This month we dig into the frenzy over China’s DeepSeek and ask whether this has punctured the narrative of US tech supremacy.
Has DeepSeek just punctured the market narrative of American tech supremacy? We think that is a stretch, but the revelation of DeepSeek’s ability to innovate in AI with shocking efficiency is a reminder that there are only two contenders in this battle to build artificial general intelligence – China and the United States.
DeepSeek has unveiled two new models – DeepSeek-V3 and DeepSeek-R1, as well as instructions called R1 Zero – that deliver performance on offerings from OpenAI and Anthropic. These models have set off a media and market frenzy, both because they appear to match or exceed the capabilities of more famous systems, and because DeepSeek is offering API access at a fraction of the cost.
Highlights:
- DeepSeek uses a method called reinforcement learning. Essentially, the models are allowed to solve the problems themselves with few guidelines and limited example solutions. Remarkably, this was accomplished using only 8,000 math problems, whereas other research groups often need millions.
- DeepSeek has managed to compress memory usage, circumventing the need for loads of expensive GPUs.
- DeepSeek has shown that AI models can work remotely and on edge computing very effectively without needing the power of data centres.
Overall, DeepSeek demonstrated that you don’t have to invest massive amounts (exactly how much is debatable) of money, hardware or human oversight to build an AI that excels at difficult tasks. The arguments about how much money they spent to get here are irrelevant: by relying on focused reinforcement learning and efficiency-boosting techniques, DeepSeek proved that powerful models can be created with fewer resources.
All training steps and code have been shared so others can also try it and change things, making concerns over “censorship” entirely moot. The result was a model that can rival Anthropic and OpenAI, even when turned into a much smaller version that can be run locally on a pair of Mac Minis! (Which use ARM architecture: the M4 Pro uses TSMC 3nm, and runs at 80W.)
The team behind DeepSeek is open about its own limitations. First, the model is akin to a brilliant scientist but would struggle to write a poem as it lacks “creativity.” Second, it doesn’t deal with languages beyond English and Chinese very well. And third, it lacks the experience in building large-scale software projects.
Implications
Anyone who has followed this story is probably now aware of Jevons Paradox. Originating from the work of economist William Stanley Jevons in 1865, the observation suggests that as technological advancements make a resource more efficient to use, the overall consumption of that resource may increase rather than decrease. This paradox occurs because increased efficiency often lowers the cost of using the resource, leading to greater demand and, ultimately, higher total consumption.
Applying Jevons Paradox to AI tools, as these technologies become more efficient and cheaper, their usage is likely to grow significantly. Just as more-efficient cars led to wider adoption over horses, more-efficient and cost-effective AI models like DeepSeek’s will encourage broader usage across various industries. This greater adoption can drive further innovation, but it also means that the demand for AI resources, such as data and computing power, will continue to rise.
As a result, businesses and developers will need to consider the implications of widespread AI deployment, including potential increases in energy consumption and the need for sustainable practices in AI development and usage.
To summarise:
- Large language models (LLM) have become commoditized. For instance, Meta’s Llama (an LLM) is open-source and therefore free. The key takeaway here is that the cost and compute requirements to run these models could potentially be reduced significantly.
- The implication is that demand for AI infrastructure including computer chips, the semiconductor supply chain and power requirements (particularly for AI training) may be lower than first thought.
- However, as highlighted by Jevons Paradox, history shows that for most technological advancements, reduced costs are almost always offset by increased demand.
What does this mean for the stocks of different global tech leaders?
It’s still early days, but how could the broad adoption of DeepSeek models impact global tech leaders’ stock prices?
Type of tech company | Stock impact |
---|---|
AI infrastructure and some semiconductor companies Jevon’s paradox will likely spur more AI applications, with the end result potentially being greater demand for compute down the line. However, the market is questioning the margins of semiconductor players and infrastructure solution providers (i.e. cooling tech). We need to see the mix of LLMs vs. “distilled models” and, more importantly, inferencing vs. training. Training requires much less compute power than inferencing. |
Unclear |
Hyperscalers On one hand, processing AI could become significantly cheaper which will reduce their cost/capex. On the other, their moat could be lowered if AI workloads can be run on less powerful data centres. Microsoft has already stated that it is prioritizing enterprise inference workload over AI training for its Azure business. That is why OpenAI went to Oracle/Softbank/Project Stargate for compute because Microsoft won’t sell them all the compute OpenAI demanded. |
Neutral/unclear |
Application-specific integrated circuit (ASIC) companies Possibly beneficial for custom ASICs as chip architecture diversifies/specialises. |
Neutral/unclear |
Applications, such as software, with access to proprietary data This is where I believe the most significant AI equity value will be created over time. Lowering AI costs is unlikely to negatively impact these companies. In fact, it could even be a positive development. The moat is in the access to data. Compute is a cost item. |
Positive/unclear |
Specialised edge computer chip companies | Positive/unclear |
In emerging markets, we believe major positions like Taiwan Semiconductor Manufacturing Company (TSMC), Mediatek and select niche names (in custom chip design and energy efficiency) remain well positioned for growth in overall demand for AI. We doubt that DeepSeek will change the demand for the highest performance chips running at the lowest possible power. In that regard, TSMC’s dominance in leading-edge production processes and advanced packaging solutions remain an intact competitive moat. We expect that their customer mix may change, but the demand for their capabilities will be resilient.
We are more cautious on data centre assemblers and memory, and see potential for an improving sentiment in software, with several high-quality names in the portfolio and on our watchlist in China, ASEAN and Latin America.