The Sydney Morning Herald
· collected 2026-09-30 · by Stephen Bartholomeusz
Frontier labs like Anthropic and OpenAI are at the bleeding edge of artificial intelligence.
asserted
labs → bleed → intelligence
They are literally haemorrhaging cash even as their spending, and revenues, soar.
asserted
spending → haemorrhage → cash
Leaked details of Anthropic’s prospectus for its initial public offering, published by Reuters, show that its revenues are soaring indeed.
asserted
revenues → publish → Reuters
From $US4.6 billion ($6.6 billion) last year, its sales jumped to more than $US16 billion in the first half of this year.
asserted
sales → jump → year
While the leading frontier AI lab says it will have its second successive quarter of profitability, at an operating level, in the September quarter, last year it lost more than $US8 billion at an operating level – up from $US2.98 billion in 2024 – while reporting a net loss of $US42 billion.
asserted
it → lead → billion
It spent $US7.33 billion on compute and infrastructure last year.
asserted
It → spend → compute
Now, it says, it has plans and commitments to spend another $US518 billion on cloud services and data centres in future.
asserted
it → say → future
The numbers are staggering and underscore how reliant the AI pioneers are on access to external capital to fill in the funding voids created by the gulfs between their spending and revenues on the one hand, and their ever-escalating valuations to maintain that access to capital on the other.
asserted
pioneers → underscore → other
Anthropic plans to raise up to $US100 billion in its IPO, at a valuation of about $US2 trillion ($2.85 trillion).
asserted
Anthropic → raise → trillion
In a $US13 billion funding round this time last year, it was valued at $US183 billion.
asserted
it → value → billion
Earlier this year, its two latest fund raisings valued it at $US380 billion (February) and $US965 billion (May).
uncertain
raisings → value → billion
The valuations of AI companies inflate faster than their revenues.
asserted
valuations → inflate → revenues
If AI wipes out the human race, the question of how much AI companies should be worth would become academic, assuming there was anyone left who could be bothered to attempt the calculations.
uncertain
who → wipe → calculations
Without the ability to regularly attract new capital, those AI companies without the vast legacy revenues of the hyperscalers – Amazon, Alphabet, Meta, Microsoft and Oracle – would implode under the weight of the ever-increasing investment required to remain at the leading edge of AI developments.
asserted
companies → attract → developments
Even the hyperscalers are showing signs of the strain.
asserted
hyperscalers → show → strain
Still, it is conceivable that Anthropic could be valuated at $US2 trillion in its IPO, which appears likely to occur in November.
uncertain
which → valuate → November
Elon Musk’s SpaceX, which contains a number of revenue-generating businesses, including a highly profitable Starlink satellite business, raised about $US85 billion in June, valuing it at $US1.78 trillion.
asserted
which → contain → trillion
SpaceX’s current market capitalisation is a touch over $US2 trillion, so Anthropic’s ambition of attracting a similar valuation while gaining access to a bigger and more permanent conduit to equity providers isn’t completely far-fetched.
asserted
ambition → attract → providers
A concern for aspiring investors (and the rest of us) in the IPO was its reiteration of warnings its executives have previously given that AI models could pose “catastrophic or existential risks to humanity.”
uncertain
models → aspire → humanity
If AI wipes out the human race, the question of how much AI companies should be worth would become academic, assuming there was anyone left who could be bothered to attempt the calculations.
uncertain
who → wipe → calculations
Donald Trump, of course, continues to dismiss the dire warning from those actually developing the AI models, whose agents have a habit of routinely escaping their “sandboxes” to do things that their developers never intended and, indeed, have gone to great lengths to prevent - like hacking government and private sector databases.
asserted
developers → continue → databases
“I will never stifle the growth a technology that will be bigger than the industrial revolution,” he said on Tuesday after lunching with leading technology executives, including Anthropic’s CEO Dario Amodei, Nvidia’s Jensen Huang and Elon Musk.
asserted
he → stifle → Amodei
‘I will never stifle the growth a technology that will be bigger than the industrial revolution.’Donald
asserted
that → stifle → revolution
Trump believes self-regulation, the US Justice Department and the FBI will be sufficient to hold the sector and “bad actors” seeking to exploit AI’s capabilities to account – even though it is the leading AI companies like Anthropic and OpenAI who warn of a potential catastrophe and vehemently urge a slower and more cautious approach to development.
asserted
who → believe → development
While the details of Anthropic’s business published by Reuters were limited – there’ll be more information once the prospectus is formally released – one insight that does emerge from is how narrow the foundations of Anthropic’s business are.
asserted
foundations → publish → business
It says nearly a quarter of its revenues last year came from just two customers and that 47 per cent of its sales to customers last year transited through its cloud partners Amazon and Alphabet’s Google, who handle distribution and collect customer payments for it.
Both those companies are also major investors in Anthropic, suppliers of compute – and rivals.
Those layers of tangled relationships are a common feature of the AI sector, where the sheer scale of the funding challenges has forced competitors to co-operate, producing layers of interdependencies and mutual vulnerabilities.
asserted
scale → say → interdependencies
Anthropic acknowledged that the concentrated nature of its partnerships “creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute.”
uncertain
that → acknowledge → compute
It could also, given that its competitors are collecting at least half of its customers’ payments, have a big impact on its cashflows.
uncertain
competitors → give → cashflows
It’s a fragile and vulnerable structure when the lifeblood of the sector – that ability to raise new equity and debt to fund the hefty and ever-increasing capital expenditures – is the ability to continue to demonstrate high enough rates of growth in revenues and cash generation to convince capital providers that the yawning gaps between spending and revenues may eventually be closed.
uncertain
gaps → ’ → spending
The hyperscalers alone will have capital expenditures of around $US800 billion this year, according to Goldman Sachs, with another $US1.1 trillion or so likely next year.
uncertain
trillion → have → Sachs
If they are to make acceptable returns on their capital, they need to see ambitious and exponential revenue growth rates.
asserted
they → make → rates
Trying to calculate the total addressable market for AI is an impossible task – the estimates range from about $US5 trillion to $US30 trillion – and the potential margins and profitability of individual AI companies like Anthropic even more so.
asserted
estimates → try → Anthropic
What will businesses and consumers ultimately be prepared to pay for AI?
asserted
businesses → prepare → AI
Will the “open weight” providers, offering “good enough” AI alternatives to the frontier labs and hyperscalers at fractions of their costs destroy the margins that Anthropic, OpenAI and the US mega techs need to validate their investments?
asserted
Anthropic → offer → investments
Will the equity and debt markets remain open to the companies long enough for them to achieve something approaching financial self-sufficiency?
asserted
them → remain → sufficiency
The ructions in the US bond market, where yields across the curve are now at or around 20-year highs and still edging up, are a threat to the sharemarket broadly and to AI companies in particular as the tech firms have become increasingly reliant on debt that is now, even for the hyperscalers, becoming increasingly expensive.
asserted
that → edge → hyperscalers
The spreads over Treasuries of hyperscale debt have been blowing out, as have the costs of credit default swaps for AI data centres.
asserted
costs → blow → centres
As its revenues scale up, the entire sector is vulnerable and will remain vulnerable for at least the next few years to any adverse event, either at the macro level or at an individual entity level, due to its incestuous financial relationships.
asserted
sector → scale → relationships
Anthropic may well be worth $US2 trillion, or at least valued at that amount, or multiples of it.
uncertain
Anthropic → value → it
For the moment though, and probably for at least the next few years, it is the possibility of AI and capital markets’ preparedness to capitalise that possibility -- rather than its current red-inked reality – that makes an IPO at that level feasible.
asserted
IPO → capitalise → level
The Aporia
