dylan patel

Inference could absorb multiple percentage points of global GDP as token consumption becomes embedded across economic activity. Continued software and model optimization is simultaneously cutting equivalent-quality inference costs by roughly 60 times per year, expanding viable use cases.

Optimizing hardware, systems software and model architecture independently can produce incremental gains, but designing them together can turn several twofold improvements into a hundredfold result. This favors vertically coordinated labs and semiconductor platforms over isolated point solutions.

Twenty gigawatts of compute should come online this year and more than thirty next year, yet model improvements are expanding the pool of economically valuable tasks even faster. The compute crunch persists while capability growth outpaces capacity deployment.

A single researcher improving training efficiency by 5% saves 5% across the entire inference fleet, making nine-figure compensation economically rational. Talent concentration creates winner-take-all dynamics where process knowledge in few hands determines competitive outcomes.

Next-gen chips (MI455X, Vera Rubin, Jaguar Shores) standardize on HBM4 with 20+ TB/s bandwidth, while Qualcomm's AI 250 pursues compute-near-memory with LPDDR6 — memory bandwidth and capacity becoming the primary differentiation vector for 2026-2027 AI accelerators.

OpenAI and Anthropic will consume 40-50% of new compute in 2025 and over half by late 2025, driven by revenue per megawatt of $50-100M that lets them outbid all other customers for scarce capacity.

Anthropic and OpenAI are accelerating their share of global incremental compute from 30% in 2024 to 40-50% in 2025 and 70-80% by 2028, driven by superior revenue per megawatt ($50-100M+) that lets them outbid all other compute consumers.

Training and inference demand drives 1-10 GW campus builds costing $50-500B each, creating a new asset class where power, land, and permitting — not just GPUs — are the binding constraints.