On January 27, 2025, Nvidia lost $589 billion in a single day. Largest one-day loss in market history.

The cause was a Chinese model called DeepSeek that cost $5.6 million to train. Wall Street did the math in nine seconds. Cheap models, fewer chips, Nvidia's a pumpkin.

Jensen Huang said the market misunderstood it.

Three weeks later the $589 billion was back. Then the business showed up: $215.9 billion in revenue, up 65%, and $97 billion in free cash flow.

He was right.

Now He's Doing It Again

Last week a Beijing startup called Moonshot released Kimi K3—2.8 trillion parameters, the largest open model ever built, benchmarking near the American frontier at a quarter of the price.

On July 27 they publish the whole thing, free. Anyone can download it and run it on their own machines with the door shut and the internet unplugged.

Washington is deciding whether to ban it. 

Agencies barred from using it. The labs behind it blacklisted. Treasury Secretary Scott Bessent has floated sanctions, claiming the Chinese copied American models to build their own.

Huang sat down with Axios and said the unthinkable: American companies should use them. 

And then he called them… excellent.

The objection came back fast, and it's the one Washington is acting on. Security.

The Security Argument

Yes, you can strip the safety guardrails off an open-weight model. Researchers at FAR AI proved it. Meaning, you can do things they’re not supposed to with them.

They also proved you can do it to OpenAI, Google, and Anthropic models—through those companies' own fine-tuning APIs, moderation systems and all. Fifteen bad examples was enough to break GPT-4.

The vulnerability belongs to fine-tuning. A closed model can at least be revoked. Open ones can never be recalled.

Which is exactly why banning Kimi accomplishes nothing. The model drops July 27. After that it sits on every hard drive on earth, permanently, whatever Washington decides in August.

The recall option expires before the debate finishes.

Ask Who's Doing The Asking

David Sacks ran White House AI policy until March. 

He said this:

"The leading closed-source labs—already a duopoly in terms of model revenue—are hoping the government will take out their open-source competitors."

Meanwhile, a source close to the administration told Axios that the labs or their allies arrive with a plan to ban open source every three to five months.

Every three to five months. That's a lobbying schedule.

The security worries are real and serious people hold them. But look at what the remedies accomplish. Procurement bans and blacklists do nothing about fine-tuning. 

They reduce competition.

Meanwhile… Nvidia has its own reasons. 

Nvidia Loves Fragmentation

Nvidia’s biggest customers are also its biggest threat.

Google, Amazon, Microsoft, Meta and OpenAI buy so many chips that it's become cheaper to design their own. All five are doing it. 

Google has TPUs. Amazon has Trainium. Microsoft has Maia. Meta has MTIA. OpenAI is building with Broadcom. Custom silicon is growing almost three times faster than the GPUs Nvidia sells.

Think of Nvidia as a landlord.

Five giant tenants rich enough to build their own house are a problem. Fifty thousand small tenants who never could are an annuity.

Free models create small tenants. When a hospital or a bank or a twelve-person firm can download something excellent for nothing, they can finally afford to run AI. And none of them will ever design a chip. They buy from Nvidia, or they rent from somebody who did.

Software people have a name for this. Commoditize your complement. 

Give away the thing next to your product, and everyone still comes to you for the part that costs money. Google gave away Android so Apple couldn't own the phone.

So the closed labs lobby Washington to kill open models. Huang defends them on television.

He's right. But he’s also selling shovels.