Meta Platforms is set to begin manufacturing its latest custom artificial intelligence chips in September, according to an internal memo reviewed by Reuters. The move marks a significant step in the social media giant’s strategy to reduce its heavy dependence on expensive GPUs from Nvidia and AMD amid a global component shortage.
The chips, part of the Meta Training and Inference Accelerator (MTIA) program and codenamed 'Iris,' are designed in collaboration with Broadcom and will be fabricated by Taiwan Semiconductor Manufacturing Co. (TSMC). Meta is also sourcing RAM from Samsung, storage from Sandisk, and fiber optic equipment from Sumitomo Electric for the supporting infrastructure.
The memo indicates that at least one of the new designs passed its testing phase in just six weeks with no major issues – a faster-than-expected outcome for a program that previously faced delays. Meta now plans to release updated chips roughly every six months through 2027, a cadence far quicker than the industry norm of annual cycles.
While the Iris chips are intended to augment, not replace, GPU purchases, the goal is to lower computing costs for training and running AI models, including ranking and recommendation algorithms. Meta’s capital expenditure is projected between $125 billion and $145 billion this year, with much of it earmarked for AI compute capacity. The company expects to deploy 7 gigawatts of compute infrastructure in 2026, doubling that figure the following year.
Analyst Cody Acree of Benchmark Research noted that overall hyperscaler spending is set to more than double, meaning Nvidia could still see absolute revenue growth even if its relative market share dips. However, the news, coupled with similar in-house chip efforts from Google and Amazon, added pressure on Nvidia’s stock, which edged lower in premarket trading.
