August 2026

By Georg Chmiel  —  August 2026
Short Video: https://youtu.be/4qJCGxjwO_8

Tulips, fibre and GPUs — why I expect a correction in AI, why it changes nothing about the destination, and why the debt is the part worth watching

Part I: Four Hundred Years of Getting This Wrong

In the winter of 1636, in the taverns of Haarlem and Alkmaar, a single bulb of Semper Augustus changed hands on paper for a sum that would have bought a canal house in Amsterdam. In the first week of February 1637, the buyers stopped turning up. Within a month the market had collapsed.

Tulip mania is one of the older bubble story, and it is almost always told wrong. The version most people carry is one of national ruin —merchants drowning themselves in the canals. The archival record, reconstructed most carefully by the historian Anne Goldgar however shows. The trade was concentrated among a few hundred merchants and skilled craftsmen who could afford to lose. Most contracts were never enforced, or were settled at a small fraction of face value. And inspite of all the drama, the Dutch Republic went on to its Golden Age. The bubble was entirely real, however the catastrophe was largely literary. What is not literary is what came afterwards.

Four hundred years on, the Netherlands produces roughly 6.5 billion tulip bulbs a year across some 13,000 hectares, handles around 90% of the global flower bulb trade, and exported €12.3 billion of floriculture in 2025 — just under 9% of total Dutch agricultural exports. Bulb exports alone run at €800–900 million annually. The industry the tulip accidentally seeded is real, industrial, and vastly larger than anything the speculators of 1637 believed they were buying. The speculation was nonsense. The flower was not.

That pattern has repeated with a consistency that ought to make us humble.

●      Britain's railway mania, 1844–1847. At its peak, proposed railway schemes absorbed capital equivalent to a substantial share of national income. When it broke, share prices fell by roughly two-thirds, a third of authorised lines were never built, and a generation of middle-class investors — including, famously, Charlotte Brontë — was wiped out. What remained was the densest rail network in the world.

●      The electrical and utility booms of the 1880s and 1920s. Enormous promoter fraud, spectacular collapses, and a national grid at the end of it.

●      The dot-com boom and bust, 1995–2001 — often misremembered as a single event in the year 2000. The Nasdaq fell 78% from its March 2000 peak to its October 2002 trough. Roughly a trillion dollars of telco debt sat behind it; WorldCom and Global Crossing filed two of the largest bankruptcies in American history to that date. The fibre laid in the late 1990s was then bought for cents on the dollar, and it became the physical substrate on which broadband, streaming video and cloud computing were built. Amazon fell about 95% and survived.

The through-line is not that bubbles are harmless. People lost their houses in every one of these episodes, and the capital destruction was real and unevenly distributed. The through-line is narrower and more useful than that.

Bubbles do not destroy the technology. They destroy the capital structure that was betting on its timing.

Part II: Where I and AI Stand

Let me remove any ambiguity before going further. I do not think AI is hype. In March of this year I argued in The Mother of All Disruptions that AI represents a structural reordering of the relationship between human effort, economic value and social stability — that it disrupts every knowledge-intensive sector at once, that there is no adjacent sector waiting to absorb the displaced, and that the macro signal will arrive faster than most models predict. Five months of evidence later, I hold that view more strongly, not less.

A market correction would not refute a word of it. The two questions are entirely separate, and they are answered by different people using different instruments:

●      Will AI change how work, value and power are organised? That is a question about capability and diffusion. I think the answer is yes, and by more than the current conversation allows.

●      Is the capital currently being deployed against that belief priced and financed correctly? That is a question about credit spreads, depreciation schedules and refinancing dates. It has almost nothing to do with model capability.

Mixing the two questions is the most common error I encounter at board level, and it runs in both directions. The enthusiast reads a falling share price as evidence that AI was overrated. The sceptic reads the same fall as vindication. Both have mistaken a financing event for a technology verdict. The railways were not a bad idea in 1848. They were a badly financed idea in 1846.

The technology is not on trial. The capital structure is.

Part III: What Genuinely Rhymes with 1999

Five things, each of which I would put in front of a board without qualification.

1. Vendor financing is back, and this time it is explicit

On 11 August 2026, Nvidia announced a financing consortium of roughly $500 billion alongside Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR. The structure is worth understanding precisely: dedicated pools of capital for Nvidia's customers, with compute itself serving as collateral, and the debt issued through special-purpose vehicles rather than on anyone's balance sheet. Jensen Huang described compute as "an investable infrastructure asset" and indicated the company might provide financing support of up to 25% of a given opportunity. Separately, Nvidia has around $100+ billion of chip financing arrangements with OpenAI, alongside reported discussions of a large backstop.

This is structurally what Lucent and Nortel did between 1997 and 2000, when they lent customers the money to buy their own switches. The reported revenue was real. Part of the cash was their own, coming back around. When credit conditions turned, the revenue and the loans went at the same time, which is why the fall was so fast. One analyst summarised the current arrangement as making "Nvidia's product cheaper without really cutting GPU prices" while making "future demand more sensitive to credit conditions." That second clause is the sentence to keep.

2. The funding mix has flipped from cash to debt

This is the single most important number in this article. Incremental debt as a share of hyperscaler capital expenditure has risen from 9% in the 2024 financial year to 32% on a trailing basis by mid-2026. Aggregate hyperscaler debt now stands at approximately $700 billion. Sitting alongside it are some $800+ billion of lease-related commitments that are not recognised as balance-sheet liabilities at all, on terms of fifteen to nineteen years commencing between 2027 and 2029.

3. Free cash flow — the strongest argument against a bubble — has broken

Twelve months ago the best rebuttal to any bubble argument was that this buildout is funded from operating cash flow rather than credit. That has now changed. Amazon's trailing free cash flow has gone negative. Alphabet's quarterly a negative free cash flow. On current guidance, full-year 2026 free cash flow of many hyperscalers moved to zero or below. Any many of them have responded, Alphabet commenced a programme to raise $80+ billion of equity in 2026. Amazon's debt has risen above $100 billion. Meta has issued around $55 billion across two offerings. For avoidance of any doubt, these are very creditworthy issuers.

4. The revenue gap is not closed

Bain's estimate is that the industry requires roughly $2 trillion of annual AI revenue by 2030 to justify the buildout, against a shortfall of about $800 billion even on generous demand forecasts. Again, I do not read that as evidence that AI fails. I read it as evidence that the schedule of capital deployment is running ahead of the schedule of monetisation. That gap is precisely what a correction.

5. The economy now depends on the capex

Goldman Sachs forecasts roughly $1 trillion of global AI investment in 2026, of which $581 billion is in the United States, and puts US AI capital expenditure at about 1.8% of GDP rising toward 2.8% by 2028. Several credible estimates attribute the majority of US GDP growth in the first quarter of 2026 to the data centre buildout alone. Goldman's own framing is that this is consistent with the 2–5% of GDP peaks observed in previous general-purpose technology buildouts. That is reassuring and alarming in equal measure: every one of those buildouts was followed by a digestion phase.

A capital expenditure cycle large enough to carry an economy could also be large enough to stall an economy.

Part IV: What Does Not Rhyme — In Both Directions

Four differences matter. Two make this cycle materially safer than 1999, but two make it more dangerous.

Safer than 1999: the buyers are solvent

In 1999 the marginal buyer of network equipment was a cash-negative competitive carrier funded by high-yield debt with no path to profitability. Today the marginal buyer is Microsoft, Alphabet, Amazon or Meta. Four of the five largest hyperscalers still carry total debt to EBITDA at or below approximately 1.0x — comfortably inside the thresholds at which rating agencies begin to act. Nvidia's quarter to July 2026 produced $96.2 billion of revenue, up 106% year on year, at a 75% gross margin, with guidance of $108 billion for the following quarter. That is collected cash from investment-grade counterparties. In 1999 the equivalent line item was page views.

Safer than 1999: the equity is not 2000

Cisco Systems traded at roughly 130 times forward earnings at the March 2000 peak, on a business whose growth then reversed. Nvidia currently trades in the mid-twenties on forward earnings. What is genuinely stretched in this market is not the multiple but the concentration: Nvidia alone represents around 9% of the S&P 500, the top ten constituents account for 35–40% of the index, and their average price-earnings ratio sits near 50. Concentration of that order is a real risk, but it produces a de-rating and a painful drawdown — not the wholesale disappearance of the underlying earnings that defined 2000 to 2002.

More dangerous than 1999: the asset ages faster than the debt

This is the defect in the structure, and it is not the capex number. Fibre laid in 1999 was overbuilt, but it was never obsolete and is still carrying traffic a quarter of a century later. A graphics processing unit has an economic life of perhaps three to six years. It is being financed against lease and bond structures running fifteen to nineteen. Whatever else is true about this cycle, that mismatch is not a detail — it is the thing the credit market will eventually have to price. It also gives boards a clean tell. When a company shortens the stated useful life of its AI hardware, that is not an accounting adjustment.

More dangerous than 1999: the loss will land where nobody can see it

In 2000 the damage was concentrated in listed equity, marked to market every day, visible to everyone at once. Brutal, but honest and fast. This time the risk is migrating steadily into private credit funds, insurance company balance sheets and special-purpose vehicles built around compute contracts — none of which is marked daily, and all of which is slower to surface. J.P. Morgan's own transaction examples give the flavour of the marginal deal: Hut 8's Beacon Point financing in June 2026 priced $4.25 billion of senior secured notes at a 95% loan-to-cost ratio. That is the edge of the market, and the edge is always where it starts.

In 1999 the losses were public and immediate. This time they will be private and late — which is probably worse, not better.

Part V: The Part I Actually Worry About

It is not Nvidia's balance sheet I am concerned about. Nvidia is net cash and carries no meaningful leverage. It is the architect of this structure, not a borrower within it — and that distinction matters enormously for the question of who is hurt. The exposure sits in three places, none of which is the company whose name is on the story.

The marginal borrower. The second-tier compute providers — the so-called neoclouds — are financing at loan-to-cost ratios in the nineties against an asset with a six-year life. They are the most sensitive point in the system and they will break first, not because they are badly run, but because they are the ones with no alternative source of funding when spreads widen.

Oracle. Its lease obligations total almost three times its own capital expenditure guidance for the coming year, driven principally by capacity commitments to OpenAI. The cost of insuring its debt has repeatedly touched record levels through 2025 and 2026, falling back each time on reassuring results. That oscillation is itself information: the credit market has not made up its mind, and it is watching one counterparty.

The credit investors. Compute-backed paper issued through special-purpose vehicles is a genuinely new asset class. It has never been through a cycle. Nobody knows the recovery rate on a partially-obsolete GPU cluster in a distressed sale, because it has not happened at scale. Price discovery in that market has not occurred yet. It will.

Which gives a plausible mechanism, and it is worth stating because it is not the one most people expect. A correction here does not begin with a model failing, or a capability disappointment, or a research result. It begins with a single refinancing that does not clear. A second-tier operator cannot roll its facility. The paper reprices. The cost of capital for every compute vehicle moves with it. Marginal orders are deferred rather than cancelled — which is how it stays invisible for a quarter. The vendor's guidance softens. Only then does the equity re-rate. That sequence takes months, not a morning, and the equity market is the last participant to be informed, not the first.

Part VI: The Bezos Test, With the Necessary Grain of Salt

Speaking at Italian Tech Week in Turin on 3 October 2025, Jeff Bezos offered the most quoted framing of this cycle: "This is a kind of industrial bubble, as opposed to financial bubbles." He went on: "It can even be good, because when the dust settles and you see who are the winners, societies benefit from those investors… This is real, the benefits to society from AI are going to be gigantic." He also observed, as reported from the same session, that investors are handing teams of six people billions of dollars with no product — something that does not usually happen.

I agree with the substance. I would still take it with a grain of salt. In June 2026, eight months after that remark, Bezos became co-chief executive of Prometheus, an industrial AI company that raised $12 billion at a $41 billion valuation. Anyone who tells you the water is fine while wading in deeper is not necessarily wrong but has an interest. Myself (and at a significantly lower level), I advise and invest in this sector, and the reader is entitled to apply the same discount to me that I have just applied to him.

What survives the discount is the distinction itself, and it is the right one. A financial bubble leaves nothing behind but losses and litigation. However, an industrial bubble leaves infrastructure. 1637 left the bulb fields of North Holland. The 1840s left the railways. 2001 left the fibre. This one will leave data centres, grid interconnection, power generation, and models that do not have to be trained twice.

But the half of Bezos's argument that travels is the optimistic half, and on its own it is incomplete. He is right that this cycle will produce an extraordinary number of AI fortunes. What the quotation does not say — and what needs saying to anyone building a business right now — is that those fortunes will accrue almost entirely to people who solved a real problem, for a real customer, at a price that customer was willing to pay. There will be AI millionaires and billionaires in large numbers. They will not be distributed evenly across everyone who was in the room.

In every previous cycle the survivors were not the companies with the best account of the technology. They were the ones whose customers would have noticed their absence. Amazon in 2001 was not saved by being an internet company; every one of its dead competitors was also an internet company. It was saved by having customers who wanted books delivered, and enough balance sheet to wait out the winter. Pets.com had the same technology, the same era, the same enthusiasm, and no answer to the question of what it was actually for.

The technology is real for everybody. The economics are only real for those solving a problem someone will pay to have solved.

Part VII: What a Board Should Actually Watch

What follows is not a prediction. It is an instrument panel — seven indicators that will move before the headlines do, and every one of them is observable from public information.

1.     Incremental debt as a share of capital expenditure. It has gone from 9% to 32% in under two years. Above 50%, the cycle is credit-dependent rather than cash-funded, and its timing passes from the boardroom to the bond market.

2.    Nvidia's receivables and inventory relative to revenue. The cleanest early signal that vendor financing is inflating reported demand. Lucent's receivables stretched well before its equity broke; the accounting told the story a year ahead of the share price.

3.    Spreads on compute-backed special-purpose vehicle paper against investment grade. The first venue in which honest price discovery on this asset class will take place.

4.    Any change to the stated useful life of AI hardware. A shortened schedule is a confession, and it will be buried in the notes rather than announced.

5.    A neocloud that fails to refinance. This is the WorldCom moment of the cycle. It will lead the equity market by months, and it will initially be reported as an idiosyncratic story about one badly run company.

6.   Lease commitments measured against capital expenditure guidance. Oracle first, because the ratio there is already close to three to one, but the disclosure is worth reading for every large issuer.

7.    Power. Grid interconnection queues are measured in years, not quarters. The constraint that finally strands capital in this cycle may turn out to be physical rather than financial — which is a risk almost no financial model contains.

Part VIII: Which Brings Me Back to the Check-Up

If the argument of this piece is right — that a correction is coming, that it will separate businesses solving real problems from businesses telling a story about a technology — then the only question that matters to a board is which of the two it is sitting on.

That is precisely what the igniteX Check-Up was built to answer. We developed it at Chmiel Global Advisory as five reality tests that surface the truth about a business quickly. What I would add here is that all five read differently, and more urgently, in a boom — because a boom is a machine for concealing exactly what they are designed to reveal.

Test 1 — The One Sentence Test. Ask the leadership team to describe the business in one sentence: we are the X of Y. In a boom, "AI-powered" becomes the answer to every question, and it is not a positioning — it is camouflage. If marketing, sales and product give three different sentences, that is not a positioning problem. It is an alignment problem, and cheap capital has been quietly paying for it.

Test 2 — Moat vs Mirage. What actually creates advantage, what are you measuring, and why those metrics and not others? A correction administers this test whether you run it or not. The specific version for this cycle: does your advantage survive your model supplier tripling its prices, or is your moat somebody else's application programming interface?

Test 3 — Forward Radar. How much of your board and management reporting looks forward rather than backward? In a credit-led correction the leading indicators are all financing indicators — refinancing dates, covenant headroom, supplier concentration, and the funding sources of your largest customers. If your board pack is 98% historical, you will read about your own exposure in the newspaper. In one case I have cited before, moving from 200 pages of history to 30–40 pages carrying 40% forward-looking insight changed how decisions were made and, over five years, took a share price from 8 cents to 42 cents.

Test 4 — The Hotpot Reality Check. Culture without the PowerPoint. Booms reward performance over candour, and the incentive to keep quiet rises with the valuation. The person who knows the pipeline is soft, or that the AI pilot quietly failed, will tell you over dinner. They will not tell you in a town hall.

Test 5 — The Investor Vote. At some point every business needs external capital, and if investors will not commit real money — not encouragement, not interest, not "let's stay in touch" — something is not working. In a boom this test is temporarily broken, because money commits to almost anything. It repairs itself violently. The businesses that will pass it in 2028 are the ones behaving in 2026 as though it were already hard.

A boom suspends the market's judgement. It does not cancel it. Run the tests on yourself, or wait and have them run on you.

A Final Word

I remain what I was in March: a strong and genuine advocate for AI. The productivity gains, the scientific acceleration, the improvements in healthcare access and educational quality are real, large, and arriving. Nothing in this article is a retreat from that.

But advocacy for a technology is not the same as underwriting the price of its financing. I expect a correction. I expect it to be led by credit rather than equity, to arrive slowly and then all at once, and to be described afterwards as having been obvious. I do not expect it to move the destination by a single degree.

The speculators of 1637 were wrong about the price and right about the flower. The railway promoters of 1846 were wrong about the returns and right about the rails. The telecommunications carriers of 1999 were wrong about the timing and right about the bandwidth — so completely right that the capacity they bankrupted themselves building became the foundation of everything we have done online since. In each case the capital was destroyed and the infrastructure survived, and it passed into the hands of whoever had cash at the moment it was for sale.

So position accordingly, and the instructions are unglamorous. Hold liquidity. Keep your obligations shorter than the useful life of your assets. Know which of your suppliers and customers depend on financing that resets, and when it resets. And be able to say, in one sentence that everyone in your company would give the same way, what problem you solve and for whom.

The people who came out of 2002 owning the fibre were not the ones who called the top. They were the ones with cash when it went on sale.