January 28, 2025

The great rethink

Hi there,

China’s DeepSeek has got the AI world in a frenzy.

To catch you up in 3 bullets:

The AI company, which emerged from Liang Wenfeng’s hedge fund High-Flyer, released an open-source reasoning model last week that’s roughly on par with OpenAI’s o1 model performance.
DeepSeek says it trained its base model with limited chips and about $5.6M in computing power — a fraction of the $100M+ US rivals have spent training similar models — thanks to some clever techniques.
Cue Silicon Valley drama (more on that below) and a sell-off in chip and energy shares, as the company has upset assumptions about what it costs to develop highly performant AI.

Here’s what you need to know.

Why is the US market in an uproar?

In the US, genAI development has raced ahead thanks to billions of dollars in funding.

Big tech companies have justified their spend on AI infrastructure based on the need for more hardware (like chips) and energy to train bigger and more powerful models.

DeepSeek’s reported ability to develop similarly performant models much more efficiently is upsetting these assumptions.

On Monday, Nvidia’s stock fell more than 15%. Other stocks linked to the AI hype train, including infrastructure suppliers like Oracle and power producers like Constellation Energy, also saw steep declines.

What does this mean for the private AI sector?

DeepSeek’s advances could undercut the vast sums of money that have gone to foundation model developers. OpenAI and Anthropic alone have raised over $30B in funding.

With DeepSeek, it may be harder for AI startups to justify raising huge funding rounds to support their infrastructure buildouts.

And it might put pressure on the valuations of closed-source model developers like OpenAI ($157B valuation) and Anthropic ($60B).

In a visual:

Source: @boldleonidas via X

Of course, there are lots of questions about DeepSeek, including the legitimacy of its $5.6M cost.

US developers maintain an advantage in compute and data, and they will likely adopt DeepSeek's engineering advances. This could widen their performance lead if they can combine their superior resources with these new efficiency gains.

What are the implications for enterprises?

DeepSeek’s advances give fuel to the open-source movement — highlighting that open, frontier models can be developed with more modest resources and computing infrastructure.

Enterprise AI strategies will need to account for more open-source options, as open models come close on performance to pricier proprietary alternatives like those from OpenAI.

More broadly, plummeting training costs would further drive down the cost of using these models.

DeepSeek’s Reasoner is nearly 30x cheaper than OpenAI’s o1 model.

Enterprises should continue to build with the assumption that AI costs will continue to decline.

TLDR — Tech loves drama, right?

Here's a roundup of recent tech drama:

Off the deep end: DeepSeek’s model releases have sparked much discussion among Silicon Valley’s who’s who. While Marc Andreessen said on X its reasoning model was “one of the most amazing and impressive breakthroughs I’ve ever seen,” others like Scale AI’s CEO Alexandr Wang and Curai CEO Neal Khosla (son of Vinod Khosla) have suggested DeepSeek “has more H100s than people think” and is “faking the cost.”
Stargate Wars: Elon Musk and OpenAI CEO Sam Altman sparred on X after OpenAI announced the $500B “Stargate” AI project to fund AI infrastructure in the US. Musk questioned the project’s funding, while Altman replied, “I realize what is great for the country isn’t always what’s optimal for your companies, but...I hope you’ll mostly put [the United States] first."

Source: X

Cloudy stargazing: Meanwhile, Salesforce’s Marc Benioff suggested Stargate could create tensions between OpenAI and its biggest investor, Microsoft, after OpenAI ended its exclusive cloud deal with Azure. Microsoft CEO Satya Nadella downplayed the drama, reaffirming their “critical partnership” and pledging $80B for data centers to boost AI development.
Fraud, but make it fancy: Former GameOn CEO Alexander Beckman and his wife, Valerie Lau, allegedly stole millions from the startup’s investors to fund a luxe lifestyle, including a Tesla, private school tuition, and a $4.2M San Francisco home… while employees went unpaid. The DOJ’s 25-count indictment accuses them of falsifying financials and faking audit reports since 2018.

I love you.

Anand

@asanwal

Co-Founder & Exec Chair

P.S. On February 11, join our analysts in discussing 2024’s biggest AI breakthroughs and what’s next for 2025. Register here.

Get started with CB Insights

Start your free trial

CB Insights' emerging technology insights platform provides all the

analysis and data from this newsletter. Our data is the easiest way to discover and respond to emerging tech.

Was this email forwarded to you? Sign up here