A simple average of GPU prices is useless because 'two H100s' can have different CPUs, RAM, and locations. A valid index requires ingesting thousands of daily prices and normalizing them against a base case, using a model that identifies key price-driving factors. This is crucial for creating a reliable hedging instrument.

Emerging cloud providers (“NeoClouds”) are sticking exclusively with NVIDIA, despite alternatives from AMD. The perceived performance risk is too high, as customers demand state-of-the-art inference speed and providers can't risk a multi-billion dollar investment on a non-NVIDIA stack that might offer lower throughput.

New AI models are designed to perform well on available, dominant hardware like NVIDIA's GPUs. This creates a self-reinforcing cycle where the incumbent hardware dictates which model architectures succeed, making it difficult for superior but incompatible chip designs to gain traction.

Publicly announcing the number of GPUs a lab possesses is "bravado" and a poor indicator of its actual power. True capability is measured by model output and performance, as compute utilization varies wildly. Focusing on inputs instead of outputs is a common mistake.

The key metric for AI chips (GPUs/TPUs) is achieving a high percentage of theoretical peak performance (e.g., 70-80%). This concept, known as "mechanical sympathy," is largely absent in the CPU world, where software performance is so inefficient that measuring against peak is considered nonsensical.

While many focus on compute metrics like FLOPS, the primary bottleneck for large AI models is memory bandwidth—the speed of loading weights into the GPU. This single metric is a better indicator of real-world performance from one GPU generation to the next than raw compute power.

Anthropic mitigates supply chain risk and optimizes cost by investing heavily in the ability to use NVIDIA, Google, and Amazon chips interchangeably for model development, internal use, and customer service. This orchestration layer is a key competitive advantage.

AI performance engineer Chris Fregley warns that developing on local machines or even consumer-grade GPUs is a waste of time. Critical differences in hardware, memory bandwidth, and drivers mean that accurate profiling and optimization can only be done on the exact production systems, like NVIDIA's Blackwell or Hopper GPUs.

A futures market for GPU compute is not viable yet because the product isn't fungible. The performance of an identical H100 chip varies significantly between cloud providers based on their proprietary software stack and operational excellence, measured by metrics like "goodput" and "MFUs."