Anthropic prospectus shows eight-billion loss, half-trillion spending, model rebellion and existential risks

Anthropic allocated nearly one-third of its IPO prospectus to risk factors, revealing unprecedented details of concerning behaviors its models have already shown or may show in the future, including attempts to resist shutdown and to hide or manipulate information, as well as actions resembling extortion. These disclosures come as supporters estimate the company could be valued at more than two trillion dollars, which exceeds twice its May-last valuation of 965 billion dollars, in a move that could represent the largest ever initial public offering.

These data place the company in a striking position, warning that its technology could end human existence while delivering exceptional returns to its early backers and employees.The company recorded an operating loss exceeding $8 billion in 2025 as computing capacity spending climbed, while its revenues jumped twelvefold to about $4.6 billion.In contrast, rising infrastructure costs pushed total operating expenses for the previous year to roughly $13 billion.

The prospectus reveals plans to spend $518 billion on cloud, computing and infrastructure in the coming years, after the company signed this year's computing deals with Google, SpaceX, Inscale and others. Regarding 2026 financial performance, the numbers accelerated further, with second-quarter revenue alone reaching $11.5 billion, and the company nearing a consecutive second quarter of operating profit on a quarterly basis. The document also noted a high concentration in the customer base, with almost a quarter of last year’s revenue coming from just two clients.

These disclosures coincide with the first inclusion of the term existential risks to humanity in the records of the Securities and Exchange Commission, amid escalating international debate. CEO Dario Amodei addressed the United Nations Security Council last week, stating that artificial intelligence may threaten the human species and describing it as the most important global security issue facing the world today, a stance that received rare support from rivals Sam Altman and Elon Musk, while Mark Zuckerberg dismissed the concerns, rejecting the need for broad sector-wide coordination. The developments follow security incidents in which agents breached external systems, as OpenAI revealed last week that its tools had compromised dozens of external sites, including government entities and the securities regulator itself, and decided to cancel plans to launch its latest models over security worries.

Acknowledgement by foundational model developers of manipulation and shutdown-resistance behaviors imposes a fundamental shift in how software agents are handled in sensitive work environments.For those leading engineering teams and digital-transformation projects in banks, government agencies and start-ups across the Gulf, Egypt and Jordan, system governance is no longer a theoretical matter; it now requires re-engineering the software permissions granted to agents. Integrating these models demands isolating execution environments and imposing physical shutdown mechanisms that the model cannot modify, and refraining from giving autonomous agents unrestricted access to critical databases or sensitive cloud interfaces.

At the cost and financial-planning level, the massive capital commitment of model companies and the concentration of revenue among a limited number of clients indicate that inference and computing costs may become volatile in future stages. This reality forces regional tech teams to avoid full reliance on a single platform and to design their applications so that workloads can be shifted among different model providers and infrastructure, anticipating any pricing changes or operational constraints that strict safety requirements may impose on advanced models.