Nicholas Mitsakos

Long-term value is being created. It’s unlikely the people financing it today will capture it.

The technology is extraordinary. The economics are not.

Artificial intelligence has triggered the largest coordinated capital commitment of the modern technology era. The builders will spend close to $800 billion this year and approach $1 trillion next year on chips, data centers, networks, and power. Total AI revenue this year will be roughly $150 to $200 billion.

AI will reshape the global economy, and much of the capital financing that transformation will be impaired. Revenue, productivity, and cash flow will all arrive, but on the wrong schedule, after much of what was built to capture them has lost its economic value.

We have seen this movie. Railroads. Telephone and wireless networks. Fiber. The commercial internet. Each transformed the economy. Each destroyed the capital of those who built it first. The infrastructure survived. The investors did not.

The question is not whether the AI opportunity is real. The question is who pays for the gap between when the capital is spent and when the value arrives.

Capital and Clocks

Every capital cycle fails because timing is mismatched: assets, financing, and demand run on different clocks. AI’s value creation – from assets to demand – is further apart than in any of these previous cycles.

1. The technology clock runs in two to three years. That is the economic life of a leading-edge AI chip before a new generation makes it uncompetitive.
2. The financing clock runs in five to twenty-plus years. Bonds, leases, joint-venture debt, and power-purchase agreements are all long-dated.
3. The adoption clock runs in a decade or more. That is how long it historically takes for a general-purpose technology to show up in economy-wide productivity.

Long-dated debt is financing short-lived assets whose payoff depends on a slow-moving economy.

That mismatch is underestimated risk. This buildout is illiquid, long-dated obligations secured by assets that reprice every eighteen months.

Industrial Mobilization, Cash Flow to Credit

These are not technology budgets. This is industrial mobilization.

Hyperscaler capital spending was roughly $387 billion in 2025. One leading ratings agency now expects about $785 billion this year and close to $1 trillion in 2027, raising that forecast by $85 billion in a single revision. Some projections put cumulative AI capital investment above $5 trillion over the next four years.

What has changed is not only the scale, but the source of capital.

The first phase was funded from the most prodigious cash flows in corporate history. That phase is ending. The group has taken on roughly $460 billion of debt to support AI spending. Its bond issuance rose from under $17 billion in 2024 to more than $190 billion so far this year.

One major investment bank estimates that more than half of the $2.9 trillion to be spent on data centers between 2025 and 2028 will come from external capital, with private credit supplying roughly $800 billion. Google, the most cash-rich company in the group, just reported its first quarterly free-cash-flow deficit since going public in 2004.

Asset-light businesses have become capital-intensive operators. Equity markets have not finished repricing that change. Credit markets have barely started.

The Math Nobody Wants

Most AI debates mix dreams, prophecies, unsubstantiated predictions, and fantasy.

Enthusiasts forecast transformation. Skeptics forecast disappointment. Neither is a business model you can value. The better question is how fast earnings must grow to justify what is already committed.

After the cost of capital, a 15% return, and depreciation, the AI sector must raise its productivity roughly 2.7x by 2030 to break even. That is roughly what the U.S. economy achieved during the information-technology boom that began in the mid-1990s. That boom took a decade. This one has four years.

No problem for the markets, apparently. Perhaps someone should use a calculator. Revenue growth from $200 billion to trillions isn’t growth so much as a leap of faith.

The music may stop

Skepticism is not analysis. But circular revenue is not sustainable.

First-quarter cloud growth across the three largest providers was the fastest since 2021, on a base three times larger. One platform’s AI revenue run rate passed $37 billion, up over 100%. The hyperscalers added about $700 billion of contracted, undelivered revenue in two quarters. Demand still exceeds supply.

But demand from whom? A large share comes from a handful of frontier model developers that are themselves losing money and funding their compute purchases with investor capital that expects the same outsized returns. Chip suppliers invest in model developers that buy their chips. Infrastructure providers extend credit to tenants that rent their capacity. Circular demand is still demand, until the funding stops.

Demand is circular, and while the backlog is real, this level of demand may not be sustainable.

A Hopeful Bet

The buildout is a bet by the capital markets and the economy.

The builders must monetize capacity.

Selling more compute is not enough if prices fall faster than volume rises. Open-weight and smaller specialized models are closing the gap with frontier systems, and for most commercial work, good enough at a fraction of the cost wins.

Usage can explode while the value captured by frontier infrastructure grows far more slowly. Volume is not pricing power. Adoption is not profit.

The customers must earn a return.

Enterprises will buy AI for a few years because they fear falling behind. But sustainable revenue must demonstrate value.

The evidence is thin.

A National Bureau of Economic Research survey of nearly 6,000 senior executives found that 89 percent saw no effect from AI on labor productivity over the prior three years. They don’t expect much over the next three years. Task-level studies show real gains in coding, customer service, research, and document work.

A tool can improve a task without improving an enterprise’s productivity. This is where the economic narrative regarding AI gets lost.

Productivity is an organizational achievement, not a software purchase. It requires redesigned workflows, clean data, accountability, and incentives. Put AI on top of a broken process, and you get the same broken process, faster. The largest gains will go to organizations built from a blank slate around what AI produces. Incumbents retrofitting legacy operations will capture far less than they spend.

Frontier models must hold their premium.

The buildout assumes the most expensive models, running in the largest facilities, remain worth their price. Every month, smaller models running on phones, laptops, and enterprise servers erode that assumption.

The public must consent.

Data centers need power, water, land, transmission, and tax concessions. The costs are local. The benefits are remote. Executives surveyed expect to raise productivity partly by cutting headcount.

Pair job losses with higher electricity bills and new gas plants serving distant shareholders, and local resistance hardens. Technical progress does not grant political permission. Social license is an economic asset. Once lost, it costs a lot to recover, if that is even possible.

The collision

If productivity comes from cheap models on modest hardware, the economy wins, and hyperscalers lose.

If productivity comes from eliminating jobs, public resistance blocks the capacity revenue depends on. Winning one component can cost another. The market prices each leg as though it doesn’t affect the others.

Five companies cannot justify this investment by earning more from one another. The returns must reach the economy: manufacturers producing more, hospitals treating patients better, drug developers shortening discovery, banks managing risk, small businesses competing, and workers doing more valuable work.

If AI merely transfers profit from customers to cloud providers, customers will resist. If savings never become outputs, products, or wages, the response and the losses will be severe.

Chips Are Not Infrastructure

A data center is not one asset. It is several assets with very different lives. The land and building last decades. The chips may be uneconomic in three years.

Accelerators and related electronics are roughly 60% of project cost. Most operators depreciate them over five to six years. Their economic life may be two to three years. A chip does not need to stop working to lose value. It only needs to become uneconomic relative to the next generation. Some estimates put the resulting understatement of depreciation above $170 billion from 2026 through 2028.

The result is a treadmill.

Operators must recover the cost of the current cluster while financing the next. Skip the upgrade and the facility becomes a stranded shell. Make it, and you finance a second buildout before the first has paid for itself. Good for AI. Dangerous for the owner of yesterday’s equipment.

A bridge never faces a competing bridge twice as productive every two years. Calling a data center infrastructure disguises the technology risk inside it.

Off Balance Sheet Risk

When the AI companies spent their own cash, a bad bet stayed with their shareholders. Debt, private credit, leases, joint ventures, supplier financing, power contracts, and residual-value guarantees now move that risk outside these companies.

One ratings agency estimates that the hyperscalers hold roughly $662 billion of signed data-center leases that have not commenced and do not appear on their balance sheets. That is nearly a year of capital spending, off balance sheet and almost invisible.

Then there’s Louisiana.

Here’s an example of the almost impenetrable, confusing structures that can potentially be alarming if the music stops.

A single campus in rural Louisiana, announced in late 2024 as a two-gigawatt, $10 billion project, now spans five gigawatts and more than $50 billion.

Meta sold 80 percent to a private-credit manager through a joint venture, which owns a landlord entity that leases the buildings back to the platform’s subsidiary. The venture issued bonds to institutional investors. The platform backstops the structure with a residual-value guarantee starting near $28 billion.

Now look at the mismatched timing.

The leases run four years at a time, a term that roughly matches the life of the GPUs inside. The platform has committed to buy power for twenty years. The state’s largest utility is building gas generation sized at about six times New Orleans’ electricity consumption.

So the technology lasts about four years, and the financing is about 20 years, and no one knows if anyone will need any of this capacity in five years.

The structure is elegant, but it may be “too clever by half.” It gives the platform flexibility and brings in outside capital. That capital may be much riskier than it’s priced. If demand shifts, chips age out, or the tenant leaves, who owns the residual value? The answer sits in a bond portfolio, a private-credit fund, a pension plan, an insurance balance sheet, and a utility rate base. Many holders do not know they own it.

Anyone who lived through 2008 recognizes the pattern. Special-purpose vehicles are back. Complexity is again being sold as sophistication. Visible risk gets priced. Invisible risk accumulates.

Financial engineering can move risk. It cannot erase it. These financial structures are far more vulnerable, and they carry far more risk than is acknowledged.

Power is Power

In every technology buildout, value migrates to the binding constraint. In AI, that constraint is not models or chips. It is power.

The International Energy Agency expects global data-center electricity use to more than double by 2030, to roughly 945 terawatt-hours. In the United States, data centers could drive nearly half of electricity demand growth through the decade.

Generation is not enough. Power must arrive through transmission lines, substations, transformers, and interconnection queues measured in years. A site with land and fiber but no firm power is not a data-center site. It is a grassy field.

That creates a different value hierarchy. GPUs depreciate. Power rights, interconnection positions, transmission access, and permitted land appreciate. The durable winners may be utilities, generators, electrical-equipment makers, engineering firms, cooling suppliers, and owners of scarce powered land—unglamorous businesses with sustainable long-term returns.

The migration does not stop at power. As good-enough intelligence becomes abundant, margins move from the model to the application, the proprietary data, the workflow, and the physical system.

Intelligence gets cheap. Power, data, and trust stay scarce.

Geopolitics

One reason this cycle will run longer than the economics justify is that governments have decided compute is strategic.

AI capacity now sits alongside semiconductors, critical minerals, and energy as an instrument of national power. Every investment decision in this cycle is simultaneously a capital decision, a national-security decision, and a civilizational choice.

Governments will subsidize, permit, and protect capacity that markets would never finance alone. No platform, and no nation, can afford to underinvest in the defining technology of the era.

Geopolitics does not replace economics.

A data center can be strategically essential to a nation and financially ruinous to the investor who owns it. Strategic value and investment return are different. Investors care about a return on capital, and geopolitical strategy does not. Investors may be severely disappointed despite glorious speeches and pronouncements.

Three Ways This Ends

Predicting when a bubble bursts is a fool’s errand. But we know what can happen in different scenarios. The market may be orderly, disrupted, or corrected.

1. Orderly absorption. Enterprise productivity arrives faster than expected, pricing holds, and demand fills capacity before the chips age out.
2. Rolling retrenchment. Spending flattens between 2027 and 2029. We write down weaker facilities, refinance them, or repurpose them. Equity holders absorb losses; the system absorbs them without crisis.
3. A “spontaneous combustion event.” A major tenant declines to renew, guarantees are called, private-credit marks fall, and ratepayers inherit stranded generation. Leverage and opacity turn an equity correction into a financial one.

Some in the industry now hope for a crash to restore discipline. Be careful what you wish for.

The dot-com collapse cost hundreds of thousands of jobs and helped push the country into recession. The 2008 crisis, built on leverage and opaque structures, was far worse. The more AI debt spreads through pensions, insurers, and utilities, the more the next bust will resemble an economic crisis rather than a sector collapse.

The Bubble and the Revolution

AI is both a bubble and a technological revolution.

The telecom boom overbuilt fiber and destroyed enormous capital; the fiber became the internet’s backbone. The dot-com crash eliminated companies and wealth; the internet transformed the economy anyway. Railroads went bankrupt repeatedly while permanently changing commerce. Technology survives bad capital allocation. Equity often does not.

But fiber was general-purpose and lasted decades. An AI campus is engineered around one hardware generation’s power density, cooling, and networking. Some facilities will remain strategic. Others will become expensive shells with excellent electrical connections.

The assumption is that more compute will keep producing intelligence valuable enough to pay for it.

That assumption has worked remarkably well, but it is not a law of nature. Progress is already shifting toward better data, inference-time reasoning, post-training, specialized architectures, and smaller models. Each path creates value.

What to Watch

Belief or skepticism is not analysis. Here are the essential components behind the AI industry.

1. Utilization, not capacity. A gigawatt under development is not a productive asset.
2. Revenue per unit of compute. Falling inference costs help adoption and hollow out infrastructure returns.
3. Customer economics. Measurable gains in revenue, cost, speed, quality, or risk.
4. Depreciation and refresh spending. Earnings may look attractive until you factor in capital reinvestment.
5. Total obligations. Leases, guarantees, power contracts, joint ventures, and supplier commitments, not just reported debt.
6. An operator dependent on one tenant, one model developer, or one financing channel is not diversified infrastructure.
7. Power liability. Who pays if contracted demand never arrives?

The Decision

Artificial intelligence will transform science, medicine, engineering, finance, and decision-making. That conviction is exactly why this capital structure matters. A transformative technology does not need a fragile financial foundation.

The danger is not that AI proves useless. The danger is that a powerful technology is attached to a capital structure that requires perfection: continuous scaling, rapid adoption, broad productivity gains, stable political support, abundant power, and trillions of dollars of high-margin revenue arriving on schedule.

That is too many assumptions, each with its own risks, and when you combine them, the industry’s risk profile is much higher than people are willing to concede.

AI will not collapse, but the industry will not be a near-term triumph either. Some assets will earn extraordinary returns while others are refinanced, repurposed, or stranded. The technology will advance, and capital will be restructured.

AI may define this era. The uncertainty is whether today’s owners, creditors, communities, and ratepayers capture enough of its value to justify what they’re building with their capital.

Selected Sources

• Goldman Sachs, global AI investment forecast and AI infrastructure scenario analysis
• Moody’s Ratings, hyperscaler capital expenditure forecasts and lease-commitment research (2026)
• Jessica Wachter et al., Wharton working paper on AI investment and required productivity growth (June 2026)
• National Bureau of Economic Research, firm-level data on AI and productivity (w34836)
• International Energy Agency, Energy and AI
• Meta Platforms SEC disclosure on the Hyperion joint venture
• MIT Technology Review, “What must happen for AI’s trillion-dollar gamble to pay off” (September 2026), including comments from Gary Gensler and Stijn Van Nieuwerburgh