AI Frontiers Weekly: Chips, Context, and Control

TL;DR: U.S. experiments with a revenue-sharing deal for chip sales to China, NSF & NVIDIA pour $152M into open science AI, Anthropic expands Claude to 1M tokens and gives it an “end abuse” safeguard, and NVIDIA drops a giant multilingual speech dataset. Access, control, and alignment are the themes of the week.

U.S. allows limited AI chip sales to China (with a revenue catch)

In a surprising twist, the U.S. has approved a deal allowing Nvidia and AMD to sell certain AI chips in China—if they share 15% of the resulting revenue with the U.S. government. Nvidia’s H20 is on the approved list, and scaled-down Blackwell chips are reportedly under consideration. While this arrangement may preserve some market access, lawmakers warn it could set a risky “pay-for-play” precedent for export controls. I get why there is interest in selling to China, and I mostly agree with maintaining that channel—but I’m not a huge fan of government over-reach here. It feels like a slippery slope. Enterprises reliant on U.S. GPUs should watch for margin shifts, supply-chain bottlenecks, and compliance overhead.

Source: Reuters

NSF & NVIDIA fund Ai2 to build open multimodal AI

The National Science Foundation ($75M) and NVIDIA ($77M) are backing a new $152M initiative with the Allen Institute for AI to build fully open multimodal models and infrastructure. Dubbed the OMAI project, this five-year push aligns with the White House AI Action Plan and aims to expand access to science-grade, open-weight models. For labs, startups, and enterprises, this could dramatically reduce reliance on closed systems while strengthening open R&D ecosystems.

Source: NSF

Claude Sonnet 4 hits 1M-token context

Anthropic rolled out a 1,000,000-token context window for its Claude Sonnet 4 model—available via API and Amazon Bedrock (with Google’s Vertex AI “coming soon”). This expansion enables analysis of entire codebases or synthesis of dozens of research papers in one go. Pricing scales after 200k tokens, and Anthropic is advising on caching/batching to manage costs. What's interesting here is that I also heard someone from Anthropic say that it was able to perform needle-in-the-haystack testing with 100% accuracy—which is impressive. They cite this will be great for coding, and it will, but I’m not sure if it’s a requirement that the AI take in the entire project source code to be effective. Still, they have it now, and it’s a larger window than GPT‑5’s. This should be a game-changer for R&D teams handling massive datasets.

Source: Anthropic

NVIDIA releases ‘Granary’ speech dataset & models

NVIDIA unveiled Granary, a multilingual speech dataset with roughly 1 million hours of audio across 25+ languages. Alongside it, they’re releasing ASR and translation models trained on the dataset. For product teams, this means faster development of multilingual voice features without building training data pipelines from scratch. I think in 2025 we're going to see an increased push with multi‑modal models as the new frontier, and datasets like this feel like the early building blocks of that shift.

Source: NVIDIA

Anthropic gives Claude models power to end abusive chats

In an unusual move, Anthropic has equipped Claude Opus 4 and 4.1 with the ability to terminate conversations if persistent abuse occurs. This is framed as exploratory “AI welfare” research, but it could ripple into enterprise deployment norms and incident-handling policies for AI assistants. If adopted widely, expect new conversations around user safety, liability, and model alignment.

Source: Anthropic

See you next week!

This week’s developments highlight a recurring theme: access and control. From U.S. chip policy experiments to open-model funding, and from massive-context AI to multilingual datasets, the tension between who can use what and under what terms is sharper than ever. Enterprises, policymakers, and researchers alike need to track these shifts closely—because today’s “pilot programs” often become tomorrow’s norms.

✌️Steven