Broadcom and Meta have agreed to co-develop several generations of Meta's in-house AI accelerator, and the headline figure in the arrangement is not a count of chips. It is a quantity of electricity. The initial commitment exceeds one gigawatt, which the two companies call the first phase of a sustained, multi-gigawatt rollout.
Ordering silicon by the gigawatt is a little like buying a car by how much fuel it will burn instead of by how many cars turn up on the driveway. It is an odd unit until you notice what it gives away, which is that the scarce input here is power and somewhere to put it, not a foundry's willingness to print wafers.
The chip at the centre of this is MTIA, short for Meta Training and Inference Accelerator and said letter by letter. Meta describes it as a purpose-built accelerator optimized for inference and recommendation at scale, which in plainer terms means it is tuned for the routine work of deciding what appears in a feed, rather than for the enormous training runs that produce a model in the first place.
A 2nm process is the specification doing the heavy lifting in the announcement, and the claim attached to it is that this will be the industry's first 2nm AI compute accelerator. That claim comes from Broadcom, and it arrives with no benchmark, no power figure and nobody independent named as having measured it, so it belongs in the column marked company statement until parts ship.
Underneath sits what Broadcom calls its XPU platform, its framework for building a custom accelerator for one customer instead of selling a single design to everybody. The argument is that logic, memory and high-speed input and output can be coupled tightly for the workload at hand, then rebuilt the same way for the generation after that.
Networking is the other half of the sale. Broadcom is supplying Ethernet switching, optical connectivity, PCIe switches and SerDes, the circuitry that pushes signals down a wire at very high speed, on the reasoning that a rack of accelerators is only as quick as the fabric keeping them fed.
"As we roll out more than 1GW of our custom silicon to start and then multiple gigawatts over time, this partnership will give us greater performance and efficiency for everything we're building," said Meta founder and chief executive Mark Zuckerberg. Strip the diplomacy and that is a cost argument: Meta expects its own chip to handle the everyday work more cheaply than silicon bought from the merchant market.
There is a governance wrinkle attached. Hock Tan, Broadcom's president and chief executive, is leaving Meta's board and moving into an advisor role, where Meta says he will provide guidance on its custom silicon roadmap. A supplier's chief executive sitting on a customer's board was always going to look uncomfortable once the customer began buying by the gigawatt.
What neither company put on the record is the money. No contract value, no unit volumes, no per-generation timetable beyond a stated plan to extend through 2029, and no word on who fabricates the 2nm part. Meta has separately said it is developing and deploying four new generations of MTIA chips within the next two years, which is a brisk cadence for custom silicon and the detail most worth watching.
Leave a Reply