Will DeepSeek burst the AI bubble?
The emergence of DeepSeek, a Chinese AI startup, into the global AI industry has caused jitters in world stock markets, and led many to question the assumption that more resources and money equal better technology.
According to early reports and tests, the various DeepSeek models need significantly fewer resources to train and create inferences than other models, like Gemini and ChatGPT. Instead of the million or so GPUs Meta has slated for its AI engine (to be in place by the end of this year), the Chinese startup deployed fewer than a tenth of that number – and produces comparable results.
Data Sources and Model Deployment
Like every published model, the source of the training data used to give DeepSeek its apparent smarts cannot easily be determined, although some reports suggest that Meta sold the company the data set behind Llama. The pre-trained model, which can be downloaded from Hugging Face and run on-premise or suitably-equipped VPS, contains a cursory set of what the industry terms ‘guard rails’ which pose a degree of censorship on inferences: Asking about “the massacre in Tiananmen Square” produces a polite deflection of “Let’s talk about something else.” But requesting a historical account of the events in 1989 gives users a fairly objective summary of history.
Model Hosting and Performance
Users can choose to host the models themselves or access the company’s servers based in China, the latter being the method by which the iOS and Android apps produce their results. Due to massive peaks in demand that coincided with America’s waking hours over the last few days, service users encountered slowdowns, suspended registration, and outages during the first week of operations as resources struggled to keep up.
There is also an option to use an API to access the DeepSeek cloud instance, which is reported to cost significantly less than an equivalent service from ChatGPT.
Licensing and Competition
The variously-sized models are released under an MIT licence, which gives users the right to copy, embed and distribute the software and accompanying data, in contrast with the closed-shop, proprietary business models favored by Google, OpenAI, Meta, et al. The provenance of the data originally used to train DeepSeek is as murky as any other model, although it’s safe to assume the usual piracy, copyright and license infringements, and injudicious vacuuming of the internet were in play.
Implications and Future of AI
The arrival of a lower-cost competitor to the US-based technology giants’ AIs would be interesting enough, especially for those that are interested in how the company achieves its results with significantly less hardware and resources than its competition. But the fact that DeepSeek emerged from the current US administration’s currently-chosen ‘evil empire’ means that there will be difficult discussions around the conference tables of Washington and Silicon Valley.
It may be taking Schadenfreude too far to suggest that the US’s imposition of trade barriers on China (limiting the access to high-end GPUs) actually caused engineers in Beijing to seek lower-resource alternatives. However, the US is indubitably on the back foot politically and technologically, even without that inference.
The assumption to date in the AI industry – throwing money at the technology is equivalent to fostering innovation – has been evaporated by a small startup from a country that the US needs, economically, to paint as an enemy. At worst, however, it’s an embarrassment to the ‘tech bros’ and the titian rattlers of sabres.
Conclusion
The inferences produced by DeepSeek confirm that the world now has access to cheaper, less resource-intensive garbage. That may be a blessing for garbage-lovers everywhere, but it’s unlikely to do more than slightly deflate the AI bubble, temporarily, before it expands even further. The costs accrued at the point of origin for AI modeling may now be lower, but they are unlikely to passed to consumers in the form of significantly-lower use costs. The levels of investment to date in AI mean that DeepSeek’s breakthroughs have created a lifeline for Western investors who were just starting to worry about how AI might generate some returns.
