal. Industry estimates put the price of designing one advanced AI chip at roughly half a billion dollars once you count elite engineering salaries and the testing needed for defect free manufacturing. Anthropic’s job listings reflect that, offering between $320,000 and $485,000 and demanding candidates who have personally shipped silicon.

Inference cost per token happens to be the single biggest lever on any foundation model‘s unit economics. So owning more of that stack directly moves gross margin.

Everybody Is Doing This Now

Anthropic is late rather than early, which is worth knowing.

Google has run TPUs for a decade. Amazon ships Trainium and Inferentia. Microsoft deployed Maia 200 in January, built on TSMC’s 3 nanometre process with over 140 billion transistors. Meta extended its MTIA partnership with Broadcom through 2029. OpenAI unveiled an inference chip codenamed Jalapeño with Broadcom in June.

Custom silicon is now the default strategy rather than a differentiator. Analysts project ASIC based AI server shipments will reach roughly 28% of the market in 2026, with custom ASIC shipments growing about 45% year over year against 16% for merchant GPUs.

Notice who keeps appearing, though. Broadcom designs for Meta, Google, and OpenAI. Marvell anchors Amazon. And every one of these chips gets fabricated by TSMC, which produces around 92% of advanced AI silicon.

So the industry is diversifying away from Nvidia by concentrating on two design partners and one foundry.

It Isn’t a Nvidia Escape Plan

But here’s the twist. Anthropic isn’t replacing Nvidia tomorrow. It’s creating leverage.

Look at what the company did just before this announcement. It committed to roughly 3.5 gigawatts of TPU capacity through Google and Broadcom, which CFO Krishna Rao called the company’s most significant compute commitment to date. That’s approximately the electricity draw of a mid sized city, dedicated to running one company’s models.

So Anthropic announced its own chip ambitions while simultaneously deepening its dependence on somebody else’s. Custom silicon redistributes spending across the chip industry rather than removing it.

More suppliers mean better negotiating power and less exposure to any single vendor. That is the actual prize here, and it arrives long before any Anthropic designed chip does.

Diagram: what changes and what doesn’t, by CVisiona. Based on Anthropic’s stated multi chip approach.

The Part Nobody Puts in the Headline

Now for the uncomfortable detail. Forming a silicon team and shipping silicon in production are separated by roughly two to three years.

Consider the precedents. Apple bought P.A. Semi in 2008 and shipped the A4 in 2010. Google started TPU work around 2013 and deployed it in 2015. Amazon acquired Annapurna Labs in 2015 and announced Inferentia in 2018. Every one of those took years, and all three had deeper hardware experience than Anthropic does today.

Worse, the schedules slip. Microsoft’s Maia 200 ran about six months late, delayed by design changes and chip team turnover. Microsoft had also scrapped a planned training chip earlier in the programme. That is a company with decades of hardware experience and effectively unlimited money.

Meanwhile Anthropic is reportedly in talks with Samsung as a manufacturing partner, which means the fabrication still belongs to someone else.

Therefore the honest read is this. The margin story investors hear today has to hold up on Nvidia and TPU pricing alone for the next few years, because custom chips will not arrive in time to help.

What It Means for People Who Buy AI

Here’s why this matters if you never touch a data centre.

Custom chips are internal efficiency tools, not products. Microsoft’s Maia 200 is not available as an Azure instance type, and there’s no public SKU for it. You benefit indirectly through capacity and latency, never by renting the hardware.

But the second order effect reaches your invoice. As labs absorb more compute cost into their own silicon, API prices should gradually decouple from Nvidia’s pricing. Analysts expect per token costs for leading models to fall materially through 2027, driven as much by ASIC economics as by model efficiency.

So the chip programmes you cannot buy are the reason the models you can buy get cheaper.

History Repeats, Slowly

Certainly, the pattern is familiar. Apple moved from Intel to Apple Silicon. Google built TPUs. Amazon built Trainium.

First you optimise the software. Then you optimise the hardware. Eventually you build both together.

Yet notice that every company on that list reached custom silicon only after achieving enormous scale. The chip follows the volume, never the other way round. Anthropic tripling revenue is the actual news here, and the silicon team is a consequence of it.

Looking Forward

In short, the AI model war started with algorithms. It’s now a supply chain war fought with negotiating leverage, power contracts, and fabrication slots.

Anthropic is not abandoning Nvidia. It’s buying itself options, and paying half a billion dollars for the privilege.

Finally, leave your thoughts in the comments below, because I’m curious. Do custom chips give AI labs real independence, or just a better seat at the negotiating table? Let me know what you think.

And if you want more of this, subscribe to our newsletter and join us on YouTube.For tech builders, check out CloudMelonVis, where we dive deeper into the technologies shaping what we build. And for your weekly insights of AI, cloud native, startups, and venture capital, follow CVisiona, where we break down what happened, why it matters, and what to watch next. See you in the next one!

Source : Anthropic custom AI chip team confirmed from TechRepublic, Anthropic enters the AI chip race with in-house chip team from Forbes, Microsoft delays production of Maia AI chip from Data Center Dynamics, and The custom AI ASIC state of play from Tom’s Hardware