1. Anthropic Quietly Builds a Physical Wet Lab for AI Drug Discovery
Anthropic has set up a wet lab in the San Francisco Bay Area to run real physical biology experiments, head of life sciences Eric Kauderer-Abrams confirmed to Reuters. The facility uses lab automation driven directly by Claude, building on the Model Hardware Standard the company published in August, with human researchers kept in the loop as a safety layer. It extends Anthropic's life-sciences push that began with its $400M acquisition of Coefficient Bio in April.
2. Crusoe Raises $3.9B Series F at a $30.9B Valuation
AI infrastructure and cloud provider Crusoe closed a $3.9 billion Series F on September 17, co-led by Atreides Management and Valor Equity Partners. The $30.9 billion valuation is roughly a 3x step up from the $10B-plus mark set by its $1.375 billion Series E in October 2025. The round underlines how much of the current AI capital cycle is flowing into compute and data-center capacity rather than models.
3. Google Ships Gemini 3.8 Live with Speaking Extended Thinking
Google released Gemini 3.8 Live, a real-time voice model whose Extended Thinking mode can talk while it is still reasoning and running tools, instead of going silent during a long turn. It scored 82.6 on Artificial Analysis' speech-to-speech benchmark, the top result on that leaderboard. Rollout is staged through the Gemini API with tiered access levels.
4. Salesforce Puts 37 Sales Workflows Directly Inside Claude
Salesforce and Anthropic launched a beta plugin that exposes 37 sales skills — account research, call prep, and CRM record updates — natively inside Claude. The integration runs under the user's existing Salesforce permissions rather than a broad service account, and every proposed write-back requires explicit user confirmation. It is one of the clearer examples of agents being scoped down to a permissioned system of record.
5. Commerce Department Reportedly Had a GPU Price Chart Pulled
Semafor reported that the Commerce Department asked prediction market Kalshi to take down a chart tracking GPU rental prices, citing national security concerns. Kalshi complied; Commerce denied making the request. The episode is a notable signal that public visibility into AI compute pricing is becoming politically sensitive, not just commercially so.
6. Moonshot Rolls Out Kimi K2.8 Preview with 1M-Token Context
Moonshot AI launched Kimi K2.8 Preview on September 11, adding vision and audio input to what had been a text-only line — images up to 4096x4096 and audio clips up to 30 seconds. Every Kimi Code membership tier now gets a 1 million token context window, and the model exposes low/high/max thinking-effort levels with max as the default. Moonshot claims performance close to K3, but that rests on its own description; independent benchmarks and the technical report are still pending.
7. SK Hynix Launches a Silicon Valley Venture Arm
SK Hynix opened SK Hynix Ventures in Silicon Valley, widening its investment mandate beyond memory into AI computing, data centers, and optical interconnect. The memory maker is one of the main beneficiaries of the HBM supply crunch, and the fund points at the interconnect layer as the next bottleneck it wants a position in.
8. D-Robotics Raises $400M Series C for Robotics Chips
Shenzhen-based D-Robotics closed a $400 million Series C on September 17, reportedly led by Mirae Asset. The company builds robotics infrastructure and AI chips, and the round lands alongside a broader push into embodied AI and robotics reasoning that has been one of 2026's consistent funding themes.
// KEY TAKEAWAYS
This week's centre of gravity was infrastructure and embodiment, not chatbots: Crusoe's $3.9B round, D-Robotics' $400M, and SK Hynix's new venture arm all bet on compute, chips, and interconnect, while Anthropic went a step further and built a physical lab for Claude to run experiments in. On the model side the theme is agents with narrower blast radius — Salesforce's 37 skills run under the user's own permissions and confirm every write. And the Kalshi episode is worth watching: when a GPU price chart becomes a national security question, the compute market itself has become strategic territory.