The Bubble and the Bisse

When you follow a bisse in summer, the water descends, inexorably, though not always at the same pace. It quickens on the slopes, slows on the flats, presses where the channel narrows, spreads where it widens. What never changes is that it keeps moving. At its surface, now and then, air bubbles appear. Some burst almost immediately, for no discernible reason. Others travel far, round a bend, and eventually dissolve into the very transparency that carried them. Still others, caught in the swirl of a small waterfall, are absorbed into the larger movement that contained them. The bisse itself expects nothing of them.

In the vocabulary of financial markets, the word has come to designate precisely this: a value sustained by its own tension, which bursts, subsides, or is swallowed by a movement greater 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 across the stock market valuations of the sector's major players. Goldman Sachs revises its projections for AI capital expenditure to 527 billion dollars for the year alone. Meta announces a range that pushes its own capex to 145 billion, and its share price falls 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 amplifies the comparison that now passes for self-evident in finance: what we are living through resembles the internet fever of 1999 and 2000.

The comparison is apt. It is also, in the way it is being used, very partially understood.

There is probably a stock market bubble in AI, and it will correct itself — perhaps soon. There is also, beneath this bubble and accompanying it without being defined by it, an economic transformation that does not depend on the fate of share prices. Both propositions coexist. More than that: all else being equal, the first is probably the condition of the second. That is what I wish to argue here.

To understand what a potential AI bubble would do, one must return to what the internet bubble actually did. And one must look at it not through its collapse figures, which are well known and have always occupied an exaggerated place in financial memory, but through what it left behind.

The NASDAQ lost seventy-eight percent in two and a half years. Hundreds of start-ups vanished. Trillions of dollars in market capitalisation evaporated. These figures are true, and they nourished a moral that held for two decades: bubbles betray the promises they carry. It is precisely this moral that needs questioning — because alongside the collapse figures there are construction figures that go largely unquoted. During the fever, and because of it, hundreds of thousands of kilometres of fibre-optic cable were laid under oceans and along continents. Data centres were built at a scale that no rational calculation, left to itself, would ever have consented to fund. Whole cohorts of engineers were trained in competencies that would not have existed without that money spent without restraint. What the bubble burned was the froth — the delirious valuations, the companies with broken business models, the promises without use cases. What it left behind was the substrate. Web 2.0, streaming, the cloud, and ultimately generative artificial intelligence itself deployed on that substrate. Amazon, founded in 1994, lost ninety-five percent of its value between the end of 1999 and early 2002. It survived to become what we know it as today. Google, founded in 1998, went public in 2004 — that is, after the collapse, on the map of an infrastructure the bubble had already left behind.

The economist Carlota Perez, in the years that followed, theorised this phenomenon with a precision that illuminates what we are living through now. Technological bubbles are not, in her view, accidents that occur alongside the revolutions they accompany. They are the operating method by which capitalism finances transformations whose horizon is too distant, whose profitability is too uncertain, whose network effects are too premature to be borne by ordinary rational calculation. The fever, in other words, is not the betrayal of the transformation. It is, in large measure, its financing condition.

Generative AI, at the end of 2025 and beginning of 2026, exhibits several of the signs that presaged the internet bubble in 1999. Private valuations exceeding those of the largest European industrial companies, without revenues to match. Circular financial structures in which Nvidia invests in OpenAI, which buys Nvidia chips, funded through capacity commitments that Nvidia guarantees. A stock market concentration where ten names represent forty percent of the capitalisation of the American index, and where Nvidia's weight alone exceeds that of several G20 countries. A press that oscillates, sometimes within the same article, between religious enthusiasm and biblical warning. Capital expenditure projections that climb, and climb, and climb — without measurable revenues keeping pace.

Everything that characterises a bubble is present.

And yet three differences from 1999 deserve to be named, because they change the nature of what would be left behind.

The first concerns who is bearing the expenditure. In 1999, the bulk of internet investment came from venture-capital-backed start-ups whose balance sheets could neither absorb a reversal nor guarantee continuity of effort. In 2026, the bulk of AI capex is carried by profitable companies — Microsoft, Alphabet, Amazon, Meta — whose cash flows allow, in theory, for the effort to be sustained through a correction. The fever is financed by solid balance sheets, not by promises. This difference does not eliminate the risk; it changes its form. A 2026 correction will probably not cause the outright disappearance of the players; it will cause their deceleration — which is not the same thing.

The second concerns the substance of what is being built. The internet capex of the late 1990s contained a heavy share of customer acquisition, marketing, and speculative software construction that left no trace after the collapse. The AI capex of 2025 and 2026 flows, nine-tenths of it, into physical material — data centres, processors, energy, cooling. A large portion of these assets will be redeployable to other uses, or simply usable for longer than the models they serve today. A graphics card amortised over five years remains useful for workloads that are not AI. A data centre built today for Anthropic can serve something else tomorrow. The physical substrate outlasts the rotation of narratives that once justified it.

The third difference is the most important, and it is the reason this note is written from a mountainside and not from Manhattan. AI use cases are already embedded in the end user's world, to a degree that the internet had not reached in 1999. At the end of the last century, most medium-sized companies were still drawing no productive benefit from the internet; it took until 2003 or 2004 for the cost-benefit ratio to show up in figures. In 2026, in my own practice, in that of a fiduciary in Sierre, in the surgery of a valley doctor, a factor of four to five is measured, banked, integrated into the rhythm of production. A recent NBER study claims that ninety percent of businesses measure no significant impact. I am willing to believe this for what concerns national statistical aggregates, which poorly capture transformations under way in very small structures and among sole traders. But on the ground, in the professions where output is counted in files processed rather than in consolidated accounting lines, individual productivities have shifted. They have shifted so fast that the aggregates have not yet had time to register them. They will eventually, or these productivities will have been superseded by other uses before they are even recorded. But they will not disappear if Nvidia loses half its market value tomorrow morning.

That is the only point worth holding onto. The AI transformation, insofar as it is taking place, does not depend on the stock market valuations of suppliers at the lowest layer. It depends on three things happening elsewhere: the availability of usable models, the capacity of actors to orchestrate them, and the execution infrastructure that runs them. None of the three evaporates with a correction in share prices.

The open models — Llama, Mistral, Qwen, DeepSeek, and their probable successors — now run on modest hardware, at collapsed marginal cost. A stock market correction at the major American laboratories will not recall them. They are published, they are available for download, they are already installed on thousands of local servers around the world. Including in Haute-Nendaz, on the local server that forms the heart of the conversational agent on this site.

Orchestration capacity, for its part, is attached to the user. A lawyer who learned in 2025 to structure his queries, to chain his prompts, to assemble the tools into a production pipeline retains that capacity whatever the stock market fate of OpenAI. This competency, like any competency, is durable. It is, incidentally, precisely what this essay designates when it speaks of the architecture and orchestration work of senior practitioners — that experienced human intelligence which, above the tools, guarantees their value.

Physical infrastructure, finally, is largely deployed, or in the process of being so. The data centres financed in 2024 and 2025 exist. The chips they contain function. The electricity that powers them is, in Europe and particularly in Switzerland, abundant and largely decarbonised. A stock market correction takes none of this back. On the contrary — and this is the reversal one must have the courage to name — it might make this infrastructure more accessible, because its owners, pressed for revenue, will be incentivised to lease it, or even to surrender it, on terms more favourable to downstream users.

Open models. Orchestrated competencies. Physical infrastructure. None of these three resources will be destroyed by a market movement. All will possibly be redistributed on different terms. The transformation will continue. The stock market narrative that accompanies it may collapse without that mattering greatly to the users.

It remains to say what this requires, from the hillside where I write it. How should a canton, an institution, a practice behave in the face of a probable stock market bubble, within a real economic transformation?

The negative answer is simple. It should not wait. Waiting for the bubble to end before equipping oneself — like waiting for the end of inflation to buy a house — is a strategy that works only for those who have no need of what they are waiting for. For everyone else — for the practice that must produce now, for the canton that must integrate its tools now, for the valley doctor who must improve his diagnoses now — waiting is a real, measurable cost: in market share lost, in patients less well served, in clients who went elsewhere.

The positive answer demands a little more precision. The bubble scenario suggests that one must, from now, distinguish the layers on which one relies. Open models deployed on sovereign infrastructure are not exposed to the correction of proprietary platforms. An institution that has 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 running locally, does not carry. The distinction is not ideological; it is operational. Practical sovereignty, in this sequence, is also a hedge against the stock market volatility of one's suppliers.

And there is, finally, a reversal to name because it changes how one can read the moment. If the bubble bursts — which is probable within a horizon of twelve to thirty-six months — the movement will have not only a financial effect. It will have a territorial one. The concentration of capital and talent in metropolitan hubs, which has accelerated since 2022 on AI money, will mechanically slow. The young engineers who left for San Francisco to join an overvalued start-up will, after a severe correction, be fewer in number. The capital that flooded urban ecosystems will redeploy elsewhere. And the territories that will have, during the fever, managed to install their own capacities — through open tools, through training of their active populations, through sovereign infrastructure — will better capture the returning tide.

A collapsed bubble is not, in this reading, the enemy of the territorial diffusion of cognitive capital. It is, perhaps, its accelerator.

It is for this reason that I continue to believe the moment to position oneself is now, not after. The bubble, if it exists, is currently financing the infrastructure from which we will benefit in 2030. It is accelerating the development of open models that, without the money pouring into them, would not exist. It is even, in its contradictory way, financing the very demonstration of its own redundancy — since every dollar spent in AI capex reduces the cost of use for the end user, and it is precisely this fall in cost that makes the suppliers' margins untenable in the long run.

A bubble, in the end, is a water that holds itself at the surface for as long as the tension sustaining it endures. A bisse is a water that flows 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 the markets believe or have ceased to believe. That is the difference between what finances itself and what installs itself.

The bubble will eventually empty. The bisse, for its part, will continue to flow.

Glossary

• Capex — abbreviation of capital expenditures. Refers to a company's investment spending on durable assets: data centres, machines, infrastructure, equipment.
• NBER — National Bureau of Economic Research. American economic research bureau, founded in 1920, which publishes the reference working papers in macroeconomics in the United States.
• API — Application Programming Interface. A software interface through which one program calls the functions of another remotely. 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 market valuations in the AI sector, March 2026.
• Goldman Sachs Research, projections of global AI capital expenditure 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 business 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.