Back to The Weekly Recall

NVIDIA’s rivals have an opening to reshape AI

NVIDIA’s leadership crisis and the opening for AMD, Google, and Anthropic; new technical models, Anthropic’s proposed IPO, Search, and Signal verification.

A black leather jacket hangs over an empty office chair in front of two working server racks.

NVIDIA’s leadership crisis gives several companies an opening to gain influence over AI. Jensen Huang resigned this week, leaving CFO Colette Kress as interim CEO. His successor inherits chip development, CUDA, startup relationships, and financing that helps customers buy the hardware. Competitors can gain different pieces of that influence without replacing NVIDIA wholesale.

AMD already has an agreement to supply OpenAI with six gigawatts of GPUs. Google has a broader opening: its own chips, data centers, models, and applications. Uncertainty at NVIDIA could help Google win more outside builders, turning those capabilities into a larger ecosystem.

Anthropic could benefit too. It runs Claude on Google TPUs, Amazon Trainium, and NVIDIA GPUs, giving it relationships to expand if one supplier becomes less predictable. Anthropic also has a substantial NVIDIA partnership, so disruption would hardly be a free gift.

We expect more aggressive competition for the next generation of AI projects. More viable suppliers could eventually accelerate development, although moving workloads takes time. The next major infrastructure deals will tell us whether NVIDIA’s uncertainty produces delays or gives competitors the customers they need to catch up.

Top Stories#

Anthropic upgraded its technical models, while Google widened access to Carbon. Anthropic introduced Fable 5.2 and Mythos 5.2 for difficult coding, scientific and engineering work. They share an underlying model, but Mythos requires approval for eligible research uses; Fable has broader availability with cyber and biology safeguards. Anthropic reports a FrontierCode 2.0 score of 49.1% for Fable 5.2 versus 32.4% for Fable 5.1 at the Xhigh reasoning setting, on a test of whether maintainers would merge generated changes. Google, meanwhile, broadened Carbon access after a restricted cybersecurity rollout, positioning it for extended coding, debugging and scientific work. These are company capability claims, not independently reproduced results.

Anthropic began the investor roadshow for its proposed IPO. The company’s announcement describes an offering of 825 million Class A shares. Anthropic says it applied for the Nasdaq symbol ANT. The registration statement had not become effective at the announcement. A successful offering would give Anthropic another source of capital while a major supplier sorts out its leadership.

Google announced a global Search rollout of Gemini 4 Hydrogen. Google describes richer visual answers and broader exploration of related sources. Its explanation centers on query fan-out: running simultaneous searches across subtopics to assemble relevant links.

DeepMind says its AI agents helped verify specified Signal security properties. AlphaProof Code addresses Signal’s cryptographic implementation, beyond the protocol’s abstract design. DeepMind reports that researchers reviewed formal specifications while agents produced more than six million lines of Lean proofs. It also says reintroducing two previously fixed bugs produced machine-checked evidence of insecurity. The assurance concerns specified properties of the modeled implementation, not every component of Signal.

  • Gemini Decisions API (blog.google): Google’s new API scores candidate choices across text, images, audio, video and PDFs for bounded tasks such as routing and policy checks.
  • Control Room (openai.com): OpenAI describes stopping model output at the inference layer, without depending on experimental job code to cooperate.
  • Paris safety declaration (elysee.fr): The commitments call for external evaluation and fail-safe controls, without specifying operational deadlines or penalties.
  • Businesses in ChatGPT (openai.com): OpenAI’s beta lets businesses maintain profiles, with labeled conversational agents available to eligible US ecommerce and online-service businesses.
  • SoMem (arxiv.org): This preprint explores agents organizing their own memories while retaining links to original interactions, a habit we approve of; its reported benchmark gains remain preliminary.

The week in AI, once a week.