NVIDIA Spending $6B on an Open-Weight Model (But Why?)

0:00 On the 20th of August, a letter went out to a few dozen investors. Not a press release, a private letter. It said Nvidia had agreed to pay $6 billion, not to buy a company, to rent the machine

0:11 inside one. 6 billion for a license, a billion more for equity, and 109 people packing up their desks. Nvidia hasn't confirmed a word of it, and neither has a startup. Both of them declined to

0:23 comment. So, everything here traces back to that letter. And to Eric Newcomer, the reporter who got hold of it first. Here's what doesn't add up. Nvidia already gives AI models away and it

0:33 gives them away free. It's the largest publisher of openw weightight models on the planet with tens of millions of downloads. What exactly does $6 billion buy a company that's already doing the

0:43 thing for nothing? Not a model. I'll show you the leaderboard where Nvidia's own flagship loses to something a fifth its size. Nvidia didn't buy intelligence with that money. It bought a clock and

0:53 the receipt is completely public. And the reason it needed a clock has been running in Chinese data centers all year. Look at what Nvidia already owns because it reframes the question.

1:03 Neatron is its open model family. Weights, training recipes, and in most cases, the training data published for anyone to download. No key, no permission, no invoice. The company said

1:15 in March that Nematron had been pulled off hugging face more than 45 million times and that Nvidia had become the single largest organization on that site with close to 4,000 people on it. That's

1:26 not a chip company dipping a toe in open source. That's the biggest openweight publisher there is and it kept shipping all year. The flagship arrived in June. 550 billion parameters, 55 billion

1:38 switched on per token, a million token context window free for commercial use. On paper, it's the strongest open model any American company has released. Two months later, Nvidia added a small one

1:49 at the other end of the range. 30 billion parameters with three billion active sized to run on a single desktop. So, the models exist. The compute obviously exist because Nvidia

1:59 manufactures the compute. The researchers exist and that's why $6 billion is confusing and why the real question isn't why Nvidia is going open. It went open years ago. The question is

2:09 what it thinks it's missing. On the 21st of July, a startup called Poolside published a benchmark table. About 150 people work there, and it's run by Jason Warner, who used to be the chief

2:20 technology officer at GitHub. The model his team builds is called Lagona, and Laguna S 2.1 is 118 billion parameters with 8 billion active. On an agentic coding test called Terminal Bench, the

2:34 kind where a model drives a real shell and has to actually finish the job, it scores 70.2. Nvidia's flagship, same table, same test, scores 56.4, four and a half times the total parameters,

2:46 nearly seven times the active ones, and 13.8 points behind. One benchmark wouldn't mean much. This isn't one. On multilingual bug fixing, it's 78.5 against 67.7.

2:59 On a tool use test, 49.7 against 34.3. Same direction every time. Now, Poolside published that table, so read it the way you'd read anybody's homework about themselves. Two things make it hard to

3:11 wave away. Poolside says it took the highest number it could find for every rival, the vendor's own claim, the benchmark author's board, or a third party board, whichever flattered the

3:21 competitor most, and it published the full trajectory of every trial. So, you can go read what the models actually did. So, the table holds up. The number that explains the check isn't on it.

3:32 Lagona started pre-training on the 22nd of May on 4,096 Nvidia chips. It shipped 60 days later, under 9 weeks from the first token of training to public weights. 4,000 chips isn't a Frontier

3:44 cluster. That's a rounding error next to what Nvidia can point at on any given afternoon. And in the 12 weeks before that, the same small team had already shipped two other models. Poolside has a

3:54 name for how it does this. It calls the thing the model factory. And in its own words, it's the internal platform that lets them quote industrialize the model development process, maximize research

4:05 iteration speed, and minimize the attention researchers pay to bookkeeping and infrastructure. So that's what $6 billion bought, not weights, a non-exclusive license to somebody else's

4:15 research loop, plus 109 of the people who know how to turn the crank. It's worth sitting with the question that raises. If the bottleneck wasn't chips and Nvidia of all companies couldn't

4:25 bypass it with chips, what does that say about the rest of us buying chips? Now, the bigger why, because Nvidia doesn't sell models, it gives them away. Why does it care who wins a race it collects

4:35 no revenue from? Jensen Hang has said the quiet part out loud in about the plainest terms available. Free AI, he said in July, should be great for hardware. Free AI should be great for

4:46 chips. That's the textbook move. Make the thing next to yours free and people spend the savings on yours. But there's a sharper version underneath it. And that one is what $6 billion is really

4:56 about. Open router is a marketplace that routes developer traffic to whichever model you pick, which makes it about the cleanest public read there is on what people actually run. In June of last

5:06 year, American models, Google, OpenAI, and Anthropic combined took roughly 70% of the tokens flowing through it. A year later, that share is around 30. Chinese openweight models took the rest and

5:18 different reads of the same data land between 51 and 61% depending where you draw the boundary, but the direction isn't in dispute. The hub at HuggingFace reports about 41% of everything

5:30 downloaded from it over the trailing year went to Chinese models. Alibaba's twin family alone has passed a billion cumulative downloads and anchors more than 150,000 derivative repositories. A

5:42 download is a soft metric though. This next one isn't. In February, a Chinese lab shipped a Frontier Open model of 745 billion parameters trained end to end on Huawei Ascend chips. No Nvidia silicon

5:54 anywhere in it. Now, that's the whole video. So, let me be precise about it. Weights are just numbers in a file. The stack around them isn't. The kernels, the quantization formats, the serving

6:05 code, the fine-tuning recipes, all of it gets written for whatever hardware the model was born on. So, if the default open model on Earth is trained on Huawei and tuned for Huawei, then every

6:15 enterprise that adopts it inherits a software stack that doesn't need CUDA. Not this year. But an ecosystem is a ratchet and it turns one way. And that's the 6 billion. It isn't a model

6:26 purchase. Nvidia is buying back the right to be the default and paying about 8 days of revenue to do it. Which brings us to what this really costs and money is the least of it. 6 billion is 8 1/2

6:37 days of Nvidia's revenue. and with the equity 10. That's roughly a third of what the company spent on research and development across its entire last fiscal year. For a business carrying a

6:47 $5.2 trillion market capitalization, it's loose change. The real cost is the shape of the deal. Nvidia isn't buying Poolside. Poolside stays independent, keeps its founders, keeps the right to

6:59 license the same technology to anybody else, and walks away with a billion dollars of fresh equity. Nvidia takes the process and 109 of the 150 people who ran it. Nvidia has done this before

7:11 twice. In September last year, it paid over $900 million for a networking startups technology and staff. 3 months after that, it did the same to an inference chip rival for a reported 20

7:22 billion, license the technology, hire the team, leave the legal shell standing. A merger gets a merger review. Three licensing agreements don't. One analysis called the price for a

7:33 non-exclusive license quote deranged then made the sharper point underneath it. Regulators are allowed to look at substance rather than labels. And there's a second cost which Nvidia's own

7:44 customers will feel. Open AAI and Anthropic buy Nvidia's chips by the warehouse. Nvidia is now openly building models pitched as the cheap customizable alternative to theirs. A supplier can

7:55 climb into its buyer's business exactly once before those buyers start funding somebody else's silicon. One more thing needs correcting because a version of this story is circulating where Nvidia

8:05 goes open while Anthropic lobbies to ban open weights. That version is wrong. On the 24th of July, Jensen Hang made his first ever post on X, an open letter warning Washington against premature

8:17 restrictions on open weights. It launched with 25 signitories and passed 150 within days. Anthropic didn't sign it. Three days later, Dario Ammedday published Anthropic's position and its

8:29 opening claim is that Anthropic has never advocated for a ban on open weights models. What it wants is chip export controls, a crackdown on industrial scale distillation and

8:39 mandatory safety testing for capable models, open and closed alike. Amade also called open models without dangerous capabilities, quote, a public good. You can disagree with all of it,

8:51 and plenty of serious people do, but declining to sign a letter isn't lobbying for a ban. And if we're going to argue against a position, it should be one somebody actually holds. So,

9:01 here's where I land, and I'd say it at twice the price. Nvidia is right to do this. Most people get one thing wrong about open weights. They treat it as a values position. It isn't. It's a

9:12 distribution strategy. And Nvidia is the one company here whose business model makes giving away the best model a dominant move. Every other lab charges for tokens. Nvidia charges for the

9:22 machine underneath every token anybody generates anywhere. The evidence says it needed the help, too. Its own 550 billion parameter flagship sits nearly 14 points behind 118 billion model from

9:35 a company a fraction of its size. On the independent index kept by artificial analysis, that same flagship scores 38, while the leading Chinese open models sit in the 50s and 60s. So, buying the

9:46 loop that produced the smaller model is the cheapest correct answer on the table. Who wins? For the next year or so, you do. Two of the largest hardware and social companies in America are

9:57 racing to hand out near frontier weights so Chinese labs don't own the default. Free models get better, they get better faster, and they keep running on hardware you already have. The

10:06 concession is real, though. This deal is a merger with the merger review surgically removed, and it's the third one in under a year. If you think that's fine because the outcome looks good,

10:16 you've just agreed outcomes should decide processes, and you don't get to apply that rule selectively. The villain here isn't a company. It's the idea that open source is charity. It wasn't. Open

10:27 source is a way of deciding whose hardware the future gets written for. And the invoice for that just arrived at $6 billion, which leaves the one thing I don't know. Nvidia bought 109 people

10:38 whose whole advantage was being small enough to ship a model in 9 weeks. What happens to a 9-week release cycle the first morning it wakes up inside a $5 trillion company?