Meta builds its own AI chip. Your inference bill has a floor.

Reuters got hold of an internal Meta memo this week, and the headline is that Meta will put its first fully in-house AI chip — code-named Iris — into production in September. The real story underneath it: the companies you rent AI from are re-architecting the cost of running a model, and they're doing it to stop paying Nvidia's margin. That has knock-on effects for your inference bill whether you use Meta's models or not.

What actually happened

Per Reuters' reporting, carried by CNBC, here's what the memo lays out:

• Iris goes into production in September. It's part of Meta's MTIA program (Meta Training and Inference Accelerator), a four-generation, in-house chip project. Iris is an inference-focused ASIC — purpose-built for the recommendation and ranking workloads that run Meta's feeds.
• Broadcom designs it, TSMC builds it. Meta co-designed the silicon with Broadcom and is manufacturing on TSMC's advanced nodes. Initial validation reportedly took about six weeks with no major issues.
• A new chip roughly every six months through 2027. Most firms ship a new AI chip on a yearly-or-slower cadence. Meta is targeting twice that pace.
• Compute doubling from 7GW to 14GW. Meta plans about seven gigawatts of compute this year and double that next year, against AI infrastructure spend of as much as $145 billion in 2026. Days earlier it committed roughly $10 billion to its first Canadian data center in Alberta.

The memo frames Iris as a supplement to the Nvidia and AMD GPUs Meta already buys by the billions — not a replacement. But the direction is unmistakable: own the inference layer, shrink the dependency.

Why it matters for your business

You are not buying Iris. But every hyperscaler doing this — Meta, Amazon with Trainium, Google with TPUs — is telling you the same thing: inference cost is a strategic lever, not a fixed input. The price you pay per token is a business decision someone upstream makes, and they're spending $145 billion to move it in their favor, not necessarily yours.

The operator takeaway is the one we push on every AI build: don't hard-wire your unit economics to today's token price on one vendor's chip. When a provider swaps the silicon under a model, or reprices to recover a $145B capex bill, your margin moves. The defense is a model layer you can re-point — route to a cheaper or open-weight model, run inference where it's cheapest, and measure cost per shipped outcome so you notice the day the number changes.

Custom silicon is also a reminder that "AI cost" isn't the API line item alone. It's the whole stack — chips, power, data centers — and the people who control that stack are consolidating it. Build so you're a tenant who can move, not one who's trapped.

Sources: CNBC, Reuters via Yahoo Finance.