Tech Weekly: DeepSeek Shatters Silicon Valley's Assumptions

It started with a jolt. Earliest this week, the release of DeepSeek’s latest AI model triggered a $1 trillion selloff in tech stocks, as investors confronted an uncomfortable truth: Silicon Valley’s dominance in artificial intelligence is not unshakable.

This wasn’t a reaction to war, recession, or regulatory crackdowns. Instead, markets were blindsided by a Chinese startup proving that cutting-edge AI doesn’t require Silicon Valley’s billions — and that the industry’s assumptions about competition might be dangerously outdated.

For years, the prevailing assumption in Silicon Valley has been that building powerful AI requires vast resources - massive investments in chips, talent, and infrastructure. As Bloomberg’s Parmy Olson pointed out, this view has become so entrenched that some tech executives see their huge funding rounds as their main competitive advantage:

Last year, the chief executive officer of a leading AI firm was asked at a private Silicon Valley dinner about how his company differentiated from others building “foundation models,” the systems underpinning chatbots like ChatGPT. Did he have a moat? Yes, he answered, according to another CEO who was there. No one else had raised the billions of dollars that he had. That was his moat.

Yet DeepSeek, founded by former hedge fund manager Liang Wenfeng, has dismantled this logic. Built with only $6 million in computing power, older-generation chips, and a fraction of the $100 million budget OpenAI expended on GPT-4, DeepSeek's R1 model now rivals GPT-4 in reasoning tasks and dominates app stores across the U.S., U.K., and Canada.

The key to DeepSeek’s breakthrough was its efficiency in using existing resources and a different approach to model building, which they have made public. Unlike OpenAI’s guarded models, the company open-sourced R1’s architecture, inviting developers worldwide to build atop its work. Founder Liang Wenfeng explained the philosophy in an interview last year:

In the face of a revolutionary technology that breaks with conventional wisdom, a competitive advantage can only be protected in a closed environment for a limited period. Although America's Open AI has kept its source code private, it can't prevent its competitors from catching up. We will grow in such a process, accumulate a lot of know-how and form our corporate structure and culture to incubate innovations. That is our strength.

This emphasis on openness and collaboration stands in stark contrast to the guarded approach of its US counterparts. DeepSeek's use of Meta's open-weight Llama model, for example, shows that the company is not seeking to control information, and instead believes that progress depends on the free sharing of knowledge and technology.

As DeepSeek’s model gains traction and other open-source models gain ground, the business model of selling access to foundation models at a premium is under challenged. Miles Kruppa of The Wall Street Journal noted how DeepSeek’s approach is going to force US firms to adapt:

The possibility that DeepSeek piggybacked on technology OpenAI and others devoted billions of dollars to building — while developing its own AI more efficiently — is upending the business models of leading U.S. tech companies. Why invest so much in creating advanced AI when it can be so easily and cheaply replicated?

In fact, this also suggests that the business of building foundation AI models is becoming a commodity business, argues Greg Jensen, Co-Chief Investment Officer of Bridgewater:

The DeepSeek results are a threat to the leading AI labs, as it is clear that close-to-state-of-the-art models can quickly be commodified. This will make it more and more difficult for the frontier labs like OpenAI and Anthropic to monetize their existing IP. It likely will lead them to be much more discreet in how they expose their IP in the future as well.

The current AI giants like OpenAI and Google may need to pivot, fast. Margins will shrink as open-source alternatives undercut their pricing. This will force them to compete on applications rather than monopolize AI models.

Quick Hits

• TikTok takeover: Perplexity AI revises TikTok merger proposal that could give the U.S. government a 50% stake.
• Azure's slowdown: Microsoft beats quarterly revenue estimates, but its Azure cloud computing business slows, compounding investor worries about the company's huge AI investment.
• AI contender: Alibaba releases AI model Qwen2.5-Max to challenge U.S. tech giants, allegedly reducing infrastructure costs by 40-60% compared to traditional large language model deployments.
• Data center boom: Chevron plans to build gas plants for data centers amid AI boom, while TikTok invests $3.7B in Thailand data center.
• iPhone sales dip: Apple's record revenue tarnished by China iPhone slump.

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.