Tuesday Jul 14
HARDWAREMeta

Meta's In-House AI Chip Finally Works

14JUL
METAIRIS6-WEEK CLEAN TEST

Meta's long-troubled chip program just had a big week. Iris enters production in September after six clean test weeks. Meta wants 14 gigawatts of compute by 2027, less Nvidia dependence.

Iris is part of Meta's four-chip MTIA family, built with Broadcom and TSMC. Past versions floundered for half a decade before this one finally worked.

Meta now plans a new chip every six months, not every year like rivals. This year's compute deploys at 7 gigawatts, doubling to 14 in 2027. That pace rivals what Nvidia ships to Meta today.

If Iris scales, Meta buys fewer Nvidia chips at list price. Nvidia's biggest customer just became a little less dependent.

full brief & sources

⚡ Why this matters

  • A working in-house chip means Meta needs fewer Nvidia GPUs at Nvidia's prices.
  • Six-month release cadence is nearly twice the industry's usual pace for custom silicon.
  • Signals Meta's chip program is past the years of stalling that plagued MTIA.

🔍 What happened

  • Meta plans to start manufacturing its Iris AI chip in September 2026.
  • Iris is part of the four-generation Meta Training and Inference Accelerator, or MTIA, chip family.
  • Testing took six weeks and found no major issues, a first for the program.
  • Meta is working with Broadcom on chip design and TSMC on manufacturing.
  • Meta plans to double compute capacity from 7 gigawatts this year to 14 gigawatts by 2027.
  • Meta expects to spend as much as $145 billion on AI infrastructure this year.

💬 Smart takes

  • Reporting: testing 'found no major issues, signaling positive momentum for an in-house effort that has floundered since its launch more than half a decade ago.'
  • Cadence read: Meta plans to launch a new chip about every six months through 2027, versus the industry's typical yearly cycle.
  • Skeptic: Google's custom TPU program is still years ahead, and passing internal tests is not the same as beating Nvidia on real workloads.

🧭 Where this goes

  1. LikelyMeta's Nvidia orders shrink as a share of total compute spend, even if they keep growing in absolute terms.
  2. LikelyMeta ships a second Iris-generation chip on the six-month cadence by early 2027.
  3. Possibleother hyperscalers accelerate their own custom-silicon timelines to match Meta's cadence.
  4. Wild CardIris underperforms at scale the way earlier MTIA chips did, and the timeline slips again.

🥄 The Spoon Take

Every hyperscaler wants off the Nvidia tax, and most have failed for years trying. Meta's chip finally passing its own tests is a small signal, not a victory lap. The real test is running real workloads at scale, not six clean weeks in a lab.

🤔 Pushback

Six weeks of clean tests is not the same as running production AI workloads at scale, where Meta's chip efforts have failed before.

·Reuters via USNews - Iris chip exclusive·Macrostream - 14GW by 2027 target·HWBusters - Iris production details