What To Expect at AI Agenda Live; Crusoe and Thinking Machines Labs Strike a Compute Deal

Happy AI Agenda Live day! For those joining us today at SFMOMA, I’m so excited to meet you all and talk about this exciting, if slightly contentious, moment in the AI industry.

Our staff and attendees will ask top researchers, founders and executives from OpenAI, Google DeepMind and Nvidia some burning questions. Here are three at the top of my mind.

1. Is consumer AI dead? This year, it seems like every AI company has doubled down on building AI for businesses, leaving consumer AI in the rearview mirror. Part of that is because of the so-called capability overhang, or the gap between consumers’ relatively dim awareness of what AI can do and the tasks AI can actually automate for them. The labs have been adding features in their consumer AI products to provide suggestions to users so they can bridge that gap, but it has yet to be seen how big of an impact that will have. I’m curious if the hype around products like Instinct and Meta’s Muse will revive the frontier labs’ interest in developing AI for consumers.

2. What happened to the AI research overhaul? At the beginning of the year, every researcher was talking about how the AI industry needed a complete overhaul, and how merely scaling up today’s large language models wouldn’t get us to superintelligence. There’s been a massive vibe shift since then, and now researchers are worried that we’re going to get to superintelligence too quickly and AI safety research won’t be able to keep up. What changed, and what’s going to happen to the spat of neolabs that emerged to try and overtake OpenAI and Anthropic by exploiting overlooked research approaches?

3. How are AI application startups and enterprises maintaining their competitive edge as Anthropic and OpenAI go after their businesses? The frontier AI developers are seemingly moving into every industry and application, from financial services to healthcare to law, and are increasingly positioning themselves as the control plane for all work. How are their customers making sure they don’t get subsumed in the process? For some customers, that might mean building their own models or using open-source AI to ensure they don’t become too dependent on the labs. For others, that might mean a hard decision between making their data and features available to outside AI chatbot users or taking a walled garden approach.

Here’s what else is going on…

Crusoe Signs $65 Million Deal To Help Run Thinking Machines Lab’s Models For Them

Fresh off a $3.9 billion funding round, data center firm Crusoe—best known as the developer of OpenAI’s Stargate project—is continuing to diversify beyond the massive AI data centers that put it on the map. The latest: Crusoe is expanding its offering to also run the AI models its customers use for them (called “managed inference”), and announced a $65 million annual contract with Mira Murati’s Thinking Machines Lab on Wednesday.

The workload will run on Nvidia’s B200 server chips largely inside Crusoe’s tiny, one-megawatt Spark data centers, said Crusoe’s senior vice president of product management Erwan Menard. Those chips will primarily run (rather than train) AI models, he said.

“ We save them time and extend their own engineering capabilities, and they see great value,” said Menard. Crusoe’s engineers will use software to help Thinking Machines run models more efficiently.

Crusoe announced a similar arrangement with Perplexity earlier this month. Crusoe’s managed inference product launched last November and has contracted more than $100 million in annualized recurring revenue, or monthly revenue multiplied by 12, the company has said.

Crusoe’s bet on managed inference is part of a broader trend towards companies that help customers run various AI models better and faster, rather than simply serving them raw (or “bare metal”) compute capacity. Together, Fireworks and Baseten are also companies in this category, which has Nvidia’s blessing. Nvidia is an investor in all four, and wants a universe of thousands of personalized AI models running on personalized clouds.

Big Number

Anthropic is in early talks to lease up to 1 gigawatt in compute capacity from Stream Data Centers, a developer majority-owned by Apollo Global Management, which could house tensor processing units designed by Broadcom and Google, The Information reported Tuesday.

Overheard

Microsoft leaders told sales staff this week that it is authorizing steeper discounts of 30% to 50% on subscriptions of its Copilot AI software for corporate customers who commit to purchase a large number of seats and commit to additional payments based on their usage of certain features, The Information reported.

Meta’s new personal AI agent, Muse, had more than 500,000 people try it out roughly a week after launch, including more than 250,000 daily active users, according to internal data viewed by The Information. Users have also submitted more than 2 million prompts.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

CoreWeave said on Tuesday it had completed a $4.2 billion convertible bond offering, an upsized amount versus what the AI cloud provider initially targeted last week.

Mirendil, an AI startup founded by former Anthropic researchers to build models that can improve themselves with little human help, is in talks to raise up to $1 billion at a $5 billion valuation, including the new investment, in a funding round Kleiner Perkins is in talks to lead, with Andreessen Horowitz also in discussions to invest, Bloomberg reported.

Ligent Technologies, a Qingdao AI infrastructure company that makes the optical transceivers and chips data centers use to network their servers together, rose as much as 19.2% in its Hong Kong stock market debut before closing up 4.6%, after raising HK5.7 billion ($727 million) in its initial public offering.

Cyera, an AI cybersecurity company, raised $400 million from Goldman Sachs as an extension of the Series G funding round that closed in June, pushing its valuation to over $12 billion.

Snorkel AI, which sells the labeled and human-graded data that AI labs need to train their models, raised $350 million in a Series E funding round led by Insight and S32 at a $3.5 billion valuation.

Heidi, an Australian startup whose AI drafts doctors’ clinical notes and paperwork for them, raised $340 million to expand its features and its business in Asia and the Middle East. The sum includes a $100 million equity round led by existing investor Blackbird Ventures, which values the company at $900 million.

Verda, a Finnish company that rents out AI computing power to compete with the big cloud providers, raised $189 million in a funding round led by Emergence Capital at a valuation the company says is “at least $1 billion.”

Micro1, a San Francisco company that sells the training data AI labs use to teach their models, has raised more than $100 million at a $4 billion valuation, up from about $500 million a year ago, Forbes reported.

Go.AI, a Chicago company that sells banks and other regulated businesses AI hardware and software they can run entirely inside their own walls, raised $85 million in a Series A funding round led by Updata Partners.

Firecrawl, which builds web-scraping tools that let AI agents pull data from websites, raised $75 million in a Series B funding round led by Smash Ventures.

Confido, a New York company whose AI ties together the finance, sales and forecasting data that consumer brands need to manage their retail business, raised $55 million in a Series B funding round led by Insight Partners.

Baselayer, a San Francisco company that helps banks verify the businesses and AI agents they work with, raised $35 million in a Series A funding round led by M13.

CellPoint, a London company whose AI helps manage airline and hotel payments, raised $34 million from Toscafund.

Planted, a California company whose robots build utility-scale solar farms on hilly land that standard equipment can't handle, raised $31.8 million in a funding round co-led by Piva Capital and RA Capital Management Planetary Health.

Biolevate, a Paris company whose AI agents turn scientific and regulatory paperwork into auditable workflows for drug and medical-device makers, raised €30 million in a Series A funding round co-led by RAISE France and Orange Ventures.

Chamelio, an Israeli company whose AI drafts, reviews and negotiates contracts so corporate lawyers only have to step in on the cases that need judgment, raised $26 million in a Series A funding round led by Entrée Capital.

Sela, a San Francisco company whose AI makes the phone calls and answers borrower questions for mortgage lenders, raised $21 million across seed and Series A funding in a round led by Costanoa.

Outerlimit, a London company building security software that checks whether an AI agent is actually authorized before it's allowed to take an action, raised $16 million in a seed funding round from AlbionVC, Evolution Equity Partners, Crane Venture Partners and a group of strategic angel investors.

Hong Kong-listed shares of Z.ai and MiniMax plunged on Wednesday, after The Information reported that China’s internet regulator is probing potential data leaks to Anthropic. Z.ai, also known as Zhipu, dropped 12.4%, MiniMax fell 4% and Alibaba declined 4.4%. The benchmark Hang Seng Index dipped 1%.

Anthropic announced the release of its newest model, Claude Opus 5.5, saying it performs as well or better than Anthropic’s previous most powerful models, Fable 5.1 and Mythos 5.1, in domains including coding, reasoning, business workflows, and safety. At the same time, it will cost 40% less to run, according to the announcement on Tuesday.

OpenAI on Tuesday announced GPT-6 Sol and GPT-6 Luna, which it describes as cheaper and faster models in the GPT-6 Astra family. The company said that the new models show significant improvements compared to their predecessors, GPT-5.6 Sol and GPT-5.6 Luna, in areas like professional work, coding, computer use and accuracy.

Anthropic and OpenEvidence, the medical-search platform free to US and European clinicians, are partnering to bring a specialized version of OpenEvidence's AI to physicians in about 100 low- and middle-income countries, including Uganda, Angola, Sudan, Haiti and Mongolia.

OpenAI and Grab, a Singapore-based transport and food delivery company, on Wednesday launched a program to train Grab’s drivers, merchants and delivery riders to use ChatGPT for work.

Known Systems AI spun out of the domain registrar Identity Digital to build a shared way of holding AI agents accountable to whoever created them — a system called DNSid that its backers describe as a “birth certificate for AI agents,” built on internet domain infrastructure and a blockchain ledger. Identity Digital remains a major investor and keeps a board seat.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.