Shockwaves in AI

Before delving into our usual Technical and Fundamental analysis, let’s explore today’s biggest price action driver across global markets - the release of a revolutionary Chinese AI model called DeepSeek-R1.

Before the announcement last week , it was a widely held belief that training cutting-edge AI models was an extraordinarily expensive endeavor. Companies like OpenAI and Anthropic spent around $6.5 billion in 2024 on computing power alone. To train their models, they rely on massive data centers filled with thousands of NVDA GPUs, each costing between $35,000 and $70.000 (B100 and GB200 series). Plus a massive team of engineers, with massive salaries meant that AI development was a game reserved for tech giants — US based in particular, since they had access to the best hardware.

This gave rise to ever mode capable models, like OpenAI's o1 model (priced at $200 / month). All was fine and dandy until DeepSeek-R1, a fully open source model, has been reported to perform on par with OpenAI's o1 model, for a fraction of the cost.

According to preliminary information, the model runs on older NVIDIA H800 processors (not top of the line by any means, but certainly capable enough) and the training cost just $6 million. In the future, with this kind of efficiency, industry analysts have speculated that a handful of gaming GPUs might be enough to run the model. If true, this might be a “Minsky Moment” for the AI industry. The paradigm appears to be shifting in such a way that advanced software engineering becomes far more valuable than ultra-powerful GPUs.

Of course, this is a major problem for NVIDIA (and other AI infrastructure companies). Their business case (and resulting valuation) relies on 2 main assumptions: sustained and accelerated revenue growth + very high margins. Below, we can see the revenue growth story (from 6.1B in Q4 2022 to 35.1B latest) paired with 72% average gross margins over the last 8 quarters (74.56% most recent). If AI development can be done effectively on gaming-grade hardware, the demand for premium data center GPUs would plummet, and so would the value of NVDA stock. This is the narrative that’s crashing markets in Europe and Asia at the moment.

The issue (and the opportunity) for investors is that US equity valuations are highly dependant on the tech sector, especially where AI and semiconductor-related businesses are concerned. For perspective, the Price to Sales of AI-related top US stocks is about twice that of non-AI related (and yes, AAPL is by no means “AI related” in our view). These valuations are not only based on sky high expectations, but on partially “gatekept” technology and internal processes (Blackwell chips are banned for sale in China).

There is both truth and some knee jerk selling associated with the market’s reaction. First of all, DeepSeek has been available since Monday, January 20, and no one batted an eye last week. It was only the recent testing over the weekend which caused a major reaction.

Second, it isn’t totally true that the model can run on gaming hardware, at least not at the moment. Minimum requirements for the full 671B parameter model include an AMD EPYC or Intel XEON processor with a LOT of threads, 48 GB of RAM, and 1.342 GB of VRAM on the GPU side (equivalent of 16 NVIDIA A100 GPUs with 80 GB each). That’s not exactly the type of hardware that even a pro-gamer has lying around in his living room.

However, renting such a setup from a cloud provider costs just $8.500 / month, a far more accessible price point than many believed possible for such performance. This suddenly makes AI research and operation far more democratic.

Where it gets interesting is the “Distillation” process. Without getting into too many specifics, the full model can train other sub-models, depending on the task at hand. On those tasks, these smaller models (distilled) outperform the established players like Chat Gpt and Sonnet. And running a distilled model can be done on a high-end gaming computer rather than a server-grade configuration.

Bottom line: this is what evolution, disruption and human ingenuity looks like. Incumbents like OpenAI, Meta and Anthropic have optimized existing processes, with the intention of keeping the lion’s share of profits for themselves. Without access to the same level of funding and with limited hardware, engineers at DeepSeek have nevertheless circumvented these barriers through innovation and efficiency. This is entirely plausible, according to our own lived experience here at Signal Sigma.

Now, for the more pressing question this morning: is a -2% selloff in the S&P 500 justified? For SPY, short term support resides at the 50-DMA ($596), about -2% below the last close. Holding this level of support by the day’s close would be a bullish sign. Breaking below it would turn the focus to the previous level of support, the 100-DMA, -3.91% lower.

After investors digest the implications of this new tech, the price action will revert to the usual drivers, namely share buyback and positive earnings announcements. There is nothing inherently “bad” in new tech, at least not on the economic or on the inflation front — which is the real problem for the current bull market.

On Wednesday, the Fed is expected to announce no change to their target rate and a new “dot plot”. We believe it’s entirely unlikely we’ll hear mr. Powell mention DeepSeek in his press conference. However, we’ll share some technical levels for SPY, NVDA, XLK and QQQ derived from our options analysis coming soon to the platform. On the latest market close day, options volume for NVDA was actually 57% short term bullish so the reversal is certainly a surprise.

Now is as good a time as any to remember that corrections are part of investing. Among the top 10 S&P 500 companies (with the exception of Apple), the mean drawdown to the 50-DMA is -5.65%. A retracement to the 50-DMA can happen for a number of different reasons — including “no reason whatsoever”.

The most likely scenario by the day’s close is a crash in sentiment to previous low values. This gives us the opportunity to analyze IF support holds and WHICH assets “stay afloat” during the sell-off. As mentioned before, new technology is rarely an issue for the economy. If anything, more accessible AI is a boon for a variety of industries and should lead to further growth.

The question now becomes where capital rotates, but we’ll need to be patient in order to see that. It’s unlikely that industries like Financials or Healthcare will see a larger than normal correction due to DeepSeek.

Our Trading Strategy

For the moment, we’ll hold allocations until the dust settles. This week will be full of earnings releases and companies like MSFT, META and AAPL will probably be grilled by analysts on their AI initiatives.

Furthermore, economic data like Core PCE will give us more information on the inflation front on Friday.

Looks like a wild ride until then, but our job as investors is pretty simple – protect our investment capital from short-term destruction so we can play the long-term investment game.

Signal Sigma PRO members will be notified by Trade Alert of any live portfolio changes (if subscribed). If you’re not on this plan yet, you can get a free trial when you join our Society Forum. If you need any help with your trading strategy (or would like to implement one on your account), feel free to reach out!