Harvard Business School will now let you pitch your startup to a synthetic venture capitalist for $699. Anthropic is lining up what could be the biggest IPO in American history. Nvidia bought a company without buying it. And Meta remembered it used to give things away.
An eight-week HBS bootcamp replaces the hardest part of business school — the twentieth practice pitch — with an AI avatar. Whether that’s the valuable half is the open question. Plus Anthropic’s run at the largest IPO in history, Nvidia’s third licensing deal that isn’t an acquisition, and what Meta open-sourced versus what it kept.
For more constantly update AI News check out the totally non-sentient SKYNet Tribune.
Harvard’s AI clones will hear your pitch now
The most talked-about AI story of the weekend is an eight-week online course called Foundry, run by Harvard Business School, in which aspiring founders practice their pitches in front of AI avatars of actual HBS instructors. It costs $699. An HBS MBA costs north of $75,000 a year, and that’s before you count two years of not building anything.
The avatars were built by HeyGen, the video-synthesis startup, and they aren’t canned lecture recordings. Students pitch to them live, sit through simulated board meetings with them, and get feedback that adapts to what they actually said. Jeff Bussgang — Flybridge Capital co-founder, longtime HBS senior lecturer, and one of the instructors who agreed to be duplicated — described watching his digital self as somewhat “creepy,” then added the line that explains why this is happening anyway: “My students love it.”
It’s worth being precise about what Foundry is and isn’t, because “Harvard replaces professors with AI” is the version of this story that will be circulating by Wednesday and it isn’t quite right. Foundry pairs the avatars with weekly live sessions from real humans. The faculty behind it are real and specific: Shikhar Ghosh, who has founded eight technology companies; Christina Wallace, the storytelling specialist and Broadway producer; Jim Matheson, the deep-tech VC. Nobody has been laid off and replaced by a rendering. The avatars are doing the thing that is genuinely hard to scale — the twentieth practice pitch at 11 p.m., the one a busy senior lecturer is never going to sit through.
That’s also the part worth sitting with. The reason elite business education is expensive isn’t the lectures, which have been free on the internet for fifteen years. It’s the repetitions with someone whose judgment you trust, plus the alumni network. Foundry is a bet that the first of those two can be synthesized. If it can, HBS can serve an effectively unlimited number of students without hiring a single additional faculty member — a proposition that Stanford, Wharton, and every other school with a brand to monetize is watching very closely.
The early evidence is mixed in an instructive way. Katharina Rings, the project director, originally imagined a free-form chatbot experience and pulled back toward something more structured after student feedback — which tracks with what most people building AI tutors have found, namely that unlimited open-ended conversation with a model is less useful than a guided path with a model in it. And a New York Times reporter who tested the system pitched Bussgang’s avatar on “Uber for bananas” and reported a “noticeably frozen smile” throughout. The avatar was not impressed by the idea, which is at least evidence the feedback isn’t purely flattering.
The open question is the one Harvard can’t answer with better rendering: whether finishing an AI-taught bootcamp signals anything to anyone. A Harvard MBA is valuable partly because it’s scarce and partly because of who else is in the room. A $699 course that anyone can take is, by design, neither. Foundry may turn out to be a genuinely good product for founders and a genuinely bad brand extension for Harvard, and those two things aren’t in tension — they’re the same fact viewed from two directions.
Applications for the October cohort are open. There’s also a pilot running with Bentley University, which suggests HBS is already thinking about Foundry as something it licenses out rather than just something it sells direct.
Anthropic is going for the biggest IPO ever
Bloomberg reported Thursday that Anthropic expects its public offering to match or exceed SpaceX’s record — $75 billion raised, $86.2 billion once the overallotment was exercised. That would make it the largest IPO in history. The company filed confidentially with the SEC back in June and could file publicly as soon as the end of this month, which is to say possibly this week.
The numbers underneath it are the part that stops you. Anthropic reported second-quarter revenue above $11.5 billion, against $787 million in the same quarter a year earlier. That’s roughly fourteen-fold growth in twelve months, at a scale where growth like that isn’t supposed to be arithmetically available anymore. Annualized revenue reportedly reached about $65 billion by the end of July. The May Series H put the company at a $965 billion post-money valuation, and the IPO will test whether public markets will underwrite a number that starts with a nine.
Context for how much weight this one deal is carrying: 2026 IPO volume stood at $160.6 billion in mid-August, against the all-time record of $195.2 billion set in 2021. Anthropic alone would clear that record and then some.
Two smaller Anthropic items from the same stretch, both of which will matter more than they read right now. The company said earlier this month it will embed watermarks in Claude’s outputs worldwide rather than only in the EU — a compliance requirement turned into a global default, which is how a lot of EU AI regulation is going to end up applying to everyone. And Bloomberg reported Thursday that Anthropic is changing its data retention policy for advanced models. Worth watching what the actual terms say when they land.
Nvidia pays $6 billion for Poolside without buying Poolside
Nvidia is paying AI coding startup Poolside $6 billion for a non-exclusive license to its “Model Factory” — the internal system Poolside built for training frontier models — plus a $1 billion investment at a $12 billion pre-money valuation, plus job offers to 109 Poolside employees. The founders stay. The company keeps operating. Investors get paid out by the end of 2027. It is not an acquisition, not an acquihire, and not obviously anything else that has a name yet.
This is Nvidia’s third deal in this shape. Groq in December 2025, roughly $20 billion for inference technology and the senior people who built it — Nvidia’s largest transaction on record — with Groq subsequently raising $350 million at a $3.5 billion valuation and Nvidia participating in the round. Enfabrica, around $900 million for networking hardware and staff. Call it $27 billion across three transactions to absorb the capability without absorbing the cap table.
The less charitable reading has already been put in writing by two U.S. senators. Elizabeth Warren and Richard Blumenthal formally questioned whether the Groq structure was an attempt to sidestep antitrust review — you don’t file for merger clearance on a licensing agreement. Poolside is the third data point in what is starting to look less like opportunism and more like a documented playbook.
Poolside’s investor letter contains the most honest sentence anyone in AI has written this year: “We had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January. We didn’t close it in time.” That’s the whole industry compressed into two sentences. Frontier model training now requires clusters more than an order of magnitude larger than the ones that got you here, the compute is spoken for before you’ve raised against it, and if you miss the window you’re not a frontier lab anymore. Poolside Infrastructure Company continues separately, building a 1.2-gigawatt data center in Texas — the picks-and-shovels half of the business turned out to be the durable half.
Markets weren’t thrilled: NVDA finished the week about 5% lower. Earnings land Wednesday, which will tell us considerably more.
Meta open-sources again, selectively
Two weeks ago Meta released Muse Glimmer, a roughly 30-billion-parameter model under an Apache 2.0 license, built specifically to run agents on hardware people already own. It’s dense rather than mixture-of-experts, 52 layers with a ~1.8B vision encoder, 100+ languages, interleaved text and images, and a context window over 131,000 tokens. The 4-bit quantized variants were engineered to fit inside 24GB and 32GB memory budgets, which is to say an RTX 4090 or an M-series Mac. With speculative decoding Meta measures 233 tokens/second on an RTX 5090, up from 75.
On Meta’s own benchmarks it leads on agentic work — 75.5 on MCP Atlas, 51.2 on SWE-Bench Pro — while Alibaba’s Qwen3.6-27B stays ahead on several multimodal and terminal tasks. Take vendor benchmarks at vendor-benchmark value, but the shape is right: this is a model built to do multi-step work locally and offline, not to top a leaderboard.
Glimmer arrived alongside a Zuckerberg manifesto arguing that broadly distributed superintelligence “has the potential to begin a new era of personal empowerment,” and the model is the argument’s exhibit A: your data stays on your machine, the weights are yours, nobody meters your usage. It’s a real position, and after the proprietary Muse Spark launch in April it’s a reversal worth noting. Zuckerberg has since signaled that Muse Spark 1.2 weights are coming, which would make it the first genuinely frontier-tier American model released openly.
The obvious catch: Spark 1.2’s weights aren’t out, and Spark itself is closed. What Meta has open-sourced is the tier below the one it keeps. That’s still more than anyone else at its scale is doing, and it’s also not the same thing as the manifesto describes. Both can be true.
Sources
- Harvard’s $699 startup bootcamp offers AI avatars of its instructors — TechCrunch
- Harvard’s $699 bootcamp deploys AI avatars as startup coaches — TechBuzz
- Foundry — Harvard Business School
- Anthropic aims for largest-ever IPO — PYMNTS, on Bloomberg’s reporting
- Anthropic to watermark Claude’s output worldwide — Euronews
- Nvidia pays Poolside $6bn to license its model factory and hire 109 staff — TNW
- Poolside strikes $6 billion licensing deal with Nvidia — Newcomer
- Warren, Blumenthal question whether Nvidia’s $20B Groq deal avoids antitrust laws — U.S. Senate
- Nvidia buying Groq’s assets for about $20 billion — CNBC
- Meta returns to open source with Muse Glimmer — VentureBeat
- Meta’s new Glimmer model and Zuckerberg’s personal intelligence vision — TechCrunch