DeepSeek’s big promises to dominate global tech and finance fade

Nearly a year ago, DeepSeek blew through global markets and triggered instant fear across tech and crypto desks.

A little-known Chinese AI lab dropped a model that forced traders to rethink U.S. control of advanced AI. Stocks tanked fast. Panic trading took over. The idea that China could match top-tier AI with cheaper chips hit hard and fast.

The fallout was brutal. Nvidia sank 17% in one session and wiped out almost $600 billion in value. Broadcom fell the same amount. ASML slid 7% in a single day. Fast forward eleven months and the picture flipped.

Nvidia pushed past a $5 trillion valuation in October. Broadcom climbed 49% across 2025. ASML gained 36%. The market moved on.

Markets price the shock and then move on

The chaos started after DeepSeek released V3 in late 2024. The lab said the open-source model ran on weaker chips and cost far less than systems from OpenAI and Google.

Weeks later came R1 in January 2025. Benchmarks showed similar or better performance than leading models. That moment reset expectations.

Haritha Khandabattu, a senior director analyst at Gartner, said January caused a broad repricing because beliefs around frontier-model costs and China’s competitiveness changed overnight. She said it directly hit the semiconductor and hyperscaler story.

Alex Platt from D.A. Davidson told CNBC the release shocked investors since the dominant view was that China lagged the U.S. by up to a year.

Brian Colello at Morningstar said fears centered on falling demand for AI hardware and weaker revenue for Nvidia, which never really showed up as spending held steady throughout 2025.

Colello said forecasts now point to faster spending in 2026 and later. Since January, DeepSeek pushed out seven updates. All were revisions to V3 and R1. None were new models. The market treated them as progress, not disruption.

Compute limits stall the next leap

Platt said DeepSeek’s compute turned into a bottleneck and there are limits to how much architecture tricks can stretch hardware. The lab delayed its R2 model, planned for May, after running into trouble training on Huawei chips.

Chinese officials encouraged the use of local processors to cut reliance on U.S. tech under export controls tied to Nvidia’s top chips.

“China’s been constrained in the amount of computing power it’s been able to access over the last couple of years, in large part because of U.S. restrictions on the sale of chips,” Chris Miller, author of “Chip War,” said. “If you want to build advanced models, you need access to advanced compute.”

DeepSeek later admitted in a research paper that it faces limits when compared with closed models like Gemini 3, including compute resources. While that played out, Western labs kept shipping. OpenAI rolled out GPT-5 in August.

Anthropic released Claude Opus 4.5. Google launched Gemini 3 in November. Gartner analyst Arun Chandrasekaran said rapid releases eased fears of sudden commoditization.

There are signs DeepSeek is not done. On New Year’s Eve, the lab published a paper on more efficient model development. Dan Ives at Wedbush said shocks will keep coming next year. He told CNBC there will be another DeepSeek moment.

If you're reading this, you’re already ahead. Stay there with our newsletter.