Nvidia currently makes the GPUs that train and serve basically every model you’ve ever touched. Something like 75% of the money spent on AI accelerators in 2026 goes to parts with Jensen Huang’s logo on them (down from around 87% in 2024), and Nvidia books a company-wide gross margin near 75% in a good quarter, most of it off that data-center silicon. So when you pay OpenAI or Anthropic for a model, a thick slice of that money is really an Nvidia invoice passed straight through to you. The industry has a half-joking name for it: the Nvidia tax.
Three of the biggest frontier labs have decided they’re done paying full price. OpenAI already has its own chip. Anthropic is quietly kicking the tires on foundries. And DeepSeek got outed this week, in a Reuters exclusive, as the newest lab drawing up silicon of its own. Three companies, two continents, one very expensive dependency they’d all love to shrink.
None of this is a fresh idea, to be clear. Google has run its own Tensor Processing Units for a decade, Amazon has Trainium and Inferentia, Meta has MTIA, Microsoft has Maia. The cloud giants worked out years ago that if you’re going to spend tens of billions on compute, you might as well design the compute. In 2026, the labs want in.
Why model labs suddenly want to be a chip companies
It (always) starts with the money.
Compute is the single largest line item at every frontier lab. Training one flagship model runs into the hundreds of millions ($$$). Then inference runs the number up firther: every ChatGPT reply, every Claude Code run, every Codex agent is on a chip somewhere drawing power on someone’s behalf. A general-purpose Nvidia GPU is a Swiss Army knife, rented at Swiss Army knife prices. But if you happen to know exactly one workload cold (and a lab knows its own model better than anyone alive), you can design a chip that does only that thing, cheaper and cooler.
Which gets to the second reason, the one nobody outside the data center debate talks about: power.
The industry long ago stopped counting these deals in chips and started counting them in gigawatts. OpenAI’s chip partnership is a 10-gigawatt commitment, not an order for X units. Electricity and cooling are the actual ceiling now, not silicon. A purpose-built inference chip that squeezes more work out of every watt is worth more than one that just posts a bigger FLOPs number. Always.
OpenAI’s and DeepSeek’s are both inference chips, not training chips (to be fair, Anthropic hasn’t said what its chip would do, but nothing points to training there either). That’s not a coincidence; training is where Nvidia’s moat runs deepest. Inference is the opposite. It’s the same operation, run a trillion times, on a model snapshot. That repetition is the natural habitat of a custom ASIC. So every lab is opening the same front (the one they can actually win), and leaving Nvidia the harder fight for now. If you want to know how serious a lab is about hardware, watch for the day it announces a training chip. Nobody has yet.
The last reason is independence. A lab that owns its silicon can cut prices and walk away from the pack. We’re already in a raging price war (Anthropic shipping Sonnet 5 as the free default, OpenAI slicing GPT-5.6 into a cheap Luna tier), but that’s still running on rented and partner chips. The owned-silicon discount is the one that compounds, and it’ll show up on your invoice as lower numbers. It’ll also show up as a widening gap between the labs that own their stack and the ones still renting it.
Lab by lab
OpenAI: already holding the chip
OpenAI is the furthest down this road by a wide margin, because it started first and spent the most. The chip is co-designed with Broadcom and, per the trade press, fabbed by TSMC. It ran under the reported project codename “Titan” in development (the “XPU” label you’ll see thrown around is just Broadcom’s generic term for its custom accelerators, not this chip’s name), and when OpenAI unveiled it on June 24 it got the public name Jalapeño. It’s inference-only, tuned specifically for the large language models OpenAI already runs, and it’s not for sale.
The partnership announced last October is a 10-gigawatt build-out running from the second half of 2026 through the end of 2029 (commitment analysts peg at somewhere around $350 billion once you count the full deployment). Broadcom hasn’t raised its 2026 revenue guidance on the back of it, though, and Hock Tan has been clear the OpenAI silicon is a 2027-and-later story. OpenAI also claims it went from initial design to tape-out in nine months, which if true is one of the faster high-end ASIC cycles anyone’s pulled off. Greg Brockman’s framing was the whole thesis in one line: “we have a deep understanding of the workload.” Sam Altman kept it diplomatic back in October, calling it a way to “add to the broader ecosystem.” A polite way of saying, “We gotta get off Nvidia silicon.”
Jalapeño is the same logic that produced Stargate. OpenAI has been desperate to control more of its own data-center stack and not just rent it off someone else. Owning the chip inside those data centers is the next logical step.
Anthropic: shopping, not building (yet)
Anthropic is the most interesting of the three precisely because it hasn’t committed. Reuters reported back in April that the company was still weighing whether to build a chip at all, and this month The Information reported it’s in early talks with Samsung about manufacturing on Samsung’s 2-nanometer process and its advanced packaging lines. That’s it. No confirmations or commitments on any of it. Samsung has also declined to comment.
Two signals suggest it’s more real than “exploratory talks” makes it sound.
- Clive Chan, an early member of the OpenAI team that built Jalapeño, left for Anthropic. You don’t poach a custom-silicon engineer from the one lab that just shipped a custom-silicon chip unless you’re planning to point him at the same problem.
- Anthropic’s $65 billion round in May pulled in Samsung, SK Hynix, and Micron as strategic partners, the three companies that make most of the world’s memory. With that lineup, “exploratory” starts to look like “aligned.”
For now Anthropic is hedging harder than anyone, smartly. Claude already runs on Amazon’s Trainium and Google’s TPUs alongside Nvidia, and the company has said all of those stay central to its compute plans. So the custom chip, if it happens, is a long-horizon bet layered on top of a stack that already leans on somebody else’s custom silicon. Anthropic doesn’t have to build a chip to escape Nvidia. It’s half-escaped already by riding its cloud partners’ accelerators.
DeepSeek: the chip it didn’t want to make
This week Reuters broke that the DeepSeek is designing its own AI chip, sourced to three people familiar with the plan. Like the other two, it’s an inference chip, and the stated goal is to lean less on both Nvidia and Huawei.
DeepSeek trains on Nvidia’s H800 GPUs (reports that it also got hold of banned H100s are unverified, and Nvidia denies them) and current US export controls now make it nearly impossible for a Chinese company to keep buying these chips. For inference it leans on Huawei’s Ascend accelerators, and it adapted its V4 model for said Ascend hardware. A Huawei-led team even post-trained the big V4 model on a thousand Ascend 910C chips to prove the domestic stack could carry a frontier model end-to-end.
So, long story short, DeepSeek is already boxed out of the best Nvidia silicon and increasingly dependent on Huawei, a supplier that’s also, awkwardly, a competitor and a national champion Beijing leans on. Designing its own inference chip is how it stops being at the mercy of either one.
You can also read this on the Handy AI Substack.
Whoever fabricates it will do so inside China, which almost certainly means SMIC (no source names a foundry, but it’s the leading domestic option and already makes Huawei’s Ascend), since the export rules also cut DeepSeek off from TSMC’s leading edge. DeepSeek (quietly) kicked off a semiconductor-design hiring push back in early 2025, so this has been coming for a while. When the world’s most talked-about open-weights lab starts drawing up silicon under a state that’s been begging its tech champions to go domestic, it’s obvious they’re both reducing a supplier risk and trying to close the loop in China’s AI stack.
The near term
Nvidia ain’t dead yet. It still takes about three-quarters of the revenue, still holds the training crown (where the real technical moat lives), and every one of these labs will keep buying its GPUs for years. What’s actually happening is narrower and slower: inference, the high-volume repetitive half of the workload, starts moving over to custom chips, and Nvidia’s share of accelerator revenue drifts from around 87% toward something closer to 75%.
The quiet winner in all this is Broadcom, and to a lesser degree Marvell. Every lab that wants a chip but doesn’t have twenty years of hardware engineering in-house needs a partner who does, and Broadcom has become the arms dealer of the custom-ASIC boom (taking a cut of OpenAI’s design and standing by for whoever’s next). Someone has to turn a model lab’s workload notes into an actual piece of silicon.
For those of us just paying the bills and counting our tokens, the near-term effect is boring but good: cheaper, faster, more energy-efficient inference. The labs that get their own chips into production will undercut the ones that can’t, and the price war we’ve already been enjoying (free frontier defaults, cheaper API tiers) will be driven down.
Enjoy it. It’s the honeymoon phase before the stack fully closes.
The long term
Zoom out and this is Apple’s playbook arriving in AI form. Apple didn’t design its own chips because it loves semiconductors. They did it because owning the silicon meant owning the experience, the performance, and the margin, and locking out anyone who couldn’t match all three.
That’s where the frontier labs are heading. The model, the chip, and the data center get co-designed as one system, each tuned to the other two, and the whole thing pulls away from anyone renting parts of the stack off a shelf.
This is great if you’re OpenAI or Anthropic. Rough if you’re a smaller lab. A custom chip lets you serve your model cheaper, cheaper serving funds more training, more training makes a better model that justifies the next chip. Labs without their own silicon (or a hyperscaler patron willing to lend theirs) get squeezed on cost against competitors. Vertical integration becomes a very efficient moat.
The geopolitics split the same way the chips do.
- One stack runs on TSMC, Broadcom, and American design
- The other runs on SMIC, Huawei, and whatever China can build behind the export wall
Two stacks, two toolchains, two sets of models trained on two sets of silicon that can’t buy from each other. A more divided industry.
Of course, there’s a real gamble buried in all of this. Silicon takes years to design and fab while models change every few weeks. Every lab building a bespoke inference chip is betting that the shape of its models (the specific math, the memory patterns, the precision) will hold still long enough for a multi-year, multi-billion-dollar chip to pay off. The labs are betting they know where their own models are going. Given how wrong everyone’s been about where AI is going, this isn’t a bet I’d personally make.
So the next time a lab brags about a cheaper, faster model, check what it’s running on. Discounts are never a kindness.