When Real Technology Meets Bad Investment: AI, Railways, Chinese EVs, and the Next Infrastructure Cycle

AI is usually discussed in two extreme ways. One side treats it as an inevitable productivity revolution. The other side calls it a bubble. I think both views miss the more useful point.

AI can be a real technology cycle and still create a bad investment cycle. Those two things are not contradictions.

Why Start With the BIS?

The Bank for International Settlements, or BIS, is often described as the bank for central banks. Its role is not to pick startup winners, review consumer products, or hype the next software platform. Its mission is to support monetary and financial stability through international cooperation.

That is why I think the BIS report is a useful place to start. When a venture capitalist talks about AI, the focus is usually product potential. When a big tech company talks about AI, the focus is usually platform strategy. When the BIS talks about AI, the focus is different: financial conditions, capital expenditure, debt, valuation, private credit, and systemic risk.

In other words, the BIS is not asking whether AI is impressive. It is asking whether the financial system is starting to price AI as if the future is already guaranteed.

The BIS Is Not Saying AI Is Fake

The important point in the BIS report is not that AI has no value. The report actually recognizes AI as a technology with potential productivity gains. The warning is about the speed and structure of the investment boom around it.

The BIS compares the current AI investment cycle with earlier historical episodes such as the canal mania of the 1830s, British railway mania in the 1840s, electrification enthusiasm in the late 1920s, and the dot-com boom in the late 1990s. The common pattern is simple: a real technological breakthrough attracts more capital than short-term commercial returns can justify.

That is the distinction I care about:

The technology can be real.
The infrastructure can be useful.
The long-term productivity effect can be meaningful.
But many companies and investors can still lose money.

Real Technology Does Not Guarantee Good Investment

Railways were real. They changed transportation, logistics, and market access. But not every railway company was a good investment.

The internet was real. It changed communication, media, commerce, and software distribution. But many dot-com companies disappeared.

Electric vehicles are real. Batteries, software-defined vehicles, smart cockpits, and assisted driving have changed the auto industry. But that does not mean every EV brand deserves to survive.

AI may follow the same pattern. The mistake is not believing that AI matters. The mistake is assuming that every AI company, every AI infrastructure project, and every AI valuation will be justified.

Why Compare AI With Chinese EVs?

I bring Chinese EVs into the discussion because the situation looks structurally similar, even though the products are completely different.

China's EV industry is not a fake technology story. It is one of the clearest examples of a real industrial transition: battery costs declined, product quality improved, software became a major differentiator, and consumers started to expect smart cockpits, driver assistance, and frequent feature updates.

But the business cycle around that transition became brutal. Too many companies chased the same direction at the same time. Capacity expanded quickly. Price competition intensified. Suppliers faced payment pressure. Regulators and industry executives started worrying about overcapacity and long-term financial health.

That is why Chinese EVs are useful as a comparison for AI. The lesson is not that EVs failed. They did not. The lesson is that a correct technology direction can still create too many companies, too much capacity, and too much capital chasing the same future.

AI has a similar setup:

Chinese EVs had factories, batteries, charging networks, and supply chains.
AI has GPUs, data centers, power systems, cloud contracts, and model infrastructure.
Chinese EV companies fought for market share through price, features, and speed.
AI companies fight for market share through model capability, API pricing, developer tools, and enterprise distribution.
Chinese EVs created real winners, but also forced consolidation.
AI may create real winners, but many labs, neoclouds, and thin AI application companies may not survive.

The Pattern: Build First, Consolidate Later

Most major technology waves seem to follow a similar pattern:

A real technical breakthrough appears.
Investors extrapolate adoption too aggressively.
Companies rush to build before demand is fully proven.
Infrastructure is duplicated across many competitors.
Costs rise faster than proven cash flow.
Weak companies fail or get acquired.
The remaining infrastructure, talent, standards, and user habits stay.

This is why a failed investment cycle does not necessarily mean a failed technology cycle. The same may happen with AI.

What Is Being Overbuilt in AI?

The AI cycle has its own version of heavy infrastructure:

GPUs and AI accelerators
Data centers
Power generation and grid capacity
Cooling systems
Cloud infrastructure
Foundation models
Model serving platforms
AI developer tools
Enterprise workflow integrations

Some of this will become durable infrastructure. Some of it may become stranded capacity.

The hard part is knowing which layer has real end-user demand and which layer only looks strong because everyone is racing to not fall behind.

This is the same measurement problem I wrote about with layoffs. When you cannot cleanly separate real value from what only looks strong, capital flows toward the appearance, and the correction shows up later.

The Risk Is in the Capital Structure

The most important part of the BIS warning has nothing to do with model quality. It is about who is paying for the buildout, and with what.

If AI infrastructure were funded entirely out of the operating cash flow of a few very large technology companies, a slow return cycle would be uncomfortable but survivable. Those companies can wait. The problem is that the spending has already outgrown that cash flow. The five largest hyperscalers are on track to commit more than a trillion dollars to AI capital expenditure across 2025 and 2026 combined, which is running ahead of their earnings and free cash flow and pushing some of them to raise debt to cover the gap. Once debt enters the picture, the buildout stops behaving like a software upgrade and starts behaving like a credit cycle.

What makes it fragile is the wiring between the players. The same companies are often each other's investors, customers, and suppliers at the same time. A cloud provider takes an equity stake in a model lab, and the lab commits to buy compute back from that provider. A chip maker extends favorable terms to a buyer whose demand is itself backed by expected future revenue. Data centers get built against long-term lease and purchase commitments rather than proven, paid usage. Capital ends up circling inside the same ecosystem, which makes the real level of risk hard to see and hard to price.

That structure has a clear failure path. If AI revenue comes in slower than the commitments assume, companies cut capital expenditure. When capex slows, the suppliers and data center operators on the other side of those commitments lose expected revenue. If those suppliers are carrying debt, the stress does not stay inside tech. It moves into private credit and the broader financing system. That chain, not the question of whether the models are good, is what a central bank is actually watching.

For Builders, the Lesson Is Different

The list of "durable" AI capabilities is easy to write and easy to fake, so let me be explicit about the test I am using. A capability survives a financing reset if someone is already paying for it out of an existing budget because it removes a specific, current cost. Not because it might pay off later. Not because it is impressive in a demo. The question is whether the spending is tied to a line item that already exists.

By that test, the things most likely to stay are the ones sitting directly on top of work people already do and already pay for. AI-assisted coding and debugging replaces engineering time. Document understanding and retrieval over private data replaces hours of manual search and review. Workflow automation removes steps a person was doing by hand. The less visible layers survive for the same reason. Evaluation, validation, cost-aware model routing, and human review exist because they make the paid usage cheaper or safer to run, so they get funded as long as the usage does. What tends not to survive is the opposite pattern, a thin wrapper around a model that has no budget line of its own and only works if adoption keeps accelerating.

I have an obvious bias here, since some of what I build lives in exactly this layer. But the test is the same regardless of who applies it: follow the existing budget, not the projected one.

The Strongest Case Against This View

I should give the other side its best shot, because it is a good one.

The bull case is that this cycle is not like the railways or the dot-coms in the way that matters most. Those manias were driven by large numbers of thin, leveraged companies. This buildout is led by a few of the most profitable companies on the planet, with real revenue engines behind them in advertising, cloud, and devices. Balance sheets like that can absorb years of overspending without breaking. The demand also looks more real than it did in 2000. The BIS itself notes measurable productivity gains at the task level, and a lot of the usage is already being paid for by enterprises, not just measured in clicks. And compute and power are not single-purpose the way a specific rail line was. Even after a correction, that capacity can run something else.

I think that case is mostly right about the technology and mostly wrong about the timing. The cash-rich argument is real, but it is the exact thing that is thinning out. The spending has already passed what current cash flow covers, which is why debt and cross-financing keep showing up. And "the productivity gains are real at the task level" is not the same claim as "the returns will arrive fast enough to justify what is being priced in today." That gap is the whole risk. So the two sides are closer than they sound. The technology can be as real as the bulls say, and the financing can still reset the way the BIS is warning about.

The Main Takeaway

A technology cycle and an investment cycle are not the same thing.

AI may still become a major productivity layer, but that does not mean today's spending pace, valuations, or company count are sustainable. If the market corrects, it may not prove that AI was useless. It may only prove that too many companies tried to build the same future at the same time.

That is why the Chinese EV comparison matters. EVs were a real technology shift, but the industry still had to deal with overcapacity, price wars, supplier pressure, and consolidation. AI can be real and still go through the same kind of financial cleanup.

The better question is not whether AI is a bubble or a revolution. It may be both: a real technology revolution wrapped inside an overheated capital cycle.

History suggests that companies may disappear, investors may lose money, and valuations may reset. But if the technology is genuinely useful, the infrastructure and product habits left behind can still reshape the next decade.

References

The sources below cover the three parts of the argument: why the BIS is relevant, the financing risk behind AI capital spending, and the comparison with China's EV industry.