DeepSeek disruption, adjustment or overreaction?

DeepSeek’s R1 model—released last week—and the launch of the company’s mobile application—which is currently #1 in the US iOS App Store—has put investment theories about Artificial Intelligence companies under the spotlight, causing a Monday of epic falls on Wall Street. Nvidia saw its stock plunge 17%, losing USD 589 billion in market value—the largest daily loss in the history of the American market.

Takeaways

The cost of AI workloads is rapidly declining, and DeepSeek marks another milestone in this trend. Lower costs are essential for AI adoption to continue accelerating. We are entering the agentic era, where AI agents perform tasks requiring reasoning and memory—capabilities that demand significant compute power. For this era to thrive, AI compute costs must decrease.

This shift presents a promising opportunity for software companies. Lower AI costs enable them to deliver greater value to customers through AI agents. We believe hyperscalers remain well positioned, with declining AI costs and rising AI usage driving sustained profitable growth. For semiconductor stocks to perform well, rapidly growing AI demand must outweigh steep price declines, and leading AI labs must consistently increase investments in developing superior models. While we remain confident in AI's transformative potential, the key uncertainty lies in the timing of AI usage growth.

What did the market do yesterday? First-order effect

If it is possible to have an advanced AI system with fewer processors/compute, as DeepSeek’s results seem to indicate, projections that data centres will have to grow exponentially are shaken—and perhaps fewer processors and less energy consumption (and everything else related to the AI infrastructure chain) will be needed than imagined.

In principle, if the cost of a product falls and demand is elastic, demand should rise. We believe this is what will happen for inference (i.e. running the AI models). However, it is harder to say the same for model training because there is a scenario in which demand was inelastic—the AI labs (OpenAI, Anthropic, xAI, etc.) were already spending everything they could and more, so spending was not the main restriction, and there may now be much less marginal gain in adding GPUs.

It comes down to whether hyperscalers and AI labs will look at DeepSeek’s approach, tweak their processes and reduce capex, or use what they have learnt to make their models even better with the capex they have already committed to. In other words, will companies try to do “the same with less” or “more with the same”?

Implications for AI enablers and adopters

For Nvidia, whose demand depends heavily on training (a bit less than 50% of their demand), the risk seems to increase, adding a question mark to the bull case. There may be a period (in 2026/27, because the next 12 months are already sold out) in which training demand slows, and the rise in inference demand is not fast enough to offset it. It may also be that these innovations make other chips sufficient—something that still seems unlikely, but which could happen if algorithmic innovations take place at the systems level. All the other companies seen as bottlenecks in the infrastructure chain (energy, construction, niche/adjacent hardware for data centres such as high-performance networking) also suffered. So, for the entire semiconductor industry, yesterday was the notoriously “Sell first, ask questions later.”

In general, for AI adopters such as Meta, Amazon, and internet/software companies, models and GPUs are the highest cost/input. Model commoditisation seems good. Satya Nadella went on record saying, “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of.” In some cases, competition with AI-native products may have become fiercer because their development and operation have become much cheaper.

We would frame the risk curve as: AI Semis, Electrical equipment & cap goods exposed to data centres, Hyperscalers, Internet/Software companies.

Is DeepSeek evolutionary or revolutionary?

Some headlines state that “DeepSeek was developed on only $6 million of hardware.” We strongly disagree with that narrative because the headline number can be misleading. Importantly, the $6 million in the DeepSeek report does not include “costs associated with prior research and ablation experiments on architectures, algorithms, and data,” which is an important caveat. By all accounts, however, it is still an efficient design.

DeepSeek’s model efficiency comes from innovative architecture and training techniques. There is an efficiency gain of around 10–11x when comparing GPU hours required for training. The application of the model also shows performance gains, requiring less memory and significantly speeding up computation.

It’s deflationary—it always has been

Not only were DeepSeek models cheaper to make (training) but also cheaper to run (inference). DeepSeek models are priced substantially below OpenAI and similar providers. Nevertheless, we draw attention to what can be achieved. V3 performs comparably to GPT-4 across most benchmarks and beats OpenAI’s model on a couple of important Maths and Science benchmarks. However, it materially lags in one important area: general knowledge and the ability to answer simple factual questions. Efficiency is all about trade-offs. The way DeepSeek was designed (MoE) is not geared for general knowledge tasks—and that is the next frontier of AI adoption with Agents. Put another way, the DeepSeek R1 model is cheaper because it is narrower and less capable in general knowledge. It is not a game changer for AI.

According to Amazon, Distilled Models on Bedrock are more than 500% faster and 75% less expensive than the original models, with less than 2% accuracy loss. AI cost-to-serve is steadily declining — Google disclosed: “Using a combination of our TPUs and GPUs, LG AI Research reduced inference processing time for its multimodal model by more than 50% and operating costs by 72%.” At Davos last week, OpenAI’s Chief Product Officer revealed that costs per token were down 99% for GPT-4o Mini compared to GPT-3.5.

Another factor in the cheaper price of DeepSeek models is that they apparently offer them with a very low gross margin—some reports indicate zero or even negative. That is a huge difference compared to the approximately 70% gross margins that OpenAI or Anthropic typically charge. We do not believe that a rational actor would be willing to run with such thin profitability for long.

Standing on the shoulders of giants

Additionally, DeepSeek was largely able to bootstrap from the compute that OpenAI used to train their models, using tokens from OpenAI. This is known as distillation—where a larger model (in this case OpenAI, but also Anthropic and Gemini) is used to train a smaller one (DeepSeek). There are queries where DeepSeek identifies itself as ChatGPT. Satya Nadella spoke about the ‘distillation’ risk on a recent podcast and compared it to ‘piracy’ and ‘reverse engineering’. Some people in the AI world believe Anthropic and OpenAI are deliberately holding back their next-gen models because they would rather ‘distil’ these models themselves into smaller, cheaper versions to avoid copycats. One important question remains unanswered: the extent of reliance on earlier, larger open-source foundation models. We believe it is crucial to validate these points before drawing conclusions, as it will indicate whether DeepSeek can catch up with leading models or leapfrog the industry entirely.

That is an important topic because it concerns how much AI models might become commoditised, if at all. It might affect how AI labs such as OpenAI, Anthropic and xAI raise money, whether externally or internally (as in the case of Llama at Meta, and Gemini at Alphabet). These AI labs are key Nvidia clients, as a substantial part of Nvidia’s business is still training models, which cascades through the entire ecosystem. If one or more of the leading AI labs reduces or entirely folds their projects, it would be very bad for the overall semiconductor space from a demand perspective, even if cheaper inference boosts elasticity.

US and big tech caught off-guard?

However, these techniques touted as breakthroughs have already been discussed by Western AI players like Microsoft and Meta in research literature. We think it is unlikely that DeepSeek has introduced performance-enhancing technology that large hyperscalers do not already know about. The idea that this is ‘revolutionary’ is somewhat overblown.

Let us not forget that last week there were four major announcements related to AI capex:

The USD 500 billion Stargate project, a substantial collaboration among OpenAI, SoftBank and Oracle.

Meta guiding for capex acceleration in 2025 towards USD 60–65 billion and announcing a massive 2GW data centre that will be almost the size of New York City. That announcement came two weeks after Zuckerberg said on a podcast, “There’s this great Chinese model that just came out, this company DeepSeek. They’re doing really good work. It’s a very advanced model.” Meta can see what DeepSeek is doing because it is based on Meta’s own model, Llama. Meta also seems confident that more compute is better.

Mukesh Ambani’s Reliance is reportedly buying Nvidia GPUs with plans to build a 3GW data centre in Jamnagar, India, making it the world’s largest.

The Bank of China is reportedly set to provide USD 137 billion to support its AI industry over the next five years, in line with China’s New AI Industry Development Action Plan.

Is OpenAI the canary in the coal mine?

About 70% or more of OpenAI’s revenue comes from ChatGPT, the consumer app. They are focusing less and less on Enterprise APIs (hence the tensions with Microsoft, but we leave that for another time). With ChatGPT, the goal is to expand use cases with big ambitions for agentic applications—‘Operator’, for example, where the app could book restaurants, travel, etc. These are large models based on general knowledge. They are also hard to “copy” through distillation, since Reinforcement Learning can handle complicated queries (like maths or code) that can be easily verified and then used to update the model if correct. DeepSeek is not playing OpenAI’s game. Therefore, many believe OpenAI will continue pushing capex higher, which is a positive for the broader space.