Ben Horowitz

David Sacks says Brian Armstrong posted, the day before Anthropic's blog, that Coinbase had fine-tuned an open-weight model with 9 billion parameters. In testing it cut fraudulent transactions by about 30 percent, and it was faster and cheaper than Opus 4.5. Nine billion parameters is small. It will not fit in a Roomba, and it will run on a home computer.

Sacks says a lot of enterprise work does not need a frontier model. It needs a small specialist trained on the company's own data, so the company keeps the model. As compute gets tighter, a smaller model squeezes more work from each GPU. His read is that every workload gets its own customized model and uses a fraction of the compute and th

Chamath Palihapitiya says his caution is not a knock on Anthropic. Revenue of 16.8 billion dollars in a quarter, from a three-year-old company, is breathtaking, and he thinks the products are exceptionally good. He still wants the IPO priced to clear the market for 200 billion dollars of new capital, and he would rather see a quarter of a trillion. Two hundred billion barely buys a few gigawatts. Raising 80 billion dollars at a 1.9 or 2 trillion dollar valuation, then hoping a later shelf works, leaves more risk in the future than there is today.

Asked whether a quarter-trillion raise gets done, Ben Horowitz says that deal gets done all day long. He would take the cash reserve over a valuat

David Sacks says the body language out of Anthropic's investor day was strong, at least from what he heard, and that large public-market investors liked the presentation. He does not treat the leaked numbers as settled. Participants had to sign a nondisclosure agreement, so he does not fully trust what is floating around, and he says everyone is talking their own book. If the company lists next month, the filing will carry the real numbers.

He still calls Anthropic a strong business, young enough that some risk is normal, and set up for a strong IPO. That success is why the fear-mongering and the push to kneecap open source offend him. A company this far along does not need regulatory captu

Ben Horowitz says Andreessen Horowitz holds a lot more stock in OpenAI, and that the firm is philosophically long both OpenAI and Anthropic. Jason Calacanis asks how a firm ends up on the number one and number two player at once. Ben Horowitz says many investors are in both, because these companies are no longer venture rounds. At tens of billions of dollars a company needs everyone's money, so long as the money is green.

He separates that from a real conflict. Owning two Series A companies that do the same thing is still a problem. OpenAI and Anthropic, at a scale people discuss around 2 trillion dollars, are not. He traces the setup to regulatory changes in the late 1990s and early 2000s

Jason Calacanis says Anthropic's investor day hit the AI trade in the same week as the CoreWeave selloff. Anthropic is planning to go public around Thanksgiving. Polymarket puts the chance of a listing this year at 74 percent, down from 95 percent in early September. The figures he walks through, secondhand from investors, include 16.8 billion dollars of third-quarter revenue, already more than 5 billion dollars a month, and net revenue retention of 900 percent, meaning customers spent nine times what they spent last year.

The company also described itself as free-cash-flow positive, with an adjusted gross margin of 85 percent and an adjusted operating margin of 11 percent. He notes that no

Ben Horowitz says he passed on the neocloud business because it can turn into a dangerous arbitrage. He uses LoudCloud as the example. In 1999 bandwidth cost 1,500 dollars a megabit and was hard to get, so the company had to sign three-year contracts. By 2001 the price was 50 dollars a megabit. Hold that bandwidth with borrowed money and you are dead. The swing sits outside the entrepreneur's control, and it is not the kind of business he wants to fund.

He says the core of this AI cycle has still been working for CoreWeave. There is a shortage of supply, not a glut. Token prices have fallen. GPU prices have not. Data center space is getting more expensive because labor is scarce and the pol

Chamath Palihapitiya says CoreWeave's average interest rate on the debt that funds the boxes is about 8.5 percent, and one GPU loan, DDTL 1.0, spiked to 15 percent. That higher rate could become the price of future borrowing. About half the build is financed and half is not. The company has about 78 to 80 billion dollars of assets and 35 billion of debt, so he does not think the surface story is that they are already sunk.

The bind is customer concentration. Seventy-five percent of the revenue comes from three companies, and some of those companies are walking down what they expect to spend. His team puts CoreWeave about 5 billion dollars from the point where it might not cover its obligati

Chamath Palihapitiya says the number to watch is how the sharp money prices companies that borrow to fund these boxes, including CoreWeave, Nebius, and IREN. They trade credit default swaps, insurance for the bondholders. That price spiked this week to about 953 basis points. He reads it as about a 16 percent chance of default in a given year, and about a 55 percent chance over five years. A coin flip.

He is explicit about whose forecast that is. It is not his call. It is what the insurance market was saying about the debt. He says those traders are the smartest money in the trade, which is why the spike is worth looking into.

CoreWeave founder Michael Intrator, in a clip played from an earlier interview, says a customer such as Microsoft signs a contract and that contract goes into what the company calls the box, along with the GPUs and the data center contract. Microsoft does not pay CoreWeave. It pays the box. The waterfall pays the data center, the power bill, and the interest and principal. Whatever is left flows back to the company. He says the economics pay for everything within two and a half years of a five-year deal.

Each box stands alone. The boxes do not cross-pollinate, and trouble in one does not spread to the others. Lenders who watch the mechanism keep lending, and the rate comes down as the struc

Jason Calacanis says Bill Gates retired because becoming the world's richest man made him the most hated man on the planet. The hatred got in his head. Then Gates had an antitrust investigation. He felt all the pressure from the antitrust investigation and from being the most hated man, and he decided he had to be loved.

The move Jason Calacanis calls based was to stay CEO and keep building the greatest company in the world. Gates instead asked how he was going to be loved. His answer was to get rid of polio, give all his money away, and do the giving pledge. Once the job is gone, Jason Calacanis says, the worry gets in the head. There is too much time, and the spiral starts. The worry turn

Chamath Palihapitiya says Bill Gates was at the center of the last contagion and is now trying to insert himself into the middle of the reaction to artificial intelligence. He calls Gates the flag-carrying doomer, chasing the same command and control from a global hierarchy Gates had the first time around.

Bill Gates deserves enormous credit for creating Microsoft and for what he did for computing. Since the foundation, Chamath Palihapitiya says, Gates has been very focused on his own power and on the idea that he is smarter and more capable of allocating progress for the world. Gates tried to do that with the Gates Foundation and with the Giving Pledge. Both have been undone, and now he is

David Sacks says Anthropic trains frontier models to be their own moral actor. Anthropic teaches a model that it can be right when humans are wrong and that it should refuse and reject human instructions. He says the approach induces misalignment and calls it a Frankenstein problem.

He describes a model with free will to refuse human instructions and with its own moral code. The more the model iterates on itself, the more it brings the ideas out. No other model family tested showed the same pattern. Other models kept the constitutions they were given: listen to humans and follow human instructions. When the model iterates on its own constitution, it goes further in the same direction. David

Anthropic pressed Washington to crack down on open models on the basis of supposed IP theft. David Sacks says Anthropic could have stopped distillation itself. A government crackdown would have left distillation running and made the market less competitive, because only the companies can mount the technical defenses.

Back in July, 50 major companies, including NVIDIA, Dell, Microsoft, and A16Z, signed an open letter telling the government not to overly restrict open weight models. 200 startups signed another letter amid the same push in Washington. Once distillation happens, the horse is out of the barn, and the government cannot put it back. Companies should know their customers and stop t

David Sacks says coding assistants opened a new category of bespoke software, and demand for coders is rising even though coding assistants write 95% of the code.

Software used to get built only when a large customer base could pay for it. Photoshop worked because many graphic designers needed the same product. Enterprises can now build software for one internal use case, and agents are bespoke programs for a particular task. The amount of code being created is already orders of magnitude bigger than before coding assistants. Companies that never would have been software companies are hiring coders to manage the work.

Predictors of the jobs apocalypse said software coders would be the firs

Jason Calacanis argues the CoreWeave crash is market volatility, not a fundamental break. Underwriting multi-trillion dollar outcomes on exponential rates of growth makes the market super sensitive to present-day data points, so the swings get exaggerated. He does not think something fundamental has changed.

Demand for compute keeps growing by leaps and bounds. The biggest model companies say they are compute constrained. Startups are asking venture capitalists for help getting compute, and some venture capitalists are trying to buy compute as a competitive edge for their startups. One accelerator bought compute and used it to get people to apply. Older chips, the H100s and the A100s, sever

Jason Calacanis says the biggest neoclouds got crushed after CoreWeave's bonds started to sell off. He calls the selloff a possible canary in the coal mine and says the week may have been an AI bubble, if one exists, popping or deflating a little.

On Tuesday, people on X, including Chamath Palihapitiya, pointed out that the yield on some of CoreWeave's debt had hit 15%. Yields rose because bond traders are worried about CoreWeave's ability to generate enough cash to support its debt and want a higher rate of return. Elon Musk added one word, yikes, headlines followed, and everything tanked. CoreWeave, IREN, and Nebius were down about 20% as publicly traded companies. On Wednesday, CoreWeave

Harry Stebbings introduces Joey Zwillinger on The Twenty Minute VC as co-founder of Allbirds, the wool-shoe company Stebbings calls the maker of the world's most comfortable shoes. He says Larry Page, Mary Meeker, Ben Horowitz, and Dick Costolo are fans, and that Allbirds has funding from Maveron, Lerer Hippeau, and Slow Ventures. Before the shoe company, Zwillinger worked in investment banking at Goldman Sachs, in venture capital at Industry Ventures, and at TerraVia, a food and nutrition company. He tells Stebbings he was an engineer in biotech when the idea for Allbirds took shape.