The AI industry presents a strange picture. Competition is intensifying, more models are entering the market, API prices keep falling, and several leading model companies are still losing money. Yet Nvidia, data-center operators, and cloud providers continue to invest on a massive scale. Capital has not retreated. It keeps pouring in.
Under the usual logic of competition, more wolves should mean less meat and thinner profits. AI does not look like that. It looks more like a market in which the number of wolves keeps rising, while the amount of meat grows even faster.
To understand why, the analysis cannot begin with Nvidia’s gross margin or the losses reported by model companies. It has to go further upstream and ask a simpler question: what did the thing AI is replacing used to cost?
1. AI Is First Replacing Expensive Human Labor
ChatGPT Plus costs $20 a month. That price looks ordinary by internet standards, but the service is no longer comparable to streaming video, music, or cloud storage. It can help with writing, translation, research, analysis, programming, and a growing range of knowledge work.
So whether $20 is expensive should not be judged against Netflix. It should be judged against labor. What does it cost to hire a programmer, lawyer, analyst, accountant, or researcher for several hours? Then compare that with the price of AI. The two are not in the same order of magnitude.
Users do not calculate how many GPUs were used in a conversation or what one million tokens cost to produce. They know only that work that once took them hours—or required paying someone else—can now be done for very little money.
Value of the human task ≫ Price of AI
That is the first economic relationship. Users feel they are getting a bargain, while AI providers still have substantial pricing room. For many users, the real constraint is no longer the price of AI. It is whether there is a worthwhile task to do and enough time to absorb the result.
That is also why the market can expand whenever AI acquires a new capability. Lower prices help, but better capability, new tasks, and deeper integration into workflows can expand demand even more.
2. Cheap for the User Does Not Mean Cheap for the AI Provider
Move one step down the chain and the picture changes. DeepSeek V4.1 Flash has been priced at about $0.30 per million uncached input tokens and $1.20 per million output tokens during peak periods.
SemiAnalysis tested the same model on H200 GPUs and converted hardware purchases, depreciation, electricity, and related infrastructure into total cost of ownership, or TCO. Under those test conditions, the inference-compute cost per million total tokens was only a few cents, and in favorable conditions even less.
Assume, for illustration, that a workload is 80% input and 20% output. At the prices above, average revenue per million total tokens is:
0.8 × $0.30 + 0.2 × $1.20 = $0.48
If inference-compute cost is $0.012 per million tokens, then:
Inference Price Multiple = Revenue per million total tokens ÷ Inference-compute cost
$0.48 ÷ $0.012 = 40
This number answers only one basic question: how far apart are the compute cost of producing one million tokens and the price charged for selling them? I call this ratio the “Inference Price Multiple.” In this example it is about 40 times. Even using a higher retail GPU-rental cost, the multiple remains large. A gap of this size would be difficult to imagine in many traditional industries.
This is not DeepSeek’s gross margin. Training, R&D, employees, idle capacity, networking, safety, and other expenses are not included. But the calculation does explain why model companies are willing to keep buying GPUs: the GPU may be expensive, but the tokens it produces can be worth much more.
3. Liang Wenfeng’s “Ten-Month Payback”
In a publicly circulated investor-meeting record, Liang Wenfeng described DeepSeek’s pricing logic in simple terms: buy a batch of equipment and, at current API prices, recover the equipment cost in roughly ten months. He also said this was not profit-maximizing pricing; even a 50% increase, or possibly a doubling, might not materially reduce token consumption.
Those two points matter. First, the equipment payback period is short. Second, users appear relatively insensitive to price within that range.
Why? Return to the starting point. Users do not compare AI with GPU cost. They compare it with human labor. As long as AI completes a task for far less than a person would cost, model providers retain significant pricing room. Even at a price designed around a ten-month equipment payback, DeepSeek can still feel cheap to the user. Leading U.S. models are often priced materially higher than DeepSeek. The remarkable feature of today’s AI market is the wide economic space between expensive human cognitive labor and increasingly cheap machine inference.
4. A Model Company Can Lose Money Without Losing Money on Every Token
This distinction is easy to miss. A company’s total loss is one account. Whether each unit of inference makes money is another.
API revenue − inference cost ≫ 0
A company may then spend heavily on the next model, researchers, data, more GPUs, new data centers, new products, and market expansion. The final income statement can still show a loss.
So a model company losing money does not mean every AI service it sells loses money. For investors, a more important question is how quickly an additional dollar of compute investment comes back. If the answer is anywhere near the ten-month period Liang described, continued capital inflows are not hard to understand.
That is very different from a business whose unit economics never worked and which survives only by burning capital.
5. Why GPUs Can Get More Expensive While Tokens Get Cheaper
A more expensive new GPU does not necessarily mean a higher cost of producing AI. The purchase price of the machine is only one part of the equation. What matters is the average cost of producing one million useful tokens over the GPU’s working life.
Cost per million tokens = Lifetime GPU cost ÷ Lifetime useful-token output
If a new GPU costs 50% more but produces more than twice as many useful tokens, the cost per token falls. SemiAnalysis tests of newer hardware point in this direction: GPUs can become more expensive while inference cost per token continues to decline.
That allows several apparently contradictory things to happen at once: GPU prices rise, AI-service prices fall, and model providers’ unit inference economics improve. Nvidia can earn very high profits while its customers still want to buy more chips. The customers are not failing to do the math. The math can simply be very good.
6. Why Do More Wolves Mean Even More Meat?
Competition in a traditional industry usually takes place inside a relatively fixed market. If ten companies are fighting over a $10 billion market and that becomes twenty companies, life gets harder for each one.
AI is different today because it is converting expensive human cognitive labor into very cheap machine cognitive labor. Whenever the models acquire a new capability, new demand appears.
A task that was not worth paying a person $100 to do may suddenly be worth giving to an AI for $1. An analysis that was previously not worth doing can now be done. A programmer who could address only a handful of problems in a day can now address many more.
The market therefore does not have to remain at $10 billion. It can become $100 billion. The number of wolves can rise from ten to fifty while the amount of meat rises from one hundred to one thousand.
Competition intensifies, but the market expands faster. Falling prices do not necessarily destroy profit because costs are falling too. New competitors do not merely divide existing demand; together they can educate the market and create new demand. Capital spending keeps rising not only because investors believe a story, but because many participants can see attractive unit economics.
This describes the market, not every company in it. More meat does not mean every wolf will eat well.
7. Where Is the Real Risk?
None of this means AI cannot contain bubbles. Local excesses are entirely possible.
The real question is not whether investment is large. It is when supply begins to expand persistently faster than real demand.
If compute prices fall, utilization drops, equipment payback periods lengthen, capital returns deteriorate, and new data centers are still being built at high speed, then a segment of the chain is genuinely moving into overcapacity.
So the AI-bubble question cannot be answered simply by looking at how much Nvidia earns or how much a model company loses. Start at the bottom: how much value does the user actually obtain from AI? Then follow the chain all the way back to compute investment.
Value of human task → AI price → Token price → Token cost → Return on compute investment
The most important relationship in that chain today is simple:
Value of human task ≫ Price of AI ≫ Cost of inference compute
As long as those gaps remain wide, the AI industry has room for many participants to make money at the same time. This is not, at least for now, a conventional market in which more wolves inevitably mean less meat. It is a market in which the wolves are multiplying, but the meat is growing faster.
8. The “Crazy” Parts of the Market Fit the Same Chain
Nvidia earns extraordinary margins while model companies still report losses. GPUs get more expensive while tokens get cheaper. API prices fall while capital expenditure rises. More competitors enter while investors keep placing larger bets. Viewed separately, each phenomenon can look abnormal.
Put them back into the same chain and the picture becomes less mysterious. End users compare AI with expensive human labor and therefore find it cheap. Model providers compare API prices with inference costs and can still see large unit-economic spreads. They then compare additional GPU investment with future token output and may find that equipment can pay back quickly. Upstream hardware suppliers can therefore capture part of the surplus, while new hardware with higher throughput continues to reduce unit token cost.
For now, this is a chain that can reinforce itself.
The thing to watch is not the absolute size of investment, but which gap in the chain begins to narrow. As long as user value remains far above the price of AI, AI prices remain well above unit inference costs, and new compute can still pay back quickly, the investment boom has a real economic foundation.
AI looks like a crazy market partly because we are accustomed to markets in which more wolves mean less meat. This may be one of those rare periods when the number of wolves keeps rising—and the meat grows even faster.
Source Note
DeepSeek API pricing, H200 inference-cost estimates, and Liang Wenfeng’s “ten-month payback” statement are based respectively on DeepSeek’s published pricing, SemiAnalysis InferenceX testing, and a publicly circulated record of Liang’s investor meeting.
References
[1] OpenAI. “What is ChatGPT Plus?” OpenAI Help Center.
[2] DeepSeek. “Models & Pricing.” DeepSeek API Docs.
[3] SemiAnalysis InferenceX. “DeepSeek V4.1 Flash on H200.”
[4] China Research Collective. “Liang Wenfeng Investor Meeting, May 2026.”
