The frontier AI model bubble is a funding problem

"We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run." – Roy Amara

Anthropic's IPO prospectus tells two stories at once.

The first story is about growth. CNBC reports 1,300% annual sales growth, 900% net revenue retention, and an implied gross margin of around 85%. The second story is about cost. The same prospectus discloses an US$80 billion net loss for the first half of the year and more than 90 pages of risk factors. Further, the amended S-1 discloses purchase commitments and lease agreements of about [US]$895.4 billion, "substantially all of which is related to compute and infrastructure."

This is all sounding eerily familiar to those of us who lived through the dot com bubble back in the day.

An 85% gross margin did not protect Anthropic from an US$80 billion loss. Both stories are true. That is why I think the US frontier model market is a bubble.

Where we started

I first wrote about the hype cycle for social media back in 2008, and more recently about why we need to stay grounded on AI. The pattern is a familiar one. Expectations overshoot, capital floods in, and then reality arrives with the bill.

Amara's Law cuts both ways here. A bubble does not mean the technology is useless. The railways were a bubble, and so was the dot-com era. Both left behind real and useful infrastructure. Further, both left behind a great many investors who paid far too much for it.

What is different this time is the scale of the fixed obligations. According to Axios, nearly US$70 billion of Anthropic's commitments falls due in the 12 months after 30 September, and the remainder stretches out to 2045.

The real question remains as to who absorbs the losses when the economics catch up.

Growth without a business model

Wharton's David Hsu asks the right question: "It's a great technology, seems to do things that we never would have imagined, but how does that then translate into a viable business?" (CNBC).

Frontier labs sell access to model capability at prices that do not cover the cost of building the next generation of models. Revenue is rising fast. However, the cost of staying at the frontier is rising faster.

A business that must keep spending more simply to stay in the race is not yet a mature business. It is a bet.

The circular flywheel

What concerns me most is how the money moves. Anthropic sources much of its compute from Amazon, Google and SpaceX, and its hyperscaler partners are also among its largest investors (CNBC). CNBC describes these companies as depending on one another to keep the "flywheel" turning.

The mechanics are simple enough:

1. The investors fund the lab.
2. The lab buys compute from the investors.
3. That spend appears as cloud revenue growth.
4. The cloud revenue growth justifies further capital expenditure.

Thus, the same handful of companies are customers, suppliers and shareholders all at once. This works while the wheel keeps spinning. Yet it makes genuine market demand very difficult to see.

When cash flow runs out

The hyperscalers started to show the strain this quarter (CNBC):

• Alphabet maintained capex guidance of up to US$205 billion, reported negative free cash flow of US$9.1 billion, and had its worst trading day ever.
• Amazon reiterated roughly US$220 billion in capex and reported negative free cash flow for a second consecutive quarter, with a US$42.4 billion outflow over the trailing 12 months.
• Microsoft trimmed its capex forecast to US$170 billion. Its free cash flow remained positive at US$18.6 billion, although significantly lower, and it announced plans to raise US$25 billion in bonds.

Futurum's Daniel Newman captures the shift in sentiment: "Suddenly, they're like, 'We can't keep funding this on free cash flow'" (CNBC).

When the richest companies in the world need debt to build, interest rates become the main risk. Rising rates make borrowing less affordable for data centre developers and neoclouds. They sit further down the chain, with far less room for error.

A market resting on one listing

Newman also observes that "Anthropic is the IPO market at this point" (CNBC). Oura, Firmus and Solidigm have recently postponed or abandoned their listings, and Nscale and SB Energy have not updated their prospectuses in weeks.

When one company carries the hopes of an entire market, that is concentration risk. A healthy market looks quite different.

What this means in Australia

In Australia we have already seen a local version of this story. The Nvidia-backed data centre operator Firmus abandoned what would have been one of Australia's biggest IPOs after failing to win over investors wary of AI valuations and rising borrowing costs. The West Australian reports that ASIC is now seeking information from the banks involved. Further, Anthropic plans to offer shares directly to Australian investors in a market still reeling from that collapse.

Thus, the frontier model bubble already reaches through into Australian superannuation funds, retail investors and the local data centre pipeline.

In AI's hidden extraction economy I argued that AI is infrastructure, with supply chains, financing structures and power relations built into it. The financial fragility of US frontier labs is therefore a governance matter for Australian organisations today.

In my view, Australian boards and governments should focus on four things:

1. Concentration risk. Where critical services depend on API access to two or three US frontier providers, we inherit their balance-sheet risk. Pricing, terms and availability could change quickly if their capital structures shift.
2. Sovereignty. Our AI capability rests on infrastructure financed by foreign debt markets and decided in foreign boardrooms. That is a supply chain dependency, and it should be managed as one.
3. Procurement discipline. Exit plans, model portability and multi-vendor architectures are basic resilience measures. As I noted when mapping the AI value chain, the model and data supply layer is the plumbing that determines whether everything above it stands up.
4. Value at today's prices. Fund AI initiatives that demonstrate value at current costs. A business case built on the assumption that subsidised frontier models will stay cheap is a business case built on someone else's losses.

Doing the work

I am not suggesting that we walk away from AI. The technology is genuinely useful, and its long-run impact will be significant. However, the current economics of US frontier models depend on cheap capital, circular investment, and a market willing to fund enormous losses.

Until a frontier lab can fund its own build-out from its own revenue, I would plan on the basis that the bubble may burst.

The organisations that come through this well will be those that did the unglamorous work early. They will understand their dependencies, keep their options open, and avoid betting the business on someone else's balance sheet.