When AI agents begin to approach or replace human knowledge labor, how should the price of computing power be revalued?
In his latest blog post, prominent tech podcaster Dwarkesh Patel presents a bold projection: if AI labs’ revenues continue to grow rapidly—such as Anthropic reaching $1 trillion by the end of next year—while computing supply increases by only about 3 times annually, the price of compute could rise more than tenfold.
The cost of silicon-based labor, H100 + AI = $250,000 Silicon Valley programmer
For decades, computing power has been viewed as a "tool"—computers became faster, servers increased in number, and cloud computing became cheaper.
But a paradigm shift is underway in the AI era: GPUs are no longer just components of machines—they may become the carriers of digital labor.
If a single H100 can support an AI agent approaching the skill level of a top-tier programmer, its value should no longer be calculated based on "server-hour pricing," but rather on the human labor it replaces.
Imagine if there were a "digital software engineer" that never sleeps, responds instantly, and possesses top-tier human-level technical skills—how much would you be willing to pay as an annual salary?
In Silicon Valley, the average market salary for software engineers exceeds $250,000 per year. However, running an agent with equivalent intellectual capacity may require only the computational power of a single NVIDIA H100 GPU.
Now, let’s do the math:
If this H100 can perfectly replace a human programmer earning $250,000 per year, then, from a purely business and economic perspective, the theoretical annual rental cost of this H100 should be set above $250,000.
But at this moment, the spot rental price for the H100, when annualized, is only about $16,000—15 times lower.
This is the largest "labor arbitrage black hole" in human history that we will face.
When the rate of intellectual evolution in the virtual world far exceeds the manufacturing limits of the physical world, conflict arises:
The rate of value growth in digital intelligence is far outpacing the expansion of physical computing infrastructure.
This 15-fold valuation gap will trigger a fierce backlash: over the coming years, the real price of computing power may face a superinflation of 10 times or more.
Naturally, there's a question: Will software engineers' annual salaries significantly decrease in the future?
Dwarkesh Patel also considered this question:
But if we accept standard economic theory regarding labor, the marginal value of computational power—and thus its marginal price—should become extremely high.
The speed at which AI generates money has surpassed the speed at which AI is being developed.
Over the past few years, the AI industry has followed a simple logic: larger models, more data, more GPUs, and greater capabilities.
But now, a new contradiction is emerging: the growth rate of AI is outpacing the growth rate of computing power supply.
Take Anthropic as an example. Over the past year, its commercial revenue has grown approximately tenfold.
At this growth rate, its annualized revenue could reach $100 billion to $150 billion by year-end. If this trend continues, next year it will face an extraordinary question: how to support a trillion-dollar potential?
There is only one answer: more intelligence. And more intelligence means: more training, more inference, more GPUs.
However, the issue is that hash rate supply has not increased at the same pace.
The physical computing power expansion of leading industry laboratories can currently sustain growth of approximately threefold per year.
Thus, a significant mismatch arises:
The digital world wants to run 10 kilometers, but the physical world can only provide a 3-kilometer road. What do we do with the remaining distance?
There are only three options: first, increase profit margins; second, raise the price of computing power; third, allocate more computing power to inference services.
The third path is actually what most AI labs would least like to see.
Because earning through reasoning is essentially just to prove: "The AI business model works."
Then continue raising funds, purchase more GPUs, and continue training the next-generation model.
The real cycle in Silicon Valley is:
Making money isn't the end goal—it's just a means to buy more computing power.
Because what all labs truly believe is that today's most advanced models may, a year from now, be like yesterday's smartphones—powerful, but outdated.
This forces the solution back to the first two options: a surge in profit margins or an increase in the cost per unit of computing power.
If this discrepancy is to be fully absorbed by profit margins alone, Frontier Lab’s profit margin must reach the mid-to-high range of approximately 95% by the end of next year.
In the history of business, this is almost fantasy, except for a very few absolute technological monopolies.
Thus, all the clues and logic ultimately converged on one conclusion: the unit price of hash power is about to experience an unprecedented surge.
GPUs are transitioning from servers to factories of the new era.
Many people haven't yet realized: the AI competition is entering the infrastructure era.
In the past, the internet war was fought over users, traffic, and entry points. In the AI era, the competition is for electricity, chips, data centers, and supercomputing clusters.
Why? Because top AI labs cannot rely on standard cloud servers.
They require: dedicated GPU clusters; a stable network infrastructure; extremely high utilization; and data and model security.
This is no longer about renting ordinary servers; it's about leasing the intelligent production lines of the future.
A typical example is: Google and Anthropic lease large-scale GPU resources from SpaceX.
Reports indicate that Google's fleet of approximately 110,000 GPUs incurs monthly rental costs of $900 million.
On average, the hourly rental cost per GPU is nearly twice the current spot market rate. And don’t forget, the current spot price itself is already 40% higher than the low point earlier this year.
Why?
Because top-tier computing power is increasingly resembling strategic resources like energy and power grids.
Imagine the Industrial Revolution: one factory has a steam engine, another does not.
The gap between them isn't 10%—it's an entire generational difference in production methods.
In the AI era, GPUs are also becoming this new "steam engine."
Why has the hash rate price spiraled out of control to this extent?
The answer lies in the physical formula of supply-side constraints—this formula has been firmly locked in by the physical limits of TSMC and lithography machine giants.
The industry’s beloved myth of “computing power tripling every year” is merely a precarious combination of three multipliers pushed to their limits:
Annual computing power growth (3x) ≈ 1.4x (Moore’s Law iteration) × 1.2x (new fab construction) × 1.8x (securing advanced process capacity).
And each of these three multipliers is approaching its ceiling.
Moore’s Law: Each node shrink is harder than the last, with leakage and atomic limits making each iteration precarious.
The construction of new fabs takes a long time, and the delivery speed of high-NA EUV lithography machines is bottlenecking the entire supply chain, with no solution expected before 2030.
Over the past two years, the only driver of computing power growth has been the scramble for production capacity—AI chips have directly competed with smartphones and PCs for TSMC’s advanced manufacturing processes.
But this multiplier is about to hit zero. By the end of 2027, AI chips will consume 86% of TSMC’s N3 capacity, up from 60%—not just growth, but absolute physical saturation with no room left.
Three multipliers; two have already peaked, and the third is about to hit a wall. The story of hash rate supply has already been told.
And the more expensive the computing power, the less willing people are to use inexpensive models.
Intuitively, when computing power is expensive, use lightweight models and open-source pruned versions to save costs.
But the economics of AI is exactly the opposite.
The core logic is the Alchian-Allen effect: when a fixed additional cost is applied to all goods, the more expensive items appear relatively more valuable.
In AI, this "additional cost" is the depreciation of computing power—regardless of which model you use, you must pay it.
Do the math: H100 costs $20 per hour.
- Top model: Done in 1 hour for $25 ($20 for compute + $5 premium).
- Average model: $0 licensing fee, but insufficient intelligence led to two hours of trial and error, costing $40.
Conclusion: Using an average model is actually more expensive—it wastes computational power and the more valuable resource: time, through ineffective attempts.
Therefore, the more expensive the computing power, the more justified the strongest models are in charging premium prices. Since you’re going to pay for computing power anyway, spending a little more on the smartest one is truly cost-effective.
Mediocre models have no future.
Ultimately, societal capital will be sucked into a black hole, drawn toward a handful of trillion-dollar giants capable of building the absolute strongest models.
Computing power is also a resource—it doesn't necessarily have to increase in price.
Of course, some people also oppose it.
In the comments, Eric Xu offered an interesting counterpoint: the logic of a $250,000 annual rent only holds in the purely digital realm. Once it enters the physical world—such as the approval timelines for clinical trials or procurement decision processes—the revenue ceiling for AI may never reach $250,000 and is more likely closer to $50,000.
Senior Software Engineer Trevor pointed out that low-quality content doesn’t require top-tier models—it will simply shift toward cheaper alternatives rather than disappear entirely. In other words, what gets cleaned up may be the price, not the content itself.
History is also not on Dwarkesh’s side. In 1980, Stanford biologist Paul Ehrlich and economist Julian Simon made a bet on whether five metals would increase in price over a decade.
Ten years later, all five had dropped; Ehrlich conceded and paid up. “Resources will always rise in price”—historically, most who bet on this lost.
Dwarkesh’s response was: The elasticity of computing power supply is far lower than that of commodities. Copper mines can increase production in response to higher prices, but EUV lithography machines cannot—TSMC itself has no excess capacity to release. Over the past few years, with supply growing by only threefold annually, there has been no substitute.
Interestingly, Dwarkesh himself admits: this prediction, like nearly all similar ones in history, has been wrong.
Perhaps computing power will one day become as cheap as sand, but until that day arrives, we will have to endure the most brutal computing power arms race and computing power inflation.
Reference materials:
https://www.dwarkesh.com/p/why-compute-might-get-10x-more-expensive
https://x.com/dwarkesh_sp/status/2082482530209411419
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by David.