AI Sector Daily Digest: July 23, 2026

In brief

1. AMD commits up to $5 billion to Anthropic's compute expansion

• AMD said it will invest up to $5 billion in Anthropic and sell the model company tens of billions of dollars' worth of AI servers. 1
• Anthropic plans to deploy up to 2 gigawatts of AMD's Instinct MI450 GPUs, with the first gigawatt scheduled to start in the first half of 2027. AMD says the equity investment is a future commitment tied to deployment milestones, not a completed cash transfer. 2
• The partnership also covers engineering: Anthropic will help optimize workloads and ROCm software with Claude, while AMD plans to use Claude across engineering and product development. 2

2. U.S. outlines a $5 billion AI-for-science program

• The Trump administration said the U.S. will spend $5 billion on AI-assisted research into chronic disease, drug discovery, longer-lasting building materials, and other scientific problems. 3
• Fifteen federal agencies are expected to participate, including Health and Human Services, Energy, Transportation, Defense, and Interior. Researchers will get access to Energy Department supercomputers, AI systems, and specialized datasets. 3
• Microsoft said it will contribute $40 million in AI computing credits over three years. The announcement did not provide a detailed allocation schedule for the $5 billion, while a related White House report calls for more support to individual scientists and more political control over federal research funding. 3

3. Gemini task automation reaches more than 40 apps on Samsung devices

• Samsung's new foldables mark the official launch of Gemini Task Automation after a beta period, with support expanding to more than 40 apps. 4
• Google says Gemini can handle tasks such as shopping, dinner reservations, travel bookings, and event tickets. Users can watch the steps, interrupt a task, take over a detail, and give final confirmation before completion. 5
• The rollout is limited by app, device, market, language, and age eligibility. Google's footnote says the current supported-app coverage includes the U.S. and South Korea, and users should supervise the automation. 5

4. Amazon cuts roles inside its AGI organization

• Amazon cut jobs in its artificial general intelligence group on July 22, continuing smaller reductions after the company eliminated 16,000 roles across the business in January. Reuters could not determine the full scope of the new cuts. 6
• An Amazon spokesperson said the company is sharpening its focus on initiatives that matter most to customers, and that the change includes eliminating roles in parts of the AGI organization. 6
• The cuts follow earlier leadership changes: former AGI chief Rohit Prasad left at the end of 2025, AGI Lab head David Luan left in February, and the work was consolidated under senior vice president Peter DeSantis in December. The report describes a reorganization signal, not an abandonment of Amazon's model work. 6

5. JANUS tests predictive guards for long-horizon agents

• A new arXiv paper proposes JANUS, a framework that trains a safety guard to anticipate where an agent's partial execution may lead, rather than waiting for an unsafe tool action to appear. Its Vanguard model combines future-risk anticipation with a second safety judgment step. 7
• Across AgentDojo, Agent-SafetyBench, AgentLAB, and LPS-Bench, Vanguard reached an average attack-success rate of 0.071, compared with 0.230 for six guard baselines and 0.397 with no guard. The authors calculate a 15.9 percentage-point protection gain over the guard-baseline average. 7
• On benign AgentDojo tasks, Vanguard's utility score was 0.680, above the guard-baseline average of 0.629. These are benchmark results from a preprint whose training pipeline uses simulated trajectories, so they do not establish performance in production environments. 7

What to watch