DeepSeek and you shall find: promise, but also perils

If you didn’t know DeepSeek before, you do now. 

Until last week, it was a little known Chinese startup founded in 2023 and bankrolled by a hedge fund. 

That’s until they caught industry watchers napping by releasing a model that seemingly outperformed those from Silicon Valley at a fraction of the cost.

Concurrent with that, Nvidia lost over a sixth - a cool $500bn - of its value. Its rival Broadcom lost nearly a fifth. 

Prevailing logic - be that around US export controls, or scaling laws - was questioned. DeepSeek was even forced to temporarily limit registrations because of "malicious attacks" on its software.

All in the space of 24 hours.

Now the dust has settled, what should we make of this? 

The geopolitics 

By matching or approaching frontier-level capabilities, DeepSeek-R1 suggests China’s AI development may be more advanced than some anticipated. And, in principle at least, private government capabilities might be even stronger.

This matters since the previous US administration tried to limit Chinese progress in AI with the Chips Act, putting strict export controls on the most advanced chips. 

While it is possible DeepSeek simply circumvented the controls, it seems likely they used more efficient training methods to reach the same level of performance on lower capability hardware. 

This does not mean the Chips Act had no impact. It’s possible DeepSeek would have achieved these results even faster without it. 

It’s also possible it has pushed Chinese AI in a more efficient direction, with faster iteration speeds and greater applicability across devices. Nonetheless, it still raises uncomfortable questions around US trade policy.

The model scores similarly to o1 on various standard metrics, which is extremely impressive given the training costs and open availability. However, we have no evidence it exceeds them in meaningful ways. 

The hyperventilating seen on Twitter that “America is now behind” is wrong. Static benchmarks may not capture real-world performance and we will likely learn more as both models are used in the real world. 

Forecasting what happens next is hard. This is being talked about as a ‘Sputnik moment,’ although the analogy isn’t quite right. This is more akin to the Americans already having a satellite in space, but the Russians doing it cheaper. 

America will respond in time, although that is complicated by dynamics in the current administration. 

AI security and performance

If DeepSeek-R1 matches OpenAI’s o1 in capability yet costs a fraction to use, it lowers the barrier to entry for state and non-state actors alike. 

It has the pro’s and con’s of all open source software. If it were to lead to software with enormous positive consequences, like making medical diagnosis available at close to zero cost, no one company could capture that. 

But, once a high-capability model exists in one place, others can make use of it to train their own models. If it were to have harmful effects, then there is no way to put the genie back in the bottle.  It also makes the development of models harder to monitor, and potentially less predictable. 

Cost significance

It’s claimed DeepSeek-R1 was trained for a fraction of the cost of o1. If true, it could reshape assumptions about the capital and hardware needed to build an advanced model - and lower-costs might threaten the market positions of companies that have invested heavily in data centers or specialized chips.

However, at the hardware level, this seems wrong. If DeepSeek were to have access to OpenAI or even Google levels of compute, it is likely they could make a higher performing model. Secondly, even if it were true, cheaper AI will likely lead to more demand for AI (the so-called Jevon’s paradox), so again this seems wrong. Although, stacking caveat on caveat, the timescales of demand and supply have to match, and mismatches here could have implications for both chip makers and clouds. 

The market movements probably relate to trade tariffs, but might be reinforced by the fact DeepSeek showed algorithmic advances can still provide very significant gains. This would not be news to experts, but perhaps shocked retail investors who priced NVIDIA on the basis that every advance in AI was going to depend only on increasingly expensive hardware. 

At the foundational model level, it does raise questions about how effectively companies can create longstanding ‘moats’. So far, we have seen that closed source models are hard to separate in capabilities, and open source models are lagging about 6 months behind. 

Given that GPUs are a depreciating asset, and the spend on these is in the tens of billions of dollars for US tech giants, it raises questions about the sustainability of foundational model companies as independent businesses. 

Future AI developments 

Lastly, DeepSeek signals important trends for the future of AI. New techniques can seemingly cut training costs and circumvent hardware shortages. These will enable faster training runs, allowing more experiments, and more experiments will allow more breakthroughs. 

DeepSeek’s model is significant because it matches a Western model in raw performance, but with surprising cost efficiency and an alternative political alignment. Geopolitically, it signals that advanced AI is no longer an obvious Western monopoly. In terms of security, it raises concerns that high-level AI can spread more quickly and cheaply than anticipated. Economically, it questions who profits from AI breakthroughs. 

Finally, this underscores that we are at the beginning of a period where new training techniques, data strategies, and hardware workarounds might allow rapid leaps forward, reshaping everything from global technological competition to how society grapples with AI’s broader implications. 

This must not be seen as another isolated breakthrough. It is the result of an enormous wave of profound technological progress that occasionally breaks through to the mainstream. Whether or not DeepSeek is the right model for them, this is another reminder for companies that the AI era is here - not as some distant point in the future.