AMD could add Anthropic as a major Instinct MI450 customer in the midst of the fight for AI infrastructure

AMD could be close to scoring another major customer in the race for artificial intelligence infrastructure. According to some rumors, Anthropic would be considering incorporating the next AMD Instinct MI450 accelerators to its servers, a move that would strengthen the company’s position in the data center market for AI and that once again shows the extent to which large model developers seek to diversify hardware suppliers.

The information comes at a particularly sensitive time for the sector. Demand for training and inference capacity continues to grow rapidly, while the supply of advanced chips remains strained. In this context, depending on a single platform is increasingly less comfortable for companies like Anthropic, which need to expand infrastructure without being completely tied to the delivery rates, prices and availability of a single manufacturer.

AMD had already hinted that it works with clients of comparable scale to OpenAI for its current and future generations of Instinct. Now the name that appears is Anthropic, which fits well with that strategy due to its need for computing and the variety of agreements it is weaving to sustain Claude’s growth. More than a simple change of supplier, the possible operation points to an increasingly visible logic. We refer to combine several architectures to cover different needs within the same data center network.

AMD enters direct competition with NVIDIA

The Instinct MI400 series will be the next major generation of AMD accelerators for AI, with variants such as MI450X and MI430X based on the CDNA 5 architecture. The company has announced that these chips will arrive with more memory capacity and bandwidthnew calculation formats oriented to artificial intelligence and a clear commitment to rack-scale interconnections supported by open standards.

On paper, AMD talks about a product capable of reaching 40 PFLOPS in FP4 and 20 PFLOPS in FP8, in addition to a significant jump in memory thanks to the use of HBM4. The MI450 would also go up to 432 GB of memory, compared to 288 GB of the previous generation, and would offer much higher bandwidth. These are figures designed to convince large cluster operators that AMD can compete not only in power, but also in complete system efficiency.

The thing is that the company is no longer limited to selling chips, but is trying to place a broader infrastructure proposal, with racks, networks and software prepared for massive AI loads. If Anthropic ends up betting on the MI450, the contract would have commercial value, but also symbolic value. In short, it would serve to validate that AMD can enter frontier deployments where until recently NVIDIA dominated almost without discussion. An NVIDIA that, by the way, sees how other important clients are beginning to diversify the origin of their chips.

Anthropic expands partners and confirms that no one wants to depend on a single ecosystem

The potential deal also fits with Anthropic’s recent moves. The company already uses NVIDIA GPUs and has access to Amazon’s Trainium, while its new alliance with Google and Broadcom will open the door to next-generation TPU capacity starting in 2027. If AMD accelerators are added to the mix, the result would be a much more flexible strategy, designed to distribute loads between several platforms depending on availability, cost and type of task.

That says a lot about the moment the industry is experiencing. The battle is no longer just about having the best model, that too, but about ensuring enough silicon, memory, power and network to run it at scale. In this scenario, AMD has found a clear opportunity: to present itself as a credible alternative for companies that do not want to be left waiting for capacity while the market remains concentrated in a few suppliers.

For now this is a rumor and there is no official confirmation from Anthropic or AMD about this supposed adoption of the MI450. Still, the simple fact that Anthropic’s name is linked to this generation makes it clear that in the next phase of AI, both models and technology matter. hardware diversification. Obviously, AMD doesn’t want to miss the party.