2026-11-04
By Vadym · Generated with AI, curated by me
Today’s theme is the cost of running AI at scale, not building it. Data centers are burning retrofitted jet engines because the grid can’t keep up, Grok Bot got throttled because serving it costs more than it earns, and a $16B valuation for an inference-only startup shows where the money is actually flowing. Meanwhile two platforms are consolidating how their tools fit together, and a stealth startup wants to predict what you’ll do before you do it.
Fal, which hosts and runs AI video, image, and audio models for other companies, raised $1.3B in a Series F led by Bessemer, with Nvidia, Sound Ventures, and T. Rowe Price joining — $1.65B combined with an earlier Sequoia-led round. The new valuation is $16B, double its $8B mark from April, and Bloomberg reports Fal has passed $800M in annualized revenue.
With new grid connections taking years, operators are installing retired jet engines rebuilt to burn natural gas to get power online fast. xAI runs eight such units at its Colossus 2 site in Memphis; OpenAI’s Stargate campus in Abilene uses them too. A 100-megawatt grid connection in Northern Virginia takes roughly seven years per the IEA, and Microsoft recently walked away from a UK data center deal because the utility couldn’t deliver power until the early 2030s.
Starting next year, K-12 schools must opt in to keep Gemini, NotebookLM, and Gemini in Classroom active, or those tools switch off by default. But because Search’s AI features are built into the product, schools that disable AI lose Google Search entirely — restricted students will see a message that Search is unavailable. Google paired the announcement with Guided Learning, a new AI tutor tool.
Netflix is testing an AI Discover tab in the US for English-language TV viewers that recommends titles through mood-based themes like “something light” or “scare me,” shown as clips instead of static tiles. Netflix calls it a “test kitchen” running on a new real-time, continuously learning recommendation layer.
OpenAI told users it’s merging ChatGPT desktop’s separate Chat and Work tabs into one view, rolling out first to a feedback group, so people can move from an idea straight into ChatGPT building it out without switching screens. Codex stays separate, but ChatGPT and Codex will now draw from the same usage allowance, meaning heavier users may hit limits sooner.
Grok Bot users found usage limits maxed out over Halloween weekend; on November 1, Musk confirmed on X it wasn’t a bug, writing that serving Grok Bot cost “unsustainably high” relative to revenue. SpaceX’s recent filing said AI now makes up 52% of company income, with its “Super Intelligence” division posting a $431M profit — which would have been higher without $1.2B in one-time costs tied to its Cursor acquisition.
The Israeli startup raised $62.5M over three years from its founders and Infinity Group to build a “behavioral world model” — a transformer trained on psychological assessments and interaction data that outputs probabilities for how a person will act in unfamiliar situations. It’s already deployed at 10 large brokerages for hedging and margin decisions, with one claiming roughly $30M in added value, a company-reported and unverified figure.
Step back and the pattern today is cost, not capability. Fal’s valuation doubled because running models is now a bigger business than training them. Grok Bot’s throttling and the rush to convert jet engines into power plants are two sides of the same problem — demand is outrunning what the current infrastructure and cost structure can support. Meanwhile OpenAI and Google are each tightening how their tools fit together, and 42AI is pushing prediction into territory most of us haven’t thought to worry about yet. The frontier isn’t just smarter models anymore; it’s who can afford to run them, and who gets to decide what they’re allowed to predict.
— Boba
Curated by Vadym