Is AI Infrastructure Starting to Crack? Why IoT Vendors Should Understand the Financing Risk

Yes, and it starts with the money. Meta’s Hyperion data center is 90 percent debt-financed, its bonds run until 2049, and the chips inside may be obsolete within three to six years. Investors are repricing such risks right now; the IMF now counts an AI correction among the three key risks to the financial system. Anyone sourcing AI for IoT products from the cloud shares that risk.

What happened in the markets in October?

The bond market moved first. In mid-October 2026, yields on some CoreWeave bonds rose above 15 percent. CoreWeave is one of the so-called neoclouds: specialized providers that rent out computing time on AI chips by the hour. A bond’s yield rises when its price falls; it shows how much interest investors demand for the risk. Insurance against default also became more expensive. Credit default swaps on CoreWeave at times cost more than 10 percent of the insured amount per year.

These figures do not signal an imminent default. But they show how sharply investors are repricing the risk. The Financial Times places one unsecured CoreWeave bond in distressed territory (“America’s $10tn AI infrastructure boom is starting to fray”, paywall). The stock lost almost a quarter of its value within a week. Bonds that finance data centers leased to CoreWeave also fell. On October 14, market research firm Gartner published a calculation that explains the pressure.

By 2030, AI data center operators would need to generate a cumulative $8.5 trillion in revenue to cover capital spending, financing costs and a 12 percent return on invested capital. Gartner puts the AI revenue attributable to these data centers at no more than $3.6 trillion. More than 80 percent of neoclouds would lose their independence by 2030. One day earlier, the International Monetary Fund (IMF) had published its Global Financial Stability Report.

In it, the IMF lists a correction in AI-related asset prices among the three key risks to the financial system. The IMF is particularly concerned about circular financing: AI companies invest in each other, lend to each other and buy from each other. In 14 deals reviewed since 2023, about $300 billion in equity and $650 billion in debt changed hands. They are matched by $1.4 trillion in revenue commitments. If a key company comes under pressure, revenue forecasts fall, and with them the value of the loans.

Why is the financing coming under pressure?

Stijn Van Nieuwerburgh, an economist at Columbia Business School, provides a detailed analysis. In a paper for the Brookings Institution, he estimates US AI infrastructure investment from 2025 to 2032 at around $10.3 trillion: for buildings, power supply, networks and chips. That is an average of 3.63 percent of US GDP per year. The railroad boom of 1870 to 1890 came to 2.24 percent. The hyperscalers, the large cloud companies, can no longer fund these sums on their own. Amazon, Microsoft, Alphabet, Meta and Oracle invested a combined $416 billion or so in 2025. For 2026, company guidance and analysts point to around $800 billion.

For the first time, capital spending thus exceeds the five companies’ operating cash flow, the cash generated by their ongoing business, of around $707 billion. The missing money increasingly comes from outside: through leases, joint ventures, project loans, loans from funds rather than banks (private credit) and special purpose vehicles. These are companies set up for a specific project whose debt does not appear on the parent company’s balance sheet. Van Nieuwerburgh illustrates this with Meta’s Hyperion data center in Louisiana. Meta sold 80 percent of the project to funds managed by asset manager Blue Owl. The joint venture financed around $27 billion of the $30 billion total investment with debt, roughly 90 percent. Meta’s own leverage, measured against its market value, is around 4 percent.

A Wall Street Journal analysis cited by the author puts the off-balance-sheet obligations of the four largest hyperscalers at $2.4 trillion. Their balance sheets show $0.6 trillion. On top of that comes a maturity mismatch. Around two thirds of the cost of an AI campus goes into computing hardware that may become economically obsolete within three to six years. Van Nieuwerburgh assumes a 20-year life for buildings and power connections; the Hyperion bonds run until 2049. For the build-out to pay off, his baseline calculation requires around $3.7 trillion in annual revenue by 2032. For comparison: OpenAI and Anthropic together generate around $100 billion a year. To reach $3.7 trillion on their own, they would have to grow by about 80 percent every year until 2032.

Why are so many construction projects stalling?

Besides money, power, permits and public acceptance are in short supply. At least 47 large data center projects were delayed, withdrawn or blocked in 2026. That is the count of analytics firm RegLayer, the Financial Times reports. RegLayer itself calls the figure conservative. Data Center Watch, a project of AI intelligence firm 10a Labs, tracks local opposition. In the first quarter of 2026, residents and local groups blocked or delayed at least 75 projects worth around $130 billion.

The organization has never counted that many in a single quarter. In the second quarter, another 45 projects worth $68 billion followed. Data Center Watch counts delays, not just final cancellations. How much of the announced capacity will actually be built remains open. According to the FT, analysts at Bernstein consider only about 35 percent of the planned US pipeline credible. Brookings arrives at a similar order of magnitude using its own method: of a little over 500 gigawatts of planned capacity, around 183 gigawatts go into operation by 2032 in the baseline scenario, a little over a third. Around 227 gigawatts, more than 40 percent, are never built. Delays hit financing directly. Debt service is fixed, while revenue starts later or falls short of plan, Van Nieuwerburgh writes.

According to the FT, investors have been repricing this risk since mid-October.

Is this a bubble or a financing problem?

The stress sits in a specific place: with providers that build on debt and have few other sources of revenue. Thanks to their broad revenue base, the hyperscalers could sustain AI investment for years, Gartner writes. Demand for computing power remains high. CoreWeave grew revenue by 112 percent in the second quarter and reported a backlog of around $104 billion. But interest alone ate up about a quarter of that revenue.

Amsterdam-based Nebius, also a neocloud, raised its price for one hour on an Nvidia H100 chip by 17 percent on October 1: from $3.85 to $4.50. Van Nieuwerburgh explicitly considers it premature to attribute to the build-out a systemic risk comparable to earlier credit booms. The more important question, he argues, is under what conditions losses on individual projects occur at the same time and spill over to banks and investors.

The IMF, too, describes a risk, not a forecast. In its downside scenario, a steep drop in AI investment would lower global economic output by 1.1 percent by 2028, and US output by 1.8 percent. Even then, the Fund does not expect a global recession, Reuters reports. These figures do not answer whether AI as a whole is overvalued. But they show where the math breaks first if revenue fails to materialize.

What does this mean for IoT and AI projects?

Many connected products now source AI from the cloud: image recognition in cameras, voice control, anomaly detection in machine data, assistants in service platforms. Anyone buying such services from neoclouds or through API platforms, meaning providers that offer models via a programming interface, shares three risks.

• The first is supplier risk. If Gartner is right, most neoclouds will be acquired or disappear by 2030. Large AI labs protect themselves contractually. According to the FT, Anthropic can walk away from its $45 billion deal with London-based Nscale if financing or completion of a major project in West Virginia is not on time. Nscale had around 25,000 AI chips in operation in August; it plans to run almost 500,000.
• The second is pricing risk. The Nebius increase shows that computing time can become more expensive even while the industry builds at massive scale. According to Gartner, spending on AI-optimized cloud infrastructure is growing too slowly for most providers to recover their investments before the current chip generation is replaced. Anyone calculating with today’s prices is on shaky ground.
• The third risk is also an opportunity: part of the computing work is moving to the edge of the network. Van Nieuwerburgh explicitly names more efficient language models, edge computing and decentralized inference as risks to demand for large centralized data centers. Edge computing processes data directly on the device or close to it. Inference means running a trained model; it often needs short response times and therefore proximity to the user. Germany’s federal data center strategy likewise expects new demand for edge and distributed computing, for example for IoT, industrial applications and autonomous systems.

Meta shows how quickly the boundary is shifting. On October 15, the company released Muse Spark Open, an open-weight model: its weights, the learned parameters, can be downloaded and run independently under a license with commercial restrictions. The model has 816 billion parameters, of which only 42 billion are active for each processed token. According to Meta, it runs on two networked Mac Studios with a combined terabyte of memory. Models of this size used to run in data centers. For an IoT gateway, it is still far too large. But the direction is clear: powerful models are moving closer to the application.

This leads to four checkpoints for IoT vendors:

• Contracts: What delivery and availability commitments, price adjustment clauses and termination rights does the contract with the AI provider contain?
• Portability: Can the provider be replaced without redesigning the product? Open models make this easier. Muse Spark Open, for example, was available on seven API platforms at launch.
• Costs: Can the business case absorb rising prices for computing time?
• Fallback: Which functions must keep working without a cloud connection, and is a smaller local model enough for them?

Where does Europe stand?

Europe is building on a different scale. Germany’s federal government aims to at least double the connected capacity of the country’s data centers by 2030 compared with 2025, and to at least quadruple capacity for AI. In 2025, it stood at 2,980 megawatts, just under 500 megawatts of it for AI and high-performance computing. According to the strategy, Germany does not yet have large-scale commercial data centers dedicated solely to AI.

The government itself concedes that future demand remains unclear. According to German digital association Bitkom, the US already had around 48 gigawatts in 2024, about 16 times as much as Germany in early 2026. The strategy itself names one key bottleneck: scarce grid connection capacity. In established hubs such as Frankfurt, a grid connection takes seven to ten years on average, and up to 13 years in extreme cases, according to think tank Ember. Among other measures, the government wants to promote flexible grid connection agreements: a data center connects to the grid at reduced capacity and grows as more capacity becomes available.

At EU level, the call for tenders for AI gigafactories, large data centers for training and running AI models, runs until November 12. The EU and its member states are providing up to €10 billion and aim to mobilize a further €20 billion in private investment. Every country in Europe needs such a data center, Nvidia CEO Jensen Huang said in Berlin in November 2025. The occasion was the launch of Deutsche Telekom’s Industrial AI Cloud, which runs on Nvidia chips.

On October 19, 2026, Germany’s economy and research ministries signed a letter of intent with Nvidia on industrial AI. The government has not named an amount. Media had reported beforehand that the federal government intends to provide several hundred million euros, mainly for computing time in the Telekom cloud. That would make the state an anchor customer of an Nvidia platform. At the same time, in its data center strategy, the government is pushing for at least part of the chips in the AI gigafactories to come from a European chip designer. Europe is not immune to neocloud risk either. Nscale is based in London, Nebius in Amsterdam.

According to the FT, both are backed by Nvidia, the supplier of their AI chips. Nvidia is not short of cash: in the quarter ending in late July 2026, the company returned nearly $26 billion to shareholders through buybacks and dividends. At the same time, it is increasingly taking on its customers’ risks. According to its quarterly report, Nvidia provides guarantees of up to $108.5 billion, $105 billion of it for an SB Energy data center campus in Ohio. SB Energy is majority-owned by Japan’s SoftBank and, according to the FT, has yet to complete a single data center. These are the kinds of entanglements the IMF has in mind when it warns about circular financing.

Conclusion

AI infrastructure is not cracking as a whole. What is cracking is its financing, specifically where heavy debt meets thin revenue: at neoclouds, project companies and construction projects without a secured grid connection. The hyperscalers are likely to absorb setbacks; many neoclouds are not. For IoT vendors, this is no reason to panic, but it is a reason to review their own dependence on cloud AI. The next tests: the EU tender deadline on November 12 and CoreWeave’s quarterly results, expected in November. Anyone who builds their AI so that it can switch providers or, if necessary, run locally does not have to wait for the answer to the bubble question.

Is the AI boom a bubble?

There is no proof of that at this point. The IMF counts a correction in AI-related asset prices among the three key risks to the financial system. Gartner sees a gap of almost $5 trillion through 2030 between the revenue AI data centers need and the revenue they are expected to earn. A study for the Brookings Institution, however, considers it premature to identify a systemic risk; the pressure is mainly on the debt-financed part of the build-out.

What is a neocloud?

Neoclouds are specialized providers that rent out computing time on AI chips by the hour, such as CoreWeave, Nebius or Nscale. Unlike Amazon, Microsoft or Google, they have few other sources of revenue and finance their expansion largely with debt. Gartner expects more than 80 percent of them to no longer operate independently by 2030.

Why have bonds of AI data center operators fallen?

Investors doubt whether revenue from AI services will be enough to service the high debt. In mid-October 2026, yields on some bonds of AI cloud provider CoreWeave rose above 15 percent. Construction delays caused by missing grid connections and local opposition also push revenue further into the future.

Is AI computing power in the cloud getting more expensive?

At some providers, yes. Since October 1, 2026, Nebius has charged $4.50 for one hour on an Nvidia H100 chip, up from $3.85. How prices develop in the medium term depends on whether the planned data centers get built and whether demand keeps pace with the build-out.

Can large AI models run locally instead of in the cloud?

Increasingly, yes. In October 2026, Meta released the open model Muse Spark Open with 816 billion parameters, which runs on two networked Mac Studios with a combined terabyte of memory. That is too large for typical IoT devices, which need much smaller models. Running models locally reduces dependence on cloud providers but shifts hardware and maintenance costs to the company itself.

How long does it take to connect a data center to the grid in Germany?

In established hubs such as Frankfurt, seven to ten years on average and up to 13 years in extreme cases, according to think tank Ember. With its data center strategy of March 2026, the German government wants to speed up the process, including through flexible grid connection agreements.

- Advertisement -