SambaNova closes $1B at $11B, lands JPMorgan as its inference partner

Good morning 👋 A regulated bank just handed its inference stack to a chip startup instead of Nvidia, and General Atlantic just wrote a check that says it's not a one-off bet.

In today's issue:

• 🔭 SambaNova closes $1B at $11B, JPMorgan bets its inference stack on it over Nvidia
• 🧠 Gemini Robotics gets a body, the Gemini API gets more autonomy
• 🔬 Doomed agent runs caught early, benchmarks that lie, compliance mapping on autopilot
• 🚀 Legal AI agents just priced in at $1.2B
• 🔗 A French HR roll-up's odd cap table, Alberta's code sprint, NVIDIA's single-thread bet

🔭 THE ONE THING

🏦 SambaNova closes $1B at $11B, lands JPMorgan as its inference partner

SambaNova closed the first tranche of a $1B Series F at an $11B valuation, led by General Atlantic, with T. Rowe Price, Capital Group, BlackRock, Qatar Investment Authority, Vista Equity Partners, and Seligman Ventures all writing checks. That's five months after a $350M Series E in February, a fast markup for a chip startup still fighting Nvidia's default gravity. JPMorgan Chase named SambaNova its inference-infrastructure partner, running SN40 and SN50 chips on-prem inside its own data centers. Banks hedge everything and move slowly, so a multi-year hardware bet from one that size is a harder signal than the valuation headline. Worth watching whether other regulated industries follow, or whether SambaNova just landed the one client willing to bet against the default.

🧠 MODELS & RELEASES

• 🤖 Google DeepMind is now training Gemini Robotics on real-world data pulled from Apptronik's Apollo 2 humanoid, as the pair expand Apptronik's Robot Park facility. A bipedal robot walking around a real facility beats another round of simulation. X post
• 🔌 Google AI widened Managed Agents in the Gemini API: background task execution and remote MCP support are in. If you've been hand-rolling long-running jobs or tool access outside your own server, some of that plumbing just got cut. Google blog
• 🦾 NVIDIA and Hugging Face are contributing new models and frameworks to LeRobot, going after the costly, fragmented datasets and sim tooling that keep physical AI work locked up outside the big labs. NVIDIA blog

🔬 RESEARCH HIGHLIGHTS

• Doomed from the Start Multi-step LLM agents routinely lock into a trajectory that's going to fail long before the failure is visible, and burn real inference budget getting there anyway. This paper finds the tell shows up early in the agent's internal representations, and a lightweight probe cascade catches it in time to abort. Run agents at any volume and that's compute you're currently spending on episodes that were never going to work. arxiv

• Your benchmark table lied to you Most LLM data-analysis benchmarks test fact retrieval on small, clean tables. This paper builds one closer to what an analyst actually opens: large multi-table joins with half-documented columns, plus questions that need outside context to answer. Scores drop hard. A model's spreadsheet-benchmark score isn't a great predictor of how it'll do on your actual warehouse. arxiv

• Compliance mapping, automated Matching cloud security controls to the technical metrics that prove you meet them is still mostly manual work. Researchers fine-tuned sentence transformers on 3,499 semantic pairs pulled from five European security standards, and the domain-adapted model does the matching on its own. Narrow use case, but it's exactly the grunt work a compliance team would happily hand off. arxiv

🚀 AI STARTUPS

• Norm AI closed a $120M Series C at a $1.2B valuation, led by Khosla Ventures, with Blackstone, Bain Capital Ventures and New York Life among the checks. The pitch: legal AI agents doing high-stakes regulated work under attorney supervision, priced on outcomes instead of the billable hour. Under three years old and already past $260M raised total, with clients managing over $30 trillion in assets. Whether "full-stack legal AI" holds up the first time one of those agents is wrong in a way that matters is the question nobody's pricing yet. Source

🔗 QUICK LINKS

• Skello landed €200M from Bridgepoint, and the odd part is the founders end up owning *more* of the company after the round, not less, off actual profit. Bridgepoint
• Anthropic had Claude scan 466 million lines of Alberta government code in 20 hours, a 6.5-year job by hand. Anthropic
• NVIDIA claims its new Vera CPU hits 1.8x the per-core throughput of x86, betting agent loops care more about single-thread speed than raw core count. NVIDIA

See you tomorrow.

Pradeep Perugu