AMD is investing up to $5 billion in Anthropic, the AI lab behind Claude, the two companies said on Wednesday. Anthropic will in turn run Claude on a large new fleet of AMD’s top-end AI chips — one of the biggest single chip commitments of the AI build-out so far, and a direct challenge to Nvidia’s grip on the silicon that trains frontier models.
The headline number is the cash. The substance is the capacity: up to two gigawatts of new AMD chips, deployed in stages from the first half of 2027.
$5bnAMD’s bet on Anthropic — and on becoming a real second supplier for frontier AI compute
What AMD is actually getting #
Anthropic will run Claude’s training and serving on AMD’s flagship chips, packaged inside Helios, AMD’s rack-scale system. The first gigawatt of capacity is scheduled to come online in the first half of 2027.
Alongside the silicon, the two companies are starting a multi-year engineering partnership. Anthropic will help AMD tune its ROCm software — the open-source toolkit that lets AI models run on AMD chips rather than Nvidia’s — and optimise GPU workloads. AMD, in return, will roll Claude out internally across its chip, software and product teams.
Tom Brown, Anthropic’s chief compute officer and a co-founder, framed the deal as a deliberate multi-vendor strategy: matching each workload to the hardware best suited to it, rather than concentrating everything on one supplier.
AMD’s bigger play #
The Anthropic agreement sits at the centre of AMD’s attempt to become the credible second supplier for frontier AI compute. AMD has now struck similar partnerships with Meta and OpenAI, each one a vote of confidence in its accelerator roadmap and the systems built around it.
Until now, the most ambitious AI labs have effectively had a single choice for serious training compute: Nvidia. A real second supplier changes the bargaining on price, on supply, and on which hardware each lab’s next model actually trains on.
For AMD, the upside is straightforward — large, multi-year purchase commitments for its top-end silicon. For Anthropic, the bet is that having two serious suppliers rather than one keeps capacity available when demand spikes and pricing pressure honest when it doesn’t. Anthropic has previously signed compute deals with Google, Amazon, Broadcom and SpaceX’s data-centre arm, and is still negotiating with Meta.
The circular financing question #
Coverage in The Decoder flags a familiar criticism of these deals: chip and cloud companies are funding the AI labs, which then turn around and spend that money on those very chips and clouds. Whether the labs can ever cover those bills from customers alone is the open question the industry still has not answered.
For UK teams paying for Claude through the Anthropic API or a Claude.ai plan, the read-through is indirect but real. Anthropic recently halved subscription limits on some Claude Fable 5 tiers — a sign that even a well-funded lab is watching its unit economics. A second big supplier of compute, in theory, gives the lab more room to keep pricing competitive and capacity available.
What to watch #
The deal only matters if the silicon arrives and the software holds up. Three shifts worth following over the next eighteen months:
Whether the chips ship on time at scale. AMD has historically lagged Nvidia on delivery dates and on the software ecosystem around its accelerators. A 2GW roadmap is only as good as the silicon landing in the rack and the ROCm stack standing up to real frontier-AI workloads.Whether the diversification actually holds. Anthropic still leans heavily on Nvidia-powered systems via Google, Amazon and Azure. The AMD deal adds capacity; it doesn’t yet replace Nvidia. Watch the mix over the next two model generations.Whether AMD’s broader software push reaches the rest of us. The same ROCm work that helps Anthropic should, over time, improve local AI on AMD hardware — including the AMD GPUs a UK small team might already own.Our piece on running local AI on AMD in 2026is a starting point.
The bigger picture is where this points: AMD is now the chip company serious challengers are publicly betting on, and what frontier AI costs the rest of us over the next two years will be shaped by whether that bet pays off.
Sources & quotes #
Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. How we verify →