Nvidia weekly big dump 10% led to the collapse of US stocks and BTC, what did DeepSeek Open Source release this week?

Written by Luke, Mars Finance

introduction

This week, the global financial markets experienced a severe shock. Bitcoin, as a barometer of high-risk assets, collapsed under the impact of the plunge in the US stock market, with prices falling to $78,000 at one point. The catalyst for this ‘avalanche’ was the collective slide of US technology stocks, with NVIDIA at the epicenter. The AI chip giant saw its stock price plummet by 10.2% in the first week, causing its market value to shrink by over $300 billion. This directly dragged down the S&P 500 index by 4.7% and the Nasdaq 100 index by 5.9%. The rise and fall of Bitcoin is just a superficial phenomenon. In order to explore the underlying causes of the collapse in the US stock market, our attention has to turn to a new variable: DeepSeek released five core technologies during ‘Open Source Week’, claiming to achieve a 3x increase in computing power efficiency on existing hardware, leaving the market full of uncertainty about the future of chip monopolies.

NVIDIA Flash Crash: The “Waterloo” of the Hardware Empire

NVIDIA’s sharp decline is not accidental, but the concentrated outbreak of multiple pressures:

Performance warning: The agency predicts that Nvidia’s data center revenue growth rate in Q1 2025 will slow from 75% to 48%, and the market’s expectations for hardware demand will cool rapidly.

The shadow of technological substitution: In the same week, the Chinese AI newcomer DeepSeek held a high-profile ‘Open Source Week’ and released five software technologies, claiming to triple the computational efficiency on existing GPUs. This not only shakes the commercial logic of ‘stacking chips’ by NVIDIA, but also makes investors smell the ‘de-NVIDIA-ization’ trend.

Behind this stock price storm, the balance of AI computing power competition is tilting - from the “brute force era” of hardware stacking to the “outwitting era” of software optimization.

DeepSeek open source week: five punches of “software-defined computing power”

DeepSeek’s five open source technologies are not simply code optimizations, but a thorough ‘software computing power revolution’, redefining the relationship between computation, communication, and storage, making AI training and inference no longer completely dependent on hardware upgrades.

1. FlashMLA: Use the GPU as an “intelligent pipeline” to squeeze out performance

Traditional AI computing is like a manual kitchen, with a large number of tasks that need to be assigned manually, resulting in waste and waiting. The FlashMLA optimization solution is more like an “intelligent pipeline”, so that text tasks of different lengths can be accurately scheduled by the GPU, short tasks can be processed quickly, and long tasks do not waste resources.

Breaking point: The computing performance of the H800 graphics card has been increased to 580 TFLOPS, compared to 220 TFLOPS with traditional methods.

Impact: The number of GPUs required for the same AI task is reduced by 60%, which directly affects the procurement requirements of cloud computing vendors.

2. DeepEP: Make GPU communication more like a ‘5G high-speed highway’

In large-scale AI model training, the bottleneck of computing power is often not computing, but communication delay. DeepEP uses FP8 compression + RDMA technology, which is equivalent to building a “5G highway” between GPUs, and the data flow is more efficient.

Performance data: The cross-node communication bandwidth is increased to 150 GB/s, and the latency is reduced by 83%.

Impact: Server cluster size can be reduced by 40%, reducing reliance on NVIDIA InfiniBand network devices.

3. DeepGEMM: The “multi-functional gas stove” of AI computing

DeepGEMM optimizes the matrix calculation method, which is equivalent to installing a “smart gas stove” on the GPU, which can dynamically adjust the firepower according to different tasks.

Efficiency improvement: The computing speed is increased by 2.3 times and the power consumption is reduced by 55% under FP8 precision.

Industry impact: Nvidia Tensor Core can be replaced in some scenarios, and some companies have replaced A100 orders with H800 + DeepGEMM combinations.

4. 3FS File System: “Intelligent Warehousing” of AI Data

Storing data is crucial for AI training, and the 3FS file system is like an ‘automated warehouse center’ that can instantly access massive data instead of manually searching for files as in traditional methods.

Throughput: The read and write speed is 6.6TB/s, which is 12 times faster than the Lustre file system.

Impact: The training data preprocessing time is shortened by 70% and the GPU investment requirement is reduced by 35%.

5. The “Snowball Effect” of Open Source Ecology

DeepSeek’s open source strategy is like a snowball. GitHub data shows that its open source library has been downloaded more than 1.2 million times a week, of which 30% are from European and American developers. This means that developers around the world are rapidly adapting to these optimizations, further eroding NVIDIA’s control over the software ecosystem.

From “stacking chips” to “picking code”: reshaping the AI industry landscape

In the past few years, the logic of AI computing power has been that “the silicon process is everything”, but DeepSeek’s Open Source Week showed that software layer optimization can be exponentially improved without upgrading hardware. This has led to a change in the valuation logic of the entire AI industry.

NVIDIA’s dilemma: The company has invested over $12 billion in the development of the Blackwell architecture GPU, but the closed CUDA ecosystem has become a problem of ‘locking in customers’, leading to a surge in software optimization expenditure from 15% to 40% in the AI budgets of companies such as Meta and Microsoft in 2025.

The rise of new forces: The parent company of DeepSeek has surged in value by 300% in three months, reaching $72 billion, surpassing Stability AI.

Global Developers’ Choice: GitHub statistics show that 27% of global AI projects use DeepSeek components, surpassing PyTorch Lightning.

Can NVIDIA be too big to fail? How will the US stock market digest this negative news?

Nvidia’s plunge is not only a turmoil in the AI industry, but also related to the stability of the entire U.S. stock market. As the leader in the market capitalization of the technology stock market, every sharp fluctuation of Nvidia triggers a chain reaction in the market. So, how should U.S. stocks digest this bearishness?

In the short term, the market may experience a period of outflow of safe-haven funds and a pullback in technology stocks, especially a reevaluation of the valuation of the AI industry.

In the medium to long term, NVIDIA remains an important pillar of the AI industry. Despite the challenge of software optimization to hardware upgrades, GPUs are still the core for training and inference of large models. As long as the market regains confidence, NVIDIA is still expected to stabilize again.

The recovery of the bitcoin market may depend on the stabilization of US stocks. Once the selling pressure on technology stocks eases, the market’s risk appetite recovers, and liquidity returns, Bitcoin, as the “digital gold”, is expected to usher in a rebound.

The future AI world is no longer just a hardware competition, but an era of software-defined computing power. In this day and age, companies that can use code to “turn stones into gold” will go further than those obsessed with building bigger hammers.

DeepSeek’s open source week is just the beginning, a whole new AI computing power landscape is quietly reshaping the world.