Sillygoose Sillygoose

The AI Bubble: What Happens When the Hype Runs Out of Capital and Compute?

Reading Time: 9 minutes

“The technology can change the world without making everyone financing it rich.”

The AI Bubble: What Happens When the Hype Runs Out of Capital and Compute?

China’s $480 Billion Warning, AI’s Financial Books, and the Expensive Business of Predicting the Future

China’s $480 Billion Warning

The first serious crack appeared where investors thought the growth was unstoppable

China’s technology market has delivered an unpleasant reminder about the difference between buying into a revolution and paying a sensible price for it. By October 26, the Shanghai STAR 50 index, home to many of the country’s semiconductor and AI-linked businesses, had fallen about 40% in less than four months. Roughly $480 billion in market value had disappeared. Reuters’ analysis of the sell-off pointed to stretched valuations, doubts about demand, tax pressures and the prospect of restrictions affecting Chinese suppliers. Some of the companies providing the equipment for the AI boom were hit harder than the market itself, with individual share-price falls running to 50%, 60% or even 70%.

That last detail matters. Investors weren’t simply turning against novelty chatbots or startups with questionable business plans. They were also selling the chipmakers and hardware suppliers meant to profit regardless of which AI application won. The assumption that selling picks and shovels guarantees a fortune turns out to have one fairly obvious weakness: you can still overpay for the shovel factory. China’s broader CSI 300 had also retreated around 20% from its June high, so the weakness was not confined to one fashionable technology index. It was a sharp regional repricing, with consequences for investors and companies hoping to raise fresh capital.

Does that prove the global AI bubble has burst? No. American and European markets have their own economics, financing structures and investor bases, and the underlying technology continues to improve. But calling the Chinese correction irrelevant would be equally silly. The market has now demonstrated how quickly enormous expectations can be revised when the returns stop looking inevitable.

Key Takeaway: China’s approximately $480 billion sell-off is evidence of a severe AI-related market correction, not proof that AI technology has failed or that an identical global crash is guaranteed.

The Forecast That Makes Investors Nervous

Trillions of dollars are being committed against a future nobody can audit yet

The scale of the international wager is easier to understand when you put the forecasts next to the actual business problem. In September, Gartner estimated worldwide AI spending at $2.67 trillion for 2026 and forecast $3.64 trillion for 2027, compared with approximately $1.79 trillion in 2025. These totals cover a broad market, including services, software, security, AI applications and infrastructure. They are not the revenue of OpenAI, Anthropic or the model industry alone. Even with that distinction, nearly a trillion dollars of additional annual spending in a single year is a spectacular expectation to place on any sector.

Worldwide AI spending Gartner estimate / forecast
2025 estimate $1.79 trillion
2026 forecast $2.67 trillion
2027 forecast $3.64 trillion

The largest single category is the physical machinery behind the magic trick. Gartner forecast approximately $1.48 trillion of AI infrastructure spending in 2026, climbing to almost $1.98 trillion the following year. Those estimates imply a remarkable appetite for servers, chips, networks and data-centre capacity. Some of it is surely justified. Companies are deploying AI into actual workflows, and better models are expanding what software can do. Yet there is still an uncomfortable gap between the ability to sell more compute and the ability to demonstrate that its eventual users will earn enough value to sustain that spending. A forecast is a map of expectations, not a receipt.

OpenAI’s Projected Burn and Anthropic’s Actual Filings

The headline numbers are alarming enough without confusing different kinds of losses

The financial documents behind the AI boom are far more interesting than another benchmark chart. In September, the Financial Times reported on an OpenAI presentation projecting roughly $278 billion in cumulative negative free cash flow from 2026 through 2030. The same reporting described around $856 billion in expected computing and infrastructure expenditure over that period, against ambitions for revenue to increase sharply. These are company projections reported by journalists, rather than audited losses already suffered. They depend on assumptions about future customers, pricing, financing, and how much it will cost to deliver increasingly capable models. The scale is still difficult to ignore: even a business growing exceptionally fast can find itself consuming more capital than it generates.

Anthropic’s numbers give us another view, this time through a public filing rather than a leaked forecast. In its IPO registration materials, the Claude developer reported approximately $33.2 billion in revenue for the first nine months of 2026 and a $104 billion net loss. That enormous loss needs explaining before somebody turns it into a social-media infographic. Roughly $94 billion reflected non-cash accounting changes associated with financial instruments; it was not $94 billion spent on electricity and GPUs. The company also disclosed approximately $895.4 billion in aggregate spending commitments, predominantly for compute and infrastructure, spread across future periods. Those contractual obligations are substantial, but they are not equivalent to an immediate operating expense.

Reading these figures carelessly makes the bubble story sound more dramatic and less convincing. A valuation is not money in the bank. A projected cash deficit is not an audited historical loss. A non-cash accounting charge is not the same as paying a supplier. And a long-term purchase commitment is not a bill due tomorrow. The more useful question concerns the relationship between growth, margins and future obligations. Can companies selling powerful AI turn remarkable demand into durable profitability before their financing requirements become a problem? The answer will depend on far more than whose chatbot gets the most enthusiastic reviews.

Key Takeaway: AI financial headlines mix actual operating performance, non-cash accounting losses, projected cash burn and future contracts. Treating them as interchangeable hides the real economic risks.

The Money-Go-Round

One company’s investment becomes another company’s future revenue promise

There is another reason the market is nervous, and it has less to do with whether the models are intelligent. Many of the biggest AI companies are tied together through investment, cloud contracts, chip orders and infrastructure financing. A technology giant invests in a model developer. The developer commits to buying computing services. The cloud provider expands its data centres and places more orders with hardware suppliers. The arrangement may be perfectly legitimate and commercially useful, but it can make several balance sheets depend on the same assumptions about tomorrow’s demand.

The International Monetary Fund examined 14 publicly disclosed circular-financing transactions since 2023 and identified approximately $300 billion in equity investment, $650 billion in debt and $1.4 trillion in future revenue commitments. These are distinct categories of financing and promises, not $2.35 trillion of sales already made. The IMF’s concern is straightforward. If a major participant struggles, the investors backing it may suffer losses just as the suppliers depending on its future purchases discover their revenue forecasts were optimistic. The same network that helped finance expansion can transmit disappointment remarkably efficiently.

It would be a mistake to suggest that all AI revenue is circular or artificial. Businesses are purchasing real services, and cloud demand is substantial. The danger lies in the concentration. When enough companies invest on the assumption that several other companies will continue spending at an extraordinary rate, a setback becomes difficult to isolate. As Reuters reported ahead of late-October technology earnings, investors were scrutinising massive capital budgets and large contracted cloud backlogs to see whether growth would translate into cash flow. Nobody is particularly eager to interrupt the money-go-round while it is accelerating. Stopping it, however, becomes much more expensive after everyone has borrowed against the next turn.

Silicon, Electricity and the Great Physical Reality Check

Software scales in minutes. Power stations have rather different schedules.

This is where the financial debate joins the technical one. In our earlier Gridlock coverage of hardware bottlenecks, we examined the simple constraint that gets lost in glossy AI presentations: chips, advanced memory, cooling equipment and reliable power cannot be summoned on demand. A model provider may announce a new architecture in an afternoon, but connecting a data centre to the grid can involve years of planning, permits and construction. Even once the infrastructure is available, it carries depreciation and maintenance costs. Every generation of increasingly efficient chips also threatens to make the previous expensive installation less competitive.

Agentic tools add another wrinkle. The software that once responded to a single prompt can now inspect repositories, run tests, call several tools and revisit its own output repeatedly. As we covered in The AI Developer Tool Explosion, that shift creates genuinely useful new workflows, but the customer-facing simplicity can conceal a surprisingly busy backend. The apparent productivity gain must be measured against the computational work required, the human review still needed and the subscription or API charges that eventually show up. An agent that saves a developer two hours may be excellent value. An agent that spends a fortune producing work someone has to redo is simply a more elaborate way to lose money.

Efficiency will improve; that’s one of the strongest arguments against the most pessimistic bubble predictions. Smaller models, better inference software and more capable hardware could reduce the cost of many AI tasks. The catch is that cheaper compute can also encourage businesses to run far more of those tasks. There is no automatic rule saying lower costs per request will lead to a lower total bill. The infrastructure market is betting not just on better intelligence, but on extraordinary growth in the amount of intelligence people choose to consume.

The Regulatory and Sovereignty Question

The cheapest cloud may not be the one businesses are allowed to use

While investors argue about valuations, governments and businesses are making their own calculations about control. Dependence on one model provider, one foreign cloud platform or one source of advanced chips creates operational risks that don’t disappear because an API is convenient. Copyright disputes over training data, privacy requirements, export controls and national-security concerns all complicate the assumption that intelligence will move freely across borders. None of this means public training data has been universally outlawed or that the open internet is legally dead. The actual legal picture varies by jurisdiction and remains contested.

China’s pursuit of domestic AI capability is one expression of that pressure, though it should not be presented as a direct consequence of October’s stock-market correction. The interest in sovereign AI predates the sell-off. In Europe, companies have their own reasons to examine data location, procurement exposure and the trade-offs between hosted frontier models and smaller systems they can run more privately. The result may be a more fragmented market in which some workloads stay with global providers while others move to regional infrastructure or specialised on-premises models. That doesn’t automatically make self-hosting cheaper; organisations still pay for hardware, security, maintenance and expertise.

For software firms, this opens a practical question that is more useful than declaring the future belongs to one kind of model. Which operations actually require frontier-level capability, which need a specialist model, and which would be better handled by conventional software? The companies answering those questions intelligently may end up with stronger products and healthier margins than competitors that insert an LLM into every possible interaction. We may be entering a period where architecture and cost discipline matter as much as access to the latest model. Frankly, that sounds like software engineering again.

Key Takeaway: A financial correction could strengthen demand for efficient, portable and sovereign AI systems, but geopolitics, privacy and vendor risk are already driving that shift independently of the market.

What a Bubble Actually Looks Like in 2027

Watch financing costs and actual margins, not just a dramatic stock chart

There are several ways the next year could go, and the least satisfying one for headline writers may be the most realistic. Some AI-linked shares could fall sharply while the industry as a whole continues growing. Valuations might stagnate for years as earnings gradually catch up. Or the most optimistic demand forecasts could materialise, especially if AI agents begin handling genuinely valuable work reliably enough for companies to budget around them. These aren’t equally likely predictions; they are different economic paths, and nobody can honestly assign them certainty from a November 2026 vantage point.

The useful indicators will be harder to dress up in a keynote. Look at operating cash flow after infrastructure spending, not just headline revenue growth. Look at the gap between committed compute capacity and paying demand. Watch refinancing costs for infrastructure-heavy companies, changes in token pricing, customer retention, and whether enterprise buyers can demonstrate measurable returns. The IMF’s October warning made the stakes clear: an AI correction could spread through concentrated, leveraged financial relationships, even while the technology continues to deliver genuine productivity benefits.

The dot-com comparison is worth keeping precisely because the internet survived. Once the speculative excitement died down, the web continued changing almost every part of life. Many investors in the first wave never recovered their money. AI could follow a similar course, although today’s largest participants have substantial existing businesses and financial resources that many 1990s startups lacked. That difference may limit some risks while introducing others, especially through the immense scale of their infrastructure commitments. Technological inevitability is a poor excuse for financial indiscipline.

Reading List: Beyond the AI Hype Cycle

Useful context for understanding the difference between a technological revolution and a good investment.

Final Thought

The technology does not need the valuation to survive

The most dangerous assumption in the AI boom is that being right about the technology means being right about the price. China has shown how violently that assumption can unravel. The latest financial disclosures from AI developers show an industry building real commercial demand while making commitments on a scale few companies have ever attempted. Meanwhile, regulators are studying the ways those commitments, investments and future revenue promises tie the wider market together. These are no longer hypothetical concerns raised by people who dislike new technology. They are questions arriving in financial statements and stock-market prices.

There is still plenty of reason to be optimistic about what AI can do. It is already making some forms of work faster, and much of its most valuable use may still be ahead of us. But history is full of inventions that changed the world while bankrupting a generation of people who financed their early expansion. Perhaps AI’s economics will improve quickly enough to justify today’s bets. Perhaps the market will demand a brutal reset before the infrastructure finds its customers. The smarter question for 2027 is not whether the machines will keep improving; it is how many companies can afford to keep improving them at today’s prices.

The future may arrive exactly as promised. The invoice is the part nobody has agreed on.

References

Was this helpful?

Thanks for your feedback!

Leave a Reply