Mistral board member and a16z VC Anjney Midha says DeepSeek won’t stop AI’s GPU hunger

Andresen Harovits General Partner and Mistral Board member Anjani “ANJ” Midh “first spying Dipsec Six months ago he told TechCrunch.

When launching DIPSEC Coder V2, which competed 4-TRBO of OpenAE for coding-specific tasks, A paper It was published last year. He said that it has been on the way to publish developed models every few months through DEPSEC R1, he said. R 1 of the new Open Source argument The model which has made up the technology industry For the supply of industrial standard performance in a fraction of expenditure.

In spite of Nvidia Stock Sales Stop SalesMidha says 1 does not mean that the AI ​​Foundational models will spend billions for GPU chips to gabble and build more data centers as soon as possible.

This means that they will do more with the counting power they can gain.

“When people are, okay. Okay, Myster has collected one billion dollars,” he said. “Does DIPSEC mean that all billion dollars are completely unnecessary? No, in fact, for them, it is extraordinary for them to see the improvement of the diploma, to internalize them and then throw a billion dollars. “

He added, “Now we can get 10 times more output from the same count” “

This does not mean that the Mistral is hopelessly the rivals are OpenAI and ethnographic behind, he argued. Each of them has raised more billion more than Myster. It is reported that the Open is in the discussion to raise another person Jaw-dropping $ 40 billion.

He says that Myster is competitive with them because it is an open source, he says. And have the qualifications of his argument. Open Source gives a company that gives us access to free technical labor from those who want to help them help using the project. Closed source rivals protect their privacy and have to pay for all labor as well as counting power.

“You don’t need $ 20 billion. You need more calculations than any other open source model app. Thus there is mistral position [well]The They have the maximum calculation of any open source suppliers, “Midha said about its portfolio company.

Mistral’s largest Western Open Source AI model rival Facebook Lama will also get a lot of investment. Wednesday CEO Mark Zuckerberg says he Still planning to spend “several hundred billion dollars” AI overall. These include capital expenditure in 2025, most data centers include $ 60 billion.

A16z’s Oxygen GPU Sharing Program “Overbuiled”

Midha, who is also a member of the AI ​​Image Generator Black Forest Labs and Board 3D Model Maker Luma (And an angle of AI decoration anthropologists, Elevenlabs and others) there is another reason because he is not seeing the hunger of the GPU at any time.

He is the leader A16z’s oxygen programThe The GPU, especially in the latest Nvidia’s sophisticated H10s, has become a rare product that the VC firm took the issues about a year and a half ago. It bought a bunch of them to use its portfolio companies.

Oxygen is now “overbooked. I can’t allocate enough,” Midha laughed. Not only does the GPUs of its startups need to be trained for AI model, but they need more to run their ongoing AI products for customers.

“Now there is this insatiable claim for this component for use,” he explains.

That is why he thinks that the engineering breakthroughs in the Deposes will not change Stargate. It Openai’s larger $ 500 billion partnership Earlier this month, AI data centers were announced with softbank and oracle for AI data centers.

The main change in the DEPSEC is that the AI ​​is recognized by the country that is the next base infrastructure like electricity and the Internet. Midha wants they to consider “infrastructure freedom”, as he said. Do they want to rely on Chinese models with their censorship and their data on their data? Or do they want Western models that follow Western law and ethics and adhere to the NATO agreements?

He is obviously using Western models like his Paris-based Mistral, advising Western countries. Hundreds of companies share that concern and Have already blocked the DEPSEC, Both are the customer app service and an open source model.

Not everyone purchases that fear of Chinese Open Source models. Companies can run it locally in their own data centers. And DEPSEC is already available as a cloud service protected from American companies Microsoft Azure FoundrySo developers do not have to use the cloud service of DIPSC.

In fact, former CEO Well known with China – Tell TechCrunch that his startup is making glue AI Chat Services in Own version of DEPSEC R1 Instead of preferences like Lama or Openai.

But if someone wants to dig their data center plans in the light of the Deposes, Midra laughs and a request: “If you have extra GPU please send them to the angle.”

There is an AI-centric newsletter on TechCrunch! Sign up here Every Wednesday gets it to your inbox.