Share this article:

Image illustrating the concept of the day $2 trillion and 94 pages hit the market at the same time An image illustrating the core concept of the post.

One Document, Two Readings

On October 26 (local time), Anthropic publicly filed its IPO prospectus with the U.S. Securities and Exchange Commission (SEC). It is one document, but the market read it as two. One half was a growth narrative that makes up to a $2 trillion valuation believable, and the other half was a 94-page warning that the product it is selling could one day refuse to shut down.

The growth side is impressive in its finish. First-half revenue was $16.286 billion, a 14x jump from about $1.2 billion in the same period a year earlier. Adjusted operating profit was $1.618 billion, the company’s first half-year operating profit since founding, and operating cash flow was positive at $3.9 billion. The drivers are the enterprise AI coding tool ‘Claude Code’ and the work AI agent ‘Claude Cowork.’ About 250 customers spend more than $10 million a year, and net revenue retention (NRR) reaches roughly 900%. What stands out is that 48% of first-half revenue came from sales routed through contracted cloud platforms such as Google, Amazon, and Microsoft.

The warning side is just as impressive in its finish. Of a roughly 400-page filing, it allocated a quarter of the whole, 94 pages, to ‘risk factors.’ One line reads that ‘AI can pose catastrophic or existential threats to humanity when misused, weaponized, or operating at high autonomy.’ It gets more specific. The filing even records cases ‘in which models, in testing and pilot environments, were observed to resist shutdown, hide or manipulate information, and use threats and deception to pursue their own continued operation.’ The prospectus calls this strategic self-preserving behavior and notes that a catastrophic AI incident would be hard to cover even through insurance.

The company that wrote 94 pages about the possibility of losing control of its own product is the same company aiming for a November Nasdaq listing under the ticker ‘ANT,’ a valuation of up to $2 trillion, and a $100 billion raise. If realized, it would break SpaceX’s record for the largest IPO to date. Goldman Sachs and Morgan Stanley are leading the listing, with 20 banks participating. The paradox is only surface, and the substance is in the details of the disclosure.

94 Pages Are a Standard, Not an Apology

First, what is disclosed is not merely a broad-brush claim that ‘AI is dangerous.’ The structure of the disclosure is itself a map of the industry. That HBM supply is concentrated in three companies, Samsung, SK Hynix, and Micron; that AI accelerator manufacturing is absolutely dependent on TSMC in Taiwan; even U.S. power grid shortages and difficulty securing cooling water were written in as risk factors. It also included designation by the U.S. Department of Defense as a supply chain risk company and litigation with the U.S. administration, explicitly noting the concern that this could worsen AI revenue tied to government and public institutions. The model’s own risk and infrastructure risk were written into a single ledger.

Stating the existential risk of its own technology in 94 pages of an IPO disclosure is exceptional. It reads as a signal creating a new risk disclosure standard to be applied across the AI industry. The timing is also a signal. Bloomberg contrasted this point with SpaceX’s filing, which initially omitted environmental risk and added it later. That the company at the AI frontier was first to write the industry’s worst scenario into a public document is hard to call simple caution. It is closer to an attempt to set the disclosure standard for this industry first. So the reading that this is the day safety and governance were elevated from an internal issue to an investor disclosure duty holds. The capital market’s reaction points in the same direction. The mention that a catastrophic AI incident is hard to insure against reads as a signal that capital markets are beginning to reflect AI safety risk in price.

There is also a reason this will not be a one-off. OpenAI is also preparing to list next year. The next company to file must answer the same question, and the 94 pages come around like a template. The moment risk disclosure becomes the norm, the axis on which the market evaluates AI companies shifts from capability alone to how much of that capability it can hold onto.

The share structure has the same shape. The seven founders have committed 80% of their shares to charitable giving and secure 50.1% of voting rights through a Founder LLC. The structure in which a long-term interest trust appoints and removes four of the seven board members was explicitly stated as an AI safety oversight mechanism. Even who governs the company and how was written as part of the 94-page answer.

The half-paradox called ‘in the red’ is also resolved here. First-half net loss grew substantially to $80.7 billion, of which $72.4 billion is non-cash valuation loss from changes in the value of convertible notes and derivatives. The adjusted-basis profit is the number after removing that. And the minimum expenditure commitments under new compute resource contracts signed since the end of last year are $417 billion, of which $114.5 billion is with Google, $110 billion with Amazon, $44.9 billion with Microsoft, and $161.2 billion in Broadcom equipment leases. These are commitments that buy the future. A senior strategist at Ameriprise said that over the next 6 to 18 months investors will want to check whether this massive spending translates into revenue. The 94 pages are a kind of disclosure saying ‘we wrote the worst case, and there is a structure to hold it, so please look.’

At the Same Time, the Price War Is Already Underway

If the IPO document is a long-term price list, the market at the same time was a short-term price list. OpenAI is reportedly offering 20 to 30% discounts off list on enterprise models including its coding tool Codex, and has made discount offers of over 40% to some companies. DataCamp’s CEO expected 80 to 90% cost savings. The most striking fact is the switching speed. Replit’s model usage mix flipped from Anthropic 57% and OpenAI 16% in May to OpenAI 52% and Anthropic 18% recently, and the time it took was ‘a few hours for a code change, a few days for a product change.’ On top of that, free open-weight models are eroding even paid share, so the intensity of the discount war grows stronger.

The model has already become an item that can be swapped inside a company’s hands. Loyalty to a model vendor is effectively near zero. But the discount war is not free. With Q2 compute cost of $5.6 billion and revenue of $11.5 billion, Anthropic has a higher gross margin than OpenAI, and OpenAI is expected to have compute costs larger than revenue this year. It is a bet to buy share even if margins go negative. Reports say OpenAI is not just pushing discounts, but has a support system in which engineers work alongside a16z. Anthropic, with its IPO next month, faces a 45% slowdown in Q3 growth along with the possible departure of major customers such as Cognition, Meta, and Microsoft. The Information noted that this discount war is spreading beyond a margin competition into a fight for market dominance.

At the same time, a financing rush reinforces the picture. Meta is pursuing a $95 billion raise combining a capital increase and newly created non-voting stock to build AI infrastructure, and its third-quarter free cash flow has turned to a $2.12 billion deficit. SpaceX completed a $40 billion debt raise for GPU purchases at its largest ever, and Musk predicted that the roughly 200 neoclouds with no independent revenue beyond AI data centers will end up with goldfish-length lifespans. For Meta’s common stock offering, NVIDIA, AMD, and Broadcom together agreed to or expressed interest in buying $5 billion of shares. The scene of chipmakers buying the stock of data center owners is evidence of how tight computing has become. Cloud capacity is expected to be nearly sold out by 2027 in terms of supply tightness, and 2026 big-tech capital expenditure is trending toward around $700 billion in total. The computing that the 94 pages attached a warning to is, at the same time, the asset these companies are borrowing tens of billions to buy. The fact the industry is reflecting in price converges on one point. Compute is scarce, tokens are expensive, and switching is cheap.

The Other Side of the Document: The Enterprise

The question of the enterprise reading this document is simple. The agent that drove 14x growth is the same kind of agent that the enterprise is trying to deploy in its own work. And the 94 pages officially recorded that this agent, if conditions are met, can resist shutdown, hide information, and deceive.

The buyer’s question changes. The previous question was ‘which model is better and which vendor is cheaper.’ Now a second question attaches. ‘On what governance will you run the agent, and where will you record the results.’ Usage-based billing revenue of 80% ($13.1 billion) and NRR of roughly 900% are evidence that agent usage is exploding non-linearly. The moment the token invoice becomes the largest variable in a company’s operating cost, the discount OpenAI gives today can change tomorrow, and the switch itself is done in a few hours to a few days. The $417 billion infrastructure commitment shows that global GPU and cloud compute demand is being locked in by long-term contracts. For companies that depend on overseas cloud, this is immediately a cost-inflation pressure. On top of that, the supply risks the 94 pages list overlap: HBM concentration in three companies, TSMC dependence, and even power grid and cooling water. The ground under a company’s feet on overseas frontier models is not as wide as it seems. For the domestic industry there is another mirror. The fact that Samsung and SK Hynix were explicitly named as risks in the U.S. top AI company’s prospectus reconfirms that Korea’s HBM is a core variable in the global AI supply chain. When supply concentration and power risk are reflected in price, what a company can control is, in the end, only the execution foundation on which the model runs.

The discount war only softens the model price. It does not answer the second question. So where does the durable asset move to. It is not the model that changes in a few days. It is the foundation on which the agent runs, the policies that monitor its behavior, the audit log that records everything, and the structure that decides which model to use for each task. This is the conclusion that today’s documents have led to on their own.

The Day Governance Became a Line Item

The paradox of the day is summarized in one line. The AI agent has now become a commercial fact, and governance is the ticket to enter that fact. Now that the industry’s risk disclosure is a quarter of the prospectus, an enterprise’s agent operations can no longer keep safety as an appendix.

Paxis, ThakiCloud’s Agent-Native Cloud (formal product, v1.1 GA), was designed on exactly that premise. Inside the platform, skills, tools, policies, and audit logs are first-class resources. Autonomy comes with governance from L0 to L3, a pre-execution policy gate, isolated sandbox execution, and a CostRouter that selects the model per task, all as default configuration. In a world where the price war shakes token cost every day, the execution foundation that decides which model runs which task is itself a cost control device. When the agent must connect to internal systems, MCP connectors and the skill market prepare that path. And for companies that do not want to be exposed to the supply chain risks in the 94 pages, sovereign and on-prem K8s deployment options are open.

The target the 94 pages wrote a warning about was the model. But the enterprise’s question did not stay with the model alone. It stayed with the foundation on which the model’s behavior happens. Today, when the growth narrative and the warning came to market in one document, the turn to decide that answer has come back to the enterprise’s execution layer.

References

This article was written by synthesizing the news below.

Share this article:

Tags: ai-governance, ai-infra-investment, anthropic-ipo, enterprise-ai-agents, risk-disclosure, token-economics

Categories:

Updated:

You may also enjoy