DeepSeek: Why are US technology stocks tumbling? US-China tech war & more

DeepSeek, a Chinese competitor to dominant US technology giants in artificial intelligence (AI), triggered a selloff in Nvidia and other US IT majors, sending Nasdaq tumbling 3 per cent overnight. Developed as an open-source LLM that matches the performance of GPT-4o using a fraction of computing power, DeepSeek's AI platform's architecture is believed to deliver high-quality parameter processing. Its development is being termed as 'Sputnik moment' for AI.  

Why US tech stocks are falling

Nvidia shares slumped 16.97 per cent in US trading and eroded $593 billion in market value, the deepest ever one-day loss for a company on Wall Street. Nasdaq 100 lost $1 trillion in value overnight. A host of foreign brokerages such as Bernstein, Goldman Sachs, Jefferies and UBS noted that overseas AI companies have been going after model improvement at all costs and the emergence of DeekSeek has prompt investors to ask hard questions about these computing power investments.

Foreign brokerage Jefferies said US AI players' management could be under more pressure to justify further raising AI capex in 2026. The AI supply chain is highly vulnerable to a de-rating, it warned.

"It would prompt the AI industry to refocus on ROI. Despite DeepSeek's amazing model efficiency, it has not accelerated any AI monetization. Re-evaluating computing power needs could cause 2026 AI capex to fall (or not grow)," Jefferies said.

To recall, Meta had on Friday revised its capex upward on Friday, with 2025 capex now expected to be around $60-65 billion. This was potentially $10 billion above consensus forecasts. Microsoft’ reiterated $80 billion of forward capex that implied the capital intensity of AI efforts may remain elevated.

"This is likely to act as an overhang on the technology and energy sectors until a broad array of company management teams provide their own insight/perspectives during the Q4’24 earnings season," Goldman Sachs said.

Decoding Deepseek model
InCred Equities said the DeepSeek discounted model pricing is disruptive, to say the least. Although users enjoy discounted prices of DeepSeek API for a few months, the list input/output costs for processing tokens is $0.55/$2.19 per million tokens against $2.50/$10 for GPT-4o. 

"Various estimates suggest this brings the training costs pricing to $6 million against $100 million-plus as the models use a 8-bit float instead of a 32-bit, leading to substantially lower memory, and read multiple tokens simultaneously," InCred said.

US China tech war

Scale AI and Tesla CEOs believe that DeepSeek has 50,000 H100 Nvidia GPUs but are not being revealed due to the US export controls. Though not verified, it could materially change the cost structure argument, InCred Equities said.

The domestic brokerage said the development could intensify the trade war between the US & China, leading to a material impact on Fortune 500 companies’ earnings and, in turn, result in a cautious environment from an IT spending perspective," it said.

China is the only market that pursues LLM efficiency owing to chip constraint. US President Donald Trump is likely recognised the risk of further restrictions is to force China to innovate faster. Jefferies therefore believes Trump will relax the prevailing AI Diffusion policy.

What's ahead?
UBS said while DeepSeek’s aggressive pricing strategy raises questions on big tech’s high capex intensity, one needs to understand that DeepSeek’s models still have many limitations compared to the frontier models from big tech coupled with limited transparency so far on training methodology. 

"Hence, we wait for more details during the upcoming tech company results before reaching any conclusions," it said.

Bernstein said it believes that DeepSeek  did not “build OpenAI for $5 million.”. It said The models look fantastic but it does not think they are miracles; and the resulting Twitterverse panic over the weekend seems overblow, it said.

InCred Equities said the recent developments are evolving, and a know-it-all scenario is sometime away, but it could also be the beginning of an accelerated AI adoption curve across enterprises.