Useful information
Prime News delivers timely, accurate news and insights on global events, politics, business, and technology
Andreessen Horowitz’s general partner and a member of the Mistrales Board, Anjney “Anj” Midha, saw for the first time Deepseek’s amazing performance six months ago, he tells TechCrunch.
It was then that Depseek introduced Coder V2, which rivaled the OpenAi GPT4-Turbo for specific coding tasks, according to A role It was launched last year. This put Deepseek on a way to free improved models every two months through R1, he said. R1 is its new open source reasoning model that has overturned to the technology industry to offer the standard performance of the industry to a cost fraction.
Despite the sale of Nvidia shares, Midha says that R1 does not mean that the fundamental models of AI will stop spending billions to engulf GPU chips and build more data centers as quickly as they can.
It means that they will do more with the calculation power they can obtain.
“When people say, that’s fine, Mistral has raised one billion dollars,” he says. “Does Depseek mean that all those billion dollars are completely unnecessary? No, in reality, it is extraordinarily valuable for them to be able to see the efficiency improvements of Deepseek, internalize them and then throw them one billion dollars. “
He adds: “Now we can get 10 times more exit of the same calculation.”
That does not mean that Mistral is irremediably behind OpenAi and Anthrope rivals, he argues. Each of them has raised many more billions than Mistral. According to reports, OpenAi is in conversations to raise another amazing $ 40 billion.
Mistral is still competitive with them because it is open source, he says. And its logic has merit. The open source provides an essentially free technical labor access from those who wish to help because they use the project. The closed code rivals protect their secrets and have to pay all the workforce and calculate power.
“You don’t need $ 20 billion. It only needs more computation than any other open source model application. Then Mistral is positioned (well). They have the greatest calculation of any open source supplier, ”said Midha on their portfolio company.
The Facebook flame, the largest rival of Western open source AI models of Mistral, will also get much more investment. The CEO Mark Zuckerberg said Wednesday that he is still planning to spend “hundreds of billions of dollars” in general. That includes $ 60 billion in 2025 in capital expenses, mainly data centers.
A16Z Oxygen Exchange Program ‘Overbooked’
Midha, who is also a member of the IA Image Generator Board, Black Forest Labs and the manufacturer of 3D Luma models (and an angel with anthropic attire, eleven, others, others) has another reason why he does not see AI hunger for GPUs decreasing in the short term.
He is the leader of the A16z oxygen program. The GPUs, particularly the NVIDIA avant -garde H100, have become so scarce merchandise that the risk capital company took the matter in its own hands about a year and a half ago. He bought a lot of them so that their portfolio companies use them.
Oxygen is “too reserved at this time.” I can’t assign enough, “Midha laughs. His new companies not only need GPU for training in AI models, but also need even more to manage their current products for customers.
“Now there is this insatiable demand for inference, for consumption,” he explains.
That is why he also believes that Deepseek engineering advances will not change Stargate either. That is the great association of $ 500 billion OpenAi announced earlier this month with SoftBank and Oracle for AI data centers.
The main change of deep changes is the recognition of national states that AI is the next fundamental infrastructure, such as electricity and the Internet. Midha wants them to consider “infrastructure independence”, as he calls. Do you want to trust Chinese models, with your censorship and claws in your data? Or do western models that follow Western laws, ethics and comply with NATO agreements?
Obviously he is advocating western nations using Western models, such as his Mistral based in Paris. Hundreds of companies share that concern and have already blocked Depseek, which is both a consumer application service and an open source model.
Not everyone buy that fear from Chinese open source models. Companies can run them locally in their own data centers. And Deepseek is now available as a safe service in the cloud of US companies such as Microsoft Azure FoundrySo developers do not have to use the Deepseek cloud service.
In fact, the former CEO of Intel, Pat Gelsinger, someone Well familiar with China – He told Techcrunch that his start Gloo is building AI chat services in his own version of Deepseek R1 instead of options as flame or openai.
But if someone wants to get rid of their plans from the Data Center in the Light of Deepseek, Midra laughs and has a request: “If you have additional GPUs, send them to ANJ.”
