Claude's Compute Appetite Is Larger Than Expected
Last week, Anthropic and SpaceX struck a deal that shocked the industry: SpaceX will provide Anthropic with $15 billion worth of AI compute annually. This deal instantly catapulted Anthropic into the ranks of super-buyers in the AI compute market.
But apparently, that's still not enough.
According to the latest report from The Information, Anthropic has already begun early-stage negotiations with Microsoft to rent Azure servers equipped with Microsoft's in-house Maia 200 chips.
What does this mean? It means Claude's compute demands have grown so large that even Google alone can no longer supply them.
From Google's "Favored Partner" to Hedging Bets Across Multiple Players
The relationship between Anthropic and Google was once one of the closest partnerships in the AI industry. Google was not only an initial investor but also Anthropic's primary compute provider—Claude's model training and inference heavily relied on Google's TPUs (Tensor Processing Units).
But now, Anthropic is actively "de-Google-ifying" its infrastructure.
This shift is driven by several factors:
Supply chain security. Betting all compute resources on a single supplier carries too much risk. If Google were to adjust TPU allocation priorities for any reason, Anthropic would be left in a vulnerable position. Diversifying compute suppliers is an inevitable choice.
Microsoft's Maia 200 is attractive. While Microsoft's in-house Maia 200 chip may not match the latest TPUs in training speed, it is "specifically designed to run existing models, like Claude." In other words, the Maia 200 could be a more cost-effective option for inference workloads.
Commercial maneuvering. Anthropic's relationship with Microsoft has always been characterized by a mix of close collaboration and strategic distance—on one hand, Anthropic needs Microsoft's compute and distribution channels; on the other, Microsoft also invests in OpenAI, creating a competitive dynamic. But in an era of compute shortages, even rivals can become suppliers.
Why Inference, Not Training?
The report specifically notes that the Maia 200 chip is "not as fast as TPUs when it comes to training new models." This suggests Anthropic may primarily use the Maia 200 for inference—that is, having already-trained Claude models answer user queries—rather than for training next-generation models.
This distinction is crucial:
- Training requires massive-scale compute and high-speed interconnects, where TPUs remain the undisputed leader.
- Inference places relatively lower demands on single-chip performance but is far more sensitive to cost and latency. The Maia 200 could offer a stronger cost advantage in inference scenarios.
As Claude's user base grows, the demand for inference compute is rising exponentially. Training a model is a one-time, massive capital expenditure, whereas inference is a continuous, daily operational cost.
The Bigger Picture: The AI Compute War Is Escalating
Anthropic's move is not an isolated incident. The entire AI industry is undergoing a compute arms race:
- OpenAI is deeply tied to Microsoft while also exploring the construction of its own compute infrastructure.
- xAI's Colossus cluster is expanding, and Musk's "cosmic-scale" compute ambitions don't stop there.
- Google continues to scale TPU production while simultaneously selling compute to third parties (including Anthropic's competitors).
- Amazon's Trainium and Inferentia chips are also catching up.
In this race, Anthropic, as a relatively "younger" player, is securing its compute supply through a diversified supply chain strategy.
Impact on the Industry
Anthropic's chip negotiations with Microsoft may look like a single company's business decision, but they actually reflect several deeper trends in the AI industry:
Compute is becoming the biggest bottleneck in AI. Innovations in model architecture and algorithmic optimization are taking a backseat to a simple physical reality: whoever has more GPUs/TPUs will move faster.
Competition among cloud providers will intensify. As top-tier AI companies like Anthropic start "shopping around" between Google and Microsoft, price wars and technological races among cloud vendors will only grow fiercer.
The strategic value of in-house chips is rising. Microsoft's Maia 200 may not be the fastest AI chip on the market, but because it is designed and controlled entirely by Microsoft, the company enjoys greater flexibility in pricing and supply prioritization. This autonomy is incredibly valuable in an era of compute scarcity.
A Question Left for Anthropic
While hedging its bets across multiple suppliers, Anthropic must answer a strategic question: As your compute sources become increasingly fragmented, can your technical roadmap remain consistent?
Compatibility across different chip architectures, migration costs, and optimization strategies could all become hidden costs for Claude's future development.
However, when faced with the questions of "whether there is enough compute" versus "where the compute comes from," the latter is clearly a "happy problem" to have.