Anthropic recently poached "employee #2" of OpenAI's in-house chip project, Clive Chan, from OpenAI. His LinkedIn position description is just one sentence: "perplexity per picojoule" — putting the model's predictive capability and the unit of energy consumption into the same optimization target.
This sentence reflects the migration of the large-model evaluation paradigm. Traditional metrics FLOPS, tokens/sec, and MMLU focus on "compute fast," while perplexity per picojoule puts energy consumption front and center. Three forces are behind it: scale is hitting the power wall — GPT-5.4 and Claude Mythos at 256K context have inference energy consumption approaching the data-center carrying capacity; hardware-software co-design is back — Anthropic is evaluating in-house ASICs plus Chan's familiarity with the OpenAI-Broadcom in-house chip project; and on-device AI forces energy efficiency first — Anemll, Ollama MLX, and the on-device models WWDC is betting on make "tokens per joule" a product-level metric.
perplexity per joule-class metrics already appeared on arXiv in 2025 (d-Matrix's roofline modeling and hardware co-design papers), not first coined by Anthropic. But when a top lab writes it into a recruiting JD and assembles a dedicated team, it means it has moved from academic discussion to industrial-scale deployment.
The future model selection criterion may shift from "how many MMLU points" or "cost per thousand tokens" to "how accurate can you run under a fixed power budget," in turn driving the co-evolution of sparse MoE, low-rank approximation, 4/2/1.58-bit quantization, and speculative decoding. It can be expected that from the second half of 2026, "perplexity per joule" will, like the cost-per-token of yesteryear, become the new benchmark for cloud-vendor comparison of LLM inference cost-performance.