Nvidia (NASDAQ:NVDA) stock rampaged across the market, rising over 900% over the last two years to hit a $3.5 trillion valuation on insatiable demand for its advanced artificial intelligence chips and CUDA software ecosystem.
Its next-generation Blackwell chips are set to upend the market once again with hyperscalers like Amazon (NASDAQ:AMZN), Alphabet (NASDAQ:GOOG)(NASDAQ:GOOGL), Meta Platforms (NASDAQ:META), and Microsoft (NASDAQ:MSFT) spending $10 billion or more each to buy up the supply for the entire year.
Yet an announcement out of China could disrupt the entire landscape. That should worry Nvidia investors as the chipmaker’s stock is premium priced based on the likelihood of continued meteoric growth. The whole investment thesis, however, could come crashing down as the latest AI development shakes the foundation upon which Nvidia was built.
24/7 Wall St. Insights:
- Although Nvidia (NVDA) is far and away the leading player in the AI market, a new development out of China could completely upend the landscape.
- Chinese AI lab DeepSeek unveiled a new LLM that is superior to what is on the market, and does it more efficiently and at lower cost.
- Because it is also open-source, it will allow users to tinker with the algorithm to fine-tune it and build on it.
- If you’re looking for some stocks with huge potential, make sure to grab a free copy of our brand-new “The Next NVIDIA” report. It features a software stock we’re confident has 10X potential.
Making do with less and doing more
In December, Chinese AI lab DeepSeek released its large language model (LLM) V3 that was equal to OpenAI’s GPT-4o and Amazon-backed Anthropic‘s Claude Sonnet 3.5. It was able to outperform Meta’s Llama 3 as well as Alibaba‘s (NYSE:BABA) Qwen2.5. Now DeepSeek released R1, and it surpassed ChatGPT’s latest model, o1.
The reason this is devastating news to Nvidia is that not only is DeepSeek R1 open source, meaning users can modify its algorithm, fine tune it, and build on it, but it is reportedly 27 times cheaper than ChatGPT.
Worse for Nvidia, the state-of-the-art V3 LLM was trained on just 2,048 of Nvidia’s H800 GPUs over two months, equivalent to about 2.8 million GPU hours, or about one-tenth the computing power that Meta’s comparable Llama 3.1 model took to train, according to Epoch AI. Meta’s LLM required 30.8 million GPU hours on 16,384 H100 GPUs.
The H800 chip differs from the H100 in that Nvidia significantly reduced chip-to-chip data transfer rates to get around U.S. export restrictions to China. That means the H800 is less powerful than the high-octane H100 chip. The H800 was subsequently banned for export, too.
An end run around restrictions
Make no mistake, DeepSeek still wants Nvidia’s full-fledged chips, but according to Alexandr Wang, CEO of Scale AI, it reportedly has access to 50,000 H100 chips. He told CNBC, “they can’t talk about it obviously because it is against the export controls that the United States has put in place,” but he also believes “they have more chips than other people expect.”
Still, a superior AI model, that’s better, faster, more efficient, and cheaper than current U.S. models is not good for Nvidia’s growth trajectory. Future AI models could very well not need to rely upon vast computational resources.
For example, last year Oracle (NYSE:ORCL) announced the world’s first zettascale AI supercomputer in the cloud, which would be powered by up to 131,072 Blackwell GPUs. But an open-source AI model that incorporated FP8 mixed precision training and leveraged low-level PTX instructions specific to Nvidia CUDA GPUs could probably make do with a lot fewer GPUs.
FP8 mixed precision training refers to a technique of deep learning that balances computational efficiency and accuracy. The PTX instructions are the fundamental building blocks of code that directly interact with the GPU hardware.
Nvidia’s outlook just got cloudy
DeepSeek R1’s efficiency could lead to a decrease in demand for Nvidia’s high-end GPUs like the H100s, especially if more entities follow DeepSeek’s approach of optimizing software and algorithms to work within hardware constraints.
However, it’s not all bad. DeepSeek’s success with open-source models might democratize AI development, potentially increasing the number of entities interested in GPU computing. This could mean more sales for Nvidia as smaller companies and even individual developers might purchase GPUs to run or fine-tune open-source models like those from DeepSeek.
Key takeaway
This new development is going to require Nvidia to innovate more in both hardware design and software optimization to maintain their AI market lead. If competitors can achieve similar results with less capable hardware, NVIDIA might need to focus on making their GPUs even more efficient or versatile for various AI workloads.
The biggest risk to NVDA stock is the market’s perception of its lofty valuation in light of what could be the dawn of much lower-cost data center projects. Nvidia may no longer be seen as worth paying 56 times earnings and 30 times sales if more can be done with less.
Travel Cards Are Getting Too Good To Ignore (sponsored)
Credit card companies are pulling out all the stops, with the issuers are offering insane travel rewards and perks.
We’re talking huge sign-up bonuses, points on every purchase, and benefits like lounge access, travel credits, and free hotel nights. For travelers, these rewards can add up to thousands of dollars in flights, upgrades, and luxury experiences every year.
It’s like getting paid to travel — and it’s available to qualified borrowers who know where to look.
We’ve rounded up some of the best travel credit cards on the market. Click here to see the list. Don’t miss these offers — they won’t be this good forever.
Thank you for reading! Have some feedback for us?
Contact the 24/7 Wall St. editorial team.
Latest from 24/7
Our expert who first called NVIDIA in 2009 is predicting 2025 will see a historic AI breakthrough.
You can follow him investing $500,000 of his own money on our top AI stocks for free.