Why Every AI Giant Now Wants Its Own Chip

OPINION — In the space of about eight months, five of the companies spending the most on AI compute have each unveiled or shipped a custom chip designed to make them less dependent on Nvidia. That is not a coincidence of timing. It is what happens when scarcity turns a shared vendor into a shared cost problem.

Start with the roll call, because the pattern only becomes obvious once you line the announcements up — and it extends well past the handful of chip makers already racing to control both training and inference. Microsoft revealed its Maia 200 inference chip in January, claiming 30% better performance per dollar than the alternatives it's competing against, according to CNBC. Google made its Ironwood TPU generally available at Cloud Next, delivering 4,614 TFLOPS per chip with 192GB of HBM3E memory. OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom silicon, in June, developed from design to manufacturing tape-out in nine months — a pace both companies call the fastest ASIC development cycle ever achieved in advanced semiconductors, per TechCrunch. Amazon's Trainium3 is now shipping at 2.52 petaflops of FP8 performance per chip. And Meta's Iris chip, the subject of an internal memo Reuters obtained in July, entered production this month as part of a plan to nearly double the company's data-center capacity to 14 gigawatts by 2027.

The economics forcing everyone's hand at once

None of these companies are walking away from Nvidia. Microsoft's own CEO said as much directly: the company won't stop buying chips from Nvidia and AMD even after launching Maia, according to TechCrunch's reporting on Satya Nadella's comments. That caveat is the whole story. Custom silicon isn't a replacement strategy — it's a negotiating position. Every hyperscaler that can point to a working, in-house alternative walks into its next Nvidia pricing conversation with leverage it didn't have three years ago, when Nvidia's GPUs were the only credible option for training or serving frontier models at scale.

That leverage matters more now than at any point in the last decade because the constraint has shifted from capital to physical supply. Building a data center is expensive, but it is a solvable problem with enough money; getting enough advanced GPUs, at a price that doesn't erase the margin on the AI product built on top of them, is not solvable by writing a bigger check when Nvidia's own manufacturing partners are the bottleneck. Custom chips convert a supply problem back into a capital problem — one every hyperscaler on this list is better equipped to solve than the alternative.

What's actually different about this wave

Two things separate 2026's chip announcements from the failed custom-silicon attempts of the previous decade. The first is speed: Jalapeño's nine-month tape-out and Meta's stated goal of shipping a new MTIA generation every six months would have been implausible timelines for application-specific silicon five years ago, and both companies are explicit that AI-assisted chip design — using models to accelerate parts of the layout and optimization process — is part of how they're compressing the cycle. The chips being used to design the chips is not a footnote; it's the mechanism that makes the whole strategy economically viable at this pace.

The second is specialization. Rather than building general-purpose GPU competitors, every one of these chips targets a narrower job than Nvidia's own hardware handles. Microsoft's Maia 200 and OpenAI's Jalapeño are both explicitly inference-optimized rather than built for training frontier models from scratch, and Meta has said the same about its newest MTIA generations. That's a tacit admission that Nvidia's GPUs remain the better tool for the hardest, most flexible workload — frontier-scale training — while everything downstream of a trained model, the inference traffic that scales with user growth rather than research budgets, is cheap enough to peel off onto purpose-built silicon.

Who benefits, and who's exposed

For the hyperscalers themselves, the payoff is a lower and more predictable marginal cost per unit of inference, which matters more every quarter that AI products scale from research demos into products serving hundreds of millions of users. For Broadcom, which is the design partner behind both Google's TPU lineage and OpenAI's Jalapeño, this wave is close to a best-case outcome: multiple billion-dollar customers paying for its chip design expertise regardless of who ends up "winning" the underlying AI race.

For Nvidia, the risk isn't obsolescence — the company's GPUs remain the default for frontier training, and every company on this list still buys them — but a slow erosion of its share of the inference market specifically, the fastest-growing and highest-volume segment of AI compute demand. Estimates already circulating put Nvidia's addressable share of accelerator compute drifting from the mid-80s percentage range toward the mid-70s as hyperscaler silicon scales, a shift that compounds every year custom chips keep shipping on schedule.

For smaller AI labs and startups without the balance sheet to design their own silicon, this wave widens the gap between companies that can absorb GPU scarcity by building around it and companies that remain fully exposed to Nvidia's pricing and allocation decisions — a two-tier compute market where the biggest players increasingly set their own hardware costs and everyone else pays list price.

The bigger pattern

What's happening across Meta, OpenAI, Google, Amazon and Microsoft this year is not a rebellion against Nvidia so much as a maturing of the AI industry's supply chain: the biggest buyers of any critical input eventually build the capability to make some of it themselves, not because they want to leave their supplier, but because the alternative is letting someone else's manufacturing constraints set the price of your entire product roadmap. The company that ends up controlling the most defensible position over the next three years won't be whichever lab trains the single best model — it will be whichever ones have quietly made themselves the least dependent on any one supplier's timeline to ship it.

Frequently Asked Questions

Which companies have announced custom AI chips in 2026?

Microsoft (Maia 200), Google (Ironwood TPU), OpenAI with Broadcom (Jalapeño), Amazon (Trainium3), and Meta (Iris) have all unveiled or begun production of custom AI silicon in 2026, each aimed at reducing dependence on Nvidia GPUs for AI workloads.

Are these companies abandoning Nvidia?

No. Every company building custom silicon has said it will continue purchasing Nvidia and AMD chips. Microsoft CEO Satya Nadella explicitly said Microsoft won't stop buying from Nvidia and AMD even after launching Maia. Custom chips supplement GPU purchases rather than replacing them.

Why are custom AI chips usually built for inference rather than training?

Training frontier AI models requires the most flexible, general-purpose compute, which is where Nvidia's GPUs remain hardest to replace. Inference — running an already-trained model to serve users — is a narrower, more predictable workload that custom silicon can handle more cheaply at scale, which is why chips like Maia 200 and Jalapeño are explicitly inference-optimized.

What role does Broadcom play in this trend?

Broadcom is the design partner behind both Google's TPU chips and OpenAI's Jalapeño chip, making it a common thread across multiple hyperscalers' custom silicon programs regardless of which AI company ultimately gains the most market share.

This is an analysis piece. Editor's note — sources: CNBC, TechCrunch, Reuters (via CNBC and DataCenterDynamics reporting on Meta), Meta Newsroom.