GPT-4-class inference cost about $30 per million tokens when it launched in 2023. By mid-2026, equal or better capability costs under $0.50 — a roughly 1,000x decline in three years. Over that same window, the share of organizations with a Chief AI Officer nearly tripled, and spending on AI governance platforms is racing toward a billion dollars a year. Letter 161 argued platforms beat models. Letter 162 argued that surviving founders build something hard to copy. This letter argues the thing worth building has quietly changed: intelligence itself is becoming a commodity, and trust is what's left that still commands a price.
When GPT-4 launched in March 2023, generating a million tokens of input cost roughly $30. By mid-2026, equal or better capability — from open-weight models, from smaller specialized systems, from aggressive new entrants — costs under $0.50 per million tokens. That's not a gradual decline. It's closer to a 1,000x collapse in three years, and one research firm tracking it calls the pattern "LLMflation": inference cost falling on the order of 10x annually, year after year. Meanwhile, DeepSeek's V4-Pro model scores within two-tenths of a point of Anthropic's Claude Opus 4.7 on SWE-bench Verified, the standard real-world coding benchmark — at roughly one twenty-eighth of the price. Raw intelligence, in other words, is turning into a commodity in real time.
If intelligence itself is collapsing in price, enterprise AI spending should be falling too. It isn't. The average enterprise AI budget grew from $1.2 million a year in 2024 to $7 million in 2026 — nearly a sixfold increase during the exact window per-token costs fell by roughly 95%. Blended frontier-model pricing fell 67% year over year between Q1 2025 and Q1 2026, from $18.40 to $6.07 per million tokens. And yet 73% of enterprises still exceeded their AI budget projections last year; one company's CTO reportedly burned an entire year's AI coding budget in four months. The FinOps Foundation now names AI and data platforms as the fastest-growing category of enterprise spend it tracks. Cheaper tokens didn't lower the bill. They funded more usage, more agents, more always-on systems — and, underneath all of it, a second cost that has nothing to do with tokens at all.
That second cost is trust, and it is growing on its own separate curve. The global AI governance and compliance market is valued at $2.54 billion in 2026. Spending on dedicated AI governance platforms alone is projected to reach $492 million this year, en route to more than $1 billion by 2030, according to Gartner. Perhaps the single sharpest data point: the share of organizations reporting a Chief AI Officer nearly tripled in one year, from 26% in 2025 to 76% in 2026. None of that spending makes a model smarter. All of it exists to answer a question the model itself cannot answer on its own — can this system be trusted to act, and can that trust be demonstrated to a regulator, a customer, or a board.
Two curves are moving in opposite directions at the same time. One measures how cheap it is to generate an answer. The other measures how expensive it has become to prove that answer can be trusted. The second curve is where the next decade of AI value actually gets captured — not because trust is harder to build than intelligence, but because, for the first time, it's the scarcer of the two.
Letter 161 argued that a tiered healthcare platform's real defensibility wasn't its AI — it was the trust already built into the Mayo Clinic Platform precedent it was measured against, the kind institutions had spent years earning before a single model was involved. Letter 162 argued that the difference between Cursor and Jasper was never the underlying model; it was whether the product became something a user trusted enough to route their actual workflow through. Neither letter used the word "trust" as its headline, but both were circling the same conclusion this letter now states directly: as the cost of generating an answer collapses toward zero, the thing that remains genuinely expensive — and genuinely valuable to own — is the layer that makes an answer safe to act on. A thick wrapper, in the language of Letter 162, is very often just a trust wrapper wearing a different name.
This reframes what "the best AI company" should mean for a founder, a country, or an investor reading these three letters together. It is not the company with the most capable model — that capability is now available to nearly anyone, at a price falling 10x a year, regardless of who trains it. It is the company whose intelligence gets trusted with the decisions that matter most: the ones with real financial, medical, legal, or safety consequences, where the cost of governance, audit, and verification is real and rising, and where a competitor cannot simply out-discount their way in with a cheaper model.
Both trends in this letter are independently well-documented, and together they describe a genuine inversion: intelligence is becoming abundant and cheap; trust is becoming scarce and expensive. That doesn't mean model quality stops mattering — a system nobody trusts is worthless regardless of price. It means model quality stops being the place where durable value accumulates, because it's no longer the constraint. The practical takeaway across this trilogy: a founder or a country deciding where to spend the next dollar of AI investment gets a better return building the governance, verification, and track record that earns trust than building yet another interface on top of a model that will be 10x cheaper to replicate within a year. The risk to this thesis: trust costs could fall too, if governance tooling matures the way inference did — several of the compliance-cost figures cited here already show automation cutting manual governance overhead by up to 40%, and a market this immature could see its own steep cost curve bend downward faster than this letter assumes.
Founder, NextGen Economics · Bangalore, India · August 2026
Sources: Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks" (2026) · ValueAdd VC, "How AI Inference Costs Have Dropped 95% in Two Years" (Jun 2026) · Medium/Predict, "AI Inference Costs Collapsed in 2026" — DeepSeek V4-Pro vs Claude Opus 4.7 SWE-bench comparison (Jun 2026) · Oplexa, "AI Inference Cost Crisis 2026" — enterprise AI budget data (Mar 2026) · PDP Spectra, "AI Token Pricing in 2026" — blended pricing YoY data (Jun 2026) · Gartner, "Global AI Regulations Fuel Billion-Dollar Market for AI Governance Platforms" (Feb 2026) · SQ Magazine, "AI Compliance Cost Statistics 2026" · Optro, "AI governance stats for 2026" — Chief AI Officer adoption data (IBM, May 2026). This letter presents a strategic framework using real market data; it is not a recommendation to build or fund any specific company.
A deeper research note applying this trust-economics framework to specific sectors, building on the healthcare platform concept in Letter 161 and the moat framework in Letter 162, is available to NGE clients beyond what's covered in this free letter.
Not investment advice. This letter evaluates a market and technology trend; it does not constitute a recommendation regarding any security.