Why compute might get 10x more expensive in coming years

AI labs could face a steep increase in compute costs over the next several years as AI systems produce more valuable work while chip supply expands at a slower pace. A human-level software engineer running on an H100-class GPU could support more than $250,000 in annual revenue at current software-engineer compensation levels, implying a compute price many times higher than today’s spot market.

The argument rests on a widening gap between AI revenue and hardware capacity. Anthropic’s revenue has grown about tenfold year over year, according to the analysis behind this article. If that pace continued, the company could reach $100 billion to $150 billion in annual revenue by the end of this year and approach $1 trillion the following year.

That outcome would require more than a larger server fleet. AI labs would need to increase margins, raise the price of compute, or devote more hardware to serving customers. The industry has pursued all three paths.

Revenue is rising faster than hardware supply

AI compute capacity has grown about threefold each year through a mix of chip improvements, new manufacturing capacity and shifts in wafer allocation toward accelerators.

The analysis estimates that Moore’s Law contributes a 1.4-times annual gain. New fabrication plants add about 1.2 times, while the transfer of leading-edge wafer capacity from other devices to AI hardware adds about 1.8 times. Those factors multiply to about three times more compute capacity each year.

Each source of growth faces a constraint. Extreme ultraviolet lithography tools limit the speed of new fab construction. AI already consumes a large share of leading-edge wafer production, leaving less capacity to redirect. Once chipmakers allocate most advanced wafers to AI accelerators, that source of growth will fade.

The supply problem becomes sharper for frontier labs. They need reliable access to large clusters, strong security for model weights and customer data, and enough flexibility to keep those systems well utilized. Spot instances cannot provide those guarantees.

Reports that Google pays about $900 million a month to rent 110,000 GPUs from SpaceX illustrate the premium for large, dependable capacity. The reported price works out to about twice the spot rate for a mix of Nvidia’s GB200 and GB300 systems. Spot prices have risen more than 40% from their February low, so the effective price paid by major labs may sit well above public hourly rates.

The Nvidia H100 product page provides a useful reference point for the class of hardware at the center of this debate. The Epoch AI research site tracks the relationship between AI progress, hardware and compute demand.

Higher margins can cover only part of the gap

AI companies can capture more revenue from each unit of compute as their models improve. Anthropic’s inference margins, according to the source analysis, rose from about 40% in 2025 to more than 80% for some API workloads this year. Blended margins may exceed 70%.

Those figures describe business-level economics, not the cost of each additional GPU-hour. A company can report high gross margins while paying a premium for scarce hardware. The marginal cost of scaling a popular model may still rise as demand fills available clusters.

Margins also depend on competition. A lab can charge more when its model outperforms the next available option. That advantage weakens when rivals offer similar capability at a lower price.

The analysis argues that margins would need to reach the mid-90% range for margin expansion to explain a tenfold revenue increase while compute grows threefold. That level would require a large and durable lead over competing labs. The economics make a higher compute price a more plausible part of the explanation.

Labs have another lever: inference. OpenAI directed about one-quarter of its 2024 compute spending to inference, according to an Epoch AI estimate. That share may now approach half as customers use models for coding, research and automated business tasks.

Inference produces revenue, while training creates future model capability. A rising inference share can fund operations, yet it leaves fewer resources for the training runs that labs use to maintain their lead. A business that spends most of its hardware budget serving current models starts to resemble a cloud provider rather than a research lab.

A human-level software engineer changes the price of compute

The strongest price argument comes from labor economics. Suppose an AI system can perform the work of a human-level software engineer and run on hardware comparable to an H100. At current software-engineer compensation levels, that GPU could support more than $250,000 in annual economic output.

A comparable H100 spot rental may cost about $17,000 per year if it runs around the clock at a rate near $2 per hour. A market price above $250,000 would represent an increase of about 15 times.

The calculation does not mean an AI company could collect the full value of every engineer it replaces. Customers would negotiate prices, utilization would vary, and competing systems would pressure margins. The calculation shows the ceiling created by useful work. As models handle higher-value tasks, hardware owners gain room to charge more.

An influx of millions of AI software engineers could reduce the value of each system through greater supply. Standard labor economics offers a counterpoint. High-skilled immigration has not produced a lasting collapse in wages because specialization creates new work and raises productivity. AI may produce a much larger supply shock at a faster pace, but the long-term outcome remains uncertain.

If software automation expands demand for complementary work, the marginal value of compute could rise with capability. Companies would pay more for a system that completes revenue-generating work, even if the hardware behind it carries a higher rental price.

Better models could capture a larger premium

Higher compute prices would change competition among AI labs. A company with no revenue would struggle to buy enough hardware to catch a frontier lab that already earns billions from model use.

Model efficiency would also matter more. A weaker system that needs twice as many tokens to complete a task would consume twice as much expensive compute. Customers would have a stronger reason to pay for a model that produces the same result with fewer tokens or fewer reasoning steps.

The Anthropic API documentation shows how model access gets priced through token consumption. As the cost of the underlying hardware rises, token efficiency becomes a larger part of a customer’s total bill.

This dynamic resembles the Alchian-Allen effect. When a fixed cost applies to goods with different quality levels, the higher-quality product becomes cheaper relative to the total purchase price. Applied to AI, a large compute bill could make the price gap between a leading model and a weaker model look small compared with the total cost of completing the task.

That shift would favor the labs with the best models and the most efficient inference systems. It would also make it harder for smaller companies to compete through low prices alone.

Consumer applications could lose access to cheap compute

Some AI applications depend on low hardware costs. Short-form video generation, image variations and other high-volume tasks can tolerate modest margins because each request uses little compute.

A large increase in GPU prices would force those products to raise prices, reduce output quality or limit usage. Consumers may continue to use them, but the economics would change when each generation competes with software development, research or customer support for the same scarce hardware.

The shift would concentrate compute in tasks that produce measurable economic value. Enterprise coding systems, scientific discovery tools and automated operations could outbid entertainment products for capacity.

That pattern resembles past warnings about resource scarcity, including the Simon-Ehrlich wager over commodity prices. Human innovation and market signals helped producers use many raw materials with greater efficiency. Compute has different constraints. Chip fabrication requires specialized tools, large factories and long construction timelines. Manufacturers cannot add capacity at the speed that software companies can add users.

The shortage would not last forever

The price pressure depends on the current supply regime. AI hardware still comes from a narrow manufacturing chain that relies on advanced lithography, high-bandwidth memory, packaging capacity and specialized accelerator designs.

Robots could reduce those constraints in a later period. Automated factories could process silica, copper and other inputs with less human labor, then turn those materials into computers at a larger scale. Commodity costs would exert more influence on compute prices once manufacturers could expand production with fewer bottlenecks.

That transition could take years. Until then, model capability may grow faster than available hardware. A threefold annual increase in compute supply would struggle to match a tenfold increase in revenue if each model upgrade lets companies monetize the same hardware more effectively.

The result would give leading AI labs strong economies of scale. Training creates a fixed cost that a lab can spread across millions of users. Human workers must learn and perform each assignment through separate interactions, while a trained model can reuse the same learned capability across customers.

That cost structure helps explain why AI revenue can grow faster than compute capacity. It also raises the risk that a small group of companies will control a large share of useful machine intelligence. The price of compute will shape both the economics of AI products and the number of companies that can afford to build the next generation of models.