The French Lab Nobody's Talking About Just Hit $400M While Everyone Watched the Big Guys

Yesterday in AI | Friday, May 1, 2026

The French Lab Nobody's Talking About Just Hit $400M While Everyone Watched the Big Guys

Big Tech just revealed what $130 billion in a single quarter actually buys you, and the answer is still not enough. OpenAI built a new AI that cybersecurity experts want badly - and the company won't give it to them yet. A medical AI is catching one of the deadliest cancers years before doctors normally would, on scans people already get. AI agents took a serious run at becoming something regular professionals can actually use. And a French AI company most people aren't watching just posted revenue growth that made a lot of people look twice. Plus, Day 2 of the Musk vs. OpenAI trial did not go the way Musk planned.

Remember to subscribe, rate, and share this podcast if you like it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence in 10 minutes or less. I'm Mike Robinson. It's Friday, May 1st, and Big Tech's Q1 numbers finally put real data behind all that AI spending. OpenAI built a cybersecurity model it's keeping out of public hands, and a European AI company quietly hit $400 million in annual revenue while the US labs were busy making headlines. Let's get into it. We hit the broad strokes of big tech's Q1 earnings yesterday, but today's deeper coverage had numbers worth talking about. The four hyperscalers combined spent roughly $130 billion on AI infrastructure in a single quarter, nearly double what they spent in Q1 2025, and demand is still outrunning supply. Google's Sundar Pachai said it directly on the earnings call. Google Cloud revenue would have been higher if they could have built data centers fast enough to meet demand. They grew cloud 63% to 20 billion. AI products on Gemini grew nearly 800% year over year. Their backlog of signed but not yet delivered cloud contracts doubled and one quarter to $462 billion. And they still left money on the table because the physical infrastructure couldn't keep up. Amazon's custom chip business, the training and graviton line, crossed a $20 billion annual run rate. That makes Amazon one of the top three data center chip businesses in the world now. Microsoft's AI business hit $37 billion annually, up 123% year over year. Microsoft says 20 million enterprise users are now paying for Microsoft 365 Copilot. Accenture alone signed up 740,000 seats. Meta raised its 2026 CapEx guidance to between 125 and 145 billion. Here's the uncomfortable number underneath all of this. Amazon's trailing 12-month free cash flow dropped 95% from about $26 billion to $1.2 billion, almost entirely because of this spending. The bet is enormous. Early revenue signals are good. But if AI usage doesn't keep growing into those costs, Amazon's cash flow picture is a preview of where the rest of big tech ends up. The next few quarters will tell us a lot. OpenAI announced a new model this week called GPT-5.5 Cyber, and they're deliberately not making it public. Sound familiar? The rollout is going to a select group of what they call critical cyber defenders, vetted organizations, and government entities. CEO Sam Altman said access is coming within days. OpenAI hasn't published technical specs, framing the stage release as a way to reduce misuse risk. Two things jump out here. One, this is a signal that Frontier Labs expect to gate their most powerful capabilities by use case going forward. If the best model for your job isn't publicly available, access becomes a real procurement problem for enterprise buyers. Two, cybersecurity is one of the clearest dual-use areas in AI. A model that finds vulnerabilities well also helps exploit them. OpenAI also released a broader cybersecurity action plan this week, aimed at helping government agencies and defenders access their models. GPT-5.5 Cyber is the first solid product coming out of that plan. While the U.S. labs are fighting in court and racing to outspend each other, a French startup is quietly building a real business. Time just put Mistral on its 2026 Time 100 Most Influential Companies list, and the numbers behind the profile are worth looking at. Mistral's annualized revenue hit 400 million in early 2026, up roughly 20fold from the prior year. More than 100 major enterprise clients, a 1.7 billion euro funding round led by ASML, and a valuation approaching 14 billion dollars. What Mistral is actually selling is different from what OpenAI and Anthropic are selling. Their pitch is that you can download and run their models on your own infrastructure. Your data stays with you. You're not dependent on a US lab's pricing, uptime, or policy decisions. CEO Arthur Minch told Time that customers need to make AI systems their own, and that open source models win long term on economic grounds. The government deals back that up. France's Armed Forces Ministry signed a framework agreement in January to deploy Mistral on national infrastructure. Several other European governments are in various stages of similar arrangements. There's a specific term for what they're building, sovereign AI. The idea that a country or organization should control the AI it depends on, not just subscribe to it. This matters for anyone watching AI geopolitics. If open weight models keep improving at this pace, enterprises and governments that don't want to depend on U.S. hyperscalers have a real alternative. Mistral is still smaller than OpenAI by a wide margin, but twenty-fold revenue growth in a year is not a rounding error. Two healthcare stories today, both built around the same idea, AI seeing things that human specialists miss. BioHub, the nonprofit backed by Mark Zuckerberg and Priscilla Chan, announced the $500 million virtual biology initiative this week. The goal is to build open data sets and AI models that can predict how human cells behave at a molecular level. $400 million goes to data generation and imaging technology. $100 million funds external research labs. Nvidia, the Allen Institute, ARC, and other major research organizations are joining as partners. The datasets stay open, which means any researcher in the world can build on them. Google DeepMinds Demas Sisabas has said he believes AI could eventually end disease. BioHub is putting half a billion dollars behind that same line of thinking. The bottleneck they're trying to break is data volume. Current AI biology datasets top out around 1 billion cells, and BioHub's researchers say they need an order of magnitude more before the models get genuinely useful for disease prediction. On the same day, Mayo Clinic published results from an AI called RedMod that reads standard CT scans and catches pancreatic cancer up to three years before a formal diagnosis would normally happen. Across nearly 2,000 routine CT scans that radiologists have originally read as clean, REDMOD caught 73% of the eventual cancer cases early. At the two-year mark before diagnosis, the AI was finding roughly three times as many early cancers as experienced radiologists reviewing the same images. Pancreatic cancer's five-year survival rate is below 15%, mostly because it's almost always found late. RedMod runs on scans patients already get as part of routine care, a validated system built to fit inside existing workflows. If it becomes a standard step in CT reading, the survival rate math changes. That's not a small thing. AI agents had a real week in Enterprise. Amazon released a desktop app for Amazon Quick, their AI work assistant, and it's worth a serious look. QUIC integrates natively with Microsoft 365, Google Workspace, Salesforce, and Zoom. It has persistent memory across sessions, proactive intelligence that runs in the background and flags emails and documents you need to see, and it can take actual actions, drafting replies, editing documents, updating Salesforce records, building dashboards. Southwest Airlines, BMW, and the NFL are already on it. A free version is available, no AWS account required. At the same time, Perplexity expanded their desktop AI agent personal computer to pro subscribers on Mac. New enterprise connectors for Databricks and Snowflakes, 50 plus pre-built workflow templates, and a one-password integration that lets the agent work inside password-protected tools without ever seeing your credentials. Perplexity says their agent has performed more than $2.8 billion in labor equivalent work for customers since launch. The common thread: agents are finely accessible without a dedicated technical team to set them up. Amazon Quick doesn't require engineering resources to get running. Perplexity's templates give you a starting point without building from scratch. The enterprise security story is being addressed for real. This is what the agent market looks like as it starts to mature. The Musk vs. OpenAI trial keeps going and day two went poorly for Musk. During cross-examination, OpenAI's lawyers used his own post to contradict his testimony. The highlight, Musk testified in court that Tesla isn't chasing artificial general intelligence. OpenAI's lawyers then put up his own post saying Tesla, quote, will be one of the companies to make AGI, end quote. Hard to unsee that in a courtroom. He also testified that he pledged approximately $1 billion to OpenAI, but actually wired $38 million. His defense was that his reputation made up the difference. The judge's face, apparently, did not convey agreement. The core of the case is whether OpenAI violated its nonprofit charter by converting to a for-profit structure. Musk says yes. OpenAI says Musk supported the for-profit idea until he didn't get the control he wanted, then filed this lawsuit once the company became extraordinarily valuable without him. Both Musk and Sam Altman are still scheduled to testify. Microsoft CEO Satya Nadella is on the witness list too. Three weeks of testimony ahead. The outcome affects OpenAI's IPO plans, how nonprofit to for-profit conversions in the AI space get regulated going forward, and ultimately who gets to shape the governance of AI's most powerful systems. There's more coming. One more thing. If you like this podcast, be sure to rate it and review it so others can find it. It really does help. Thanks. That's all for this edition of yesterday in AI. Stay curious, and I'll see you tomorrow.