Starting from Jensen Huang’s “AI Factory” Intuition
24 September 2026
The market has been arguing for a long time over whether AI is a bubble.
The investment reality is increasingly visible in very large numbers. Microsoft, Amazon, Google, Meta and other companies are still raising capital expenditure; data centers are getting larger, power contracts longer, and financing structures more ambitious. In September 2026, Reuters cited market expectations that spending by five major AI infrastructure investors could exceed $795 billion this year and surpass $1 trillion in 2027.[2]
The evidence on the other side is equally real: AI usage is still expanding rapidly. Models are getting cheaper and more capable. AI is moving into office work, programming, search, advertising, design and enterprise management, while agents are pushing it from answering questions toward executing tasks.
The opportunity looks enormous. It also looks like a very large bubble.
There is an interesting divide in the debate. The strongest advocates of long-term demand tend to be industrial leaders committing enormous amounts of capital to AI infrastructure; the strongest bubble warnings often come from capital markets. Neither side is proved right by its institutional position, but the split tells us that industry and finance are looking at the same phenomenon through different time horizons.
Our current view has three parts. Long-term demand for machine intelligence may be seriously underestimated by traditional user-based models. Large industrial firms would rather accept some overbuilding than risk being pushed off the compute table. And the more plausible danger is not that the entire machine-intelligence system loses value at once, but that bubbles emerge locally—in assets, financing structures and valuations.
This essay does not attempt to assign one bubble score to AI. It asks a different question: if excess begins to surface, where should we expect the first verifiable signals?
1. AI factories are productive capacity
Jensen Huang has repeatedly used the term “AI Factory.” At GTC Taipei 2026, he described AI infrastructure as a new kind of factory, argued that tokens had become revenue-generating units, and emphasized that agents were entering real workflows—raising demand for computation.[1]
The idea deserves to be taken seriously. Traditional data centers were built primarily to store, transmit and process information. Large AI factories combine chips, electricity, networks, data, models and software into the computational capability required to perform intelligent tasks. Building AI infrastructure is therefore not quite the same thing as building an ordinary internet data center.
When I say that an AI factory is productive capacity, I do not mean that an empty facility without tasks, models, software or electricity automatically creates value. The change is that traditional tools usually require a human to remain inside the operating loop. Once goals, tasks and permissions are specified, an AI system can continue reasoning, searching, coding, testing and executing without continuous human participation. Human labor may stop; machine cognitive labor does not have to stop with it.
That is why I am willing to keep the stronger formulation: an ordinary data center is primarily a production tool; an AI factory increasingly becomes productive capacity in its own right. It is not merely waiting to be operated by a person. It is directly taking over part of the cognitive production process that previously required human labor.
This changes the demand question. Internet products were usually judged by user counts and user growth. Apply the same logic to AI and you risk understating the phenomenon.
One person using AI cannot simply be counted as “one user.” We also need to ask how much work that person delegates to AI each day, how complex that work is, how deeply AI participates, and whether one person is beginning to orchestrate several AI systems at once. If AI doubles a person’s effective capacity in a particular class of tasks, the demand represented by that person is plainly larger than one account.
In other words, AI demand is not only about how many people use it. It is about how much cognitive work people are handing over to machines.
That means machine cognitive services can keep expanding even if the number of AI users eventually stops growing rapidly. This is one of the easiest things to miss when judging whether AI infrastructure is excessive.
Visible capital expenditure alone cannot prove overinvestment. We still lack a mature ruler for measuring how quickly machine cognitive labor may grow.
AI has properties of both a tool and productive capacity. When the mode of production changes faster than the measurement system, a predictable error appears: investment is easy to see, while newly created capability and demand are only partially visible.
The internet era trained us to watch users, activity, conversion and revenue. Industrial investment is usually judged by capex, capacity utilization and labor productivity. Those measures remain useful, but they do not fully describe AI. A person who asks a few questions and a person who delegates research, coding and analysis to several continuously running agents can look identical in a “one user” statistic. Likewise, fewer hires, shorter R&D cycles or entirely new tasks enabled by AI may never show up as a separate AI subscription line.
An old ruler can therefore misread the bubble in either direction: it may exaggerate the scale of excess, or locate the excess in the wrong part of the system.
2. The key question is whether investment or demand is running faster
AI is no longer confined to laboratories. Individuals use it. Enterprises use it. Software firms, programmers, advertising systems and search systems use it. Many AI services are free or very cheap, so paid subscription revenue alone can materially understate actual usage.
The central uncertainty is therefore not whether there is demand. It is whether demand for machine cognitive services can expand fast enough to match today’s growth in infrastructure investment.
If people continue transferring more work to machines, and agents further multiply the effective capacity of each person, today’s enormous data-center investment may turn out to be less excessive than it looks—perhaps even insufficient.
But if machine-cognition demand fails to grow to the level embedded in current capital plans, some assets will inevitably be overbuilt. The important word is some.
By a “local bubble,” I do not mean one single phenomenon. It may appear as excessive valuation, overly aggressive financing, capacity built too early, or cash flows that never cover the cost of capital. This essay focuses mainly on AI infrastructure: once assumptions about future demand, price and utilization are capitalized too aggressively, already-built or committed compute capacity will begin to show return pressure.
There is another source of optical distortion. Investment is highly visible. We can count GPUs, data centers and power contracts. Demand is dispersed across millions of daily workflows, internal corporate processes and agent tasks, much of which is not separately priced and will never appear as a new line called “AI revenue.”
Comparing highly visible capex with only partially measured machine cognitive demand naturally makes supply look faster. That does not prove a bubble. It tells us that “investment is growing faster than demand measured by old indicators” is not yet sufficient evidence.
3. AI is not entering the same world as the early internet
The dot-com bust left a simple story in public memory: a technological revolution arrived, capital rushed in, and the bubble burst.
That makes the year 2000 an obvious analogy whenever AI capex rises. But the analogy can blur two different questions.
The collapse of the internet bubble did not mean the internet failed. The internet transformed the world. What failed were many specific companies, business models and valuations. A correct technological direction did not make every company correct, and certainly did not make every purchase price correct.
The same distinction is becoming increasingly important in AI. AI may be a durable new productive force, while many investments built around it can still be wrong.
We may therefore be facing not one undifferentiated bubble but a genuine technological revolution containing many local bubbles.
The diffusion environment is also different. The early internet first had to build PCs, connectivity, payment systems and software distribution. AI does not have to recreate those foundations. It sits on top of the internet, smartphones, cloud computing, digital payments and mature software ecosystems.
For that reason, AI can plausibly diffuse far faster than the early internet. Simply mapping the dot-com template onto AI can therefore be misleading.
4. Chinese competition has to be taken seriously
AI usage can rise while model prices fall. Token volume can increase tenfold while unit prices fall even faster. A company that leads today is not guaranteed to keep the same share five years from now.
This is why Chinese competition matters. China is developing a distinctive combination: rapidly improving models, very low prices and extensive use of open weights. DeepSeek, Qwen, Zhipu and Kimi are not only used domestically; they are increasingly part of global developer workflows.
OpenRouter provides a useful but clearly bounded observation window. In 2025, roughly three quarters of the tokens on its platform went to U.S. models. By early June 2026, Chinese models had surpassed U.S. models in token share on OpenRouter.[3] That is not the same as saying China leads all AI usage worldwide—large amounts of proprietary app traffic, private deployments and cloud usage never pass through OpenRouter. But it does show that cheaper, capable Chinese models are changing real developer choices.
Open weights change the economics in a way that a simple API price cut does not. A closed model keeps the user inside one provider’s platform. An open model can be run on a company’s own servers or on providers in the United States, Europe, the Middle East, China or elsewhere. OpenRouter reported in September 2026 that open-weight models already accounted for more than half of requests originating in both the United States and Europe, with Chinese models representing a substantial share of that volume.[4]
The larger threat from Chinese open models may therefore not be that they are simply “smarter than American models.” It may be that machine intelligence becomes cheaper while U.S. frontier-model companies capture a smaller share of the resulting economic value.
This matters directly to the bubble question. The hundreds of billions of dollars being invested in U.S. AI infrastructure eventually have to be justified by future cash flow. AI demand could rise tenfold, yet if prices fall sharply and competition intensifies, social value may soar while investor returns do not.
That is a classic location for a local bubble: the direction is right and the demand is real, but investors mistakenly extrapolate today’s margins, market share and scarcity into the distant future.
There is another side. Open models compress prices and value capture at the model layer; they do not necessarily reduce global compute demand. If lower prices stimulate usage even faster, local deployment expands and more agents run continuously, total compute demand can rise. The primary U.S. exposure is therefore profit capture and platform control—not a simple disappearance of global compute demand.
5. Why the large technology companies may be less fragile than they look
At this point another distinction becomes clear: the same level of AI spending can create very different risks depending on who is doing the spending.
Microsoft, Amazon and Google have more than deep pockets. They also have internal absorption capacity. If external demand underperforms, compute can be redirected toward their own models, cloud products, office software, search, advertising, recommendation systems, coding tools and internal R&D.
That does not make the investment automatically correct. Internal absorption lowers the risk that unused compute immediately becomes a liquidity crisis; it does not eliminate capital misallocation. Overinvestment can still appear as higher depreciation, margin compression and weaker returns on invested capital.
For these firms, a mistake is therefore more likely to show up first as a return problem rather than a survival problem.
Model companies possess a different option: the option to innovate. A weak model can be replaced; an API can be repriced; a foundation-model business can move toward agents, applications, open weights or enterprise services.
Talent can leave, algorithms can diffuse and organizations can become rigid, so innovation is no guarantee of survival. But human capital, algorithms and organizations retain more room for recombination than assets already frozen into facilities, power contracts and debt.
The more exposed group is the pure compute lessor.
6. Why compute factories may be the first place trouble appears
The business model of a pure compute provider is straightforward: raise large amounts of capital, buy GPUs, lease or build data centers, secure electricity, and rent the compute to model companies and other customers.
When AI compute is scarce, this can be an excellent business. The problem is that once the assets are built, they are difficult to reshape.
A weak model can be changed. An empty GPU hall still depreciates. Power contracts still have to be honored. Interest still has to be paid.
The main risk is therefore not that AI disappears. It is that AI demand remains real but compute is no longer as scarce as investors assumed.
Small changes in supply and demand can then move economics quickly. Hourly GPU prices soften; competitors appear; customers negotiate harder; five-year commitments become one-year commitments; discounts on long contracts widen. Revenue may still be growing while risk is already increasing. If GPUs eventually sit idle, the pressure reaches cash flow and debt.
The core capability of a compute lessor is not simply owning many GPUs. It is the ability to deliver useful compute that customers can actually use and are willing to pay for, at the lowest possible cost in capital, electricity and time. “Useful” matters: not theoretical FLOPS on a specification sheet, but reliable compute after networking, cooling, scheduling and operational losses.
The best operators secure power faster, build faster, run denser and more reliable clusters, maintain higher utilization and finance themselves more cheaply.
But these advantages may be less durable than algorithmic advantage. Power can be competed for. Liquid cooling can be copied. Network designs diffuse. GPUs are sold to many buyers. When capital is abundant, competitors arrive.
Once supply catches demand, a scarce asset can become an ordinary infrastructure business very quickly.
That is why the pure compute-rental layer may be the best place to look first for a local AI bubble. Not because all bubbles must appear there, but because this layer is unusually sensitive to pricing, utilization and financing.
7. Three leading indicators of an AI infrastructure bubble
If that thesis is right, monitoring AI excess cannot begin with NVIDIA orders. By the time upstream orders visibly weaken, the demand shift may already be well advanced. Earlier signals should come from the firms selling marginal compute every day.
We currently use three observation windows. They are not three stock recommendations, and they are not guaranteed to trigger in a strict sequence. They measure different parts of the chain—price, contract realization and financial stress. Persistence and co-movement matter more than a single datapoint.
Nebius is a price thermometer. On September 17, 2026, the company announced that it would raise pay-as-you-go prices for certain NVIDIA GPUs effective October 1, its second increase in three months, citing strong demand for AI compute.[5] This is not a comprehensive market index, but it is a useful check on whether marginal high-end compute remains scarce. The warning would be sustained declines in comparable spot and on-demand prices, wider discounts, and easier GPU availability.
Nscale is a contract-to-capital test. As of August 31, 2026, it reported approximately $103.4 billion of active and contracted total contract value supporting about 461,000 active or contracted GPUs. Its Anthropic agreements alone provide for up to roughly $44.6 billion of payments, yet the IPO filing also stated that binding financing commitments for the GPU and data-center infrastructure required to fulfill those agreements had not yet been obtained.[6] This is the real-world version of “a contract is not revenue.” The variables to watch are financing closure, delivery timing and any renegotiation of terms.
CoreWeave is the financial transmission gauge. In the second quarter of 2026, 98% of its revenue came from committed contracts; its top three customers accounted for 36%, 26% and 10% of quarterly revenue. The company also warns that a shift from take-or-pay commitments toward more pay-as-you-go consumption would reduce cash-flow visibility and could pressure margins and financial results.[7] The key is therefore not whether quarterly revenue is still growing, but whether contract duration and quality, customer concentration, cash flow, interest burden and refinancing costs begin to weaken together.
Taken together, the three windows ask: is compute pricing softening; can future contracts still attract capital and reach delivery; and has that pressure reached the income statement and balance sheet?
We deliberately do not declare that a 5% price decline is a warning and 10% is a bubble. GPU generations, regional power prices, contract duration and customer credit differ too much for one universal threshold. A hard number would create false precision. What matters is sustained deterioration over months or quarters and confirmation across several indicators.
As of late September 2026, that chain has not yet appeared. Nebius is still raising prices, Nscale is still attracting substantial capital, and committed contracts still account for the great majority of CoreWeave’s revenue. At the same time, capital markets are demanding more compensation for AI-linked debt. Reuters reported on September 22 that spreads on AI-related corporate bonds were around 115 basis points, versus roughly 78 basis points for the broader investment-grade market.[8]
The better description today is therefore: financial markets have begun repricing AI expansion, while the physical compute market remains strong.
8. What if we are still underestimating AI’s potential?
At this point the analysis has to turn around and ask the opposite question. Markets worry that investment is running too fast. But what if we are still underestimating how far machine-intelligence demand can expand?
First, AI may not be one S-curve. Individual products move from explosion to adoption to slower growth, but AI contains several curves at once: chat, coding, multimodality, enterprise workflows and now agents that can execute continuously. Earlier applications are not fully mature before later ones begin to accelerate. Overlapping S-curves can keep aggregate demand growing faster for longer than any single product curve.
Second, AI expands labor supply in a different way from previous technologies. It does not add people; it increases the amount of equivalent cognitive labor each person can mobilize. Traditional knowledge work faces a hard limit: there are only so many people and hours. With agents, a person can set goals and permissions while machines continue searching, analyzing, coding, testing and organizing. One person can orchestrate several AIs, and machine work does not have to stop when the person stops.
This means the AI market is not limited to replacing existing work. When analysis, experimentation, translation or software modification becomes dramatically cheaper, tasks that were previously too expensive or too slow to attempt can be created. Cheaper machine cognitive labor can produce more versions, more verification, deeper research and more parallel attempts.
Third, AI does more than increase quantity. Our earlier work separates four effects: doing more in the same time; doing the same task better; doing things that were previously beyond reach; and allowing one person to organize multiple AIs, extending individual capacity toward what once required a team. The first two effects fit reasonably well into traditional productivity measures. The latter two are much harder to capture with one metric.
The result may be a structural shift that is easy to underestimate: humans are no longer the only bottleneck in the supply of cognitive labor. As machine cognitive labor expands, physical constraints—chips, electricity, networks and compute centers—can re-emerge as the binding limits. That helps explain why Microsoft, Amazon, Google and Meta may rationally accept some overbuilding rather than discover, after demand takes off, that they lack compute.
China’s low-price and open-model strategy may reinforce the same process. Cheaper intelligence does not necessarily mean lower total compute demand. If usage grows faster than efficiency improves, lower cost can push AI into more tasks and more environments. The important variable is not compute per task, but how much additional work society delegates to machines after the price falls.
No statistical system yet captures this demand perfectly, but it is not completely invisible. Useful proxies include token and inference volumes, the duration of continuous agent execution inside enterprise workflows, penetration of automation in coding, customer service and research, and the elasticity of task volume after the price of intelligence falls. None is complete; together they are more informative than user counts alone.
This remains a hypothesis that must be allowed to fail. If agents do not create durable economic value, if enterprises remain in perpetual trial mode rather than becoming dependent on AI, or if compute costs fall faster than new task creation rises, today’s infrastructure buildout can still prove excessive.
But by this point, “capex is very large” is no longer an adequate bubble test. The comparison that matters is between the expansion of physical compute supply and the amount of new cognitive work humanity is actually handing to machines. The first is easy to count. The second is only beginning to become visible.
9. We know where to look first
After this research, I am less interested in answering the question “Is AI a bubble?” The question is too coarse.
The more useful question is: if investment really outruns demand, who feels the pain first?
For now, the best answer is the pure compute-rental layer. Not because it is certain to fail, and not because bubbles can only appear there, but because it combines heavy fixed assets with limited flexibility and direct exposure to marginal compute prices, utilization, customer credit and financing costs.
Our three judgments therefore remain intact. Long-term demand for machine intelligence may be underestimated by traditional user models. Large industrial firms may rationally accept some overbuilding rather than risk a compute shortage. Capital markets are already differentiating among AI assets, so the more useful risk to watch is local excess in assets, financing and valuation rather than treating the entire AI system as one bubble.
The thesis also has a clear direction in which it can fail. If actual AI usage stops expanding materially, agents fail to create durable economic value, enterprises reduce deployment, and compute prices, contract quality, utilization and financing conditions deteriorate persistently together, then the thesis of rapidly expanding machine cognitive labor has to be re-examined.
That is the difference between judgment and belief. We do not trust every AI investment because AI is powerful. Nor do we declare the whole sector a bubble because capex is large.
The job is to identify where pressure should appear first, and then watch the facts.
The internet eventually proved that a great technological revolution and enormous local bubbles can coexist.
AI may be similar. The difference is that this time, we already know where to look first.
References
[1] NVIDIA, “GTC Taipei 2026 Keynote with Jensen Huang,” 31 May 2026. https://www.nvidia.com/en-us/on-demand/session/gtctaipei26-stw61044/
[2] Reuters, “Investors nervous about AI spending slowdown after industry warnings,” 15 September 2026. https://www.reuters.com/legal/transactional/investors-nervous-about-ai-spending-slowdown-after-industry-warnings-2026-09-15/
[3] OpenRouter, “DeepSeek V4 Is Earning Agentic Token Share,” 30 June 2026. https://openrouter.ai/blog/insights/deepseek-v4-adoption/
[4] OpenRouter, “In-Region Routing: keep your data in the US or EU,” September 2026. https://openrouter.ai/blog/announcements/us-in-region-routing/
[5] Reuters, “Nebius hikes AI cloud prices again as demand for computing power soars,” 17 September 2026. https://www.reuters.com/technology/nebius-hikes-ai-cloud-prices-again-demand-computing-power-soars-2026-09-17/
[6] Nscale, Registration Statement / IPO Prospectus, filed 18 September 2026, SEC. https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm
[7] CoreWeave, Quarterly Report for the quarter ended 30 June 2026, SEC. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm
[8] Reuters, “Corporate bond buyers get picky with flood of AI debt,” 22 September 2026. https://www.reuters.com/legal/transactional/corporate-bond-buyers-get-picky-with-flood-ai-debt-2026-09-22/
Sources current as of 24 September 2026. OpenRouter figures refer only to traffic observed on that platform and do not represent all global AI usage.
