Meta just gave the clearest sign yet that it's tired of standing in line for Nvidia's chips. According to an internal memo reviewed by Reuters, the company plans to start manufacturing its own custom AI processor, code-named Iris, this September. The chip breezed through six weeks of bug testing without a major hiccup, and it's central to a plan to double Meta's computing capacity from 7 gigawatts this year to 14 gigawatts in 2027.
But the more interesting story isn't really about Meta. It's about the fact that every major AI company — Google, Amazon, Microsoft, OpenAI, and now even Anthropic — is quietly building or buying its way into the chip design business. Nvidia still dominates, but the ground underneath it is shifting fast. Here's what's happening, why it's happening now, and why it'll eventually show up in how much you pay for AI features.
What Meta just announced
Iris falls under Meta's MTIA program — short for Meta Training and Inference Accelerators — which has been shipping inference chips with little fanfare since 2023. It's designed with Broadcom and manufactured by TSMC, and it's built to supplement, not replace, the Nvidia and AMD GPUs Meta already spends tens of billions of dollars buying.
A few numbers put the scale in perspective:
- 7 → 14 gigawatts — Meta's planned computing capacity growth from 2026 to 2027. For reference, a single gigawatt is roughly enough electricity to power a mid-sized city.
- $125–145 billion — Meta's projected 2026 capital expenditure, most of it aimed at AI infrastructure.
- Four generations — Iris is just the first of four planned MTIA chip generations Meta has mapped out through 2027, with the newest expected to use TSMC's 2-nanometer process.
Meta has also locked in long-term supply deals with Samsung for memory, SanDisk for flash storage, and Sumitomo Electric for fiber-optic components — a sign that the company is trying to secure its entire supply chain, not just the chip itself, ahead of an industry-wide component crunch. Its stock jumped more than 7% on the news, as investors read the memo as a credible path toward needing Nvidia somewhat less over time.
One line from the internal memo is unusually candid for a company that rarely admits weakness: adopting the latest Nvidia GPUs at Meta's scale "has been a heavy lift, and it has cost us time." That's the whole story in one sentence — even the companies with the deepest pockets on Earth are tired of waiting for someone else's silicon.
Why every AI company is suddenly building its own chip
Meta isn't an outlier. It's the latest name on a list that already includes Google, Amazon, Microsoft, and OpenAI — and now Anthropic is reportedly in early, exploratory talks with Samsung about a custom accelerator of its own. This isn't a coincidence; it's the predictable result of three pressures hitting every major AI company at once.
1. Nvidia's GPUs are excellent — and expensive
Nvidia still controls roughly 70% of the AI chip market, and for good reason: its hardware is fast, and its CUDA software ecosystem has a two-decade head start that nothing else fully matches. But that dominance comes with a steep price tag and thin room to negotiate when everyone on the planet wants the same chips at the same time. Amazon, for one, has said its Trainium chips deliver 30–40% better price-performance than comparable hardware from outside vendors — the kind of margin that can justify years of in-house chip development for a company operating at hyperscale.
2. Inference, not training, is where the real money goes
Training a model happens a handful of times a year. Inference — the actual process of answering a prompt — happens billions of times a day, every time someone opens ChatGPT, asks Meta AI a question, or scrolls a feed that's quietly being ranked by a recommendation model. At that scale, even small efficiency gains compound into enormous savings, which is exactly why most of these new chips, including Iris and OpenAI's newly unveiled Jalapeño chip, are built specifically for inference rather than training.
3. Supply security has become a strategic issue, not just a cost one
When every major AI company is competing for the same limited pool of advanced chips, delivery schedules become a bottleneck that can slow down an entire product roadmap. Add in export controls, geopolitical risk around Taiwan, and the sheer size of the capital being committed, and no board wants its entire AI strategy to depend on a single outside vendor — however good that vendor's hardware is.
The players: who's building what
Here's where things stand across the industry as of mid-2026. All of it, notably, still gets manufactured by the same company: TSMC, which fabricates roughly 92% of the world's most advanced AI chips.
| Company | Chip | Design Partner | Status |
|---|---|---|---|
| TPU (Ironwood / v7) | Broadcom | Most mature program; long a Google-Cloud exclusive, now expanding externally through a new commercial partnership with Blackstone | |
| Amazon | Trainium 3 | Annapurna Labs (in-house) | Over 1 million Trainium chips deployed; powers Anthropic's training clusters in Indiana |
| Microsoft | Maia 200 | In-house, TSMC 3nm | Live in Azure data centers, handling OpenAI and Microsoft 365 Copilot traffic |
| OpenAI | Jalapeño | Broadcom | Unveiled June 2026; built specifically for LLM inference; deploying by end of 2026 |
| Meta | Iris (MTIA) | Broadcom | Enters production September 2026; first of four planned generations |
| Anthropic | Unnamed / exploratory | Reported early talks with Samsung | Runs primarily on Nvidia GPUs and AWS Trainium; reportedly evaluating Google TPU and Microsoft Maia capacity |
Look closely and a pattern jumps out: Broadcom is the design partner behind three of these five major programs — Google, Meta, and OpenAI all route through the same company. Amazon and Microsoft instead lean on their own in-house teams (Amazon acquired chip-design house Annapurna Labs back in 2015). Broadcom has become the AI industry's quiet kingmaker, reportedly carrying a $73 billion AI order backlog and targeting $100 billion in annual AI chip revenue by 2027.
Nvidia isn't going anywhere — but its share of the pie is shrinking
None of this means Nvidia is in trouble. Every company on that list is still buying enormous quantities of Nvidia GPUs, and most describe their custom chips as a supplement rather than a replacement. Meta's memo makes that explicit: Iris is meant to work alongside Nvidia and AMD hardware, not instead of it.
But the trend line is unmistakable. Custom AI chips — ASICs, in industry terms — are projected to make up close to 28% of AI server shipments in 2026, growing at nearly triple the rate of general-purpose GPUs. Nvidia's revenue is still climbing too, so this is less a story of decline and more a story of a market that's splitting into two lanes: flexible, general-purpose Nvidia GPUs for the workloads that need them, and narrower, cheaper custom silicon for the massive, repetitive workloads that don't.
My Take
As someone who spends most of my day thinking about how fast a page loads and how responsive an interface feels, this is the part of the AI story I find genuinely useful to track — because it's the part that eventually reaches the front end, even if it never shows up in a changelog.
Every AI feature we bolt onto a product — a chat widget, an autocomplete, a smart search bar — has a hidden cost sitting behind it: the price of the inference call that powers it. Right now, a lot of that cost is subsidized by venture capital and absorbed into subscription pricing that hasn't fully caught up with reality yet. Chips like Iris, Jalapeño, and Maia exist specifically to bring that per-query cost down. If they work as intended, AI features get cheaper and more generous over time — higher usage caps, fewer paywalls, faster responses. If the custom-silicon bets stumble, expect the opposite: tighter rate limits, more aggressive tiering, and "AI credits" creeping into more products the way they already have in a few coding tools.
It's also worth remembering how young this all is. Google's TPU program has had a decade-plus head start; OpenAI's Jalapeño is a few weeks old. Meta's own in-house chip effort, meanwhile, has struggled to gain traction for more than five years before this latest attempt. Custom silicon is hard, and the graveyard of abandoned in-house chip projects (Intel's Nervana being the most famous example) is a reminder that announcing a chip and shipping one at scale are very different things. September will be a real test for Meta.
Frequently Asked Questions
Is Meta replacing Nvidia with its own chips?
No. Meta has been explicit that Iris is designed to supplement its existing Nvidia and AMD GPUs, not replace them. The company is still one of Nvidia's largest customers and is expected to remain so for the foreseeable future.
Why are AI companies building their own chips instead of just buying more from Nvidia?
At the scale these companies operate, even modest efficiency gains from a purpose-built chip can translate into billions of dollars in savings, plus more control over supply and delivery timelines that aren't shared with every other AI company competing for the same hardware.
Will this make AI products cheaper for regular users?
Potentially, over time. Custom inference chips are built to lower the cost of answering each individual query. If those savings materialize at scale, they tend to show up eventually as better free tiers, higher usage limits, or slower price increases — though there's no guarantee companies pass the savings on rather than pocketing them.
Who makes these custom chips?
Design and manufacturing are usually split. Companies like Broadcom and Marvell handle chip design and architecture for clients including Google, Meta, OpenAI, Amazon, and Microsoft, while Taiwan Semiconductor Manufacturing Company (TSMC) does the actual manufacturing for nearly all of them, regardless of whose name is on the chip.
Is Nvidia losing the AI chip market?
Not in any immediate sense. Nvidia still holds roughly 70% of the AI chip market and continues to post record revenue. What's changing is the growth rate: custom ASICs are expanding faster than general-purpose GPU sales, gradually chipping away at Nvidia's share rather than displacing it outright.
