OpenAI raises $11B at a $1.25T valuation as Anthropic files for IPO and Microsoft retreats from Oracle compute.
Capital and compute dominate the day. OpenAI closed roughly $11B at a $1.25T valuation, per Bloomberg, and the round could reach $20B. Anthropic's IPO filing means both leading labs are now pricing themselves against public-market scrutiny. Meanwhile Microsoft is unwinding its Oracle deal to rent AI cloud capacity, which suggests the biggest buyer of third-party compute is rethinking how it provisions GPUs.
Bloomberg reports OpenAI has closed on roughly $11B in new funding at a $1.25T valuation. The round could grow to $20B in total. This was the most widely covered story in the set, at three sources.
A $1.25T price tag sets the reference valuation for every frontier-lab fundraise and IPO. It also commits OpenAI to growth that justifies it.
Latent Space's AINews leads with Anthropic's IPO filing. The same issue covers ARC Prize benchmarking Inkling 2 Small and Thinking Machines reporting that guardrails could be bypassed in early checkpoints. The source summary doesn't give filing terms.
An Anthropic S-1 would force public disclosure of compute commitments, margins, and safety claims. The Thinking Machines result is a reminder that guardrail robustness is weakest early in training.
Microsoft is ending its arrangement to rent AI cloud capacity from Oracle, according to Bloomberg via Reuters. The report ties the move to a shift in Microsoft's AI infrastructure strategy and its internal GPU provisioning.
Oracle's AI cloud story depends on large anchor tenants, and losing Microsoft hurts that narrative. It also signals that even the largest hyperscaler is trimming rented capacity.
Apiiro found that the "Don't Ask Again" feature in Anthropic's Claude Code cached credentials. Those credentials ended up in 1,122 public GitHub files. The flaw sits in Anthropic's IDE extension.
Coding agents write to repos, so a convenience setting becomes a secrets-leak pipeline at scale. Teams running agents need secret scanning and permission audits before commit, not after.
The Verge reports that China's MSS allegedly recruited an executive at Megaspeed to gather intelligence on Nvidia's AI chip strategy and leadership. Two sources covered the story.
It shows how much Beijing values visibility into Nvidia's roadmap and export posture. Nvidia partners and distributors are now a counterintelligence surface.
Latent Space reports that Inkling 2 Small, a 17A-594B-parameter model, trained in 68 hours on 12T tokens and outperforms Muse Spark. The same issue notes METR published a study on Anthropic's AI research.
If the training-time figure holds, it points to a large drop in time-to-frontier-class model. ARC Prize benchmarking, mentioned in the same digest, is where the claim gets tested.
OpenAI hired Hacktron to probe its systems, Apiiro showed Claude Code leaking credentials, and AWS published an Okta plus Bedrock AgentCore pattern for governing agent access.
OpenAI's $1.25T round, Microsoft exiting Oracle capacity, and China targeting Nvidia's leadership all point to compute as the strategic chokepoint.
GitHub added Copilot code review APIs, an expanded impact dashboard, and new JetBrains and Xcode integrations.
AI scraping bots are swarming university library servers and threatening public access, a real infrastructure cost that AI companies are externalizing onto public institutions.
A commentary piece claiming AI is "dramatically accelerating AI research" offers no concrete findings, metrics, or new evidence.
Watch for Anthropic's IPO filing details and independent benchmark results on Inkling 2 Small, which will show whether the 68-hour training claim and the capital narrative hold up.