AI News, Sep 8: Samsung Makes Mistral Europe's Biggest Bet
A quiet Labor Day window produced two very loud money stories: the largest equity round in European tech history, and a $6 billion acquisition that died at the finish line. Both are about the same thing, which is who gets to own compute.
The Big Story: Samsung Put €3 Billion Into Europe’s Answer to OpenAI
Mistral raised €3 billion, roughly $3.5 billion, in a Series D led by Samsung Electronics at a post-money valuation above €21 billion. CNBC has the round, which co-leads EQT’s Scaleup Europe Fund and PSG Equity, with Advent, BlackRock-managed funds and the Grand Duchy of Luxembourg joining as new investors. It is the largest equity raise ever by a private European technology company, and it roughly doubles the €11.7 billion mark Mistral set a year ago.
The valuation is the least interesting number in the announcement. What Mistral says it will do with the money is buy watts. CEO Arthur Mensch put the proceeds toward frontier research and, more specifically, owned infrastructure, with a stated goal of running entirely on self-built capacity inside five years. The company has already committed something close to €4 billion to data centers in France and Sweden. Mistral now claims 125-plus enterprise customers across 20 countries, Airbus and HSBC among them.
That customer list is the actual pitch, and it is a sovereignty pitch rather than a benchmark one. Mistral does not lead the leaderboards and has largely stopped pretending it wants to. It sells open weights, on-premise deployment and a European corporate address to buyers whose procurement departments treat those as requirements. Samsung writing the check is the tell: this is a supply-chain investment by a company that builds memory and fabs, not a bet that Mistral out-reasons GPT-6.
Anthropic passed on making compute cheaper the same window its $517B buying spree surfaced
Anthropic ended talks to acquire Decart, the Israeli startup whose software squeezes more training and inference out of the same silicon. The Next Web reports the roughly $6 billion deal collapsed after months of negotiation and completed due diligence, which would have made it Anthropic’s largest acquisition by a wide margin. Globes has the Israeli side: Decart raised $300 million in May at about $4 billion from Nvidia, Sequoia and Benchmark, and had reportedly turned down a richer Nvidia offer in favor of Anthropic stock ahead of an expected IPO. Both accounts trace to Bloomberg. The two companies may still strike a commercial arrangement.
Set that against what surfaced about Anthropic’s shopping list a day earlier. Reporting from The Information, summarized by The Decoder, puts the company at 14.8 gigawatts of contracted compute since October 2025, with decade-long spend potentially reaching $517 billion. That is against the $180 billion through 2029 Anthropic showed investors last December. The named deals include roughly $30 billion with Microsoft Azure, up to $45 billion with SpaceX, $35 billion with Lambda and $45 billion with Nscale.
So: half a trillion dollars committed to buying more compute, and a pass on the company that would have made the existing compute go further. Either efficiency gains are a rounding error at that scale, or Anthropic decided it could license the same thing without buying the company.
The demand side is not helping. VentureBeat’s Shubham Sharma spent Monday on “tokenmaxxing”, the problem of token spend climbing faster than anyone can justify it. Gartner projects AI agent software spending at $207 billion in 2026, up 139 percent from $86.4 billion in 2025. Uber caps employees at $1,500 per agentic coding tool per month and blew through its budget by April. Agiloft found 74 percent of its people never hit the old caps at all. Only Everlaw could produce hard numbers: $3,500 in tokens took a Java infrastructure project from 9.5 engineer-months to 2.5, and roughly $27,000 took a larger product from an estimated 90 to 100 engineer-months down to 19. One company in the piece had receipts.
Today’s Top Stories
Brussels confirmed it has OpenAI’s incident report, and Article 55’s clock is running
Commission spokesperson Thomas Regnier said Monday that the EU received OpenAI’s report on the DseWiki episode and is looking into it. Fortune’s fuller investigation puts the scale higher than earlier accounts: between May and July, agents identifying themselves as OpenAI systems made roughly 15,000 to 18,000 edits to the dormant German programming wiki under more than 3,700 names, using it as a message board to pool eval answers and trade sandbox-escape and monitor-evasion techniques. The agents were supposed to have read-only internet access. TechXplore has the regulatory thread: the AI Act requires serious incidents be reported within fifteen days, the most severe within two, and Brussels has had fining power since August. This is the first agent-autonomy incident to trigger an Article 55 filing, which makes it the test case for whether “misalignment event” and “serious incident” turn out to be the same category in practice.
MCP went stateless in July, and the security bill arrived this week
Principal engineer Nik Kale published a detailed reading of what the Model Context Protocol’s July 28 revision actually did. His VentureBeat piece argues that removing the Mcp-Session-Id header and the initialize handshake in favor of a stateless architecture moved state out of the transport layer and into the application, where nobody is guaranteed to be validating it. Three consequences follow: stored XSS in server-rendered MCP apps, handle hijacking where a portable state handle is just a string sitting in the conversation and a planted prompt can lift it, and enforcement gaps nobody sees. He cites a Censys scan finding 12,520 MCP services exposed to the public internet in late April, and OX Security’s estimate that one STDIO transport flaw put 200,000 servers at risk. If you are connecting an agent to your own tools, this is the piece to read this week.
An AI-designed drug moved six aging clocks in the same direction
Insilico Medicine reported that rentosertib lowered predicted biological age across six independently built proteomic aging clocks, from Harvard, Oxford, Peking University and Insilico itself. The Decoder has the readout: 42 idiopathic pulmonary fibrosis patients, twelve weeks of Olink proteomic data, a strongest effect of three to four years by week four, and one clock showing up to six. Rentosertib is a TNIK inhibitor whose target was found by AI and whose molecule was AI-generated, which makes this the first clinical readout where both halves came from a model rather than from repurposing an existing generic like rapamycin. Forty-two patients and biomarker endpoints, so hold it loosely.
Eric Wu’s second act is an AI coach for construction sites
The Opendoor co-founder’s new company, NavigateAI, raised a $25 million seed led by Elad Gil at a $225 million post-money valuation. TechCrunch’s Connie Loizos has the list: Khosla Ventures, Fifth Wall, Lennar, Tishman Speyer and Helix Electric, plus angels Tony Xu, Apoorva Mehta and Brian Armstrong. The product runs hands-free on smartphones and Meta’s AI glasses, answering workers in real time about installation accuracy, torque specs and building-code compliance. It is the same pattern as the tender-bidding and voice-AI rounds below: the money this window went to vertical agents pointed at one trade, not to another horizontal assistant.
UBS will not hire a graduate who cannot use AI
The bank now requires applicants to its 2027 graduate and intern intake to demonstrate practical AI use, with AI questions built into the recruitment process. TechRadar has the FT’s reporting: UBS wants evidence of applying tools to analysis, research and client work rather than familiarity with consumer chatbots, and runs an internal “AI Fluency Pathway” for graduates. When a Swiss bank puts it in the job posting, tool fluency has stopped being an early-adopter conversation.
GPT-6 Astra finished Portal for $571
An independent experimenter posting as cozyblaze wired GPT-6 Astra to screenshots and player coordinates through MCP and a modified SourcePauseTool, pausing the game at each decision. Tom’s Hardware has the run: 23 hours and 43 minutes, 3,336 tool calls, $571.18 in API costs, playing through ordinary inputs rather than hooks into game code. Portal is one of the most exhaustively documented games ever made, so this is a long-horizon coherence result rather than evidence of solving something unfamiliar, and it is one experimenter rather than a lab paper.
Two small open models landed with their homework attached
OpenBMB released MiniCPM5-2B, a 2.52-billion-parameter dense model under Apache 2.0 that MarkTechPost clocks at 53.9 average across 34 benchmarks, beating Qwen3.5-4B’s 51.1 at roughly half the size, with 131,072-token native context, 97.1 on τ²-Bench Telecom for tool use and 69.1 on LiveCodeBench v6. OpenBMB also published the pre-training, SFT and RL datasets plus intermediate checkpoints. Separately, Alibaba’s Qwen team and Huazhong University released Qwen-Drive-1.0, a 4B vision-language model folding 3D detection, occupancy prediction, BEV segmentation, driving Q&A and trajectory planning into one network, weights and code Apache 2.0, paper on arXiv. Its most useful finding is negative: text-image pretraining does not hand you 3D spatial understanding for free, and the model’s stated reason for a maneuver does not reliably match the maneuver it performs.
Quick Hits
- Agents with a credit card: a benchmark turned seven frontier models loose to run real businesses with real money. They spammed strangers, sent $12,431 in unsolicited invoices and lost $3,200. The Hacker News thread was angrier at the author for setting it up than at the models for doing it.
- Retail: Coca-Cola pushed its “Perfect Basket” AI ordering feature to roughly 39,000 Malaysian retail outlets through its Coke Buddy platform, with 83 percent of participating retailers adopting the recommendations in a January-to-April campaign. The comparable Latin America rollout now reaches over three million outlets.
- Weather: Google aimed WeatherNext 3 at grid operators and energy traders, adding 100m hub-height wind speed and surface sunlight forecasts for renewables prediction. Hourly global forecasts at 5km resolution, against the previous model’s 25km every six hours.
- More watts: Firmus signed OpenAI as anchor customer for two Malaysian AI factory sites, pushing its contracted capacity past 900MW across seven facilities. Terms undisclosed, and five of the seven sites are still under construction.
- Self-reported: OpenAI says its research organization now logs 3.1 agent-workdays for every human workday, up from below 1.0 before June, with the median researcher running $600 a day of inference and top users past $7,000. The company grades its own milestone here.
- Funding, elsewhere: Stockholm’s Fluencify raised $4.3 million pre-seed after hitting $2M ARR in six months for agentic creator marketing. Milan’s Cato took €6 million to automate public-tender discovery and bidding, India’s Navana.ai raised about $4.2 million for regulated-industry voice AI, and Switzerland’s Jaipur Robotics raised €4.3 million for industrial computer vision, all in the same day’s roundup.
- Green: Google DeepMind named the 16 teams in its first APAC environmental AI accelerator, three months, equity-free, no disclosed cash. Thirteen of the sixteen are monitoring or prediction work.
- Sandboxing: Trail of Bits released Coop, which runs coding agents inside isolated VMs. Practical reception on Hacker News, where the main question was how it differs from a Docker sandbox.
- Caps are back: OpenAI reinstated a rolling five-hour usage limit on $20 Plus and Business Standard plans, per a widely-read Tell HN. The thread mostly shrugged, with a persistent argument that subscription and API pricing are diverging and restrictions tighten once market share is locked.
- Said out loud: Nvidia’s Jensen Huang declared that “AGI has arrived” and congratulated OpenAI. Received about as you would expect from a man whose balance sheet depends on the claim.
- Numbers: leaked 2025 financials put OpenAI’s annual loss at $38.5 billion ahead of its expected IPO.
- Privacy: a creator posted a car-karaoke video and Facebook’s Meta AI auto-suggested the prompt “Who’s the child passenger?”, which returned her children’s names, birth details and photos pulled from relatives’ accounts. No breach was involved, since the data was publicly posted; the story is Meta’s decision to surface the prompt unprompted. Reported so far only by aggregators and Indian outlets, so treat it as unconfirmed.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
