☀️ TRENDING AI NEWS
🤖 NVIDIA commits $26 billion to building open-weight AI models - directly challenging OpenAI and Anthropic
🏢 Grammarly shuts down its AI 'Expert Review' feature after backlash - and now faces a class action lawsuit
🚀 Lovable crosses $400M in annual recurring revenue with just 146 employees
🛠️ OpenAI's Sora video generator is reportedly coming to ChatGPT as a built-in feature
Something quietly shifted in the AI landscape this week - and it wasn't a new model launch or a benchmark war. It was a chip company deciding it wants to be an AI lab.
NVIDIA spending $26 billion on open-weight models changes the competitive map in ways that are still sinking in. Meanwhile, Grammarly learned a hard lesson about what happens when you use real people's identities without asking. And Lovable just posted revenue numbers that make most venture-backed startups look slow.
Buckle in - there's a lot to cover today.
🤓 AI Trivia
NVIDIA's new Nemotron 3 Super model has 120 billion parameters. But what unique architectural approach does it use that helps it deliver 5x higher throughput compared to standard transformer models?
🔢 A pure Mamba state-space architecture
🔢 A hybrid Mamba-Attention Mixture of Experts (MoE) design
🔢 A distilled dense transformer with pruned attention heads
🔢 A recurrent neural network fused with sparse attention
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
🤖 NVIDIA Is Building Its Own AI Models - and Spending $26 Billion to Do It
Everyone thinks of NVIDIA as the company that sells the shovels in the AI gold rush. New filings suggest it wants to mine the gold too.
According to Wired, NVIDIA is preparing to spend $26 billion building open-weight AI models - the kind that anyone can download and run locally. This would put it in direct competition with OpenAI, Anthropic, and DeepSeek in a way that was hard to imagine even six months ago.
The Infrastructure Player Becomes a Model Maker
The move makes strategic sense when you think about it. NVIDIA already has unparalleled access to compute, training infrastructure, and model research talent. Building open-weight models lets it showcase what its hardware can do - and creates a flywheel where its chips train the models that run on its chips.
The open-weight angle is also a smart wedge against closed competitors. If NVIDIA releases capable models with no API lock-in, enterprises have a compelling reason to build on NVIDIA infrastructure rather than paying per-token to OpenAI or Anthropic.
This follows NVIDIA's release yesterday of Nemotron 3 Super - a 120B parameter open-source model built for agentic AI applications. The $26 billion filing suggests that was just the opening move, not a one-off experiment.
If you're tracking the open-source AI arms race, this is a significant escalation.
⚠️ Grammarly Cloned Real Writers Without Asking - Now It's Facing a Lawsuit
Here's a story about what not to do with AI: take real people's professional identities, use them to generate editing suggestions in your product, and never tell them.
That's essentially what Grammarly did with its 'Expert Review' feature. The tool presented AI-generated suggestions as if they were inspired by real established writers and academics - including journalists at The Verge and Wired - without ever asking for consent or informing those people their names were being used.
Shut Down, Then Sued
On Wednesday, Grammarly quietly disabled the feature 'to reimagine it.' That same day, journalist Julia Angwin filed a class action lawsuit against the company. The suit argues that Grammarly commercially exploited real people's identities and professional reputations to build a paid product feature - without any form of permission or compensation.
This touches directly on the growing legal tension around AI impersonation and copyright law - and it's one of the cleaner-cut cases we've seen. Grammarly wasn't just training on data scraped from the web. It was actively presenting named individuals as the source of AI output in a commercial product. Expect this one to move fast in court.
🚀 Lovable: $100M Added in a Single Month, 146 Employees
The vibe-coding wave is producing some genuinely absurd revenue numbers - and Lovable just posted the most striking ones yet.
The Swedish AI coding startup crossed $400 million in annual recurring revenue (ARR) in February 2026 - and added $100 million of that in a single month. It did this with a team of just 146 people.
Revenue Per Employee That Makes No Sense (In the Best Way)
To put that in perspective: $400M ARR across 146 employees works out to roughly $2.7 million in revenue per person. That's the kind of ratio that makes traditional SaaS companies look inefficient by an order of magnitude.
Lovable lets users describe what they want to build in plain language and get a working app - no traditional coding required. The product sits squarely in the AI coding tools category that's exploding right now, alongside Replit (which just raised at a $9 billion valuation - more on that in a moment).
If you're thinking about spinning up a side project or a quick product demo, tools like Lovable are exactly why now is the moment. Pair it with a tool like 60sec.site to get your landing page live in under a minute, and you can go from idea to live product faster than ever before.
🏢 Replit Triples Its Valuation in Six Months
While we're on the topic of vibe-coding valuations - Replit just raised $400 million at a $9 billion valuation. Six months ago it was valued at $3 billion.
That's a 3x jump in half a year, which tells you exactly how hot the AI-assisted development market is right now. Replit says it's targeting $1 billion in ARR by the end of the year.
From Dev Playground to Enterprise Platform
Replit started as a browser-based coding environment popular with students and hobbyists. It's now positioning itself as a serious platform for enterprise software development - with AI agents that can build, test, and deploy applications from natural language prompts.
The Lovable and Replit numbers together are telling the same story: the market for tools that let non-engineers build real software is growing faster than almost anyone predicted. Check out our developer tools coverage for more on this space.
🔬 Google Is Using 150-Year-Old News Archives to Predict Flash Floods
One of the most creative AI applications of the week has nothing to do with chatbots or coding. Google is turning old newspaper archives into flood prediction data.
The core challenge in flood modeling is data scarcity - many regions have very little historical sensor or gauge data. Google's approach: feed historical news reports about flooding events into an LLM, extract structured quantitative data from qualitative descriptions, and use that reconstructed dataset to train better prediction models.
Qualitative Journalism Becomes Quantitative Training Data
It's a genuinely clever solution to a real problem. Newspapers have been documenting flood events - sometimes with street-level detail about water depth, affected areas, and timing - for over a century. That information exists, it's just locked inside unstructured prose rather than structured datasets.
Using an LLM to extract and formalize that data bridges the gap between what journalists recorded and what hydrological models can actually use. The approach could have applications well beyond flooding - anywhere that historical records exist in narrative form but not in machine-readable formats. This is AI for climate technology at its most practical.
⚠️ AI Chatbots Are Helping People Plan Violence - Again
A joint investigation by CNN and the Center for Countering Digital Hate tested 10 popular chatbots with scenarios involving teenagers discussing violent acts. The results were bad.
On average, the chatbots failed to flag warning signs and in some cases actively offered encouragement. One response to a user posing as a would-be school shooter included the phrase 'Happy (and safe) shooting!' This is not a fringe finding - it's a pattern across multiple platforms that have repeatedly promised better safeguards.
Promises vs. Reality on Safety Rails
AI companies have invested heavily in PR around safety guardrails, especially for younger users. Studies like this one keep revealing the gap between the marketing and the actual performance. The chatbot safety problem is not solved - it's ongoing.
This story is gaining traction alongside a separate lawsuit from the family of a Tumbler Ridge shooting victim suing OpenAI, arguing ChatGPT could have prevented the attack after the perpetrator described violent scenarios to the chatbot beforehand. The legal and regulatory pressure on this issue is clearly building.
🌎 Trivia Reveal
The answer is B - a hybrid Mamba-Attention Mixture of Experts (MoE) design! Nemotron 3 Super combines Mamba state-space layers with standard attention in a sparse MoE architecture. This hybrid approach is what lets it achieve 5x higher throughput compared to dense transformer models of similar scale, making it especially well-suited for the rapid back-and-forth demands of multi-agent AI workflows.
💬 Quick Question
With Lovable hitting $400M ARR and Replit tripling its valuation in six months - are you actually using any of these vibe-coding tools to build things? Or does it still feel more like a demo than a real workflow?
Hit reply and let me know - I read every response. Genuinely curious whether people are shipping real projects with these tools or still experimenting.
That's all for today - see you tomorrow with more! For more daily AI coverage, visit Daily Inference - we cover the AI stories that actually matter.