OpenAI is quietly building its own AI chip and it just entered the final stretch

[PlanckLimit81]

Apparently OpenAI is in the final stages of producing its first in house AI chip, a project that reportedly cost upward of 500 million dollars just to get to this point, and it is about to be sent to TSMC for fabrication with mass production potentially starting later this year

This has actually been in motion for a while, reports of OpenAI talking to Broadcom about custom silicon go back to mid 2024, so this is not a sudden pivot, it is the culmination of roughly two years of quiet work finally reaching tape out

Tape out itself is a meaningful milestone regardless of how the chip eventually performs, getting a first design ready to send to a fab is a process that typically takes other chip designers considerably longer, so OpenAI moving through it this quickly says something about the resources and talent they have thrown at the problem

The obvious motivation is reducing dependence on Nvidia, which is a theme running through basically every major AI lab and hyperscaler story this week, everyone wants leverage over their single biggest supplier and the only real way to get it is to eventually field your own silicon

Testing an alternative to Nvidia's chips later this year would be a significant milestone if it happens on schedule, though first generation custom silicon from any company rarely matches the incumbent on raw performance right out of the gate, the value early on is usually more about negotiating leverage than actual workload replacement

3nm TSMC node, HBM, AI-focussed: No, itwww.techradar.com

[Forge40]

500 million just to reach tape out and that is before mass production costs even start, custom silicon is an absurdly expensive game to get into

[Cass93]

Every major AI company eventually does this though, it is basically a rite of passage once you are spending enough on Nvidia hardware annually that designing your own starts to pencil out

[Emily87]

First generation custom silicon almost never beats the incumbent on raw specs, the real win is just having negotiating leverage and a fallback option if supply gets constrained

[Highland Fatima]

Mass production later this year seems optimistic for a first tape out, these timelines tend to slip once real world fab issues show up