The standard for measuring large compute deals has shifted from number of GPUs to gigawatts of power. This provides a normalized, apples-to-apples comparison across different chip generations and manufacturers, acknowledging that energy is the primary bottleneck for building AI data centers.

Andreessen asserts that the AI models we use daily are intentionally limited versions of what labs have developed. The primary constraint is not research progress but the severe shortage of GPU capacity. If compute were plentiful, current models would be significantly more powerful.

The relationship between computing power and AI model capability is not linear. According to established 'scaling laws,' a tenfold increase in the compute used for training large language models (LLMs) results in roughly a doubling of the model's capabilities, highlighting the immense resources required for incremental progress.

An AI model's operating environment—its "harness"—is now the primary driver of capability. Benchmarks show the same model achieves vastly different results in different harnesses, proving that the runtime, tools, and state management are as critical as the model's internal weights for achieving results.

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.

Judging an AI's capability by its base model alone is misleading. Its effectiveness is significantly amplified by surrounding tooling and frameworks, like developer environments. A good tool harness can make a decent model outperform a superior model that lacks such support.

The narrative of insatiable AI compute demand is partially a bubble. It's fueled by inefficient early models ("token maxing") and a culture where tech executives brag about their AI spending as a status symbol, a behavior not seen with traditional cloud costs. This suggests demand could normalize.

The report of XAI's low GPU utilization reveals a critical, non-obvious bottleneck in AI: it's not just about acquiring compute, but using it efficiently. This 'FLOPS utilization' problem, caused by architectural and load-balancing issues, means billions in hardware sits underused, creating an opportunity for companies that can optimize the compute stack.