🧠 Food for thought

Implications, context, and why it matters.

The new hire’s experience outlines Anthropic’s infrastructure plan

• Eric Boyd previously served as corporate vice president for Microsoft’s AI Platform. He described Microsoft deploying tens of thousands of Nvidia graphics processing units (GPUs) tied together with a high-throughput, low-latency InfiniBand network, a high-speed system used to link servers in AI data centers, while it built Azure AI infrastructure for OpenAI 1.
• Boyd said the job required system-level tuning to keep workloads running across thousands of GPUs for weeks. He also said Microsoft learned through close work with OpenAI on training setups as OpenAI released GPT-4 1.
• One analyst framed Anthropic’s approach as a shift from being “investment light” to building its own infrastructure 2. Anthropic hired Boyd while it pushes a first large effort to build custom infrastructure and announced plans to spend US$50 billion on U.S. AI infrastructure with Fluidstack, a cloud provider focused on supplying large-scale AI computing clusters 3.

This hire marks a move from renting clouds to building digital factories

• The AI contest now includes physical buildouts plus the engineers who can run them at scale.
• Top AI companies are moving from standard cloud rentals to custom data centers for lower costs and faster performance, which an analyst described as a “means of saving money” 2.
• This shift is lifting “neocloud” providers that sell AI computing capacity. Fluidstack supplies large-scale GPU clusters to Meta, Midjourney, and Mistral, a French AI startup 4.
• In the third fiscal quarter, Fluidstack ranked third for leased data center capacity behind Oracle and Google. It placed ahead of Meta, Amazon, CoreWeave, and Microsoft, according to a TD Cowen report cited by Yahoo Finance 5.