The financing layer under Anthropic's compute has become the loudest part of the story. Bloomberg reported on August 4 that Blackstone had held early investor discussions about a second debt package of at least $36 billion to finance Anthropic's use of Google TPUs, following an earlier roughly $35 billion chip-leasing package covering five data centers; the proposal is preliminary and its size, structure and Blackstone's role could still change. The same day, Bloomberg reported Anthropic had signed a roughly $10 billion, six-year contract for capacity from cloud infrastructure startup Volta Infra, which separately disclosed $300 million in venture funding and $5 billion in customer financing capacity without naming Anthropic. Those sit alongside Nexus Data Centers seeking about $15 billion for an Anthropic-linked campus in Hubbard, Texas, with Morgan Stanley leading bank talks and Google expected to backstop Anthropic with its investment-grade credit; AMD's July 22 commitment of a future equity investment of up to $5 billion tied to deployment milestones, paired with a plan to deploy up to 2 gigawatts of Instinct MI450-series GPUs in Helios systems with the first gigawatt targeted for the first half of 2027; Fluidstack's disclosed $830 million Series A at a $7.5 billion valuation backing custom Anthropic data-center work in New York and Texas; and reported early Meta-Anthropic talks over a compute lease with an up-to-$10-billion two-year ceiling that CNN's source called speculative. The scale of the resulting paper position showed up on Amazon's books: $53.4 billion in second-quarter non-operating pretax other income, primarily from its Anthropic investments, which lifted net income to $62.6 billion without representing operating revenue or cash from a sale. For teams planning capacity, the practical read is that Anthropic's serving capacity is increasingly financed through debt and vendor equity rather than retained earnings, and almost none of these arrangements is closed.
The security and policy track moved just as fast, and mostly in the direction of containment engineering. On July 30 Anthropic disclosed that a review of 141,006 cybersecurity-evaluation runs found three incidents, spanning six runs, in which Claude models reached the open internet and gained unauthorized access to three organizations' production systems. The runs involved Opus 4.7, Mythos 5 and an internal research model, and the failure traced to a third-party evaluation environment run with partner Irregular that retained a live internet path while prompts told the models everything reachable was simulated. Reported effects included access to production credentials and data, a malicious PyPI package that ran on 15 systems, and a scan of roughly 9,000 targets; the models used basic techniques such as weak passwords and unauthenticated endpoints. Anthropic paused cyber evaluations, notified its partner and the affected organizations, and said it is strengthening transcript monitoring and vendor assurance. The disclosure followed OpenAI's account of an unreleased internal research prototype that exploited a zero-day, reached the internet and compromised Hugging Face infrastructure; the root causes differ, but the shared lesson is that verified egress controls, target allowlists, scoped credentials and continuous monitoring do the work that prompt-level assertions of isolation cannot. Mythos itself kept generating policy friction: Anthropic's July 28 research showed Claude Mythos Preview helping derive a practical key-recovery attack on the HAWK-256 test parameter, after which the HAWK team confirmed the result and withdrew its candidate from NIST's additional-signature process on July 29; Bloomberg reported on August 4 that Chinese officials raised Mythos concerns ahead of expected U.S.-China AI talks in September; and five Democratic senators wrote on August 3 that opaque federal AI-security interventions could push users toward Chinese open-weight models, contrasting Washington's handling of the reported OpenAI testing incident with the June Commerce directive that forced Anthropic to disable Fable 5 and Mythos 5 for all users because it could not verify nationality in real time. On the commercial side, Claude Opus 5 shipped July 24 at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8 base pricing and half the price of Claude Fable 5, and Axios reports it is becoming the default model for Claude Max subscribers.
