August 2026 Performance Report

The AI Data Centre Boom: Follow the GPU Price, Then Follow the Debt

I have been trying to find the right leading indicator for the end of the AI data centre boom.

We know there will eventually be an overbuild. The amount of compute capacity being added is simply too large for supply and demand to remain perfectly balanced indefinitely.

But that doesn't tell us whether the turning point is six months away or six years away.

As I discussed last month, the problem is made harder by the extraordinary improvement in technology. New data centres are not just adding more megawatts. Each new generation of chips produces vastly more useful compute from each megawatt.

That means supply measured in tokens is growing much faster than supply measured in gigawatts. And it means that charts promising doom with falling token prices are probably not relevant. 

So what should investors watch?

I think one of the best indicators is simple:

How much does it cost to rent a GPU for an hour?

And the second indicator is increasingly important:

Who is financing all of these GPUs and data centres, and on what terms?

Don't watch token prices

There is a lot of misinformation around falling AI token prices.

Token prices are collapsing. But that does not necessarily mean that returns on data centres are collapsing.

The cost of producing tokens is also collapsing.

A Blackwell GPU can generate substantially more useful compute than a Hopper GPU. Software is getting better. Quantisation, speculative decoding, caching, model architecture and utilisation are all improving.

If the cost of producing a token falls 90% and the price charged to the customer falls 80%, the economics have actually improved.

And if lower prices encourage customers to consume ten or twenty times as many tokens, the total market grows rather than shrinks.

So $/token is not the number I want to watch.

$/GPU-hour is much closer to the underlying economics of the asset.

H100 is NVIDIA's 4-year-old chip and is the most common benchmark.  One-year pricing around $3.25.

The economics are still extremely good

My rough estimates for an H100 look something like this:

These are deliberately rough numbers. There will be substantial differences depending upon purchase price, utilisation, financing, electricity costs, data-centre configuration and the assumed residual value of the GPU.

But the broad conclusion is difficult to escape.

At $3.25 an hour, the economics are fantastic.

Taking the middle of the ranges, an H100 is generating around $3.25 of revenue against perhaps $0.80 of cash operating costs and around $1.70 of total economic costs.

That leaves an enormous margin.

$2 is where it starts getting interesting

There are actually two different break-even points.

The first is the investment break-even point.

If an H100 rental price falls to around $2/hour, a brand-new investment starts looking considerably less appealing. Against total costs around $1.55-$1.85, there is not much left to compensate investors for downtime, financing risk, execution risk or return on equity.

It does not mean you turn the server off.

It means you probably stop rushing to build another one.

That distinction is crucial.

At current prices, using a midpoint all-in cost of around $1.70/hour, the GPU only needs to be rented for roughly 52% of available hours to recover its economic costs.

At a rental price of $2, required utilisation rises to about 85%.

Suddenly, an interruption, weaker demand or a customer renegotiating its contract matters enormously.

$0.80 is a completely different threshold

Once the data centre has been built and the GPU purchased, capital expenditure is sunk.

The decision becomes:

Do I make more money running the GPU than switching it off?

For an H100, I estimate that cash operating costs are somewhere around $0.65-$0.95/hour.

Call it roughly $0.80.

That is the really ugly number.

At $2/hour, companies might stop constructing marginal new capacity.

At around $0.80/hour, existing capacity itself starts becoming uneconomic to operate.

The important implication is that there is an enormous distance between today's approximately $3.25 rental rate and the price at which servers would actually start getting switched off.

That does not look like an industry suffering from massive overcapacity.

At least not yet.

The six-year-old A100 is perhaps even more revealing

Another useful test is to look at old GPUs.

The Nvidia A100 was launched in 2020. By technology standards it is ancient.

Yet A100 pricing is around $1.80 per GPU-hour.

I estimate cash costs at roughly $0.45/hour.

So a six-year-old GPU is still generating more than four times its approximate cash operating cost.

That is important.

In a genuine compute glut, I would expect older equipment to get crushed first.

The newest GPUs will always command a premium because they offer better performance, more memory and greater energy efficiency. But obsolete or semi-obsolete equipment should increasingly become the marginal capacity.

If A100 prices were collapsing toward cash cost, I would be considerably more worried.

They aren't.

But there is another way the boom can end

So far, this sounds very bullish.

There is, however, an increasingly important complication.

Debt.

You don't need GPU prices to fall all the way to $0.80 to create a financial crisis.

A data centre may still be cash-flow positive at $1.50 per GPU-hour, but the company that owns it might be insolvent.

That is because the electricity bill falls when the machine gets turned off.

The debt doesn't.

And the amount of debt now flowing into AI infrastructure is becoming extraordinary.

The funding structures now span corporate bonds, project finance, private credit, joint ventures, equipment financing, leases and equity.

The Bank for International Settlements has highlighted the growth of what it calls "shadow borrowing": special-purpose vehicles that own data-centre assets and borrow against long-term leases or capacity agreements from hyperscalers. Economically the hyperscaler has made a long-term financing commitment, but much of the debt sits elsewhere.

This matters because leverage changes the shape of the cycle.

CoreWeave shows both sides of the story

CoreWeave is probably the clearest example.

Its operating economics remain exceptionally strong.

In the June 2026 quarter CoreWeave generated $2.575 billion of revenue and $1.51 billion of adjusted EBITDA — a 59% EBITDA margin.

That hardly looks like an industry in distress.

But CoreWeave also had $35.6 billion of debt at the end of June. Some of its major borrowing facilities carry effective interest rates of around 9%-15%.

Its quarterly interest expense was $640 million.

This is where the risk lies.

If GPU rental prices remain high and utilisation remains strong, leverage is wonderful. Huge operating cash flows accrue to a relatively small equity base.

But suppose rental prices fall 30%.

Electricity costs don't fall 30%.

Data-centre leases don't fall 30%.

And interest payments certainly don't fall 30%.

The effect on equity earnings can therefore be dramatically larger than the decline in GPU prices.

Debt turns a normal cyclical slowdown into something potentially much more violent.

Vendor financing is both an accelerant and a warning signal

This is an old story in capital-intensive industries.

During the early phase of a boom, customers finance expansion themselves.

Then competition increases.

Suppliers start helping customers obtain finance.

Eventually suppliers may guarantee loans, support residual values, provide extended payment terms or effectively finance customers' purchases.

None of those things proves there is a bubble.

But increasingly generous vendor finance is a useful indicator of where we are in the cycle.

It allows supply to keep growing even when standalone customer balance sheets would otherwise constrain expansion.

And that makes the system increasingly pro-cyclical.

It increases the boom on the way up.

It also increases the eventual bust.

Financing is still remarkably available

There is a counterpoint.

Credit markets are certainly not closed.

In June, Hut 8 financed its Beacon Point data centre with $4.25 billion of senior secured notes, against a facility that was completely pre-leased for 15 years to an investment-grade customer.

The remarkable part was the leverage: 95% loan-to-cost.

That is an extraordinary level of debt finance for a new project.

So I wouldn't interpret the rise in debt as evidence that the boom is about to end.

If anything, abundant financing can keep the boom going for much longer than fundamental analysis might suggest.

What we need to watch is the terms on which that money is available.

There are some early warning signs. Recent reporting suggests lenders are becoming more selective, demanding more contractual protection and stronger pre-leasing before financing projects. AI-related debt issuance has meanwhile become a substantial part of US credit markets.

That isn't the bust.

But it may be where the first cracks eventually appear.

What the end of the boom probably looks like

I don't think we will wake up one morning and discover that nobody wants AI anymore.

The more likely progression looks something like this:

This is why I think looking solely at Nvidia revenue is too late.

By the time Nvidia's order book is clearly falling, the deterioration could have been visible for months in GPU prices, contract terms and credit markets.

The long-term GPU contract price may be the single best indicator

Spot prices matter.

But I suspect the most useful number of all will ultimately be the one-to-three-year GPU rental price.

A one-hour rental tells you whether somebody somewhere wants compute today.

A multi-year contract tells you what sophisticated buyers and sellers think the capacity will be worth over the economic life of the hardware.

H100 one-year contract prices weakened through 2024 and 2025, but moved back up during 2026. That is not what I would expect to see immediately before a severe compute glut.

My rough rule of thumb would be:

Above $2.50: economics remain very strong.

Around $2.00: start paying close attention. New H100 investments become much harder to justify unless utilisation is extremely high or acquisition costs are low.

$1.50-$2.00: the economics of new capacity become questionable and leveraged operators start becoming interesting for the wrong reasons.

Around $1.00: expect serious financial stress, collapsing residual values and cancelled projects.

Around $0.80: we are approaching the point where even running some existing H100 capacity may cease to make economic sense.

These aren't precise lines. Different operators have different costs and different contractual structures.

But I think they are a much better framework than looking at token prices.

We are not there yet

My previous article argued that there will eventually be an AI infrastructure overbuild.

I still believe that.

The supply response is simply too enormous, and technological progress is increasing effective compute supply much faster than the growth in physical data-centre capacity.

But the evidence today doesn't suggest we have reached that point.

H100 rental economics remain extremely profitable.

Six-year-old A100s still rent for multiples of their estimated cash running costs.

New-generation GPU rental prices remain high.

Long-term financing is still available.

Major projects can still obtain extraordinary leverage when backed by strong customers.

The red flags are instead appearing in the financial architecture around the boom.

Debt is growing enormously. More risk is moving into project-finance vehicles and private credit. Nvidia itself is increasingly involved in securing capacity, supporting its cloud partners and bringing outside capital into the ecosystem.

None of those things tells me the boom ends tomorrow.

They tell me what to watch.

The most important signal won't be that token prices have fallen another 80%.

It will be when GPU rental prices fall toward the economic cost of providing the capacity at the same time that utilisation weakens and lenders start demanding substantially better terms.

That is when the virtuous cycle can reverse.

Lower compute prices reduce returns.

Lower returns make debt more difficult to service.

Tighter financing means fewer new data centres.

Fewer new data centres mean fewer GPU orders.

And eventually that reaches Nvidia.

For now, we are still a long way from the cash-cost threshold.

But with the amount of money being borrowed to fund this expansion, we won't need to reach it before things start breaking.