Cover Image Made With AI

Why We Bet Against Nvidia Last Friday

In our post last Friday, we laid out our bearish thesis for Nvidia, Inc. (NASDAQ:NVDA):

A Chinese start-up created its own competitor to Open AI's ChatGPT, called DeepSeek, and some of the smartest people in tech, like Steve Hsu and Marc Andreessen are extremely impressed with it. The reason this seems bearish for Nvidia is the Chinese apparently accomplished this with only a few million dollars in capital, meaning they didn't need lots of expensive chips of the sort Nvidia sells. And if they don't need them, maybe everyone else doesn't need so many of them either.

What About Jevons Paradox?

Over the weekend, others came to the same conclusion, but some, including Microsoft Corporation (MSFT) CEO Satya Nadella, brought up Jevons Paradox.

You know already that Nvidia shares tanked 17% on Monday (which enabled us to exit half of our puts from Friday for a 2,633% gain), and then the stock bounced back almost 9% on Tuesday. On Tuesday night, Steve Hsu addressed Jevons Paradox on X:

Re “Jevons paradox” and NVDA valuation

This is a glib response which does not take into account timescale mismatch. The people buying NVDA chips today only have a few years before the value of the chips depreciates by 50% or more. They have to make ROI on their chips in the near term. But actual revenues flowing to genAI are not very big right now, and if much of this is captured by free open source models, which don’t necessarily require NVDA chips to run (see Groq or even legacy hardware for distilled models), and are ~30x more efficient, then the investment today in NVDA chips may turn out to have low ROI. Similarly, if everyone is as efficient as DS in model training the NVDA chips bought today may be more than enough for this training.

I can imagine that in the next few years DS V3 and R1-like models (including other highly optimized Chinese models like Qwen or from ByteDance) are widely used in AI applications. But these models need ~1/30 the compute for inference, so the demand over the next few years for NVDA chips sold today could be much less than people anticipate, even with healthy growth in adoption of genAI in schools, workplaces, web search, etc. (30x is hard to make up!). On top of this if model training is also much more efficient (a la DS), and maybe data-limited for pretraining, the model training component of demand for NVDA chips could also be much smaller than expected.

I’ll be shocked if in a few years the NVDA advantage in genAI hasn’t mostly disappeared. It isn’t that hard to design competitive chips for LLM-transformer computations. Many entities from GOOG to AMD to Huawei etc. are doing it. NVDA have the CUDA software library lock-in (ex-PRC, where people may be forced to switch to the HW Mindspore ecosystem), but that is also going away over time.

I’m at an AI for CX (customer service, call centers, etc.) meeting rn. Altho CX is one of the best targets for genAI replacement of human labor, there is tons of friction and hesitance to deploy from human decision makers. It will take years before ~50% of human FTEs are replaced by AI, even tho it’s clearly possible to do it.

That last paragraph goes to why Steve Hsu is worth paying attention to here. He understands both the science behind AI, as a Caltech-trained, PhD physicist, and he also understands the business aspect of it, as the founder of an AI business.

Portfolio Armor's Take

All else equal, our automated system gives preference to stocks that have suffered a short term decline, because, historically those names have outperformed. Despite that, Nvidia wasn't one of our system's top names on Tuesday or Wednesday. Not only that, but gauge of options market sentiment was neutral on it, rather than bullish. Not a good sign.

Timing Our Trades

If Steve Hsu is right about declining demand for Nvidia GPUs, we might get a hint of that in next month's earnings call, but we'll probably see a bigger impact two quarters out.

• February 2025 Guidance: Possible early hints (e.g., softer Q1 FY2026 outlook, margin concerns), but not a full collapse.
• Stock Reaction: If investors believe Hsu's thesis, NVDA could sell off even on slightly cautious guidance, as the market prices in longer-term risks.
• Inflection Point: The August 2025 earnings call (FY2026 Q2) is when Hsu's bear case would more clearly hit guidance.

In short: The February 2025 call could be the canary in the coal mine, but not the full explosion.

So our plan here is to place a moderately bearish bet on Nvidia expiring at the end of February, and then a more aggressive bearish bet against it expiring in September, after Nvidia's August earnings report. If you are subscribed to our trading Substack, you can check your email for the specific trades. If not, you can subscribe below. 

And if you just want to limit your risk, you can download our optimal hedging app by aiming your iPhone camera at the QR code below, or by tapping here, if you are reading this on your iPhone. 

If you’d like to stay in touch

You can scan for optimal hedges for individual securities, find our current top ten names, and create hedged portfolios on our website. You can also follow Portfolio Armor on X here, or become a free subscriber to our trading Substack using the link below (we’re using that for our occasional emails now).