DeepSeek disrupted the AI disruptors, but don't count out Big Tech just yet

NBC News

28-01-2025

OpenAI CEO Sam Altman posted to X toward the end of the most momentous day for AI since his company released ChatGPT in December 2022.
His message on Monday was a confirmation from AI's hierarch that the disruptors had been disrupted.
'deepseek's r1 is an impressive model, particularly around what they're able to deliver for the price,' he wrote, a hat tip to a rival nobody saw coming as recently as a few weeks ago. 'we will obviously deliver much better models and also it's legit invigorating to have a new competitor! we will pull up some releases.'
DeepSeek, a suddenly popular Chinese AI model that upended U.S. markets Monday, shot to prominence in the last several days, with its AI assistant topping Apple's App Store and inspiring reactions from President Donald Trump and tech's top leaders.
The abrupt appearance of DeepSeek has rattled Silicon Valley in a way that Silicon Valley used to rattle everyone else. A handful of talented programmers operating out of China upended the assumptions that have dominated Big Tech regarding the future of AI — which to many in tech is the only future worth thinking about.
Those assumptions centered on money, energy and models. The thinking went something like this: The best way to build an AI model is to train it over and over and over again, using the most powerful computer chips to hone its mathematical intricacies. Want more AI horsepower? Build a bigger AI engine (aka a data center that could take up a space the size of most of Manhattan). It's an expensive and resource-intensive proposition.
And just about everyone in tech accepted that proposition. That's why Goldman Sachs projected more than $1 trillion in AI investment in coming years and Mark Zuckerberg touted $60 billion just this year in AI spending, while Elon Musk and Altman sparred over a $500 billion project called Stargate.
DeepSeek went another way. Using advanced (but known) concepts for improving efficiency, its crew of scrappy kids — the CEO has said he looked to hire straight out of Chinese universities — went about making a model without the fancy chips or energy-intensive training that could compete with the big U.S. models.
That alternative throws into question a variety of Big Tech assumptions, upon which billions of dollars have already been invested.
What if AI models aren't that hard to build? What if AI models aren't that expensive or difficult to run, even on something as small as a smartphone? What if AI chips aren't quite as important as once thought? How did China sidestep the export controls meant to stymie its AI development?
Did Silicon Valley — and the tech market more broadly — see this coming?
Some did.
The idea of frontier, or the most-advanced, AI models becoming less differentiated has been a growing topic in the AI world, with the co-founder of tech consultancy Infosys telling CNBC in September that they were on their way to becoming commoditized and interchangeable.
Faint buzz about DeepSeek can be traced back to 2023, when the company's coding model was tested out by people in the AI community. Some discussion picked up in 2024 as the company released more models. By the end of the year4, some were warning that DeepSeek was already showing signs of throwing the industry into chaos.
'The crucial significance of DeepSeek's cheaper training methods lies in the path it opens up for wider adoption and innovation — particularly outside established tech hubs,' wrote Azeem Azhar, a technology entrepreneur, on Dec. 31. 'If cutting-edge LLMs can be built on smaller budgets, then Western efforts to control or slow AI development may prove futile. Constraints breed ingenuity; ironically, export controls appear to have spurred Chinese AI teams to engineer leaner, more efficient solutions.'
What's less clear is whether this is truly a gut punch to Big Tech. Yes, Nvidia lost a record-setting amount of market value, but it's still worth almost $3 trillion (and its stock bounced back almost 9% on Tuesday). Tech analysts have noted that making AI a cheaper technology has a major upside, including for the companies working on developing consumer platforms.
Ben Thompson, a business and technology analyst, wrote in his blog Stratechery that the emergence of DeepSeek and its initial shock could give way to a future in which Big Tech ends up benefitting.
'In the long run, model commoditization and cheaper inference — which DeepSeek has also demonstrated — is great for Big Tech,' Thompson wrote. 'A world where Microsoft gets to provide inference to its customers for a fraction of the cost means that Microsoft has to spend less on data centers and GPUs, or, just as likely, sees dramatically higher usage given that inference is so much cheaper.'
'In the long run, model commoditization and cheaper inference — which DeepSeek has also demonstrated — is great for Big Tech.'
Ben Thompson
On Monday evening, Trump agreed with that assessment in his own statements about DeepSeek, saying that 'instead of spending billions and billions, you'll spend less, and you'll come up with, hopefully, the same solution.'
Christopher Ackerman, an independent AI researcher who used to work for Google, said in a message that companies focused on building frontier AI models may be in for a challenging run, particularly OpenAI and Anthropic, two of the biggest AI startups, but that chips will still be necessary even with more efficient models, offering some security to Nvidia. DeepSeek has said that its own models were trained using less advanced Nvidia chips.
'The market doesn't agree right now, but the market is reactive,' he wrote.
Some are even less concerned. Dan Ives, an analyst for the financial services firm Wedbush Securities, wrote in an analyst note Tuesday that the reaction to DeepSeek has been excessive and that Jevons paradox — also mentioned recently by Microsoft CEO Satya Nadella in relation to DeepSeek — means this episode will be looked back on as a positive for Big Tech.
'We expect more innovation in AI and LLM model costs to come down,' Ives wrote. 'That is ultimately a great thing for computing power, use cases and where the tech world is going in this 4th Industrial Revolution.'