One gigawatt. That’s the initial computing capacity commitment baked into Meta’s extended chip deal with Broadcom alone — enough power to run a small city, now pointed squarely at feeding Mark Zuckerberg’s AI ambitions. And that’s just one piece of a supplier puzzle that keeps getting more crowded.
Meta has now signed a deal with Amazon to use millions of AWS Graviton chips for its AI workloads. This comes on the heels of a multi-billion dollar agreement with Google to rent AI chips, plus the freshly extended Broadcom partnership locked in through 2029. For a company that has spent years telling the world it’s building its own silicon future, Meta is doing a lot of shopping at other people’s hardware stores.
What’s Actually Happening Here
Let’s be clear about what AWS Graviton is. These are Amazon’s custom ARM-based CPUs, designed for general compute efficiency in cloud environments. They’re solid chips — fast, power-efficient, cost-effective at scale. But they are not the GPU-class accelerators you typically associate with training large AI models. Graviton chips are better suited for inference workloads, running models after they’ve already been trained, rather than the heavy lifting of building them from scratch.
So when Amazon announced this deal, the framing of “AI chips” deserves a bit of scrutiny. Meta isn’t using Graviton to train its next Llama model. What they’re likely doing is deploying inference at massive scale — serving AI features to billions of users across Facebook, Instagram, WhatsApp, and Threads. At that volume, CPU-based inference on efficient chips makes a lot of economic sense, even if it doesn’t carry the same headline weight as an H100 cluster.
The Multi-Supplier Strategy Is a Feature, Not a Bug
Some analysts will read Meta’s growing list of chip partners as a sign of desperation or disorganization. I’d push back on that. What Meta is actually building is a deliberate multi-supplier strategy, and for a company operating at its scale, that’s one of the smarter moves you can make in 2025’s chip market.
NVIDIA still dominates AI training hardware, and Meta is almost certainly still buying and using GPUs in large quantities. But NVIDIA supply is constrained, prices are high, and depending on a single vendor for your entire AI infrastructure is a liability. By spreading workloads across Amazon’s Graviton for inference, Google’s chips for certain AI development tasks, and Broadcom’s custom silicon for longer-term needs, Meta is hedging against supply shocks and negotiating from a position of distributed use — without being held hostage by any one partner.
The Broadcom Deal Is the One Worth Watching
Of the three partnerships, the Broadcom extension is the most strategically interesting. Custom silicon deals of this nature typically involve co-designed ASICs — chips built specifically around a company’s AI workloads rather than general-purpose hardware. The fact that Meta extended this deal through 2029 with a commitment exceeding one gigawatt of compute suggests they’re serious about owning more of their AI stack over time, even if they’re renting capacity in the short term.
This is the same playbook Google ran with its TPUs and Apple ran with its Neural Engine. You start by buying off-the-shelf, you learn what your workloads actually need, and then you build something purpose-fit. Meta appears to be in the middle chapters of that story.
What This Means for the AI Chip Space
Meta’s shopping spree is a signal to the broader market that demand for AI compute — across training, fine-tuning, and inference — is not slowing down. Every major cloud provider and chip designer is now a potential partner for the hyperscalers, not just a vendor. Amazon, Google, and Broadcom all benefit from Meta’s willingness to distribute its workloads, and that competition is likely pushing prices and terms in Meta’s favor.
For smaller AI companies and developers watching from the sidelines, there’s a practical takeaway here too. If Meta — with its enormous internal engineering team and years of custom chip investment — is still renting compute from multiple external sources, the idea that you need to own your own hardware to be serious about AI is increasingly hard to defend.
Meta is building one of the most ambitious AI infrastructures on the planet, and it’s doing it by being strategically promiscuous with its chip partnerships. Whether that approach produces better AI products for users is still an open question. But as a supply chain strategy, it’s hard to argue with the logic.
🕒 Published: