Anthropic, the firm responsible for the AI model Claude, is currently negotiating with Microsoft to utilize Microsoft's specialized AI chips for inference workloads. This development points to a significant shift in how major AI laboratories assess the hardware that supports their models, impacting the industry beyond just Silicon Valley.
The Information first reported these discussions, which come shortly after a significant partnership involving Anthropic, Microsoft, and NVIDIA, committing a substantial $30 billion for Anthropic to acquire Azure computing resources.
What does the deal entail?
Running large language models involves high costs. While training these models is expensive, it is the inference—the process of the model responding to user queries—that incurs the ongoing major expenses. Each time users engage with Claude to draft emails or summarize documents, it involves inference, and scaling this leads to increased costs.
Anthropic's current focus on Microsoft’s custom silicon is aimed specifically at reducing these inference costs. The goal is clear: to lower the expenses linked with running Claude for millions of users.
It's important to note that Anthropic isn't turning its back on NVIDIA. The company will continue to utilize NVIDIA’s advanced computing infrastructure, optimizing models for efficient performance on this existing hardware. However, depending heavily on a single supplier like NVIDIA poses risks, and Anthropic seems to be hedging its bets against that.
Microsoft's CEO has framed their evolving partnership as mutually beneficial. Claude will be integrated into Microsoft's product ecosystem, giving Microsoft access to a leading AI model while allowing Anthropic to leverage Azure's extensive cloud resources and now possibly its proprietary silicon.
Why is diversifying silicon important?
Anthropic is exploring not only Microsoft's chips but also AI inference technology from Fractile, a UK startup focused on specialized hardware. This trend highlights Anthropic's desire for flexibility. Other major tech companies, referred to as hyperscalers, have established their own chips to minimize dependency on NVIDIA's GPUs. Microsoft is creating its AI accelerators to provide viable alternatives to conventional GPU costs.
For Anthropic, this strategy is straightforward. The high cost of top-tier NVIDIA GPUs, combined with tight supply, has led to fierce competition for these chips. By integrating Microsoft’s custom silicon and potentially Fractile's offerings, Anthropic strengthens its bargaining position while enhancing its infrastructure.
The company's commitment of $30 billion to Azure underscores the seriousness of its procurement, marking a strategic move in chip allocation discussions.
How does this affect the cryptocurrency space?
On the surface, this negotiation between two AI companies might seem irrelevant to the crypto sector. However, the ripple effects are significant. The competition between AI infrastructure influences the economics of decentralized computing networks. Initiatives like Render, Akash, and io.net, which offer cheaper GPU computing options, may find their advantages eroded if Microsoft and Anthropic successfully drive down inference costs through their chip development.
This situation will challenge decentralized networks not only to remain cost-competitive against NVIDIA's GPUs but also to navigate the landscape of specially designed inference chips.
Additionally, Microsoft's Azure is increasingly seen as foundational for numerous Web3 and AI-crypto hybrid projects. Anthropic’s integration into Azure, distributing Claude via Microsoft's solutions while operating on Azure’s infrastructure, enhances Azure’s position as a primary provider for enterprise-level AI workloads. Any Web3 endeavor integrating with Azure-based AI should closely monitor the progression of this collaboration.
Investors interested in AI-related crypto tokens should focus on trends in inference costs. If companies like Microsoft, Google, and Amazon continue to diminish these costs through their custom chips, tokenomics for projects relying on GPU compute margins may need recalibration. Successful ventures will likely be those providing unique services, such as censorship resistance or decentralized access, that large cloud providers overlook.
In conclusion, while Anthropic’s diversification of chip suppliers reflects a rational choice, it remains a strategy confined within the frameworks of major tech corporations. This situation creates both threats and opportunities for decentralized alternatives depending on their ability to cater to unmet needs in the market.