DeepSeek-R1 crashed Nvidia's stock and forced everyone to rethink AI economics
On January 20, 2025, a Chinese AI lab called DeepSeek published a model called R1 on GitHub, free for anyone to download. By January 27th, it had become the most downloaded free app on the US App Store, and Nvidia had lost more than 17% of its market value in a single trading session. That is roughly $593 billion in market capitalisation erased in one day. For context, that is more than the entire value of most Fortune 100 companies.
The reaction was not irrational panic. It was the market repricing a core assumption it had held for three years: that building frontier AI requires extraordinary amounts of money, proprietary hardware, and years of infrastructure investment. DeepSeek-R1 appeared to challenge all three of those assumptions simultaneously.
What DeepSeek-R1 actually is
R1 is a reasoning model, meaning it is designed to think through problems step by step before producing an answer rather than generating a response in a single forward pass. It was trained using reinforcement learning, with the model receiving rewards for producing correct answers and exploring different reasoning paths during training.
What surprised researchers was not just its performance but its cost. DeepSeek reported training the model for approximately six million dollars. For comparison, GPT-4 was estimated to have cost more than one hundred million dollars to train. OpenAI and Anthropic had built entire narratives around the idea that frontier AI requires nine-figure compute budgets. DeepSeek published a technical report showing you could get within striking distance of those models for a tiny fraction of the cost.
R1 uses mixture-of-experts: only a fraction of parameters fire per token, cutting inference cost:
DeepSeek also squeezed memory bandwidth during training — the real bottleneck under export-controlled GPU access.
The numbers that shook the market
Benchmark comparison
R1 performed at or near the level of OpenAI o1 across standard reasoning benchmarks. It scored particularly well on mathematical problem solving and competitive coding tasks.
Why the Nvidia drop made sense
Nvidia's entire premium valuation rested on a simple story: training frontier AI requires enormous amounts of GPU compute, and Nvidia makes the best GPUs. If the cost of training frontier AI compresses by 10 to 20 times, you need fewer chips per model. The addressable market for Nvidia's H100 and H200 data centre cards shrinks significantly, at least for training runs.
This does not mean demand for GPUs disappears. Inference at scale still requires hardware, and more capable models deployed to billions of users consume enormous amounts of compute. But the story that every AI lab needs to spend hundreds of millions on training hardware every few months was suddenly less convincing.
The Stargate Project, a joint venture between OpenAI, SoftBank, and Oracle, was announced by President Trump on January 21st with a commitment of up to $500 billion in AI infrastructure investment. One week later, DeepSeek made the case that you might not need it.
The geopolitical dimension
DeepSeek is a Chinese company operating under export controls that restrict access to the most advanced Nvidia chips. The US government had assumed that chip restrictions would constrain Chinese AI development. R1 suggested the opposite: restrictions forced DeepSeek to optimise harder than labs with unlimited hardware budgets, producing more efficient training methods that then benefited the global open-source community.
This is the version of events most cited in policy circles, though it is worth noting that DeepSeek likely had access to older H800 chips in significant quantities and has not disclosed its full hardware inventory. The efficiency gains are real, but the exact extent of the resource constraint is not fully confirmed.
What it means for AI access
Because R1 is open weights and available for download, any developer with a reasonable server can run it. Research labs in lower-income countries that could not afford $10 per million tokens for API access suddenly had a model that rivals the best closed systems. Universities could run it on their own compute. Companies could fine-tune it on proprietary data without sending that data to an external API.
The immediate practical consequence is that the cost of integrating frontier-quality reasoning into applications dropped dramatically. The longer-term consequence is a shift in how the industry thinks about the relationship between training compute and model capability. January 2025 was the month that the assumption of inevitable scaling costs started to crack.