One gigawatt. That’s the scale of compute Meta has committed to deploying through its expanded chip partnership with Broadcom — and if that number doesn’t immediately register, think about it in terms of what it takes to power a small city. Meta is essentially building a small city’s worth of AI processing capacity, and it wants Broadcom to help design the silicon that runs it.
The two companies announced an extended multiyear deal this week that runs through at least 2029. The agreement covers custom AI chips and networking technology, co-designed between Meta and Broadcom, destined for Meta’s AI data centers. This isn’t a pilot program or a hedge. It’s a long-term structural commitment to building AI infrastructure on Meta’s own terms.
Why This Deal Actually Matters
Let’s be direct about what’s happening here. Meta has been developing its own in-house AI chip — the MTIA — for a while now. This deal is the clearest signal yet that Meta is serious about scaling that effort rather than staying dependent on third-party silicon at hyperscale volumes. When you’re committing to 1 GW of custom chips, you’re not dabbling. You’re making a strategic bet that owning more of your compute stack is worth the engineering investment and the long-term partnership obligations.
For Broadcom, this is a significant win. The company has quietly built a strong position as the go-to partner for hyperscalers who want custom silicon without standing up an entire chip design operation from scratch. Google has used Broadcom for its TPUs. Now Meta is deepening that same kind of relationship. Broadcom gets a locked-in customer with enormous volume commitments through 2029 — that’s not a bad place to be.
The Bigger Spending Picture
This deal doesn’t exist in a vacuum. Hyperscalers — the Metas, Googles, Microsofts, and Amazons of the world — are projected to spend somewhere between $635 billion and $665 billion on AI infrastructure in 2026 alone. That’s a 67% jump from 2025 spending levels. The money flowing into AI compute right now is staggering, and every major player is trying to figure out how to spend it efficiently.
Custom silicon is a big part of that efficiency equation. Off-the-shelf GPUs from Nvidia are powerful, but they’re also expensive, supply-constrained, and designed to serve a broad market. When you’re operating at Meta’s scale, a chip tuned specifically to your workloads — your models, your inference patterns, your data center architecture — can deliver meaningfully better performance per watt and per dollar. That’s the thesis behind MTIA, and the Broadcom partnership is how Meta gets there at scale.
What This Means for the AI Chip Space
The narrative around AI chips has been dominated by Nvidia for the past few years, and for good reason — their hardware has been the default choice for training large models. But the space is shifting. Inference workloads, which is where a lot of the day-to-day AI compute actually happens, have different requirements than training. Custom chips can be optimized specifically for inference in ways that general-purpose GPUs aren’t.
Meta running billions of daily interactions across Facebook, Instagram, WhatsApp, and its various AI products generates an enormous inference workload. If MTIA chips — co-designed with Broadcom — can handle that more efficiently than buying Nvidia hardware at scale, the economics become very compelling very fast.
This also puts pressure on other hyperscalers to accelerate their own custom silicon efforts. Amazon has Trainium and Inferentia. Google has its TPU line. Microsoft has been investing in custom silicon too. Meta locking in a 2029 partnership signals that this isn’t a short-term experiment — it’s a long-term infrastructure strategy, and everyone else in the space is watching.
The Honest Take
From where I sit reviewing AI tools and infrastructure, this deal is a smart move by Meta — not because custom silicon is some magic solution, but because at 1 GW of deployment scale, even marginal efficiency gains translate into enormous cost savings and competitive advantages. The risk is real: custom chip programs are expensive, slow, and can fail. But Meta has the engineering talent and the financial resources to absorb that risk in a way most companies can’t.
Broadcom, meanwhile, gets to play a quieter but very profitable role in the AI infrastructure buildout without needing to own the spotlight. That’s a solid business to be in right now.
The 2029 timeline is what stands out most to me. That’s not a short-term hedge — it’s a declaration that Meta is building its AI future on its own silicon, and it’s willing to commit years of engineering and capital to prove it out.
🕒 Published: