This article was first published on AI Business on February 26, 2025, and was written by Dean Tsopanides, Senior Analyst at Third Bridge.
The sudden arrival of DeepSeek earlier this month caused panic in the tech world and sent stocks tumbling.
Beyond this, DeepSeek’s arrival has served to challenge assumptions about US dominance in AI and chipmaking. This has sparked fears about an AI arms race, a tech-driven proxy for wider geopolitical struggles between the US and China.
Much of the anxiety around DeepSeek results from it apparently being delivered at a fraction of the cost of rival LLMs and apparently without needing the support of US chipmakers.
How true are these claims however?
Cheap as chips?
The Chinese state is famously opaque but nevertheless, DeepSeek’s low-cost claims may not hold up.
Analysis suggests that DeepSeek may have trained on OpenAI’s ChatGPT via distillation, specifically ChatGPT-01. Early reports suggest the model even mistakenly identified itself as ChatGPT.
There are also questions about its claim of using only 2,000 GPUs given it has access to over 50,000. Additionally, the reported $6 million training cost excludes earlier development expenses, making direct spending comparisons more complicated. As a reasoning model, DeepSeek R1 should require more processing power and higher inferencing costs, challenging claims of cost-efficiency.
However, if DeepSeek does rely on distillation from frontier models like ChatGPT, it could point to the continued need for large-scale AI training sites. While distillation may improve efficiency, it doesn’t eliminate the need for powerful foundation models and hyperscale data centers.
The mother of necessity
One way in which DeepSeek has undoubtedly broken new ground is by innovating a solution to the chip restrictions imposed upon the project. This led to the development of Multi-Head Latent Attention (MLA), an approach to AI which reduces memory usage per inference by up to 90% – making high-level reasoning more efficient and scalable.
While this is no doubt a cause for alarm in Silicon Valley, it is also something other developers can emulate given DeepSeek’s open-source platform. Time will tell if MLA becomes the go-to approach for LLMs.
DeepSeek, Big Tech
While DeepSeek has made leaps forwards in cost and processing efficiency, as the dust settled it seems the markets over-reacted.
AI labs including Google, OpenAI and Meta were already on track to achieve similar efficiencies. So, while DeepSeek’s disruptive arrival was a PR victory, in reality DeepSeek’s progress is more incremental than disruptive. Since these breakthroughs remain within transformer models too, there is no reason to forecast a meaningful shift in hardware demand away from GPUs.
The reality is DeepSeek will likely drive greater chip demand as more companies enter the AI space, echoing former Intel CEO Pat Gelsinger’s belief that creating AI cheaply would actually increase chip demand.
AI hyperscaler CAPEX is expected to continue its upward trajectory through 2026, with major cloud providers such as Google, Microsoft, and Meta maintaining their aggressive investments in AI capacity.
The anticipated efficiency gains from DeepSeek’s advancements are unlikely to deter these investments, as demand for AI-driven computing remains insatiable.
The information used in compiling this document has been obtained by Third Bridge from experts participating in Forum Interviews. Third Bridge does not warrant the accuracy of the information and has not independently verified it. It should not be regarded as a trade recommendation or form the basis of any investment decision.
For any enquiries, please contact sales@thirdbridge.com
