AI News: September 10, 2026

1. Harvey Raised $550 Million at $15.5 Billion and Bought Guardrails AI

Harvey. The legal AI company closed a round co-led by Diffusion and Lightspeed at a $15.5 billion valuation, up from $11 billion a few months earlier, bringing total raised past $1.55 billion. Sequoia, Kleiner Perkins, a16z, Coatue, Conviction, Elad Gil, GIC, and Goldman Sachs Alternatives participated, with Sapphire Ventures and Whale Rock joining as new investors. Annual recurring revenue is above $400 million. Harvey also acquired Guardrails AI, its fourth acquisition of 2026, and the raise follows the release of Harvey Tenet, its first post-trained open-weight model, and the Harvey LAB legal agent benchmark. Source

2. Suno Retrained From Scratch on Licensed Music and Retired Its Old Models

Suno. Suno v6 ships in three variants: a base model for controlled output, v6-wild for experimental generation, and v6-mini for speed across all tiers. It was built with licensed data from Warner Music Group, BMG, and Believe, and Suno states plainly that v6 is not trained on the data used for previous versions. New capabilities include editing songs by prompt or lyric change, conditioning on text, image, or video references, and instrument separation, with remixing planned for artists who opt in. Older models are being retired. Sony and Universal lawsuits remain outstanding, which is the context for retraining rather than patching. Source

3. An Anthropic Pretraining Researcher Resigned Over Self-Improving AI

Anthropic. Jacob Coxon, who spent three years on pretraining research across OpenAI and Anthropic, posted his resignation publicly, writing that labs “are racing straight to self-improving superintelligence and gambling with our lives” and that accepting the race is “a hubristic gamble.” He called for pacing agreements between labs and for researchers to demand different conditions rather than treat the trajectory as inevitable, including temporary capability bans. Separately, Anthropic alignment researcher Evan Hubinger put the odds of catastrophic AI misalignment this decade above 10%. Source

4. Massachusetts Became the Third State in Three Months to Restrict Data Centers

Massachusetts. Governor Maura Healey’s executive order requires data centers above 25 megawatts to supply clean power or pay into a ratepayer protection fund, holding them to 100% clean energy against a statewide standard of 40% by 2030. On-site generation is preferred; otherwise developers must fund nearby clean generation. The order also advises communities against signing NDAs with developers and pauses new data center sales tax exemption applications. Texas mandated audits for new facilities in August and New York halted construction of anything at or above 50 megawatts in July. Source

5. AI Spend Per Employee Fell Nearly 10% at the Heaviest-Spending Firms

Ramp. Data from Ramp covering 70,000 businesses shows 56% of its customers paid for AI products in August, up only 0.4% from July, while spend per employee among the top 1% of users dropped almost 10% to $7,205. Average token cost fell to $0.68 per million, down from a March 2026 peak of $1.15. Economist Ara Kharazian attributes the drop to competition between OpenAI and Anthropic making AI cheaper faster than volume grows, with companies increasingly picking older models over frontier releases. Census Bureau data still puts overall US business AI use at 22%, and only 6.4% of AI-spending businesses touched open-weight model platforms. Source

6. Listen Labs Walked Away From a Signed $125 Million Round to Talk to Salesforce

Listen Labs. The three-year-old AI customer research startup abandoned a signed Series C with Menlo Ventures at a $1.5 billion valuation after Salesforce reportedly offered roughly $2 billion to acquire it. Listen Labs runs automated customer interviews and generates research reports, with about $30 million in annualized revenue and Microsoft, Canva, and Anthropic as customers. At $2 billion that is roughly 67 times revenue. Reneging on a signed term sheet is unusual enough that VCs quoted in the piece flagged it; if the Salesforce talks fail, the expectation is a return to market at $2 billion or above. Source

7. Cymphony Raised $30 Million to Track AI Agents as Enterprise Identities

Cymphony. Sequoia Capital and SMBC Fin Atlas Beyond Fund co-led a Series A at a valuation above $100 million for a startup building a workforce graph that combines identity, data, and activity signals across employees, AI agents, and other nonhuman identities. The technical argument is that agents reach the same systems as human workers but at machine speed, can take different routes to the same task, acquire new capabilities, and spawn other agents, which breaks access governance designed around stable human roles. At one US public company Cymphony found roughly 85,000 files that had become reachable by AI tools. Source

8. A Former OpenAI Staffer Is Manufacturing Chip Precursors in Orbit

Besxar. Ashley Pilipiszyn’s startup launches wafer precursor production in small containers called fabships aboard SpaceX Falcon 9 boosters, using the vacuum of space instead of building terrestrial clean rooms. Early missions established that wafers can survive launch without contamination and tolerate exposure. The roadmap adds capability incrementally, heating wafers and then depositing materials, across roughly a dozen Falcon 9 flights before scaling to Starship. Besxar has raised about $14 million including a $9 million seed led by Dauntless Ventures and Overture VC, and targets chips for data centers, robotics, and EVs. Source

9. Instacart Shipped a Conversational Shopping Agent Called Clementine

Instacart. Clementine turns conversations, grocery lists, recipes, and photos into ready-to-buy carts, generates personalized recipes, substitutes for dietary constraints such as gluten-free and vegetarian, surfaces deals, and reads handwritten lists. It is live in the US and Canada. Shipt announced a comparable assistant the same day, making cart construction the current competitive surface in delivery apps. Source

10. IBM Released a 385M-Parameter Time Series Model Under Apache 2.0

IBM. Granite Time Series PatchTST-FM-r2 handles context lengths up to 8,192 steps and ranked second among replicable zero-shot models on the GIFT-Eval leaderboard as of September 8, with a geometric-mean CRPS of 0.467 and MASE of 0.6846. Against all models including those trained on benchmark data it places third on CRPS and fourth on MASE. The architecture is conformer-based, mixing self-attention with temporal convolution, and the model does zero-shot forecasting, 99-quantile probabilistic output, and missing value imputation. It is dual-licensed under Apache 2.0 and OpenMDW 1.0. Source