Anthropic is in discussions to purchase AI chips from a U.K.-based startup, according to The Information. The report signals another major lab moving to diversify its silicon supply beyond the dominant Nvidia stack, and it lands at a moment when compute access is the single biggest constraint on AI progress.
The Information broke the story, framing it as part of a broader push by frontier labs to lock in alternative hardware sources. Anthropic already runs Claude on a mix of Nvidia GPUs, Google TPUs, and AWS Trainium chips. Adding a U.K. supplier to that list would extend the pattern: hedge against any single vendor, any single supply chain, any single geography.
Why this matters
Compute is the new oil, and right now Nvidia controls most of the wells. Every major lab is scrambling to reduce that dependency. Anthropic has been the most aggressive about it on the cloud side, splitting workloads across three providers. A direct chip deal with a smaller startup is a different play: it’s a bet on alternative architectures, not just alternative buyers of the same Nvidia silicon.
The move also fits Anthropic’s recent posture. The company has been signing massive compute deals all year, including a reported multi-billion-dollar arrangement with Google for TPU capacity and an expanded partnership with Amazon. Layering in a U.K. startup adds optionality without betting the company on it.
The U.K. AI chip scene
The U.K. has a small but credible pool of AI silicon companies. Graphcore, once valued at $2.8 billion, was acquired by SoftBank in 2024 and has been retooling its IPU architecture for the transformer era. Other names include Myrtle.ai, Blaize, and a handful of stealth-stage startups working on inference-specific designs.
The Information didn’t name the specific company in the early reporting, but the industry context is clear:
- Inference is the new battleground. Training runs grab headlines, but inference now eats the majority of compute budgets at deployed AI labs. Specialized inference chips can deliver dramatically better cost-per-token than general-purpose GPUs.
- Custom silicon partnerships are heating up. OpenAI is reportedly working with Broadcom on a custom chip. Meta has its MTIA program. Google has TPUs. A startup deal lets Anthropic get into the custom-silicon game without building an internal chip team from scratch.
- Geopolitics matters. U.S. labs sourcing from U.K. firms gives them a hedge against any future export-control or supply-chain disruption tied to Asia-based fabrication.
What stands out
This is significant because it suggests Anthropic sees real performance or economic upside outside the Nvidia ecosystem. Smaller chip startups have struggled for years to win meaningful contracts from frontier labs. Most labs default to Nvidia because the software stack works, the supply is (mostly) reliable, and engineering teams already know CUDA. A startup chip has to clear a high bar to be worth the integration cost.
If Anthropic is willing to do that integration work, it implies the U.K. chip offers something the current options don’t. That could be inference cost, memory bandwidth, energy efficiency, or simply better availability at a moment when every Nvidia H100 and B200 is spoken for.
What to watch next
- Timing. Talks reported in the press often take months to close. Watch for an official announcement, ideally with deployment specifics.
- Workload type. If the chips are earmarked for inference, that confirms the cost-optimization angle. If they’re for training, that’s a much bigger bet.
- Volume. A pilot order is a hedge. A nine-figure commitment would reshape the U.K. AI hardware landscape.
- Software stack. Whatever lab Anthropic picks will need to port models off the Nvidia/CUDA path. The amount of porting effort tells you how serious this is.
Practitioners building on Claude shouldn’t expect immediate changes. Anthropic’s API, pricing, and latency are unlikely to shift on the back of a single supply deal. But over the next 12 to 24 months, expect more of these stories. The race to escape Nvidia’s gravitational pull is on, and Anthropic is moving faster than most.
For the full breakdown, see the original report at The Information.