Lambda shops $3 billion. Sports is the AI hedge.
Nvidia-backed Lambda is in talks for as much as $3 billion at $12 billion-plus. Chris Isidore says the sports-team buying spree is an AI-proof hedge.
Nvidia-backed Lambda, the neocloud that rents chips and other AI infrastructure, is in talks for as much as $3 billion at a valuation of as much as $12 billion or more, and the same tape has Chris Isidore at CNN Business arguing that the people who made those fortunes are buying sports teams as a hedge against the technology they built.
1. Lambda talks $3 billion at $12 billion-plus
Bloomberg reports that Lambda, the Nvidia-backed neocloud that rents chips and other AI infrastructure, is in talks to raise as much as $3 billion in a round that could tee the company up for a public listing next year. Those talks are still open, and Bloomberg is explicit that the terms have not been finalized.
A Bloomberg rewrite of the same wire, not the original body, says Lambda has already received multiple term sheets, expects 2026 revenue of more than $1.5 billion, and raised more than $1.5 billion last November in a TWG Global-led round whose backers include Andra Capital, Scott Hassan’s SGW family office, OpenAI co-founder Andrej Karpathy, Cathie Wood’s ARK Invest, and Nvidia. That rewrite is why the print matters this morning: a neocloud with a $1.5 billion revenue run-rate is shopping a pre-IPO without a named lead or a closed check, so the move is to watch the term sheets rather than assume the round is done.
2. Sports teams are the AI hedge
Chris Isidore at CNN Business writes that the sudden buying spree in sports franchises is not only scarce trophies meeting more billionaires, and that the surprising reason the experts give is artificial intelligence — live sports as a relatively AI-proof investment that is not likely to be upended the way other assets might be. The gold-rush print in that same story is specific: former Disney chief Bob Iger and venture investor Josh Kushner agreed this month to a controlling Lakers stake at a record $12.5 billion, the NFL is closer to a pending $9.6 billion Seahawks sale, MLB approved a $3.9 billion Padres sale, the Timberwolves and WNBA Lynx sold Friday at $4.5 billion, the Yankees took a $2.6 billion Apollo Global Management minority injection, and Fenway Sports Group sold 40 percent of Liverpool to a consortium that includes Amazon founder Jeff Bezos.
Sal Galatioto, the investment banker who has spent thirty years selling teams and says he has never been busier, tells Isidore he would bet the Yankees are still here in a hundred years compared with IBM, and he now describes a franchise as a hedge against technology disruption rather than a cash-return play. Victor Matheson, the Holy Cross economist who studies sports business, says utilities may also be AI-proof but that sports is sexier because no one dreams of running National Grid the way they dream of owning the Yankees, and he adds that legalized betting could mean as much as a billion dollars a year handed to teams and leagues.
3. Hugging Face ARR jumps 50 percent to $150 million
Laura Bratton and Stephanie Palazzolo at The Information report that Hugging Face, the decade-old open-model hub where developers host and share models, has seen annualized recurring revenue, the last twelve months of subscription-like sales expressed as a yearly run rate, jump 50 percent to $150 million. The Information’s public teaser says the company is nearing a deal to sell itself, which is why a $150 million run-rate now reads as a sale process rather than a growth update, and why the move is to wait for a buyer and a price that The Information has not printed.
4. Meta’s Hatch agent lands in weeks; Watermelon in October
Jyoti Mann at The Information, citing internal documents, reports that Meta Platforms plans to launch Hatch, its consumer version of an OpenClaw-style AI agent — software that can take actions for you rather than only answer a chat box — as soon as the next several weeks, and that it is targeting October for Watermelon, its latest AI model. Hatch sits inside Mark Zuckerberg’s push to monetize Meta’s enormous AI investments and diversify revenue beyond advertising, and The Information says Meta has considered a tiered pricing system, though the public teaser cuts off before any sticker price.
5. Generalist raises about $200 million, led by 8VC
Dan Primack at Axios reports that Generalist, the robotics-AI startup that builds models rather than robot hardware, has quietly raised around $200 million just two months after a $400 million round, with 8VC leading and other existing investors joining. Axios does not print a new specific valuation except that it sits above the $2 billion mark from June, and the company, founded by alumni of Google DeepMind and Boston Dynamics, listed the round in a federal filing after not responding to Axios.
The backers already on the cap table include Nvidia, Bezos Expeditions, Radical Ventures, Union Square Ventures, Hanabi, Spark Capital, and Fei-Fei Li, which is why another $200 million ten weeks after $400 million still reads as physical-AI cash clearing. Those checks belong to Generalist, the model lab, and not to General Intuition, a separate company.
6. Unitree slumps 45 percent from a $66 billion debut
Reuters reports that shares of Unitree, China’s best-known humanoid-robot maker and a competitor to Tesla and Hyundai-owned Boston Dynamics, have slumped about 45 percent since a more than fivefold jump on its Shanghai debut, a slide that has raised bubble fears, retail-investor losses, and questions about the IPO system. The valuation soared to $66 billion at one point and later plunged by about $30 billion, after the stock finished up 460 percent on day one versus a 226 percent average first-day gain for newly listed China stocks over three years, then spent three consecutive sessions falling before steadying Tuesday.
The prospectus in that same wire shows adjusted net profit for the first three months of 2026 fell 53 percent to 40 million yuan, about $5.95 million, which is the earnings print sitting under the $66 billion peak. Dong Baozhen of Lingtong Shengtai said all bubbles are doomed to burst, which is the tight version of the fear Reuters is naming.
7. Chinese state-linked groups use DeepSeek as attacks double
Mark Anderson at Bloomberg, citing TeamT5, a Taiwanese research firm, reports that Chinese state-affiliated cyber groups have more than doubled their attacks after handing mundane tasks to AI and using it to develop more advanced malware, with DeepSeek as the model of choice. Charles Li, TeamT5’s chief analyst, says DeepSeek wins because it is relatively powerful with very low cyber guardrails, while Kimi K3 is more powerful but has no recorded incident in this research and is believed to be prohibitively expensive to run.
The named groups in that reporting include Grimfengxi, Huapi, and Teleboyi, which is enough to see the pattern TeamT5 is describing. That pattern is the cheap attack surface of open-weight models with weak refusal layers, not a side note to the lab race.
Watch
Nate B Jones — Stripe paid $7.5 billion for OpenRouter
Nate B Jones walks through Stripe’s reported $7.5 billion acquisition of OpenRouter, the routing layer that lets a developer send one request across more than 400 models from 80-plus providers. He treats the check as a measurement rather than a mood, because OpenRouter was marked at $1.3 billion in May, weekly token volume has doubled about every eleven weeks for three years, and the week of August 10 closed at 75 trillion tokens with the next week pacing above 87 trillion.
In the same investor letter, Stripe told limited partners that what it calls the singularity — agents doing real economic work, not just answering chat boxes — began on January 1, 2026, which is why Nate says a payments company that does not sell GPUs paid five times last quarter’s price. He describes a Stripe demo in which one agent reviewed public APIs, deployed an app, provisioned hosting and billing from the command line, and then charged another agent two dollars for a review, which is the agent-to-agent commerce stack Stripe is assembling around OpenRouter.
That is why the scarce move is no longer waiting for a GPU narrative, and is instead noticing which of your own curves have stopped behaving like 2025. The one move is to pick a workflow you actually know, assume a tiny team can now buy intelligence and sell to both people and agents, and ship a thin product against that workflow before an unmapped competitor shows up.
Matthew Berman — Thibault Sottiaux on Codex and the next personal agent
Matthew Berman is joined by Thibault Sottiaux, known as Tibo, the OpenAI product leader behind Codex, for a deep dive into the next wave of personal AI agents. Tibo, who previously built research infrastructure at DeepMind, starts with the year Google had an internal LM Chat, a large-language-model chat product, roughly twelve months before ChatGPT and was too nervous to ship it, which is the cultural lesson he took to OpenAI: research and product sit together, and you disrupt yourself even when a cash cow is printing.
He says Codex has reached 20 million users, ChatGPT and Codex are merging because future models want one multimodal, voice-first harness, and the end state is a personal AGI that tailors itself to you and your mom without a product dropdown. Ultra-fast inference, stated around 10 to 14 times prior “fast” speeds, flips the bottleneck from waiting 30 to 45 minutes on a swarm of agents to staying in flow with a few concurrent threads, while recursive self-improvement is already showing up as frontier models rewriting the inference stack so Luna, OpenAI’s cheaper model tier, could drop about 80 percent in price.
The one move is to run one real workstream through the merged Codex and ChatGPT path this week, brief it by voice or dictation the way Tibo does in the morning, and notice whether you still want fifteen parallel agents or three that keep you in the loop.
How I AI — Ryan Carson manages 15 agents as a solo founder
Claire Vo is joined by Ryan Carson, the solo founder of Untangle, a B2B platform for family-law firms, for a deep dive into managing about fifteen cloud coding agents around the clock. Carson, a five-time founder who previously co-founded the coding school Treehouse, lives in Devin, Cognition’s cloud software engineer: ten to fifteen threads sorted into priority folders, a paper list of weekly P0s — the must-ship bugs and features — and a Watchdog playbook, a repeating skill that inspects every law-firm account for errors, Sentry noise, and whether the latest of his roughly forty daily pull requests already fixed the issue.
He spent as much as $20,000 in a month on Devin before Cognition covered credits, still reaches for OpenAI’s Codex when a front-end feature needs low-latency pair programming at his machine, and closes the loop with a Land PR playbook that runs a Devin review, records a narrated browser walkthrough, and merges only after he signs the video. The product-market-fit story is the anti-hype, because consumers did not want an AI divorce app, so he sat with lawyer Renee Bower, learned discovery was the nightmare, and rebuilt for firms, which is why both hosts say more generated code is the wrong goal.
The one move is to pick one cloud agent, give it a named folder and a weekly P0 list on paper, and write a Watchdog-style skill that returns the top three problems per customer instead of a dump of logs.