Meta approved production of its own custom AI chip.

Meta approved production of its own custom AI chip.

Meta announced its custom silicon rollout weeks after signing enormous chip deals with Nvidia and AMD.

Not before. Not instead of. Weeks after.

Meta's own framing is the tell: it is deploying more than a gigawatt of its own custom silicon, plus a significant quantity of AMD chips, to complement the new Nvidia systems it is rolling out. Complement. That word is carrying the entire story, and almost nobody quoting this news has read it.

What actually got approved

The chip is codenamed Iris. It enters production in September 2026, cleared its bug-testing phase in roughly six weeks without significant problems, and is designed with Broadcom and fabricated by TSMC.

Iris is one of four chip generations Meta has planned under MTIA — Meta Training and Inference Accelerators — a program Meta detailed publicly back in March. So this isn't Meta entering custom silicon. It's Meta's newest generation reaching production in an effort already years old.

Which matters, because the framing of "Meta joins the club, the last piece clicks into place" gets the standings wrong.

Meta isn't completing the set. It's in fifth place.

Roughly 1.9 million custom AI accelerators were deployed across the industry in 2026:

Units deployed, 2026

Google TPU ~900,000

AWS Trainium / Inferentia ~600,000

Microsoft Maia ~250,000

Meta MTIA ~180,000

Google has been shipping TPUs since 2016 and deploys five times what Meta does. Meta is the newest and smallest of the four, entering a race that Google effectively started a decade ago.

(One precision point: Microsoft's custom chip is Maia, developed in-house and fabbed at TSMC. Microsoft also buys AMD's merchant GPUs — but purchasing AMD silicon is procurement, not custom design. Different thing.)

"This isn't about saving money" — it's almost entirely about money

I'd push back hard here, because the money framing is more interesting than the control framing, not less.

Nvidia's gross margin runs around 75%.

Sit with that. For every $100 a hyperscaler spends on Nvidia accelerators, roughly $75 is Nvidia's gross profit. Nvidia posted $215.9 billion in revenue for fiscal 2026, up 65%, and captures something like 40% of hyperscaler capex.

Now look at Meta's side of the ledger. Its 2026 capex guidance is $130–145 billion, raised twice, against $72.2 billion actually spent in 2025 — more than a doubling. Meta is targeting 7 gigawatts of AI compute capacity this year and 14 gigawatts in 2027, across 32 data centres operating or under construction.

At that scale, tens of billions of dollars of someone else's gross margin sits permanently inside Meta's cost base. Custom silicon is a margin recapture play. Control is the press release. The P&L is the reason — and there is nothing embarrassing about that.

What custom silicon actually takes

Here's the part that makes the whole picture legible.

Custom chips are not coming for Nvidia's business. They're coming for inference — which now represents about two-thirds of all AI compute.

Inference is repetitive, predictable and well understood. You know the model, you know the shape of the workload, you're running it billions of times. That is precisely the condition under which a fixed-function ASIC beats a general-purpose GPU on performance per watt and per dollar.

Frontier training is the opposite: the workload changes every few months, the research lives in CUDA, and flexibility is worth more than efficiency. So it stays on Nvidia.

The split isn't Nvidia versus custom. It's experimentation versus production. Every hyperscaler builds its own silicon for the part of the workload that has stopped changing, and keeps buying Nvidia for the part that hasn't.

If that pattern sounds familiar, it should — it's the same shape as Microsoft moving email summarisation to its in-house MAI models while leaving hard reasoning outsourced. The commodity tier goes in-house first. The frontier stays rented.

The number that settles it

Nvidia's share of the data centre AI accelerator market by revenue has fallen from roughly 92% in 2023 to 80–85% in 2026.

Over that same window, Nvidia's revenue grew 65%, to $215.9 billion.

Losing share in a market expanding this fast is not losing. Custom silicon is forecast to reach perhaps 15–25% of the accelerator market by 2030, growing at a 44.6% CAGR — real, significant, and still a minority of a much larger pie.

My Read

The strategic logic is sound and I'd do the same in Zuckerberg's position. Owning your inference stack lowers unit costs, improves perf-per-watt in your own data centres, and gives you a credible negotiating alternative — which is worth something even if you never fully switch.

But "owning your destiny" overstates it. Meta still depends on TSMC to fabricate, Broadcom to co-design, and Nvidia for everything at the frontier. It has swapped one dependency for three, and gained leverage rather than independence.

That's still a good trade. It just isn't liberation.

Question for you: if custom silicon only wins where the workload has stopped changing, what does it say about the state of AI that hyperscalers are now confident enough to bake two-thirds of their compute into fixed-function chips? I'd argue that's a bigger signal about maturity than anything in the earnings calls.

Address

Boston, Massachusets

Contacts

news@techproduit.com

Get your weekly newspaper