Three cards from two operators. Two themes connect them: how the AI cost curve is reshaping where demand concentrates, and where the new bottlenecks appear as adoption scales.
The AI market is not a pyramid
Tom Tunguz published "The Most Important Market in AI Is the Middle" on September 23, 2026. The core claim in AI intelligence demand follows a normal distribution, not a pyramid; mid-tier models dominate commercial spending because enterprise requirements are stable while AI costs fall exponentially is a structural reframe of where AI revenue lives.
The common picture is a pyramid: a small group pays frontier prices, most pay less, a long tail pays almost nothing. Tunguz's data shows a different shape. Mid-tier models captured 40% of enterprise AI spending and 30% of tokens. Claude Fable 5.1 captured only 3.7% of gateway spending in its first 12 days. Frontier consumption fell from 53% of gateway volume in early August to 45% in September.
The mechanism is clean. Enterprise requirements for AI are fixed: a sales classification workflow needs to be "smart enough," not frontier-level. AI costs fall exponentially. The tier that satisfies the fixed requirement at the lowest cost holds the volume.
"Demand for intelligence is not a pyramid with a small, wealthy peak paying for everything beneath it. It is a normal distribution with a fat middle."
"Intelligence costs keep plummeting. What enterprises demand from AI does not change nearly as fast."
This card builds on Enterprise AI workloads run overwhelmingly on non-frontier models because buyers optimize for price-over-performance, not benchmark scores (August 14, 2026), which documented 84% of tokens running on non-SOTA models. The September post extends that observation with a structural theory and newer data showing the same pattern accelerating.
- Tom Tunguz, AI intelligence demand follows a normal distribution, not a pyramid; mid-tier models dominate commercial spending because enterprise requirements are stable while AI costs fall exponentially. AI demand is normally distributed; mid-tier models capture 40% of enterprise spending because stable enterprise requirements meet exponentially falling AI costs at the middle tier.
What engineers do now
"Thinking in Systems, Shipping in Loops" appeared September 24, 2026. The claim in AI shifted the engineer's core task from writing correct code to designing feedback loops that let AI write correct code at scale is about a structural shift in the engineering job.
Dan Shiebler's Artemis Security is the clearest evidence. The company merged 16 pull requests daily by August. That number was 2 in January, 6 in May. Total in eight months: 30,000.
"Every line of code in our platform is written by AI agents. Our engineers design systems, set constraints, and review outputs."
Lauren Tan ships 2,000 pull requests monthly at SpaceX's Grok team, approximately 100 per working day. The multiplier is not faster typing. It is verification loops that let AI validate its own work before human review.
Tunguz frames the shift as three properties of good feedback systems: resilience (tests that let AI catch its own errors), self-organization (loops that capture failure modes), and hierarchy (scoped component boundaries). Teams that design these systems compound. Teams that skip them hit a ceiling where output volume grows but output quality doesn't.
David Heinemeier Hansson states the economic case plainly:
"Writing code by hand is no longer economically productive for most programmers at most companies"
This connects to As AI handles coding reliably, most software bugs shift from implementation errors to specification errors. Ball showed that specification bugs dominate as the primary failure class when AI handles implementation. Feedback loops that verify technical correctness don't catch specification errors. Systems design is also the answer: loops that force specification clarity before AI writes code.
- Tom Tunguz, AI shifted the engineer's core task from writing correct code to designing feedback loops that let AI write correct code at scale. AI-era engineering shifted from writing correct code to designing the feedback systems that let AI write correct code; the three system properties are resilience, self-organization, and hierarchy.
Decision models as evaluator judges
The LangChain engineering team published "Jev-as-a-Judge for Agent Evals" on September 20, 2026. The finding in Decision models outperform LLM judges in agent evals at 1,000 to 80,000 times lower cost because they output typed verdicts directly rather than generating reasoning text extends the decision model mechanism from production classification to the eval pipeline.
The Sep 22 log captured Once an AI system is hardened and repeats at scale, specialized decision models replace frontier models at a fraction of the cost and higher accuracy (Tunguz, September 21), which established that decision models outperform frontier LLM classifiers at 76 to 209 times lower cost. The September 20 LangChain post applies the same mechanism one layer up: to evaluation itself.
Evaluation is a classification problem, not a generation problem. An LLM judge generates a reasoning chain token-by-token before embedding a verdict in text. Two sources of variance enter: the stochastic generation process, and the extraction step that parses a verdict from unstructured output. A decision model skips both.
"Autoregressive models...reach a judgment through token-by-token generation. In our experiment, that decision-first design coincided with lower latency, lower cost, and lower variance."
Cost per evaluation call: Jev at $0.00035, versus $0.34 to $28.17 for frontier LLM judges. At that cost gap, teams can evaluate all traces rather than sampling. Full-trace evaluation becomes economically viable for the first time.
The tradeoff Decision models that output typed numbers instead of text remove the reasoning trace, making decisions fast and cheap but unexplainable identified still applies. Decision models return a verdict without a reasoning trace. For teams that need to understand WHY a response failed, the typed output alone is not enough. For teams that need to know WHETHER it failed at scale, it is sufficient.
- Harrison Chase (with Daniel Shea and Seán Roche), Decision models outperform LLM judges in agent evals at 1,000 to 80,000 times lower cost because they output typed verdicts directly rather than generating reasoning text. Decision models outperform LLM judges in agent evals at 1,000 to 80,000 times lower cost; typed output eliminates the variance from autoregressive generation and verdict extraction.