Editor's Note: The following research note has been made available to the public at no cost. If you're interested in subscribing for full access to this research, click here.
********
We're removing our NVDA Long. We went Long in April 2024 at $87. NVDA has soared 63% since then and outperformed the QQQ index by 40% over the same time frame.
Deepseek is the main reason why we're less confident on Nvidia's growth prospects and on the AI revolution. The list of concerns for NVDA has been growing since late 2024, but Deepseek is the turning point, in our opinion. Deepseek is the surprising and mysterious Chinese AI company founded by Liang Wenfeng, who was formerly a portfolio manager managing $8bn in assets at his quant fund, High-Flyer. His team started using NVDA GPUs back in 2021. Deepseek spent only ~$50m on 2,048 Nvidia H800s (not even H100s) and only $5.576m in training costs on its DeepSeek V3 model, which outperforms the best LLM models in the world, including those from OpenAI, in some categories (e.g. quantitative reasoning, coding). Put simply, Deepseek defies the logic that only hyperscalers can achieve the best AI models in the world by spending >$600bn, including ~$325bn in 2025 capex alone.
This is bad news for NVDA, OpenAI and AI names. It suggests NVDA's highest performance, most expensive chips ≠ best AI model. Will NVDA's customers' torrid pace of investments in NVDA's AI chips change following Deepseek's announcement? As for OpenAI, why pay OpenAI 50x more to run on its o1 reasoning model when Deepseek can achieve as robust of a LLM? The focus will turn to low-cost innovation, as we transition to inference. Fears of commoditization of GPUs and LLMs will gain traction.
V3 Comparison:
- Deepseek V3's biggest model is 671B parameters (GPT uses 1.8T parameter).
- V3 training uses 30% less memory and achieved 14.8 trillion token pre-training in just 2.788m GPU hours (vs Llama 3.1's 30.8m hours)
Deepseek is free and open-sourced and available on the app stores. On Friday, the company introduced a reasoning model codenamed DeepThink (R1), which outperformed OpenAI's o1 reasoning models , according to some benchmarks. We think R1 really ignited the Deepseek debate/conversation, and consequently, most tech stocks reacted negatively on Friday.
Supposedly, Deepseek learned and trained from a Llama (META) model via a distillation technique (when a smaller model learns from a larger model). Supposedly, Meta's engineers are currently working around the clock to learn from Deepseek, according to Blind (see below).
Deepseek's popularity is also growing - it was the #1 downloaded app in both the US and China last night.
One of the key distinctive features we noticed while using DeepThink R1 is that the model lays out step by step to the user how it thinks (like someone explaining each step of his/her reasoning process). This transparency may increase the app's stickiness because it gives the user helpful context on the output generation.
Deepthink has received praise from key leaders in the AI world including:
- Marc Andreessen (key investor in OpenAI): "Deepseek RI is one of the most astonishing and impressive breakthroughs I have ever seen, and it is open-source, a great gift to the world"
- Jack Clark (former head of policy at OpenAI): "Deepseek hired a bunch of (young) enigmatic wizards
- Meta's chief AI scientist, Yann LeCun: "Open-source models > proprietary models"
- Nadella (CEO of MSFT): "Super impressive in terms of both how they have really effectively done an open-source model that does inference time-compute, and is super-compute efficient"
- Support from a flurry of research scientists and developers around the world
With Deepseek's incredible breakthrough, questions (and conspiracy theories) are popping everywhere: 1) How can you prove the $6m training cost - is it made up?; 2) Are the US chip restrictions backfiring?; 3) Are the US export control rules failing?; 4) Did Deepseek receive help or guidance from the government or other Chinese tech giants?; 5) Will Trump ban Deepseek and/or ban all chip exports from NVDA to China? 6) Was the Deepseek timing a coincidence given it was right after the Stargate announcement?
While most investors are focused on NVDA's earnings in late February and GTC in mid-March, we think Deepseek will be the #1 concern in the near-term. We're resetting our 2025/2026 growth and multiple expectations for NVDA and will look further into Deepseek and other debate points soon. Stay tuned!