The bubble and the bisse

The word bubble is back in the spring of 2026. A bubble, looked at closely, is a promise made of surface: a thin skin that seems to hold a volume greater than itself. What the Internet sequence of 2000 has to teach us about the one now unfolding before our eyes — and why the bisse, for its part, keeps flowing.

Follow a bisse in summer and the water moves downhill, relentlessly, though not always at the same pace. It speeds up on the steep stretches, slows in the flats, presses forward where the channel narrows, spreads out where it widens. What never changes is that it keeps going. On its surface, now and then, air bubbles form. Some burst almost at once, for no obvious reason. Others travel a long way, round a bend, and eventually dissolve into the same transparency that carried them. Others still, caught in the swirl of a small fall, are swallowed by the larger movement that held them. The bisse itself expects nothing from them.

In the vocabulary of financial markets, the word has come to mean exactly that: a value sustained by its own tension, which bursts, dissipates, or is absorbed by a movement larger than itself.

In the spring of 2026, the word is back. Capital Economics writes, at the end of March, that the artificial intelligence bubble has already burst on the stock valuations of the sector's major players. Goldman Sachs revises its projections for global AI capital spending to 527 billion dollars for the year alone. Meta announces a range that puts its own capex at 145 billion, and its stock loses five percent in the following session. The Bank of England and the International Monetary Fund, within days of each other, warn of a correction that would weigh on global growth. Ray Dalio publicly repeats the comparison that has by now become received wisdom in finance: what we are living through resembles the Internet fever of 1999 and 2000.

The comparison is accurate. It is also, in the way it tends to be used, only partly understood.

There is probably a stock market bubble around AI, and it will likely correct, perhaps soon. There is also, underneath that bubble, an economic transformation that the bubble accompanies without defining, and which does not hinge on the fate of share prices. The two propositions coexist. More than that: the first is probably the condition for the second.

To understand what a possible AI bubble would actually do, we need to go back to what the Internet bubble really did, and look at it not through its collapse figures — well known, and always overrepresented in financial memory — but through what it left behind.

The NASDAQ lost seventy-eight percent over two and a half years. Hundreds of start-ups vanished. Trillions of dollars in market value evaporated. These figures are true, and they fed a moral that held for two decades: bubbles betray the promises they carry. That moral deserves scrutiny, because alongside the collapse figures sit construction figures that get cited far less. During the fever, and because of it, hundreds of thousands of kilometers of fiber-optic cable were laid under the oceans and along the continents; data centers were built at a scale no rational calculation, left to itself, would ever have agreed to fund; entire cohorts of engineers were trained in skills that would not have existed without that money spent without counting the cost. The bubble burned off the froth — the delirious valuations, the companies with broken business models, the promises with no use behind them — and left the substrate on which Web 2.0, streaming, the cloud, and eventually generative AI itself would later be built. Amazon, founded in 1994, lost ninety-five percent of its value between late 1999 and early 2002; it survived to become what we know today. Google, founded in 1998, went public in 2004 — after the collapse — riding an infrastructure the bubble had already left in place.

The economist Carlota Perez, in the years that followed, theorized this phenomenon with a precision that illuminates what we are living through now. Technology bubbles, in her account, are not accidents that happen alongside the revolutions they accompany; they are the mechanism by which capitalism finances transformations whose horizon is too distant, whose returns are too uncertain, whose network effects arrive too early to be carried by ordinary rational calculation. The fever, in other words, is not a betrayal of the transformation. It is, to a large extent, how that transformation gets financed.

Generative AI, in late 2025 and early 2026, shows several of the signs that preceded the Internet bubble in 1999: private valuations that exceed those of Europe's largest industrial companies, with no significant revenue to match; circular financing arrangements in which Nvidia invests in OpenAI, which buys Nvidia chips, which it finances through capacity commitments that Nvidia itself guarantees; a concentration in the stock market where ten companies account for forty percent of the American index's capitalization, and where Nvidia's weight alone exceeds that of several G20 economies; a press that swings, sometimes within the same article, between religious enthusiasm and biblical warning; capex projections that keep climbing, climbing, climbing, without measurable revenue following at the same pace.

Everything that defines a bubble is there.

Three things, however, set 2026 apart from 1999, and they change the nature of what a bursting bubble would leave behind.

The first concerns who is carrying the spending. In 1999, most Internet investment came from venture-backed start-ups whose balance sheets could neither absorb a downturn nor guarantee the effort would continue. In 2026, most AI capex is carried by profitable companies — Microsoft, Alphabet, Amazon, Meta — whose cash flows can, in theory, sustain the effort through a correction. The fever this time is financed by solid balance sheets rather than by promises. This difference changes the shape of the risk without eliminating it: a correction in 2026 will probably not cause these players to simply disappear, only to slow down, which is not the same thing.

The second concerns the substance of what is being built. Internet capex in the late 1990s carried a heavy share of customer acquisition, marketing, and speculative software builds that left no trace once the bubble collapsed. AI capex in 2025 and 2026 goes, nine-tenths of it, into physical hardware: data centers, processors, power, cooling. A large share of these assets will be redeployable to other uses, or simply usable for longer than the models they currently serve. A graphics card depreciated over five years remains useful for workloads that have nothing to do with AI; a data center built today for Anthropic can serve something else tomorrow. The physical substrate outlives the turnover of the narratives that justified it.

The third is the most important, and it is why this piece is being written from the versant rather than from Manhattan. AI use is already embedded at the end-user level, to a degree the Internet had not reached by 1999. At the end of the last century, most mid-sized companies still drew no productive benefit from the Internet; it took until 2003 or 2004 for the cost-benefit ratio to show up in the numbers. In 2026, in my own practice, in a trust company's offices, in the practice of a doctor in a mountain valley, a factor of four to five is being measured, banked, and built into the rhythm of production. A recent NBER study claims that ninety percent of companies register no measurable impact; I am willing to believe that, as far as national statistical aggregates go, since they capture poorly what is happening inside very small structures and among the self-employed. On the ground, in professions where output is counted in files handled rather than in consolidated accounting lines, individual productivity has shifted, and so fast that the aggregates have not yet caught up with it. They will catch up eventually, or these gains in productivity will have been overtaken by other uses before they are even recorded. Either way, they will not vanish if Nvidia loses half its market value tomorrow morning.

This is the one point worth holding onto. The AI transformation, to the extent it is happening, plays out somewhere other than in the stock valuations of the providers sitting at the bottom layer. It depends on the availability of usable models, on the capacity of the workforce to orchestrate them, and on the execution infrastructure that runs them. None of the three evaporates with a market correction.

Open models — Llama, Mistral, Qwen, DeepSeek, and their likely successors — now run on modest hardware, at a marginal cost that has collapsed. A stock market correction at the big American labs will not take them back: they are published, downloadable, already installed on thousands of local servers around the world. Including in Haute-Nendaz, on the local server that serves as the backbone of this site's conversational agent.

Orchestration capacity, for its part, belongs to the user. A lawyer who learned, in 2025, to structure requests, chain prompts, and link tools into a production sequence keeps that ability regardless of what happens to OpenAI's share price. That skill, like any skill, is durable. It is, incidentally, exactly what this essay means when it speaks of the architecture and orchestration work carried out by senior professionals: that seasoned human judgment which, standing above the tools, is what guarantees their value.

The physical infrastructure, finally, is largely built, or well on its way to being built. The data centers financed in 2024 and 2025 exist, the chips inside them work, and the electricity that powers them is, in Europe and especially in Switzerland, abundant and largely decarbonized. A stock market correction will take none of that back. It might even, and this is the most counterintuitive point in the whole sequence, make this infrastructure more accessible: its owners, short of revenue, would have every incentive to lease it out, or even sell it off, on terms more favorable to downstream users.

Open models, orchestrated skills, physical infrastructure: none of these three resources will be destroyed by a market movement. All of them may well be redistributed on different terms. The transformation will continue, and the market narrative that accompanies it can collapse without that mattering much to the people actually using the technology.

What remains is to say what this calls for, from the versant where I write this. How should a canton, an institution, a firm behave in the face of a probable stock market bubble sitting on top of a real economic transformation?

The negative answer is simple: do not wait. Waiting for the bubble to end before equipping yourself is like waiting for inflation to end before buying a house — a strategy that only works for those who do not actually need what they are waiting for. For everyone else — the firm that has to produce now, the canton that has to integrate its tools, the doctor in a mountain valley who has to improve diagnoses — waiting carries a real, measurable cost: market share lost, patients less well served, clients who went elsewhere.

The positive answer calls for a bit more precision. The bubble scenario suggests distinguishing, starting now, between the layers one relies on. Open models, run on sovereign infrastructure, are not exposed to a correction hitting proprietary platforms. An institution that built itself, in 2025 and 2026, around the APIs of OpenAI or Anthropic carries a risk that an equivalent institution, built around Llama or Mistral run locally, does not carry. The distinction is operational before it is ideological. Practical sovereignty, in this sequence, also functions as protection against the market volatility of the providers.

One last reversal changes how this moment can be read. If the bubble bursts, which seems likely on a horizon of twelve to thirty-six months, the effect will not be only financial; it will also be territorial. The concentration of capital and talent in metropolitan hubs, which has accelerated since 2022 on the back of AI money, will mechanically slow down. The young engineers who used to move to San Francisco to join an overvalued start-up will do so less often after a severe correction, and the capital that had been flooding urban ecosystems will redeploy elsewhere. The regions that, during the fever, managed to build their own capabilities — open tools, workforce training, sovereign infrastructure — will be better placed to catch that returning flow. A bursting bubble, read this way, might actually accelerate the territorial spread of cognitive capital rather than derail it.

This is why I continue to think the moment to position oneself is now, not later. The bubble, if it exists, is financing the infrastructure we will draw on in 2030; it is accelerating the development of open models that, without the money pouring into them, would not exist; it is even financing, in its own contradictory way, the demonstration of its own uselessness, since every dollar spent on AI capex lowers the cost of use for the end user, and that falling cost is precisely what makes providers' margins unsustainable over time.

A bubble, in the end, is water held at the surface for as long as the tension holding it up lasts; a bisse is water moving downhill. The first can vanish from one moment to the next without anything around it changing; the second keeps flowing, in the channel it has carved for itself, indifferent to what markets believe or stop believing. The whole difference between what gets financed and what takes root lies there, and it is visible to the naked eye, in summer, when you follow the water. Bubbles burst, dissipate, or get absorbed. The water keeps moving downhill.

Glossary

• capex: short for capital expenditures. A company's spending on durable assets: data centers, machinery, infrastructure, equipment.
• NBER: National Bureau of Economic Research. The American economic research bureau, founded in 1920, which publishes the reference working papers on macroeconomics in the United States.
• API: Application Programming Interface. A software interface through which one program remotely calls the functions of another. In the case of AI, it is through an API that an application accesses a model hosted on its provider's servers.

References

• Capital Economics, note on the correction of stock valuations in the AI sector, March 2026.
• Goldman Sachs Research, global AI capital expenditure projections for 2026, April 2026.
• Meta Platforms, first-quarter 2026 financial communication and 2026 capex guidance, April 2026.
• National Bureau of Economic Research, working paper on the impact of generative AI on enterprise productivity, February 2026.
• Bank of England, Financial Stability Report, spring 2026.
• International Monetary Fund, Global Financial Stability Report, April 2026.
• Carlota Perez, Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages, Edward Elgar Publishing, 2002.