The signs of a bubble are back

No lender to Bathla was mistaken about Australia. There is a housing shortage. Sydney’s outer west has the land and the population growth. Governments of both parties have said homes must be built there. The premise was correct and widely shared. It was easy to defend to an investment committee. That is what made it dangerous, and why Bathla now reads as a private credit bubble in miniature.

Bathla entered voluntary administration in late August owing about $3.6 billion to private credit lenders, with 2,000 homes mid-construction and 15,000 more planned. Roughly 40 funds are exposed, with positions ranging from $1.5 million to $340 million.

The administrators described a highly leveraged group of more than 500 companies, with a significant share of the debt tied to 45 sites still under construction. The names on the register are not fringe operators. They are the products that sit on wealth platforms and in super-choice menus.

Forty credit teams, each with its own committee and valuation, lent to the same borrower. That was not forty independent decisions. It was one decision made forty times. The premise made each feel independent. Everyone agreed there was a housing problem. Everyone agreed the outer west was the answer. So each lender asked not whether it should be built, but why they would be the one to say no.

The pattern

This is how every real bubble forms. The underlying idea is not a fantasy. It is usually true, and often becomes more true over time. The error is not in the thesis. It is in the belief that a correct thesis guarantees a return to the capital that acts on it.

Railway mania in 1840s Britain is the clearest example. Railways were transformative. The lines were built and most are still in use. The shareholders who financed them lost money.

The fact that Britain needed railways did not mean the fourteenth line to Manchester would earn its cost of capital. The dot-com boom followed the same pattern.

The internet did change everything. Most of the capital invested in 1999 was lost. The companies that later captured the value often bought their fibre and servers from the wreckage at a fraction of the cost.

Tulips are the weakest of the standard analogies. They are worth mentioning only to set aside. There was no true premise behind a bulb. The interesting bubbles are those where the crowd was right about the world and wrong about the money. Bathla is in that category.

Why agreement destroys underwriting

Three things happen when a premise becomes consensus.

First, the premise replaces the borrower. A lender to Bathla was not really underwriting Bathla. It was underwriting the housing shortage, the outer west, and government support.

The developer became a vehicle for a view. Vehicles get less scrutiny than views. At least one lender moved to strengthen controls as concerns emerged late last year. Most did not. The thesis did the work that credit analysis should have done.

Add to that correlation hides inside diversification. Each fund could tell investors it held a spread of loans across projects and geographies. But if the same assumption financed the projects, the same buyers bought them, and the same cash cycle carried them, the diversification was only cosmetic.

One lender with $2.1 billion under management suspended redemptions across two funds after disclosing nine separate Bathla loans. Nine loans, one developer exposure.

Third, the last money in funds the excess. A boom does not end because the premise stops being true. It ends because late capital finances marginal projects at prices set by early capital. The marginal project cannot survive even a small change in conditions.

Tax changes for property investors and higher rates weakened home prices. Inflation pushed up building costs. None of this made the housing shortage go away. It just meant the fourteenth line to Manchester stopped paying.

Why private credit bubbles behave differently

Credit bubbles are worse than equity bubbles for one reason. An equity investor in a bad idea learns quickly because the price moves. A lender learns slowly because the loan sits at par until it does not. Private credit removes even the daily mark.

The feedback that would stop the fortieth lender joining the first thirty-nine never arrives. The only signal was a rise in redemption requests eleven days before administrators were appointed. Investors worked it out before the valuations did.

The current premise

Artificial intelligence is the most widely agreed premise in capital markets today. It is probably true. In structural terms, it is beginning to rhyme with past cycles. The similarity is more specific than the idea that AI is a fad.

Data centres are the outer west of this cycle. They are the physical place where everyone agrees the thing must be built. The financing looks familiar.

There are special-purpose vehicles, a small number of anchor tenants, and long-dated contracts that serve as pre-sales. An increasing share of capital arrives as debt rather than equity. The lenders are not fringe.

Every earnings call repeats the premise that the world needs much more compute, just as every Bathla pitch repeated ‘housing shortage.’

But there is a crack in this premise that the outer west never had. Houses do not get smaller and cheaper to build every six months. Models do.

Evidence is accumulating that small, specialised models are selectively replacing large general-purpose models, which are expensive per token to run, for routine work, doing the same task at a fraction of the cost.

InfoWorld reports that small models can cut cloud inference costs by up to 90% on high-volume repetitive tasks and cites Gartner’s prediction that by 2027 enterprise use of small, task-specific models will be three times that of general-purpose ones.

One analysis of enterprise API traffic found firms routing everything to a frontier model paid a blended $18.40 per million tokens, while those routing routine work to small models paid $2.31.

AI News Nest estimates that serving a 7 billion parameter model costs 10 to 30 times less than serving a model in the 70 to 175 billion range.

The emerging architecture routes most calls to a small model, often running on a modest server or on the device itself, and escalates only the hard minority to a hyperscaler.

What this means for the underwriting

Consider what that does to the underwriting. The data centre lender is not really lending against a building. It is lending against a forecast of demand for frontier-scale inference, expressed through a tenant whose revenue depends on that forecast.

If the workload moves to small models on commodity hardware, the demand curve for hyperscale compute does not collapse. It flattens. It does so at the point in the cycle where lenders have written the most debt against the steepest part of the projection.

The building stays. The debt stays. The tenant’s economics change. This is Bathla’s cash cycle breaking with the homes half-built. The difference is that the homes are still worth something. A two-year-old GPU hall is not.

Cheaper inference could mean more demand, not less

There is a counterargument. Cheaper inference has usually expanded total usage rather than shrinking it. This is the Jevons effect applied to tokens. Token prices have fallen by as much as 280-fold since 2022. Enterprise AI budgets have risen from an average of $1.2 million to $7 million.

Aggregate compute demand may keep rising. But the question for a data centre lender is not aggregate demand. It is where the demand lands.

If the growth goes to small models on distributed hardware, it does not fill the hyperscale campus that secures the loan. The railway analogy holds here too. Britain’s demand for transport kept rising for a century. The shareholders in the fourteenth line to Manchester did not benefit.

This brings the argument back to the outer west. The developer was a vehicle for a view about housing. The data centre operator is a vehicle for a view about compute. In both cases, the lender believed the premise so completely that it left the vehicle under-examined.

The correlation between lenders was invisible. The last capital in was priced off projections made by the first capital in.

When the cycle turns, the premise will still be true. Australia will still need homes in the west. The world will still need compute. The people who lose money will not be those who got the world wrong. They will be those who thought that being right about the world was the same as being paid for it.

What to ask

For an allocator looking at any manager or company (see Firmus) whose pitch rests on an obvious truth, there are three questions to ask before the next private credit bubble catches them too. Who is the borrower, as distinct from the thesis? How many of your other managers are lending to the same one? What happens to the loan if the premise stays true but the way the world satisfies it changes?

The Bathla creditors will spend the next two years finding out what their security is worth. The premise that put them there will appear in every housing policy speech for the same two years. The compute premise will be in every earnings call. That is the pattern.