One year after the DeepSeek freak, the AI industry has adjusted and roared back

A year ago this week, Chinese startup DeepSeek shook the AI world with the release of its DeepSeek-R1 model. Since then, everything has changed.

Before the release of DeepSeek, pretty much all of the big players in AI were following the same playbook for building large frontier models — more GPUs plus more data gets you a smarter model — which worked for a while.

But while OpenAI, Meta, and xAI were hoarding Nvidia H100 GPUs to train their next models, DeepSeek was trying a different tactic. Constrained by export controls that denied it access to the latest and most powerful GPUs, DeepSeek hobbled together a small cluster of slower, older Nvidia H800 GPUs and for about $6 million, it was able to train its open-source, open-weight DeepSeek-R1 model, which bested the state-of-the-art models from OpenAI and Meta at the time in some key benchmarks.

By using smaller, specialized models to work together, using a technique known as “mixture of experts,” DeepSeek’s reasoning model sent a shock through the industry — maybe everyone was going about it wrong. Does every AI company really need to buy tens of thousands of Nvidia’s latest GPUs? Do all the Big Tech companies really need to be spending hundreds of billions in capital expenditure on bigger and bigger data centers?

The stock market was throttled as the industry wrapped its collective head around the news. About $1 trillion in market value was wiped out, including eye-popping 17% single-day drops in Nvidia and Broadcom and a 4% decline in Google.

The breakthrough caused AI leaders to point fingers and question DeepSeek’s transparency. Elon Musk joined Meta’s new AI wunderkind, Alexandr Wang, in accusing DeepSeek of using banned GPUs.

Some tech execs, like Microsoft’s Satya Nadella, took comfort in the Jevons Paradox, welcoming lower costs and greater efficiencies to turn cheap AI computing into “a commodity we just can’t get enough of.” But reports emerged that Microsoft and its partner OpenAI had evidence that DeepSeek used ChatGPT to train its R1 model.

At Meta, DeepSeek-R1’s performance reportedly sent execs into a panic, fearing that their in-development Llama 4 model would not perform as well as DeepSeek, prompting an internal race to implement “reasoning” in Llama, like DeepSeek-R1. Months later, after a bungled, incomplete launch of Llama 4, CEO Mark Zuckerberg would bet the farm on building a new AI all-star “superintelligence” team from scratch.

In the months that followed, the “Sputnik moment” of DeepSeek-R1’s release caused the startup’s much larger competitors to adjust their strategies — despite the fact that the model parroted Chinese communist propaganda.

OpenAI had released its o1 model in September 2024, well before DeepSeek-R1’s release. But shortly after DeepSeek came out, the company said its new gpt-4.5 would be its last non-reasoning model. Since then, OpenAI has followed DeepSeek’s lead and released its own open-weight model, gpt-oss.

Musk’s xAI quickly announced it would be adding DeepSeek-like reasoning to its Grok 3 model. And Google’s Gemini 2.5 Pro was positioned as its reasoning model, as the entire industry adopted the approach.

DeepSeek also cemented the importance of capable, free, open-source, open-weight models. These kinds of models are being distilled down to create specialized, smaller models, which may take the place of behemoth frontier models going forward.

The markets, for their part, recovered quickly: after falling hard on January 27, many of the affected tech stocks made up for lost ground, with the Nasdaq 100 having erased its losses just days later. And it wasn’t just a recovery — since the DeepSeek freak-out, Nvidia is up 60%, Broadcom is up 65%, and Google is up 75%. OpenAI’s valuation has swelled to $500 billion as of December.

Looking back, it’s clear the release of DeepSeek-R1 really did change everything. It put enormous competitive pressure on a fast-moving industry, and it forced tech juggernauts to scramble plans, reimagine their AI offerings, and rethink how models could be trained.