I’m writing this while waiting for Nvidia’s earnings report. I’m interested to see if the company addresses the DeepSeek challenge.
China’s AI disruptor DeepSeek is preparing to introduce a new model. Reuters had initially reported that DeepSeek would launch its next‑generation model “V4,” focused on coding, in mid‑February 2026. Rumors now peg the expected release window as “Q1–Q2 2026.” The mid‑February window has passed but context‑window changes and internal benchmark leaks signal that V4 is close.
And it will be a BIG DEAL for AI competitors and AI chipmakers such as Nvidia (NVDA).
A big enough deal that the V4 release will move the entire tech sector and quite probably the stock market as a whole.
DeepSeek V4 is shaping up as a multi‑layered challenge to Nvidia and the broader U.S. AI ecosystem. With significant improvements in compute efficiency expected, V4 could call into question current projections for future chip demand. And the economics of AI infrastructure spending. If rumors prove out that DeepSeek was trained using Nvidia’s newest Blackwell chips–export of those chips to China is supposed to be prohibited–it will call into question the efficacy of U.S. export controls.
Reports say that DeepSeek’s novel architectures like “Engram conditional memory” and “Manifold‑Constrained Hyper‑Connections” allow V4 to cut training costs to a fraction of what U.S. frontier AI labs spend.
If advanced AI models can be trained with dramatically lower computing budgets, that will call into question the Wall Street bull case for Nvidia that capital spending on high‑end GPUs must keep rising. Even a modest downward revision in spending expectations would hit AI chip stocks hard. And could savage U.S. AI model-builder and datacenter stocks. Wall Street is already nervous about huge capital spending plans from these companies.
A senior Trump administration official confirmed to Bloomberg that DeepSeek’s upcoming model was trained on Nvidia’s most advanced Blackwell AI chips at a data center in Inner Mongolia, despite US export controls explicitly banning Blackwell shipments to China.
DeepSeek is expected to remove technical markers that would reveal the use of American chips, making enforcement harder.
This revelation undermines confidence in the export‑control regime and could trigger tighter restrictions that hurt Nvidia’s revenue from the China market.
OpenAI, Anthropic and Alphabet/Google have all accused DeepSeek of using distillation techniques that effectively use outputs from their models to train DeepSeek’s own model. Anthropic found that DeepSeek (along with Moonshot and MiniMax) created an estimated 24,000 fake accounts and extracted over 16 million conversation records from Claude.
Even if true–and I think it’s extremely likely–I don’t see how that helps these companies to meet the DeepSeek challenge. It’s not like there’s any effective international tribunal that could judge these claims or enforce findings or penalties.
This means DeepSeek may be building competitive models on the back of billions of dollars of U.S. research, at a tiny fraction of the cost, threatening the theoretical competitive advantage that justifies massive valuations for OpenAI, Anthropic, and Google’s AI division.
A V4 launch could trigger another Nasdaq selloff reminiscent of the January 2025 “DeepSeek shock,” when Nvidia lost hundreds of billions in market cap in a single session.
Depending on what numbers Nvidia reports today, and on what the company says about guidance, and how the market reacts tomorrow, it might be worth stepping aside from stocks in the AI sector until we see the fallout from V4.
Photo by John Cameron on Unsplash.
