Historical Behavior

OpenAI's historical behavior in model optimization shows an evolution from "scale-first to efficiency-first". In 2022, GPT-3.5 relied on brute force expansion (175 billion parameters), in 2023, GPT-4 introduced a mixture of experts architecture (MoE), in 2024, GPT-4o optimized multimodal inference, and in 2025, GPT-5 achieved "self-optimization" for the first time (the model improves its own inference code). In 2026, GPT-5.6 Sol is the first model with "self-iterative efficiency"—optimizing its own service stack without relying on manual engineering. Compared to Anthropic (which still relies on manual optimization in 2026) and Google (which launched AutoML optimization in 2025), OpenAI's "self-optimization" path is more aggressive.

Capital Path

The capital impact of efficiency optimization is clear: OpenAI's inference costs in Q2 2026 are projected at $1.2 billion, and if GPT-5.6 Sol optimization is fully deployed, it could save $960 million annually (20% × $1.2 billion × 4 quarters). A 15% increase in token generation efficiency means that the same GPU cluster can handle 15% more requests or reduce 15% of new purchases (with plans to purchase 500,000 H200 units in 2026, this could be reduced to 425,000 units, saving $5.1 billion). Funding is shifting from a "computational arms race" to an "efficiency arms race", challenging NVIDIA's narrative of "sales growth" while the narrative of "unit computational efficiency" rises.

Analogy + Industry Position

Similar cases include: in the 2010s, data center PUE (Power Usage Effectiveness) dropped from 2.0 to 1.2, saving 40% in electricity costs, but server sales growth slowed. In the 2020s, chip processes have shrunk from 14nm to 3nm, improving performance by 5 times, but transistor density growth has slowed. Currently, AI inference is in an "efficiency optimization explosion period"—compared to 2023 (where model scale increased 10 times annually), in 2026, scale is expected to increase 2 times annually, but efficiency is expected to increase 3 times annually. The industry is in a "post-scale era"—the market is shifting from a "parameter quantity competition" to a "unit cost intelligence" competition. OpenAI's "cost-intelligence curve" essentially means "buying more intelligence with $1 of computational power", rather than "buying the same intelligence with more computational power".

Structural Judgment

This fundamentally represents a technological substitution leading to a reconstruction of the value chain. The traditional AI value chain is "chip → cloud service → model → application", with value concentrated in chips (NVIDIA gross margin 75%) and cloud services (AWS gross margin 30%). After efficiency optimization, the value chain is reconstructed as "chip → efficiency tools → model → application", with value shifting towards "efficiency tools" (compilers, inference engines, model compression). This explains why OpenAI is self-optimizing—not to reduce costs, but to "vertically integrate the efficiency stack". When models can optimize their own inference, the value of manual optimization decreases by over 50% (OpenAI's inference engineers are projected to be 1,200 in 2026, which could reduce to 600 by 2027). This drives the "AI self-optimization paradigm"—models are not only products but also production tools, with the marginal cost of using AI to optimize AI approaching zero.

ABAB News · Cognitive Laws

Scale is the past, efficiency is the future
The marginal cost of AI optimizing AI approaches zero
The cost-intelligence curve is the new competitive barrier