Tech giants around the globe were rattled on Jan. 27 after Chinese AI start-up DeepSeek unveiled an impressive, low-cost artificial intelligence (AI) model, sparking widespread concerns about the scale of investment being poured into expensive hardware and data centers.
DeepSeek released a free AI assistant last week that quickly saw its downloads overtake OpenAI's ChatGPT on the Apple App Store in the U.S. The Chinese start-up claims the free, open-source large language model known as V3 took just two months to train and cost less than $6 million. What's more, DeepSeek claims its latest cost-effective model, the R1, can match the performance of OpenAI's o1 reasoning model on certain benchmarks.
The company also claims R1 is 20 to 50 times cheaper to use than o1. The advent of this cheaper model hammered the shares of tech giants that have been investing billions of dollars in AI infrastructure.
AI stalwarts have a forgettable day
Nvidia (NVDA 3.17%), which is the leading provider of graphics processing units (GPUs) used for AI model training and inferencing, saw its stock price plunge 17% on Jan. 27. The company has been one of the biggest beneficiaries of the massive spending on AI infrastructure as it controls an estimated 85% of the AI chip market.
NASDAQ: NVDA
Key Data Points
Nvidia's top-of-the-line data center GPUs are reportedly priced between $30,000 to $40,000 per unit. Tech giants such as Microsoft and Meta Platforms have been purchasing hundreds of thousands of these chips from Nvidia to train and deploy AI models. These two companies reportedly spent $9 billion in 2023 to purchase a total of 300,000 AI chips from Nvidia. That investment most likely shot up last year as Microsoft was reportedly planning to get its hands on 1.8 million Nvidia GPUs, while Meta hoped to procure 350,000 chips last year.
DeepSeek, meanwhile, claims it used 2,048 of Nvidia's performance-capped H800 GPUs (a downgraded version of the chipmaker's popular H100 AI GPU) to train its model. For comparison, OpenAI reportedly used more than 10,000 flagship H100 chips to train its GPT-4 and GPT-4o models.
If DeepSeek's claims are true, then it means competitive AI models can be trained at a fraction of the cost of what the likes of Microsoft, Meta, and others have been spending, putting in doubt the entire tech industry's huge investment plans. It was no surprise that the entire AI ecosystem was deep in the red following this development.
Broadcom (AVGO 2.92%), which supplies custom AI processors to Microsoft, Alphabet, Meta, and OpenAI to develop in-house chips, saw its stock plunge 17% as well. Foundry giant Taiwan Semiconductor Manufacturing (TSM 2.79%), popularly known as TSMC, fell more than 13%.
Both Broadcom and TSMC have also been riding the surge in AI chip demand with their share prices soaring in the past couple of years. Broadcom dominates the custom AI processor market with an estimated market share of 55% to 60%, according to JPMorgan. The company sees a massive addressable market worth $60 billion to $90 billion for its custom processors and networking chips over the next three years. That points toward a major increase in its AI revenue from last year's level of just over $12 billion.
NASDAQ: AVGO
Key Data Points
TSMC, meanwhile, is the go-to manufacturer of chips for fabless chipmakers such as Nvidia and Broadcom. It is the dominant player in the global foundry market by quite some distance, and it has built a solid clientele that allows it to benefit from the adoption of AI in multiple industries. The foundry giant reported impressive revenue growth of 37% in 2024 and is expecting solid performance in 2025 as well.
Decoding the bigger picture
Wall Street analysts believe the DeepSeek-inspired fears about a slowdown in AI infrastructure spending are overblown. The compute capacity that will be freed up by the launch of more efficient AI models is likely to be absorbed elsewhere, according to Bernstein Research. Others share a similar view, believing that the use of AI should only expand, sustaining demand for chips from Nvidia, Broadcom, and TSMC.
At the same time, there are suspicions that DeepSeek may be downplaying how much it spent training its models. According to one source, it may actually have access to 50,000 of Nvidia's powerful H100 GPUs. Given the uncertainty of the situation, investors would do well to avoid a knee-jerk reaction as seen with the recent sell-off.
After all, spending by major tech giants on AI infrastructure is still likely to head higher once again in 2025, as evidenced by the capital spending plans laid out by the likes of Microsoft and Meta, as well as the announcement of the Stargate project. Investors who remain bullish on AI can now buy many top AI stocks at a modest discount.
