Meta ships Iris AI chip in September — 7GW to 14GW in a year, prying inference off Nvidia
TL;DR
Reuters: Meta will mass-produce its Iris AI chip (MTIA series) in September, co-designed with Broadcom on TSMC 2nm, targeting a 7GW → 14GW compute doubling by 2027.
Reuters broke an internal Meta memo on July 9: the codenamed Iris in-house AI chip enters mass production this September. It is the next-gen flagship in the MTIA family, and the vehicle for lifting data-center compute from 7 GW in 2026 to 14 GW in 2027 — a doubling in one year.
Iris is designed inside Meta, co-engineered with Broadcom, fabbed by TSMC, with the newer MTIA line moving to 2nm. Silicon validation ran only 6 weeks with no major issues — an unusually smooth curve against the industry's 3–6 month norm.
The release cadence is changing too. In March Meta laid out MTIA 300 / 400 / 450 / 500 in one shot and committed to a new generation every 6 months, far ahead of the 12-month industry default. MTIA 300 already runs ranking and recommendation in production; 450 / 500 target generative image and video inference through 2027. The Broadcom partnership now covers through 2029.
Strategically, Meta remains one of Nvidia's largest customers — training stays on GB200 / GB300. But inference — Reels ranking, meta.ai chat, Instagram image gen — migrates onto MTIA every 6 months, clawing back Nvidia's inference margin. 2026 Meta capex guidance sits at $125–145 billion, mostly data centers, GPUs and custom silicon.
Meta is the second hyperscaler after Google TPU to prove a 6-month chip cadence works. That squeezes AWS Trainium / Inferentia, Microsoft Maia, OpenAI's in-house — all of them now have to match the tempo.
Win the bet, and half of Meta's 14 GW 2027 compute runs on Iris silicon, with inference cost low enough to worry Nvidia. Lose it, and a bug missed in the 6-week validation blows up post-launch and MTIA's fast lane snaps back to 12-month cycles.
via Reuters via TNW / CNBC / Sina Finance
Iris is designed inside Meta, co-engineered with Broadcom, fabbed by TSMC, with the newer MTIA line moving to 2nm. Silicon validation ran only 6 weeks with no major issues — an unusually smooth curve against the industry's 3–6 month norm.
The release cadence is changing too. In March Meta laid out MTIA 300 / 400 / 450 / 500 in one shot and committed to a new generation every 6 months, far ahead of the 12-month industry default. MTIA 300 already runs ranking and recommendation in production; 450 / 500 target generative image and video inference through 2027. The Broadcom partnership now covers through 2029.
Strategically, Meta remains one of Nvidia's largest customers — training stays on GB200 / GB300. But inference — Reels ranking, meta.ai chat, Instagram image gen — migrates onto MTIA every 6 months, clawing back Nvidia's inference margin. 2026 Meta capex guidance sits at $125–145 billion, mostly data centers, GPUs and custom silicon.
Meta is the second hyperscaler after Google TPU to prove a 6-month chip cadence works. That squeezes AWS Trainium / Inferentia, Microsoft Maia, OpenAI's in-house — all of them now have to match the tempo.
Win the bet, and half of Meta's 14 GW 2027 compute runs on Iris silicon, with inference cost low enough to worry Nvidia. Lose it, and a bug missed in the 6-week validation blows up post-launch and MTIA's fast lane snaps back to 12-month cycles.
via Reuters via TNW / CNBC / Sina Finance
