DeepSeek’s Breakthrough: A Game-Changer for Nvidia’s Dominance in the AI Arms Race

DeepSeek’s Breakthrough: A Game-Changer for Nvidia’s Dominance in the AI Arms Race

DeepSeek, a Chinese startup, claims to have found a way to train models at GPT-4-level performance for just a fraction of the usual cost—anywhere from 1/20th to 1/45th of what others spend. They’ve reportedly reached this level with just $5-6 million, while competitors typically spend over $100 million. They’ve made all of this public, allowing others to assess their approach. The key to their low-cost success seems to be innovations like Mixture of Experts (MoE), multi-token steps, compression, and chain-of-thought methods, which reduce the need for extensive GPU resources during training and inference. If this holds true, it could completely change the AI landscape.

DeepSeek was created in less than 2 months for under $10 million and is now the top app on the App Store. It was built with outdated chips and a small team of under 200 people. In contrast, the US is investing $500 billion in AI, and OpenAI has about 4,500 employees as of November 2024. DeepSeek managed to build something almost as advanced as OpenAI, even though OpenAI has 22 times more staff.

DeepSeek’s Breakthrough: A Game-Changer for Nvidia’s Dominance in the AI Arms Race

So, why does this matter for Nvidia? Nvidia’s bull case has been that AI’s explosive growth will drive an insatiable demand for GPUs, especially with massive investments coming from major companies. Nvidia has capitalized on this trend, with profit margins over 70%, thanks to its high-end hardware and exclusive software ecosystem (like CUDA and drivers) that no one else can match at scale. But if DeepSeek’s approach drastically reduces the GPU power needed—by as much as 10 times—it could raise serious questions about whether companies are over-investing in GPUs. If big tech giants like Meta or Microsoft start cutting back on their GPU orders, even a small reduction in capital expenditure could hurt Nvidia’s stock price, which is currently valued at a high multiple.

Some are calling DeepSeek a “black swan” event, meaning it could unexpectedly disrupt Nvidia’s growth trajectory. If it raises doubts about the ROI of AI or causes companies to scale back their AI investments, it could severely impact Nvidia. On the other hand, some argue that these cost-saving innovations will make AI more affordable, which could fuel greater AI adoption and ultimately boost GPU demand. This is where Jevons Paradox comes into play—lowering costs might lead to higher overall GPU demand as AI becomes more widely used.

The question of AI ROI is crucial at this point. We’ve seen massive spending on GPUs without clear signs of returns. If DeepSeek’s approach proves successful, it could force a reevaluation of whether we need to spend $200 billion or more on new hardware to achieve similar results at a much lower cost. This could either puncture the current hype around AI or, ironically, lead to even more widespread AI adoption. At the very least, it will force the industry to rethink just how much capital is truly needed and whether Nvidia can continue to maintain its near-monopoly margins.

Liang Wenfeng: The Founder of DeepSeek

Early Career: Liang Wenfeng studied machine vision at Zhejiang University.

Hedge Fund Success: At 30, he started High-Flyer, a quant hedge fund, in 2015. The fund now manages $8 billion, making him wealthy.

AI Ambitions: Liang aimed to create human-level AI as a side project, though his partners were initially doubtful.

Building the Foundation: In 2021, he purchased 10,000 Nvidia H800 GPUs and hired top experts from his hedge fund to work on Nvidia hardware optimization.

Launching DeepSeek: In 2023, he launched DeepSeek, hiring dozens of PhDs from China’s top universities like Peking, Tsinghua, and Beihang, offering competitive salaries similar to Bytedance.

Overcoming Restrictions: Due to U.S. export bans, DeepSeek innovated by developing affordable AI training methods that compete with OpenAI, Anthropic, and others at 1/20th the cost.

Efficiency & Controversy: Their training methods are highly efficient, but there are rumors about potential copying, undisclosed GPU clusters, or possible government support.

Open Source: DeepSeek shares its methods publicly, with a research paper on R1 co-authored by over 200 researchers.

Big Achievement: DeepSeek’s AI app is now the #1 free app on the Apple App Store.

After DeepSeek’s announcement, the Nasdaq fell more than 550 points. In premarket trading, NVIDIA (NVDA) is down by 8.06%, followed by MicroStrategy (MSTR) down 7.12%, and Coinbase (COIN) falling 5.52%. AMD is down 4.86%, while Meta (META) drops 4.28%. Microsoft (MSFT) sees a 3.46% decline, and Amazon (AMZN) falls by 3.43%. Tesla (TSLA) is down 3.30%, Google (GOOGL) drops 3.22%, and Apple (AAPL) declines 1.81%.

Cantor Fitzgerald on DeepSeek V3 and $NVDA:

“The launch of DeepSeek’s V3 LLM has sparked concerns over GPU demand and fears of peak compute spending. We believe these worries are unfounded. The announcement brings AGI closer to reality, with Jevons Paradox likely driving higher demand for compute. Any NVDA weakness is a buying opportunity.”

Update

DeepSeek has limited registration to its services, allowing only users with a mainland China mobile phone number to sign up.

3 thoughts on “DeepSeek’s Breakthrough: A Game-Changer for Nvidia’s Dominance in the AI Arms Race”

Leave a Reply

Your email address will not be published. Required fields are marked *