Silicon Valley raises the stakes after the DeepSeek challenge: multi-million dollar offers for programmers, more internal security, and new models.

Last January, the ground shook in Silicon Valley , and not because of the tectonic movements of the San Andreas Fault, the usual source of the area's numerous earthquakes. Sixty-five thousand miles west, in the Chinese city of Hangzhou , a little-known company had just unveiled an open-source artificial intelligence that they claimed was capable of rivaling the most advanced models in the United States.

Dubbed DeepSeek R1 , this intelligence had been trained at a fraction of the cost of the most advanced models from OpenAI or Google , but offered comparable, if not better, results and was distributed under open source.

Alarm bells rang on both sides of the country. In New York, for example, the stock market severely punished the valuations of many technology companies, which had invested billions of dollars in developing much more complex artificial intelligence tools. The first victim was Nvidia , whose market capitalization fell by more than 17%, wiping out nearly $600 billion in valuation in just a few days. DeepSeek proved, after all, that it wasn't necessary to rely exclusively on its expensive H100 GPUs to achieve efficient and competitive models.

But it was at the venture capital firms on Sand Hill Road, the lubricant that keeps Californian startups lubricated, where the panic became palpable. Since 2022, with the launch of ChatGPT, the consensus among these companies was that artificial intelligence would become the region's new driving force, the next big revolution after the dot-com boom and the app economy. Suddenly, all those hundreds of billions invested in financing the high development and personnel costs of dozens of AI-dedicated companies seemed very difficult to recover.

DeepSeek's biggest innovation was its frugal approach . The company—founded in 2023 by Liang Wenfeng, linked to the investment fund HighFlyer—launched R1 after a training program that required only $6 million and a mere 2,000 GPUs . To give you an idea, this is less than a tenth of what it cost to train GPT-4, the language model that, until recently, powered the popular ChatGPT tool and has now been replaced by GPT-5 .

Despite its low training and " inference " costs (which is what the process of running a long language model for use is known as), R1 offers reasoning, mathematical and code generation capabilities - currently the most promising skill of these tools - comparable to the reference models.

Grudgingly, the industry's leading figures acknowledged the challenge. Marc Andreessen , founding partner of Andreessen Horowitz, called DeepSeek R1 "one of the most impactful revolutions I've ever seen." Even Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman had to call it "impressive."

Analysts were quick to compare the announcement to the launch of Sputnik , the Soviet satellite that put the US space program on the ropes and shattered the narrative of its technological superiority.

MORE INVESTMENT

Six months later, the sense of surprise has transformed into urgency. Meta , led by a Mark Zuckerberg who seems obsessed with keeping up, has accelerated an already massive plan. It will invest between $60 billion and $72 billion in AI infrastructure over the coming years, including massive data centers capable of scaling up to 5 gigawatts and with a surface area comparable to that of Manhattan.

It has also gone on a talent hunt , with offers described internally as "explosive," offering bonuses and salaries that can reach hundreds of millions of dollars for top-level researchers or engineers, something unprecedented even in recent technological history.

Behind this strategy lies a reasonable concern. Until last January, Meta's language model, Llama 3 , was considered the best open-source alternative, less powerful than OpenAI's but easy to adapt to all kinds of tools and projects. The arrival of DeepSeek and the somewhat disappointing launch of its evolution, Llama 4 , have forced Zuckerberg to take a decisive stance.

For its part, OpenAI, the company that so far appears to be leading the artificial intelligence race, has gone on the defensive. Its CEO, Sam Altman, had no choice but to publicly acknowledge DeepSeek-R1's progress, calling it surprising, although he clarified that it has yet to prove its long-term sustainability.

Since then, Altman, like much of the United States, has preferred to ignore its existence, keep his head down, and redouble his efforts and investment. OpenAI has launched new tools, such as ChatGPT Agent , which allows users to delegate complex tasks to an artificial intelligence capable of operating autonomously, and now speaks more freely about future open-source models based on its technology, an option that until recently seemed ruled out. Furthermore, it has replaced the language engine that enables its operation with GPT-5, its most advanced artificial intelligence system to date. The executive compared using ChatGPT now to conversing with a PhD student.

The company has also strengthened security in its internal processes, concerned about the possibility that part of DeepSeek's training was based on leaked data from its organization. The prevailing narrative among Silicon Valley experts is that the reason DeepSeek has managed to launch such a competitive model with so few resources is that it has done so by "distilling" the existing technology in GPT-4, leveraging OpenAI's model to teach its own to operate in the same way.

To protect themselves, several investment funds are also now considering a change in their investment strategy, prioritizing companies that focus on less infrastructure , with a greater focus on software and algorithms, and based on open source models instead of proprietary ones, which are more expensive and difficult to replace.

After the initial panic, exorbitant valuations have once again become the norm in the region, and the fight for talent has intensified. But the feeling that something fundamental has changed—the fear that Silicon Valley is no longer the center of the tech world—remains very present.