Broadcom Locks Google TPU Deal Through 2031, Bets 3.5 GW on Anthropic

TL;DR

• Broadcom inked a multi-year deal with Google to develop and supply custom TPUs and AI hardware through 2031.
• The company’s expanding its Anthropic partnership to deliver 3.5 gigawatts of next-gen TPU-based AI compute starting in 2027.
• The move — formalized via SEC Form 8-K — positions Broadcom against NVIDIA and AMD in the custom silicon arms race for frontier AI labs.
• Secures massive compute capacity for Google and Anthropic amid global chip shortages and sustained hyperscaler investment in bespoke hardware.

Broadcom and Google Extend TPU Partnership Through 2031

Broadcom announced a multi-year agreement with Google to develop and supply custom Tensor Processing Units and AI hardware through 2031. The deal cements Broadcom’s role as a key supplier in Google’s AI infrastructure buildout, extending a collaboration that’s already underpinned multiple generations of TPU silicon.

The company disclosed the agreement in an SEC Form 8-K filing, signaling the materiality of the commitment. Financial terms weren’t disclosed, but the timeline — stretching six years into the future — reflects the kind of long-term capacity planning hyperscalers need to keep pace with AI workload growth.

Broadcom also revealed it’s expanding its partnership with Anthropic, the AI safety-focused lab behind Claude. Starting in 2027, Broadcom will supply 3.5 gigawatts of next-generation TPU-based AI compute to Anthropic. That’s not a typo — 3.5 gigawatts, enough power to run a small city, now earmarked for training and inference workloads.

Why Broadcom’s 3.5-Gigawatt Bet Reshapes AI Infrastructure

This isn’t just a supply contract. It’s a statement about where the AI hardware market is heading — and who’s building the rails.

For years, NVIDIA has dominated AI accelerator sales with its GPUs, racking up market share while hyperscalers scrambled for H100s and H200s. But Google never fully bought into that dependency. Instead, it’s spent the better part of a decade designing its own TPUs, partnering with Broadcom to fabricate the silicon and integrate it into custom AI systems.

Now Broadcom’s doubling down on that strategy, locking Google in through 2031 and throwing 3.5 gigawatts of capacity at Anthropic. The message? Custom silicon isn’t a side bet anymore. It’s the main event.

And here’s the thing: Anthropic isn’t some scrappy startup tinkering in a garage. It’s a frontier AI lab reportedly valued north of $18 billion, backed by Google, and racing OpenAI to build the next generation of large language models. Securing 3.5 gigawatts of dedicated compute gives Anthropic the kind of infrastructure certainty that’s nearly impossible to find in today’s supply-constrained market.

Think of it like this — it’s the difference between renting a hotel room every night and owning the building. Anthropic now owns the building, metaphorically speaking, and Broadcom’s the contractor who just agreed to keep expanding it for the next three years.

The competitive stakes are obvious. NVIDIA’s still the 800-pound gorilla, but AMD’s clawing for share with its Instinct accelerators, and now Broadcom’s carving out a lane supplying custom silicon to the hyperscalers and AI labs that don’t want to depend on off-the-shelf GPUs. If you’re Google or Anthropic, you’re betting that bespoke hardware — optimized for your specific workloads — beats general-purpose chips in the long run.

I’d argue they’re right. General-purpose GPUs are incredible, but they’re also designed to do everything reasonably well, not one thing exceptionally well. TPUs, by contrast, are built from the ground up for tensor math and transformer architectures. That specialization pays dividends when you’re burning through petaflops training Claude 4 or running inference at scale.

But there’s a flip side. Custom silicon locks you into a vendor and a roadmap. If Broadcom stumbles on a node transition or Google’s TPU architecture hits a performance wall, Anthropic can’t just swap in NVIDIA chips overnight. That’s the trade-off — more control, less flexibility.

Hyperscaler Custom Silicon Strategies Accelerate

Broadcom’s deals with Google and Anthropic fit into a broader pattern reshaping the AI hardware landscape. Hyperscalers aren’t just buying chips anymore — they’re designing them.

Google pioneered this approach with TPUs, but it’s no longer alone. Amazon’s built its Trainium and Inferentia chips for AWS. Microsoft’s working on its own AI accelerators, reportedly codenamed Athena. Meta’s designed custom silicon for inference workloads. Even startups like Cerebras and SambaNova are pitching purpose-built AI systems.

The logic is straightforward: if you’re spending billions on compute, you want hardware tailored to your workloads, not a general-purpose chip that does a hundred things you don’t need. Custom silicon also insulates you from NVIDIA’s pricing power and supply constraints, both of which have been pain points for AI labs over the past two years.

Broadcom’s role in this shift is critical. It doesn’t design the chips — Google and Anthropic do that — but it handles the fabrication, packaging, and integration. That makes Broadcom a linchpin in the custom silicon supply chain, and these multi-year deals lock in revenue streams that are harder to disrupt than one-off chip sales.

The SEC filing underscores just how seriously Broadcom’s taking multi-gigawatt AI workloads. A Form 8-K is reserved for material events — things investors need to know about immediately. That Broadcom filed one for these deals signals they’re not incremental expansions. They’re strategic bets.

What Broadcom’s 2027 Deadline Reveals About AI Compute Timelines

The 2027 start date for Anthropic’s 3.5-gigawatt TPU capacity is worth unpacking. That’s not next quarter or even next year — it’s three years out. Why the lag?

Building multi-gigawatt AI infrastructure isn’t like ordering servers off a rack. It requires custom chip designs, advanced packaging, data center construction, power grid upgrades, and cooling systems that can handle densities most facilities weren’t built for. Three years is actually aggressive for a buildout of this scale.

It also tells you something about Anthropic’s roadmap. The company’s planning for workloads in 2027 and beyond that will dwarf what it’s running today. That kind of foresight — and capital commitment — is only possible when you’ve got deep-pocketed backers and a clear vision of where model scaling is headed.

For Broadcom, the timeline spreads risk and revenue. It’s not delivering 3.5 gigawatts overnight — it’s ramping production and deployment over multiple years, which smooths out manufacturing challenges and capital expenditure.

But it also creates a window for competitors. If NVIDIA or AMD can deliver comparable performance and availability before 2027, they could chip away at Anthropic’s reliance on TPUs. If Broadcom hits delays or yield issues, the whole timeline slips. Custom silicon is powerful, but it’s also unforgiving when things go sideways.

Three things to monitor as this plays out. First, watch whether other AI labs follow Anthropic’s lead and lock in multi-year custom silicon deals. If Cohere, Mistral, or xAI ink similar agreements with Broadcom or competitors, it signals the industry’s moving away from NVIDIA’s GPU monoculture faster than expected.

Second, keep an eye on NVIDIA’s response. The company’s not going to cede the custom silicon market without a fight, and it’s already offering more customization options through its Grace Hopper Superchip and NVLink partnerships. If NVIDIA starts undercutting Broadcom on price or performance, the calculus for custom silicon shifts.

Third, track Broadcom’s execution. Delivering 3.5 gigawatts of TPU capacity by 2027 is a massive operational challenge, and any stumbles — yield issues, packaging bottlenecks, power delivery problems — will ripple through Anthropic’s roadmap. The SEC filing commits Broadcom publicly, which means investors and customers will be watching closely.

Source: minichart.com.sg