A Chinese startup's release of an open-source reasoning model in January 2025 set off one of the sharpest single-day shocks in stock-market history, erasing roughly $589 billion from Nvidia's market value in one session and forcing a reckoning over how much computing power advanced AI truly requires.
January 2025 · DeepSeek-R1
A $589 Billion AI Reckoning
A Chinese startup's open-source reasoning model matched leading US systems at a fraction of the cost — triggering the largest single-day loss of value for any US company on record.
$589B
Nvidia market value erased in one session — a record single-day loss
-17%
Nvidia share price drop on January 27, 2025
$5.6M
Reported GPU-time cost to train the underlying V3 model
Price per million tokens — undercutting the field
R1's listed API pricing on input vs. output tokens.
Output tokens priced ~4× input — but the whole schedule sits far below comparable offerings.
Efficiency over brute force
Mixture-of-Experts
Only a portion of the model activates per input, cutting compute needs.
RL-heavy pipeline
Reinforcement learning used to elicit multi-step reasoning behavior.
Export-limited H800s
Trained on performance-crippled chips built for the Chinese market.
Distilled versions ship from 1.5B to 70B parameters , letting developers run smaller models on modest hardware.
✓ Impressed
Praised for math, coding and multi-step reasoning; open MIT-licensed weights welcomed by teams wary of closed APIs. Matched or exceeded OpenAI's o1 on several benchmarks.
! Caveats
Early R1-Zero showed repetition, poor readability and mixed-language output. Some tests found it trailing o1 or producing overly long reasoning chains; the largest models still demand heavy GPUs.
The open debate
Do efficiency gains reduce demand for accelerators — or merely redirect it?
Nvidia called R1 an "excellent AI advancement," arguing inference-heavy reasoning models will ultimately require substantial computing power. Either way, a well-executed open-weight release from outside the US moved markets and shifted competitive assumptions overnight.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…