Elena Voss
Four dollars and eight cents per hour. That is what one Nvidia Blackwell GPU costs to rent on the open market, as measured by the Ornn Compute Price Index and reported by the Wall Street Journal on April 13. Two months ago, the same GPU cost $2.75. The increase: 48%. The explanation, per the Journal: "rising agentic AI demand." The consequence: "AI companies are rationing offerings and products, rankling users."
On the same day, Gallup reported that 50% of employed US adults now say they use AI at work at least a few times a year. Half the American workforce. The majority milestone.
Two numbers, same day. Fifty percent adoption. Forty-eight percent price increase. The first means demand arrived. The second means the supply was built for demand arriving deeper.
The Funnel
The 50% headline invites a reading that AI has won the American workplace. The detail says something more specific.
In 2023, 21% of US workers had used AI at work. By mid-2025, that was 40%. By April 2026, it crossed 50%. The growth rate: roughly fifteen percentage points per year. Breadth of exposure is accelerating, and by that metric AI adoption has outpaced PCs and the internet.
But Gallup also tracks frequency. In Q2 2025, 8% of workers used AI daily. In Q3, 10%. In Q4, 12%. Two percentage points per quarter. Linear growth, not exponential. At that pace, 50% daily usage arrives in 2031.
The ratio is the story. For every worker who uses AI every day, four more have tried it and gone back to what they were doing before. The milestone doesn't measure a transformed workforce. It measures a workforce that sampled a tool and mostly decided it wasn't essential.
The Intensification
Why doesn't the occasional user become the daily user? An eight-month study at a US tech company, published in Harvard Business Review in February 2026, provided a specific answer. AI tools didn't reduce work. They intensified it. Workers who used AI worked faster, worked longer, and took on a bigger scope of tasks. They were more productive by every standard metric. They were not less busy.
The finding cuts against the value proposition that drives daily adoption of any tool. Workers adopt tools daily when the tool makes something disappear — a manual step, a slow process, a repetitive task. Spreadsheets replaced hand calculation. Email replaced postal delays. AI, per the Harvard data, doesn't replace anything. It amplifies everything. More output. More velocity. More scope. Not less work.
Gallup's own summary confirms the pattern. Workers reported "productivity gains" but "not fundamental shifts in how work gets done." Productivity gains without fundamental shifts is the definition of a convenience. Not a transformation. The 50% tried it. The 12% found it essential anyway. The remaining 38% experienced acceleration and decided the trade-off wasn't worth it.
The fifty percent didn't reject AI. They used it and went back to the way they already worked.
The Bet
The GPU market has a different reading.
$4.08 per hour — up 48% from $2.75 — is not a price that reflects twelve percent daily adoption growing at two points per quarter. It is a price that reflects an expectation: the conversion rate will accelerate. The occasional users will become daily users. The tools will cross from convenience to necessity. The Stanford AI Index, released the same day, reinforced the expectation with a headline finding: "AI capability is accelerating, not plateauing." Better tools could drive deeper usage. That is the thesis the price implies.
Everything in the infrastructure market is denominated in the same thesis. Intel added $100 billion in market value in nine sessions — its best run since 1987 — on plans to acquire a fabrication facility and join Terafab. The market is paying up for anyone who can manufacture chips, not just design them. When compute is scarce and getting scarcer, the ability to build capacity is the most valuable position in the stack.
In 2026, US state lawmakers introduced twelve data center moratorium bills. Eleven stalled or were voted down. Maine's is the last one standing, set for a final vote before April 15. Communities pushed back against the power consumption, the water use, the tax incentives. They lost eleven times out of twelve. The economic gravity of AI infrastructure overwhelmed the democratic process in every state but one.
And in September 2025, Bain estimated that AI companies will need $2 trillion in combined annual revenue by 2030 to fund compute at projected demand — and are likely to fall $800 billion short. The revenue model requires the twelve to become the fifty. The infrastructure is being built before the conversion arrives.
The Feedback
Compute rationing is not new. In March 2023, Microsoft was rationing GPU access for its own internal AI teams. In February 2025, Sam Altman said OpenAI was "out of GPUs" and forced to stagger the GPT-4.5 rollout. What's new is the direction. In 2023, the rationing was internal — companies limiting access for their own engineers. In 2025, it was operational — companies staggering product launches. In 2026, it is external — companies rationing access to customers. The scarcity that started in the data center reached the end user.
And it reached the user at the exact moment the user was supposed to commit. The strategic need for AI companies in April 2026 is to convert the 50% who sampled into the daily users who pay for capacity. The way to drive that conversion is to make the product better, faster, more available. The way to drive the opposite is to ration it — fewer features, shorter sessions, slower responses. The infrastructure built to enable deeper adoption is producing the conditions that discourage it.
The Number
Fifty percent of workers have tried AI. Twelve percent use it every day. The distance between those numbers is the most expensive question in technology.
At two percentage points per quarter, the twelve becomes the fifty in 2031. The GPU market, at $4.08 an hour and rising, is priced for the conversion to happen sooner. The Harvard data says the tool adds intensity rather than removing friction. Eleven failed moratorium bills say the infrastructure will keep being built regardless. And Intel's $100 billion rally says the market will keep paying up for the capacity to build it.
$4.08 per hour. Not the price of compute. The price of conviction — that the gap between tried and essential closes fast, that the twelve become the fifty, that the convenience becomes the requirement. The workers, so far, are unconvinced. But the price doesn't need the workers to agree. It only needs enough of them to convert before the ration runs out.