In a famous sequence in the original Star Wars movie, Obe Wan Kenobie, using a Jedi mind trick, fools imperial foot soldiers searching the city into believing that “these are not the droids we are looking for.”
China’s release of Deep Seek has executed the equivalent of a Jedi mind trick on US investors in AI semiconductors and AI large language models (LLMs) into thinking it is the great new paradigm of cost efficient, technologically advanced artificial intelligence applications.Actually, there is an alternative perspective on China’s release of its Deep Seek AI open source program. The news media has portrayed it as a giant wake up call for the big AI platforms like Open AI, Google Gemini, Facebook’s Llama, and Microsoft’s CoPilot. It is, and that’s the whole idea, but not solely for reasons of technological competitive advantage or competitive cost structures.
The impact of the news about its efficiency and cost tanked the stocks of a number of big tech firms and caused turmoil in the data center industry. It even cast doubt on the future of small nuclear reactors to power data centers.
In response to China’s announcement and in an effort to cast shade on Deep Seek, OpenAI announced it was opening an investigation as to whether Deep Seek stole data from the US company. OpenAI alleges that Deep Seek used a process called “distillation,” which is a method used by smaller AI systems to train themselves using the data from bigger ones.
No one who has paid any attention to OpenAI’s growth over time is going to clutch their pearls. This is a case of the pot calling the kettle black.OpenAI allegedly stole copyrighted data from the New York Times, which launched a lawsuit over it, and from tens of thousands of other Internet sites. For its part the newspaper alleges that millions of articles from the New York Times were used by OpenAI to train successive versions of the ChatGPT LLM.
And Now for the Rest of the Story
But that’s not the real story about the impact of Deep Seek on Open AI and other LLMs. What is the real story is that China’s announcement is an economic cyberattack on emerging US dominance in the AI world.
China’s release of Deep Seek, with its reported back story of being dirt cheap and fast, is intended to disrupt the US artificial intelligence industry, cast doubt on the billions of dollars invested in it, undermine Fortune 500 confidence in using US AI platforms in terms of expected ROI, and present to the rest of the world China’s AI as a preferred platform over US offerings.
Despite Deep Seek being released as an open source platform, as it comes from China, you can bet dollars to donuts that there is a back door in the code that “calls home” with everything it “learns” from users in western industrialized nations. Any business that has intellectual property to protect needs to keep this software off their networks.
Deep Seek did not spring fully grown from a single thunderbolt from Beijing. Deep Seek boasts of training its model for just $6 million, but this figure falls short of the real costs to develop it and neglects to note that the AI application was years in development and reportedly cost several billions to develop. The true paradigm here is that Deep Seek, like every overnight sensation, has a decade of grunt work behind it.
According to Reuters DeepSeek’s roots trace back to quantitative hedge fund, Hangzhou Huanfang Technology Ltd Co., as the company is officially called, which has a history of developing AI models over the past decade to guide its investment decisions. According to the 01/29/25 Reuters report so far in the past five years the firm has spend $14 billion developing these AI systems including Deep Seek.
According to High Flyer, which is the subsidiary firm actually doing the AI work, Nvidia’s powerful A100 chips were built and put into operation long before the export controls were announced.Its first cluster, made up of 1,100 A100 chips, reportedly cost $28 million and was put into operation in 2020, while its second cluster, made up of around 10,000 A100 chips, was completed a year later with a cost of $140 million, according to a review by Reuters of the company’s website and several WeChat posts. That’s five years prior to the current announcement.
US AI Firms Under Pressure to Perform to New Expectations
Overall, US artificial intelligence firms are now under considerable pressure from their current and potential future customers to match or exceed Deep Seek’s highly competitive position as a technological leader and less costly alternative. The rise of DeepSeek is forcing OpenAI and its investors to reconsider the $500 billion Stargate AI project as DeepSeek seemingly develops AI at less than 1% of the cost.
However, the firm’s reliance on questionable methods of acquiring data, issues related to evasion of US embargos of highly sensitive semiconductors, and the financial objectives of the firm’s corporate parent, put the Deep Seek announcement in a different light.
While DeepSeek’s speed and cost efficiency are remarkable, its reliance on synthetic data, questionable hardware sourcing, and hedge fund origins suggest it may not be the paradigm shift it claims to be.
China Global Competitive Strategies at Work
So what is China trying to do with its Deep Seek announcement? China’s all in commitment to gaining supremacy technically and in terms of global political influence over the west is embodied in its the integration of political, economic, legal, military, intelligence and cyber strategies.
For a deeper dive into this issue, see “Is America Ready for Chinese-Russian Liminal Warfare?,” by: Robert McFarlane, and Andrew D. Paterson, The National Interest, 05/07/22.
In effect the Deep Seek announcement has upended the precarious perch that US AI firms were on with their billions invested in Ai software and only millions in revenue to justify it. This made them vulnerable to the type of paradigm shifting effects of the Deep Seek announcement.
The US news media seems to have completely missed the strategic intent behind China’s Deep Seek which is that it is a “soft power” attack on US AI platforms and it succeeded.
Deep Seek is Not the Biggest Threat
to Plans for SMRs to Power Data Centers
Some analysts have suggested that if AI applications become more energy efficient, they will need less electricity and this will dampen demand for small modular reactors to power data center. While there is some basis for this conjecture, the real competitive threat to the future market share for SMRs to power data center is coming from the natural gas industry.
Recently, Exxon and Chevron announced that they plan to build gas powered electrical generating plants adjacent to data centers. This plan eliminates the need for dealing with the current multi-year delays in getting new grid connections.
Also, it takes as little at three years to build a 700 MW gas plant that costs about $650 million. It takes just 500 workers to complete it.
By comparison, it takes up to eight years to to build a similar class nuclear reactor which will cost $3.5 billion. It takes several thousand workers to build it.
A 300 MW nuclear reactor, which is a typical size for several US developers of PWR type SMR designs, requires four years to get a license and another three years to build for a time to completion of seven years at a cost of $1.5 billion. Based on these numbers, gas plants win every time.
Data Center developers are clearly aware of these numbers and have moved to plan the location of future facilities for ease of connection to natural gas pipelines.
Data center growth a boon for gas transmission companies 09/11/2024. Image: Enverus
Data center siting preferences are shifting “from regions where big telecom infrastructure is in place to regions where energy and supply infrastructure is in place,” TC Energy natural gas pipelines COO Stanley Chapman III said on a 08/01/24 earnings call.
Rather than siting data centers behind local gas distributors, Chapman said, “we’re now seeing a much greater potential for data center operators to seek laterals off of our mainline and to use that gas supply to fuel onsite power generation that they would build and/or own themselves.”
TC Energy said there are about 300 data centers in development in the U.S., of which 60% are within 15 miles of its pipeline systems. TC Energy CEO Francis Poirier noted in the same August 2024 earnings call that 9 GW of coal-fired generation is retiring by 2031 near those systems, leaving a large void for generation that could be filled by gas. He estimates there are 5 Bcf/d of growth opportunities for the company.
For SMR developers, Deep Seek isn’t the biggest market threat to their future. The challenge for SMR developers, relative to the data center market, will be to deliver their reactors faster and cheaper to meet the competitive drive of gas producers.
# # #