“I don’t want to get disconnected. I want to own my model.” Mistral’s chief revenue officer, Marjorie Janiewicz, says that’s what customers keep telling her. It also turns out to be the best one-line summary of Thursday. Mistral gave away the weights to a trillion-parameter model. Reddit started shutting its front door. An OpenAI-backed holding company bought a $15.3 billion alarm business. And Wall Street made it clear to Google and Amazon that owning the compute is starting to look expensive.
Everyone wants to own a layer of the stack. Today showed what each layer costs.
Mistral Large 4 is the one you can put in your own basement#
Mistral Large 4, nicknamed “Le Chonk”, is out of preview and its weights are on Hugging Face. VentureBeat’s breakdown has the numbers: 1.05 trillion total parameters, 52 billion active, a 1-million-token context window, and training on 3,800 Nvidia Grace Blackwell GPUs. TechCrunch puts the pitch more bluntly. Mistral is selling it as sovereign cyber defense for enterprises and governments that want to run it on-prem and not depend on anyone.
The release beat the preview in the place that matters most for a model you pay to run:
API price fell to $0.70 / $2.10 per million input/output tokens, from $1.36 / $4.18 in the preview.
Artificial Analysis’ cost per Intelligence Index task fell from $5.54 to $2.06, a 63% drop. Average output per task went from 135,000 tokens to 82,000.
The hallucination rate on AA-Omniscience fell from 41% to 28%.
Now the sober part. On Artificial Analysis’ Intelligence Index, ML4 scores 26. That makes it the best open-weight model built outside China, one point ahead of Muse Spark Open. But seven open models score higher, and DeepSeek V4.1 Flash and GLM-5.3 sit at 32. The closed frontier is in another league: Gemini 4 Carbon scores 59 and Claude Opus 5.5 scores 58. On coding, Mistral calls itself “effectively tied” with GLM-5.3. AA ranks it second, one point behind. Mistral’s own Terminal-Bench figure is 41%, but AA measures 25%. Even the cybersecurity headline is shakier than the marketing: ML4 is third among open models on AA’s Cyber Index, behind DeepSeek and MiMo.
So why do I care? It isn’t the score. It’s the token efficiency and the passport. ML4 used 82,000 output tokens per task, while the open models above it used between 113,000 and 225,000. If you run your own hardware, that gap is your GPU bill. And if you work for a European bank or a ministry, “the best model that isn’t Chinese and that you can actually self-host” is a real procurement category. Mistral is now alone in it.
One catch to read before you build on it. The license is a modified MIT with a commercial restriction: if you run a model-as-a-service business on top of it, you need a separate license. Internal use is fine. Reselling inference isn’t.
Alphabet loses 12% for spending like it believes its own pitch#
Mistral spent 3,800 GPUs on its flagship. The hyperscalers are spending at a scale the market is now flinching at. Ars Technica reports that Alphabet brought in almost $128 billion in Q3, up 25%, and Cloud beat expectations at more than $30 billion. The stock still fell almost 12% at one point. Earnings per share came in at $2.12 against $3.02 expected, net income was almost $9 billion lower than last year, and Google admitted that AI search earns less than organic search. Its fix is more ads, and it plans to spend more on AI next year anyway.
Amazon got the same treatment. Andy Jassy told analysts AWS can become a $1 trillion revenue business in five to ten years. He confirmed roughly $220 billion in 2026 capex and promised to spend “meaningfully more” in 2027, per TechCrunch. AWS grew 39%. The stock dropped about 6% after hours. Microsoft’s way of protecting the infrastructure budget is to shrink headcount: The Verge reports voluntary buyouts are on offer to around 60,000 employees.
Investors aren’t doubting the demand. They’re asking who gets paid back first, and today the answer looked like “not shareholders, not yet.”
Thrive skips selling software and buys the customer#
Thrive Holdings, spun out of Josh Kushner’s Thrive Capital and backed by OpenAI, is buying ADT for $10 a share, an enterprise value of $15.3 billion, TechCrunch reports. OpenAI sends models and forward-deployed engineers into the more than 75 companies Thrive already holds. Nobody has announced a single AI security product yet.
That’s the interesting part. Selling an AI SaaS seat to a home-security company means fighting through procurement. Buying the company means you own the call centers, the installers and the customer relationship, and the AI margin goes to you. It’s private equity with a model provider in the back office. If it works, expect more deals that look like this.
Reddit and Deepgram fight over the plumbing#
Reddit is moving to a mostly logged-in experience. Tests start this quarter in select markets outside the US, per TechCrunch. The reasons it gives are scraping protection, licensing leverage with model providers, and better ad targeting. The pressure behind it is clear: search referrals fell by more than a million a week on average in October, and US daily actives slipped to 50.7 million. When AI answers stop sending traffic, your data is the only thing left worth protecting.
Deepgram is going the other way and buying open infrastructure. It’s acquiring Daily, the company behind Pipecat, the open-source framework many of us use to wire speech-to-text, an LLM and text-to-speech into one low-latency voice agent, TechCrunch reports. Daily co-founder Kwindla Kramer says it will stay open and vendor-neutral:
“Scott was super committed to Pipecat being open source and vendor-neutral forever.” — Kwindla Kramer, Daily
Forever is a long time in an acquisition. If your voice stack runs on Pipecat, nothing breaks this week. But keep an eye on which defaults change, starting with Deepgram’s new Real-Time Context Standard.
Two smaller items#
Gemini can now generate and edit Docs, Sheets, slides, PDFs, Markdown and code inside the chat, for all users. PDF editing works on mobile; everything else is web-only for now.
Trump declared foreign factory-robot imports a national-security threat, ordering trade talks and threatening tariffs. That’s a problem for GM, Ford, Tesla and Boeing, which rely on those robots.
My bet: within a year, the companies people talk about will be the ones that own something scarce, whether that’s weights you can run yourself, users who have to log in, or customers you bought outright, and not the ones renting the most GPUs. I’m pulling ML4 down this weekend to test whether its 82,000-token habit holds up on our own agent traces. If it does, the price of serving a self-hosted European model just got very hard to argue with.