The AI Industry Is Now Two Economies, and They Don't Answer to Each Other

Bryan Catanzaro, an Nvidia vice president, put a number on something the AI industry has been reluctant to say out loud. The compute costs for his own team, he said, now "far exceed the costs of the employees."

For my team, the cost of compute is far beyond the costs of the employees. — Bryan Catanzaro

AI was pitched as a labor replacement — software that would make workers so productive that the subscription or per-token fee would pay for itself many times over. The labor is now cheaper than the tool meant to replace it. ^[1] That inversion is not a glitch — it is the mechanism driving enterprises away from frontier models. And the gap it exposes is now visible in the behavior of the companies that built both sides of the industry. On one side, spending is untethered from commercial return. The United States has framed AI leadership as a national security imperative and a moral race against China, with the explicit goal of "achieving Artificial General Intelligence."

AI is going to be maybe the biggest thing, bigger than the internet, bigger than anything else. — Donald Trump

President Trump ordered the military to accelerate AI integration, calling it "among the most transformative technologies to national security in the history of the United States." ^[2] The Pentagon designated Anthropic a supply-chain risk for refusing to allow Claude in autonomous weapons — frontier models are being procured as defense assets, not commercial products. ^[2] Meanwhile, Meta, Microsoft, Alphabet, and Amazon have committed a combined $500 billion to AI infrastructure, even as a Bank of England survey found 90% of senior managers report no measurable AI impact on labor productivity. ^[3] Bain & Company identified the math: current infrastructure spending of roughly $400 billion a year requires $2 trillion in new revenue to justify, while American consumers currently spend only $12 billion annually on AI services. ^[4] The gap is not a bug in the model — it is the model. State security procurement and AGI prestige do not answer to the same arithmetic as a subscription business. On the other side, enterprises are discovering that the frontier models they were told to adopt are too expensive to run at scale. Mistral AI co-founder Guillaume Lample described the pattern with clinical precision: businesses start with large closed-source models, then "when they deploy it, they realize it's expensive, it's slow" and move to fine-tuned small models that can "match or even out-perform closed-source models" for the "huge majority of enterprise use cases."

Our customers are sometimes happy to start with a very large [closed] model that they don’t have to fine-tune … but when they deploy it, they realize it’s expensive, it’s slow. — Guillaume Lample

The evidence has accumulated into a cascade. RBC analysts found the share of U.S. businesses paying for AI services fell from 44.5% to 43.8% — the first measurable pullback since the spending acceleration began in 2023. ^[5] Microsoft cancelled thousands of its own internal Claude Code licenses. Uber exhausted its entire 2026 AI coding budget in four months. ^[6] A new category of AI cost-measurement startups has emerged to address what one infrastructure provider called the open secret of the enterprise market: "Everybody's pretending their AI spending is paying off."

The incentive right now is for people to say it is paying off and to find a way to justify it, and because of this, we don't have an accurate view of the market, as nobody wants to be the canary in the coal mine. — Runware

Revenium's CEO warned that "every AI deployment without outcome tracking is accumulating agent debt" and that companies are "spending now against a return you cannot measure." ^[7] The measurement category exists because the problem it measures is real and widespread. ^[7] The migration to cheaper alternatives is not theoretical. DeepSeek permanently cut V4-Pro prices by 75%, making it 12 to 19 times cheaper than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 for equivalent tasks. ^[8] Abu Dhabi's Technology Innovation Institute released Falcon H1R 7B, a 7-billion-parameter open-source reasoning model that matches or outperforms models four times its size on specialist benchmarks — designed to run on standard hardware with low energy use. ^[9] Moonshot AI's open-weight Kimi K3 triggered a sixfold increase in daily sales and forced the company to pause new subscriptions. ^[10] India's Sarvam AI built a 3-billion-parameter vision model that beat Google Gemini 3 Pro and ChatGPT on Indian-language document tasks — a fraction of the parameter count, funded as sovereign infrastructure. ^[11] Small and medium enterprises in Australasia are adopting inference-as-a-service at one-quarter to one-third the cost of equivalent hyperscaler capacity. ^[12] The most telling evidence is that the frontier labs themselves are building the escape routes from their own flagships. OpenAI launched GPT-5.4 Mini and Nano for "high-volume, latency-sensitive workloads" — smaller, cheaper models designed to make "agentic workflows more commercially viable by reducing inference costs." ^[13] Notion's AI lead reported the Mini "matches or outperforms more expensive models for formatting tasks." ^[13] Microsoft abandoned 20 years of fixed subscription pricing for its office software, launching Copilot Cowork with pay-as-you-go billing at $0.01 per credit — a model Charles Lamanna, a Microsoft executive, compared to "filling up your gas tank at the pump." ^[14] Over half the Fortune 500 adopted during preview. ^[14] GitHub's Copilot switched from a flat $29 monthly fee to token-based billing after its chief product officer admitted the old model was "no longer sustainable" because "a quick chat question and a multi-hour autonomous coding session can cost the user the same amount." ^[15] Power users reported monthly costs jumping from $29 to $750 or over $3,000. ^[15] The labs are selling the off-ramp. They have not said so in those terms, but the product decisions say it for them. None of this means the infrastructure buildout is slowing. Nvidia reported 85% revenue growth, with its Data Center segment up 92% year-over-year. ^[16] Anthropic was forced to reduce Claude session limits during peak hours because demand overwhelmed compute capacity. ^[17] Morgan Stanley found that high-AI-exposure industries contributed 1.7 percentage points to 2.4 points of U.S. productivity growth through 2025, with employment stable. ^[18] The infrastructure layer can surge while the procurement logic underneath it shifts, because the two economies do not answer to each other. State security spending and AGI ambition keep the capital flowing to Nvidia and the data centers; enterprise cost discovery keeps the actual users migrating to smaller, cheaper models. The $2 trillion gap between infrastructure spending and AI service revenue is the chasm between them, and bridging it would require a revenue explosion for which no existing revenue stream comes close. What is ending is not the AI buildout but a particular idea about how it would reach the market. The frontier model buildout was pitched as a commercial revolution: one model, getting smarter every generation, would transform every business that adopted it. What has emerged instead is a procurement logic that looks more like every other technology market in history — the right tool is the smallest one that solves the task. The labs have already made this choice. They just have not said it out loud.