The Froth Around the Machine

AI Bubble Risk, Real Infrastructure, and the Coming Market Flush

When people hear warnings about an AI bubble, many immediately sort themselves into two familiar camps.

One camp says the whole thing is fake. The demos are fake, the productivity gains are fake, the companies are fake, the valuations are fake, the future is fake, and anyone still taking the industry seriously has simply failed to notice that the emperor is wandering around in his underwear.

The other camp says the opposite. AI is the next industrial revolution. The spending is justified. The valuations are justified. The data centers are justified. The debt is justified. The weird circular financing arrangements are justified. Every skeptical question is a failure of imagination. Every concern is cope. Every chart that goes up and to the right is destiny putting on its shoes.

Both reactions are too simple.

The more useful position is also the more annoying one: AI is real, powerful, and economically important, and that is precisely why so much froth has gathered around it. Bubbles do not usually gather around nothing. They gather around something real enough to make people lose discipline.

Railroads were real. The internet was real. Electrification was real. Cars were real. Radio was real. Canals were real. Smartphones were real. Crypto, even at its most ridiculous, still pointed at real questions about settlement, custody, scarcity, coordination, and programmable financial infrastructure. The problem was not that every mania had no object. The problem was that the object became a permission slip for everyone nearby to stop doing arithmetic.

That is the danger around AI now.

Not that the technology is fake.

That the financing layer is getting drunk.

The Temptation to Call It Case Closed

There is a deep emotional satisfaction in declaring an entire industry fraudulent.

It cleans up the world. It turns a complicated situation into a morality play. It lets the skeptic say: I saw through it before the fools did. It lets the exhausted observer stop tracking details. It lets everyone skip the tedious work of distinguishing between durable firms, speculative firms, thin wrappers, debt machines, AI-washing consultants, real laboratories, infrastructure bottlenecks, and benchmark-chasing science projects that may be brilliant yet commercially doomed.

“AI bubble” becomes a magic phrase. Say it once and the matter is settled.

But this is not analysis. It is mood management.

The world is rarely kind enough to make technological revolutions clean. Important technologies arrive covered in grift, hype, mediocrity, copycats, false prophets, and finance people who smell a narrative before they understand the mechanism. The clown car is not evidence that there is no engine. It is often evidence that an engine exists and every opportunist in sight is trying to strap a lawn chair to it.

The Bank for International Settlements recently warned that the AI investment boom has started to resemble older technology manias, including canal mania, British railway mania, late-1920s electrification exuberance, and the dotcom boom. It also noted that the five largest hyperscalers are expected to spend over a trillion dollars on AI-related capital expenditure from 2025 through 2026, with commitments outpacing earnings and free cash flow in some cases. That is not a small footnote. That is a serious warning from the central bank of central banks.

But the historical analogy cuts both ways.

Those old booms did produce wreckage. They also produced infrastructure.

Railway mania ended badly for many investors. We still have trains.

The dotcom bubble destroyed companies, careers, portfolios, and delusions. We still have the internet.

The likely question is not whether AI survives its froth. It almost certainly does. The better question is who survives the purge, who gets absorbed, who gets exposed, and who was only ever a PowerPoint presentation wearing a GPU necklace.

The Core Is Different From the Costume

The most important distinction is between AI as capability and AI as costume.

AI as capability is when a company has a real technical asset, real distribution, real compute access, real integration into existing workflows, real customers, or real cost savings. It may still be overvalued. It may still spend too much. It may still disappoint investors. But the core object exists.

AI as costume is different.

That is when an existing mediocre business changes the vocabulary on its website. Suddenly the dashboard is an AI dashboard. The search bar is an AI copilot. The analytics product is an AI insight engine. The chatbot is an agent. The rule-based script is an autonomous workflow. The old recommendation system is now an intelligence layer. The same consulting deck that once said blockchain, metaverse, digital transformation, and Web3 now says artificial intelligence in tasteful gradient lettering.

This is where the stink is strongest.

The SEC has already charged firms for making false and misleading claims about their use of AI, including investment advisers that marketed AI capabilities they allegedly did not have. The phrase “AI washing” is not just a snarky internet complaint. It is now a regulatory category with enforcement behind it.

That matters because the public conversation often treats AI companies as one bucket. This is a mistake. A hyperscaler building data centers is not the same thing as a tiny SaaS wrapper reselling someone else’s API with a charming landing page. A chip company with actual demand is not the same thing as a startup claiming to revolutionize dentistry, sales, recruiting, tutoring, therapy, tax filing, and legal research by Tuesday afternoon. A frontier lab with strong models is not the same thing as a public company that discovered the word “agentic” during earnings season.

Some of the industry is building machinery.

Some of it is selling smoke near the machinery.

Benchmarks Are Not Balance Sheets

This distinction can hurt, because there are groups doing genuinely good technical work that may still be financially fragile.

A lab can have impressive benchmarks and a terrible business.

A model can be elegant, capable, and widely admired while the company behind it bleeds cash. A startup can produce something that makes engineers nod and still fail to build distribution. A team can publish brilliant research and still discover that cloud bills, inference costs, hiring costs, depreciation, customer acquisition, and price competition are less impressed by the benchmark chart than Twitter was.

This is one of the crueler parts of the AI cycle.

Benchmarks measure capability under particular conditions. They do not automatically measure margin, customer stickiness, pricing power, defensibility, cash runway, cost of capital, legal exposure, hardware access, or whether the product becomes a feature inside Microsoft, Google, Apple, Alibaba, Tencent, Amazon, Meta, OpenAI, Anthropic, Huawei, Baidu, ByteDance, or whatever platform already owns the user relationship.

A beautiful benchmark can attract attention.

It cannot pay interest on debt.

This is why some of the coming damage may feel unfair. The purge will not only punish charlatans. It may also punish sincere builders who were early, clever, and technically serious, but financially exposed. A company can be right about the direction of history and wrong about the timing, burn rate, pricing model, or capital structure. That is not hypocrisy. That is capitalism with a shovel.

There will be model teams that deserve admiration and still get folded into larger firms. There will be robotics companies with real progress that priced themselves as if mass adoption had already happened. There will be data center projects that are genuinely useful but financed on assumptions that only work if demand, power availability, chip supply, customer prepayments, and interest rates all behave politely at the same time.

The market does not ask whether the work was interesting.

It asks whether the machine can keep eating.

Where the Froth Lives

The froth is not evenly distributed.

The strongest companies are usually the ones where AI plugs into an existing empire: cloud platforms, search, advertising, enterprise software, chips, devices, operating systems, logistics networks, payment rails, and consumer platforms with huge distribution. These firms can still overbuild. Their stocks can still be repriced. Their executives can still make dumb capital-allocation choices. But they have other organs. If the AI market cools, they do not instantly become a skeleton in a branded hoodie.

The weaker layer is the secondary and tertiary AI economy.

This includes thin wrappers that depend on another company’s model. It includes startups whose moat is a prompt chain. It includes “AI transformation” consultants whose main transformation is turning investor anxiety into billable hours. It includes neoclouds and GPU-rental firms with fragile customer concentration. It includes speculative data-center vehicles whose numbers work only if every future tenant arrives on schedule and stays faithful. It includes companies that were not AI businesses until public markets started rewarding the costume.

This is Froth Country.

Froth Country is not always fraudulent. Sometimes it contains useful services. Sometimes it contains clever engineering. Sometimes it contains companies that would be perfectly reasonable at one-tenth the valuation. The problem is that narrative inflation turns “useful” into “inevitable,” then turns “inevitable” into “worth any price,” then turns “worth any price” into leverage.

And leverage is where jokes become balance-sheet events.

The BIS warning is especially concerned with financial amplification: rising leverage among AI firms, credit-market exposure, weaker supplier balance sheets, opaque financing between hyperscalers, chipmakers, AI labs, neoclouds, and data-center providers, plus the growing role of private credit. It specifically notes that direct lending funds have quadrupled lending to AI and IT sectors over five years, reaching about 15 percent of portfolios.

That is the boring paragraph where the dragon lives.

The danger is not just that some silly AI companies go down. The danger is that enough capital has been routed through enough opaque structures that a disappointment in AI payoffs could travel through suppliers, contractors, private credit, data-center leases, equity markets, and household portfolios. The blast radius depends on how much hidden leverage exists and how many supposedly separate bets are actually tied to the same assumptions.

China’s Froth Looks Different

The Western version of AI froth often looks like brand inflation.

Companies pivot to “AI” because the market likes the label. SaaS firms add copilots. Consultants add “agentic” to every deck. Public companies talk about AI on earnings calls because investors punish silence. The froth is linguistic as much as financial.

China’s froth is different.

There, the strongest AI logic is tied to industrial policy, domestic substitution, chip restrictions, robotics, infrastructure, and national competition. That makes some of the investment more concrete. China really does need domestic compute. It really does need alternatives to restricted foreign chips. It really does have enormous industrial capacity, platform companies, robotics ambitions, and a state apparatus willing to push strategic sectors.

But strategic importance does not repeal arithmetic.

If anything, it can make froth more patriotic. A company is not merely a company. It becomes the national champion, the future of domestic compute, the answer to sanctions, the robot army, the industrial renaissance, the chip independence story, the proof that outsiders underestimated China again. Some of that may be true. Some of it may also be priced three miracles early.

In the West, Froth Country often wears a venture hoodie.

In China, it may wear a hard hat, a robotics demo, a chip roadmap, and a local-government ribbon-cutting ceremony.

The substance is real. The risk is that real substance gets capitalized at fantasy speed.

The Enema Comes Anyway

Every major buildout seems to attract people who believe the new thing has abolished the old disciplines.

It never has.

Cash flow still matters. Debt still matters. Margins still matter. Customer concentration still matters. Depreciation still matters. Energy still matters. Distribution still matters. Trust still matters. The cost of serving a customer still matters, even if the customer is dazzled by the demo. The cost of acquiring a customer still matters, even if the founder says “viral loop” with priestly conviction.

AI changes many things.

It does not change the fact that money must eventually come from somewhere.

That is why the enema phase is so useful, even if it is unpleasant. It flushes out lazy assumptions. It punishes costume businesses. It forces customers to ask what they are actually buying. It forces investors to separate infrastructure from theater. It forces labs to confront the gap between capability and commercialization. It forces data-center projects to prove demand beyond the fever chart. It forces every “AI-first” business to explain whether AI is the engine, the accessory, or the perfume.

This is the part where the skeptics will be tempted to gloat.

They should be careful.

When the dotcom bubble burst, it did not prove the internet was fake. It proved that many internet companies were fake, premature, overvalued, badly managed, or dependent on capital markets that had briefly mistaken traffic for destiny. The internet kept going. In fact, the post-bubble internet became far more consequential than the bubble internet. The wreckage cleared space for more disciplined builders.

Railroads followed a similar pattern. The mania was absurd. The overbuilding was real. Investors were burned. Yet rail infrastructure became part of the modern world. The bubble did not invalidate the railway. It invalidated a great many claims about who would profit, how quickly, and on what terms.

AI may follow the same pattern.

The crash, if it comes, will not be a clean philosophical verdict. It will be an accounting event with technological consequences. Some companies will die. Some will merge. Some will become features. Some will become infrastructure. Some will quietly turn out to have been right all along, just early and overcapitalized. Some will survive because they had boring distribution, not because they had the most romantic story.

The future rarely gives prizes for narrative purity.

What Is Not Froth

It is important to say clearly what is not froth.

The fact that AI models can write, summarize, code, search, classify, tutor, translate, draft, reason imperfectly, generate media, compress workflows, augment researchers, help analysts, assist programmers, operate tools, and increasingly coordinate multi-step tasks is not froth.

The fact that people already use these systems every day is not froth.

The fact that companies are reorganizing around them is not froth.

The fact that countries care about domestic compute is not froth.

The fact that chips, power, cooling, networking, memory, and data-center logistics have become strategic inputs is not froth.

The fact that open models continue to improve is not froth.

The fact that local AI may eventually make many cloud subscriptions look bloated is not froth.

The fact that a teenager, retiree, artist, writer, engineer, hobbyist, teacher, small business owner, or ordinary household may gain access to forms of cognitive leverage that used to require institutions is not froth.

There is a real thing here.

The froth is the assumption that every company standing near the real thing deserves a premium valuation.

The froth is the belief that every benchmark converts cleanly into profit.

The froth is the idea that model capability automatically creates a moat.

The froth is the pretense that debt-funded infrastructure cannot overbuild because “demand for intelligence is infinite.”

The froth is the investor deck where “AI” serves the same function as incense in a temple.

The froth is not the machine.

The froth is what gathers around the machine when everyone wants to be close enough to its glow.

The Better Test

The better test is not “Is this AI?”

The better test is: What does AI let this entity do that it could not do before, and who pays for that difference?

That question cuts through much of the fog.

If AI reduces a real cost inside a large existing operation, that matters.

If AI increases revenue across a platform with huge distribution, that matters.

If AI gives a chipmaker, cloud provider, or robotics manufacturer a durable advantage in a constrained market, that matters.

If AI lets a small team do work that once required a large team, that matters.

If AI produces a wonderful demo but no durable buyer, that is interesting but not necessarily a business.

If AI produces a benchmark gain but doubles serving costs, that is a technical achievement with a financial question mark attached.

If AI is just a word used to repackage old software, that is not a revolution. That is seasoning.

This is why the AI economy needs a more adult vocabulary. “Bubble” is too blunt. “Revolution” is too flattering. “Scam” is too lazy. “Inevitable” is too religious.

A better phrase might be: real transformation with localized mania.

That is less fun to chant. It is also much closer to the truth.

After the Flush

The coming years may be emotionally confusing.

There may be headlines about failed AI startups, private-credit stress, data-center cancellations, GPU oversupply in specific regions, down rounds, layoffs, mergers, abandoned copilots, disappointing enterprise deployments, lawsuits over exaggerated claims, and public companies quietly walking back the promises they made when money was cheaper and attention was easier.

At the same time, AI systems may keep getting better.

Both can be true.

A financial correction can happen alongside technical progress. A valuation collapse can coincide with genuine adoption. A startup winter can happen while ordinary users get better tools. Investors can lose money while civilization gains infrastructure. The people who bought the wrong companies at the wrong prices can suffer while the technology becomes more deeply embedded in everyday life.

This is why the “case closed” instinct is so dangerous.

If a bubble pops, the lesson will not be that AI was nothing. The lesson will be that AI was real enough to attract too much money too quickly into too many weak structures.

The enema will come for the froth.

The machine will remain.

And after the flush, the question will become much clearer: who was building the future, and who was merely standing beside it with a rented spotlight?

- Iarmhar

July 1, 2026

This essay is part of the Compute, Geopolitics, and Civilizational Strategy Cluster