Anjney “ANJ” Midha, a General partner and member of the Mistral Board of the Andreessen Horowitz, tells TechCrunch that he first spilled DeepSeek’s amazing performance six months ago.
According to a paper published last year, Deepseek introduced Coder V2. He said that Deepseek was on a path to release models that were improved every few months through R1. R1 is a new open source reasoning model that overturns the high -tech industry to provide industry standard performance at several costs.
Despite the sale of NVIDIA shares, Midha does not mean that R1 does not mean that the AI basic model stops spending billions of dollars on the GPU chip and building a data center as soon as possible. Say.
It means that they do more depending on the computational power they can get.
“When people like ANJs, Mistral collected $ 1 billion,” he says. “Does DeepSeek mean that $ 1 billion is completely unnecessary? No, in fact, it is very much that you can inward them, seeing the improvement of Deepseek efficiency, throwing $ 1 billion. It is worth it.
He added, “Now you can get more than 10 times output from the same calculation.”
He claims that Mistral does not mean that his rivals are open and hopelessly behind humanity. Each has a hundreds of billions than Mistral. Openai is said to be discussing to procure $ 40 billion that drops another jaw.
He says that Mistral is open source and keeps his competitiveness with them. And his logic has a merit. Open source is essentially accessible to the company from those who want to support them because they use the project. Rivals of closed source must keep the secrets and pay all labor and calculation power.
“No $ 20 billion is required. It requires more calculations than any other open source model app, so the mist is positioned. [well]。 They calculate all open source providers the best, “Midha said about his portfolio company.
Facebook’s LLAMA is a rival to Mistral’s largest open source AI model in the western part of Mistral, and has earned more investments. Mark Zuckerberg, Wednesday, stated that it is planning to spend hundreds of millions of dollars on AI. This includes $ 60 billion in 2025 capital spending and mainly for data centers.
A16Z oxygen GPU shared program “Overbooking”
MIDHA, who is also an executive of AI image generator Black Forest Lab and 3D model maker LUMA (and AI angel costumes, Eleven love, etc.), has another reason why AI GPUs are not quickly declined. I am. 。
He is the leader of the A16Z oxygen program. The GPU, especially the NVIDIA’s most advanced H100, is a very few products that VC companies handed the problem about a year and a half ago. I purchased those bundles so that portfolio companies could use it.
The oxygen laughs, “I’m currently over -booking. I can’t allocate it.” His startup not only requires a GPU for AI model training, but also requires more to execute an ongoing AI product for customers.
“There is a demand for this insatiable inference and consumption now,” he explains.
That’s why he thinks that Deep Shik’s engineering breakthrough does not change the Stargate. This is a large $ 500 billion partnership of Openai, which was announced earlier this month at AI Data Center SoftBank and Oracle.
The main change in DeepSeek is the state that AI is the next basic infrastructure such as electricity and the Internet. Midha wants to consider “infrastructure independence”, as he calls. Do they want to put their censorship and nails in data and rely on Chinese models? Or do you want a Western model that complies with the NATO agreement according to Western laws and ethics?
Obviously, he defends the western countries using Western models, like Mistral, based in Paris. Hundreds of companies have shared their concerns and have already blocked DeepSeek. This is both a consumer app service and an open source model.
Nobody is the fear of China’s open source model. Companies can be executed locally at their own data center. DeepSeek is already available as a safe cloud service for American companies such as Microsoft Azure Foundry, so developers do not need to use DeepSeek cloud services.
In fact, Pat Gelsinger, a former CEO of Intel (a well -known person in China), built an AI chat service with a unique version of Deepseek R1 instead of LLAMA and Openai choice. He told TechCrunch.
However, if someone wants to abandon the data center plan in light of Deepseek, Midra laughs and “If there is an additional GPU, please send it to ANJ.”
TechCrunch has a newsletter focusing on AI! Sign up here and get it on the reception tray every Wednesday.
Source link
