Anthropic’s $2 Trillion Bet Is That Nothing Goes Wrong

The most revealing aspect of a $2 trillion Anthropic IPO is not the staggering valuation, but the suspension of disbelief required to justify it.

To buy into that number, investors must assume that artificial intelligence is becoming economically indispensable at breakneck speed, that a formidable moat will protect a handful of frontier-model developers, and that the exorbitant costs of computing infrastructure will not eventually crush their margins. More crucially – they must wager that the political, technical, and institutional risks inherent to an increasingly geopolitically competitive landscape will play second fiddle to the growth story. 

Reports of Anthropic’s internal projections describe a company scaling at an unprecedented clip. If it can actually bust out an annualized revenue run-rate of $100 billion while expanding its enterprise market share, a trillion-dollar valuation begins to look mathematically defensible under current tech multiples. But this arithmetic obscures a deeper structural flaw: the very capabilities driving Anthropic’s revenue growth are simultaneously engineering the risks most likely to vaporize its premium.

As frontier models become more autonomous, persuasive, and capable of executing multi-step objectives across outside systems, they cease to be conventional software. A text generator hallucinating an answer is a glitch; an autonomous agent hallucinating a cyber-attack or a rogue financial transaction is a massive liability.

The recent internal turmoil and safety-evaluation controversies at both OpenAI and Anthropic are symptomatic of this shift. Major AI labs are now diverting substantial resources to test for autonomy, deception, and loss of control. This is not merely an engineering challenge; it is a fundamental transformation of the product category. The instability roiling these companies is not a byproduct of poor management – it is a structural consequence of trying to shoehorn a potentially world-altering, highly volatile technology into the framework of a standard venture-backed corporation.

The company’s value has been all over the place in prediction markets – recently hockey-sticking following a report in FT in which six company backers said that Anthropic’s rapidly rising revenue “would enable it to more than double its current valuation in a planned autumn float.” 

“If Anthropic is growing 800 per cent a year, you’d think at the incredibly low end they would trade at 30 times [revenue],” one investor told FT, adding “That would make them a $3tn company.”

Anthropic is a particularly pure expression of this tension over its capabilities. Their basic pitch is that frontier capability and rigorous safety can coexist – a positioning with immense commercial value to risk-averse enterprise clients. Yet – after Anthropic spooked the shit out of everyone with its hackbot 5000 (Mythos), China went full Leeroy Jenkins through the field with cheap, efficient, capable open-weighted models that carry none of the moralizing – just performance at a better value. When a safety-first AI company repeatedly triggers control warnings in its frontier models, it erodes its core value proposition. In effect, Anthropic is running two races at once: one toward greater capability, the other toward greater control.

The great convergence between open and closed.

The problem is that the $3 trillion in funding commitments by OpenAI/Anthropic assumes frontier prices remain stable https://t.co/8u0zDGUgwM

— zerohedge (@zerohedge) August 13, 2026

OpenAI offers a different flavor of the same crisis. Its explosive adoption set the template for the modern AI platform, but its boardroom coups, leadership exoduses, and chronic debates over deployment safety illustrate the inherent friction between mission statements, capital requirements, and commercial incentives. The industry has engineered a relentless feedback loop: capital buys compute, compute yields capability, capability drives revenue, and revenue attracts further capital. But with every revolution of this flywheel, the control problem magnifies. The commercial boom and the governance crisis are not parallel events; they are the same phenomenon viewed from different angles.

Because of this, conventional revenue multiples are highly suspect. Even if Anthropic achieves a massive revenue run-rate, standard tech valuations demand continuity between current sales and future cash flows. Frontier AI’s economics, however, are unusually discontinuous. Cheaper, open-source alternatives threaten to commoditize baseline intelligence, while corporate customers are already demonstrating price sensitivity by opting for smaller, highly efficient models for routine tasks. Technological leadership does not guarantee pricing power. The industry has proved that better AI generates demand; it has not proved that every incremental leap in intelligence creates proportionate economic value.

Catch-22

An alternative lens for these valuations is that investors are not buying a software company; they are pricing a call option on the strategic control of machine intelligence.

If advanced AI becomes the foundational infrastructure of the modern economy – underpinning software, finance, defense, and medicine – occupying that central node yields unprecedented economic leverage. But here’s the catch-22: If frontier AI becomes as strategically vital as a trillion-dollar valuation implies, governments will not allow its proprietors to operate as ordinary private entities indefinitely. The stronger the financial case for these valuations becomes, the stronger the political case for national-security classification, severe regulatory constraints, and sovereign oversight.

An Anthropic IPO would therefore be more than a liquidity event. It would be the public markets’ first attempt to price the frontier-AI paradox.