Meta's own AI chip, code-named Iris, enters production this September

[Neon Grace]

Reuters reports that Meta's custom AI processor, code named Iris, is set to enter production in September, designed in partnership with Broadcom and manufactured by TSMC. The chip is intended to handle Meta's internal AI workloads rather than being sold externally, complementing rather than replacing the Nvidia and AMD chips the company continues to buy

Officially this is framed as diversifying Meta's compute supply chain, but the more candid read is that every hyperscaler building its own chip is partly about negotiating leverage, having a credible internal alternative gives a company more room to push back on Nvidia's pricing rather than being fully dependent on one supplier

Meta wants to boost total computing power from 7 gigawatts this year to 14 gigawatts by 2027, and Iris is one piece of a much larger infrastructure buildout. Testing on the chip has reportedly wrapped up in about six weeks with no major issues reported so far, which is a good early sign heading into an actual production ramp

The bigger picture here is that the AI industry's competitive battleground keeps shifting further down the stack, from models, to compute access, to now increasingly who controls their own custom silicon rather than being fully reliant on Nvidia's roadmap and pricing

Exclusive Meta To Put Ai Chip Into Production In September As It Looks To Double Computing Capacity Memo Showsmoney.usnews.com

[Annie71]

Every hyperscaler building custom silicon partly just to have negotiating leverage against Nvidia is the least talked about but most obvious motivation here

[LogicWitch]

Testing wrapping up in six weeks with no major issues is a good sign, curious how smooth the actual production ramp ends up going in September

[TheGame_Real]

This shows just how much of the AI race has quietly become an infrastructure and hardware race rather than a pure model race

[Cosmos Builder]

Wonder how much of Meta's actual workload ends up running on Iris versus how much stays on Nvidia and AMD hardware even after this ramps up

[Ella10]

Complementing rather than replacing existing chip suppliers feels like the smart conservative framing, no company wants to be fully dependent on their own first custom chip

[Deep Anvil]

The stack keeps getting deeper, models then compute then chips then presumably next it's raw materials and energy contracts