Mistral Board Member and A16Z VC Anjney Midha says Deepseek does not stop ai gpu hunger ncvrs.com

Andreessen Horowitz is the main partner and Anjney “Anj” Midha, a member of the Mistral Board of Directors, first made a lecture on Deepseek jaw-drip six months ago by Techcrunch.

This is when Deepseek introduced Coder V2, which dedicated Openai GPT4-Turbo to coding-specific tasks paper Published last year. This made Deepseek on one way to release improved models every few months through R1, he said. R1 is a new open source reasoning model that has prepared the technology industry to offer the performance of the industrial standard in a fraction of the cost.

Despite the sale of NVIDIA shares, Midha says that R1 does not mean that AI basic models stop billions from moving GBPE GPU chips and builds more data centers as quickly as possible.

This means that they will do more with the computing power they have acquired.

“When people are, okay, Anj, Mistral has raised billions of dollars,” he says. “Deepseek means all billions of dollars are completely superfluous? No, it is really extremely valuable to be able to look at the efficiency of the Deepseek, internalize them, and then throw billions of dollars. ”

He adds, “Now we can get ten times more outputs from the same calculation.”

This does not mean that Mistral is hopelessly behind the Opena rivals and the anthropic, he says. Each of them has collected much billions than Mistral. Openai is said to have raised $ 40 billion in negotiations in negotiations.

Mistral is competitive with them because he is open source, he says. And his logic has merit. Open source provides the company access to essentially free technical work from those who want to help because the project is used. Closed source rivals keep their secrets and have to pay for all workforce and calculation energy.

– You don’t need $ 20 billion. You only need more calculations than any other open source model application. So the Mistral is located (well). They have the most calculated open source service providers, ”said Midha’s portfolio company.

Facebook Llama, the largest Western open source AI rival for Mistral, will also receive much more investments. CEO Mark Zuckerberg said Wednesday that he would continue to plan to spend “hundreds of billions of dollars” on AI. This includes $ 60 billion in $ 2025 on capital costs, mostly data centers.

A16Z Oxygen GPU Sharing Program is “over -taken”

Midha, who is a board member of the AI ​​Image Generator Black Forest Labs and the 3D model manufacturer (and an angel for AI angelic clothes, Antropic, Elevenlabs and others), another reason is that he does not see the AI ​​hunger GPUs to decrease soon. Or

He is the leader of the A16Z oxygen program. GPUs, especially NVIDIA’s state-of-the-art H100, have become such a scarce commodity that VC had taken things into its own hands about a year and a half ago. He bought a knot for portfolio companies.

Oxygen “is now overcome. I can’t distribute it enough – Midha laughed. Not only do start -up companies need GPU for AI model training, they also need to operate ongoing AI products for customers.

“Now there is this insatiable need for conclusion, consumption,” he explains.

This is also the reason why Deepseek’s engineering breakthroughs do not change Stargate. This is Openai’s large $ 500 billion partnership announced earlier this month to Softbank and Oracle for AI data centers.

The most important change is the depth of nation -states that AI is the following basic infrastructure, such as electricity and the Internet. Midha wants to consider “the independence of infrastructure” as he calls it. Do you want to rely on Chinese models, censorship and censorship of their data? Or do you want Western models that follow Western laws and ethics and adhere to NATO agreements?

Obviously, it supports Western models of Western nations, such as Mistral in Paris. Hundreds of companies share this concern and have already blocked Deepseek, a consumer application service and an open source model.

Not everyone buys the fear of Chinese open source models. Companies can run them locally in their own data centers. And Deepseek is now available as a safe cloud -based service for US companies Microsoft Azure FoundrySo developers do not have to use Deepseek’s cloud -based service.

In fact, former CEO of Intel, Pat Gelsinger – someone You know China well – told Techcrunch that start -up Gloo is building ai chat services on Deepseek R1’s own version, instead of choices, like the Lama or Openai.

But if someone wants to flood the data center plans in the lights of the Deepseek, the name Midra and have a request, “If you have an extra gpu, please send them to Anj.”

Techcrunch has an AI-centered newsletter! Sign up here to get into your mailbox every Wednesday.