DeepSeek and Elon Musk deliver lessons on technology and power.
A Chinese champion ripped up the rulebook on model engineering—and as the United States musters a response, Elon Musk is ripping up the “nerve systems” of the federal government.
What’s happening? Two major flashpoints in the geopolitics of technology are still flooding the space with noise this week—DeepSeek and Elon. Both have sent shock waves through the technological plumbing of global power.
DeepSeek. In the past week, the unfolding impacts of DeepSeek’s R1 model on open-source innovation and the geopolitical dynamics of technology have been staggering to behold. In extreme brief, the Hangzhou-based company:
used a technique for collating data scraped from the internet called “Common Crawl” to build an effective set for coding and mathematical reasoning;
pruned away some reinforcement-learning steps in model pre-training, limiting the amount of costly human-labelled data needed;
deployed a lightweight algorithm—Group Relative Policy Optimization (GRPO)—to tune responses, instead of a more resource-intensive large model;
got the most out of a relatively small number of H800 GPUs using Infiniband and NVLink optimizations;
Then? DeepSeek open-sourced R1, with freely available model weights, and comparable if not better performance overall than closed models from competitor labs, all reportedly for a fraction of the development and end-use cost.
US technology stocks shed over $1 trillion in a huge confidence slump; with Nvidia, the leading chip supplier to the US foundational model developers, being the worst hit.
So what? There might be a hole in the high fence leaving the yard is exposed. Suspicions have mounted that export controls aimed at limiting China's access to advanced technology may not work.
Jeffery Ding, for example, argues that R1 shows the controls have been “ineffective at preventing other countries from developing frontier models".
Others argue that DeepSeek is an exception that proves the rule; having acquired access to 10,000 of Nvidia's older generation GPUs and some cutting-edge H100s before the controls came into full effect.
Can DeepSeek's claims be believed? The claim that R1 was trained at roughly 5 percent of the cost of OpenAI's o1 is difficult to verify outright. Some experts have speculated that undisclosed costs and clusters could have played a part in R1's development.
Proving replicability would go a long way to showing that China's wider ecosystem has really closed the gap on American frontier model developers.
Candidates are already emerging. Paul Triolo points to Moonshot AI's Kimi k1.5, as a case in point, out of the $3.3 billion-valued Alibaba-backed startup, as well as Alibaba’s Qwen 2.5-Max.
What're the implications for policy makers in the US? DeepSeek's explosive success has, predictably, emboldened China hawks in Washington to argue for even tougher controls and policing, to reduce the risk of more Chinese firms replicating DeepSeek's R1.
Why? China is clearly catching up in the race to deliver value. As we wrote a few months ago, China's approach of targeting lower cost and less energy-intensive inferencing could also sidestep the challenge of continually raising the ceiling of model performance, by making them more cost-effective and useful for specific applications.
What about the global tech stack? If Chinese companies can reliably produce other high-performance, low cost models despite US efforts to curb their access to the highest-end chips, it could attract more companies and countries towards building with Chinese models and software layers.
DeepSeek’s success with R1 is likewise a signal to other countries that, despite existing labs having built a considerable advantage in terms of raw compute, the innovation game is still very much afoot.
If a small laboratory in China can produce a model that performs comparably to the best at a fraction of the cost, then so might an enterprising firm with access to sufficient risk capital in the UK, Germany, India or elsewhere.
What to watch next? Firstly, for the emergence of subsequent models from Chinese firms that have been more decisively hit by US chip controls, as well as the outcomes from a reported White House investigation into the provenance of DeepSeek’s chips used in training R1.
Secondly, the realization that constraints on the availability of advanced chips had not fully bitten during R1's development, as Lennart Heim points out; public benchmarks lag the actual frontier of capabilities on both sides of the geopolitical fault line. In other words, America's ecosystem has yet to show its best.
Thirdly, watch for the resurgence of firms that are not overexposed to the huge capex risk of scaling clusters of cutting edge chips, but still have potential for small-i innovation, through more efficient inferencing on devices. This piece makes a clear and compelling case for Apple as a company that will benefit from this dynamic.
Now for Elon. The DOGE of Washington is methodically taking control of the US government’s “nerve systems.” Musk and his Department of Government Efficiency team appear to be replicating the smash mouth approach that Musk took at Twitter as he transformed it into X, this time inside the federal government.
How? Musk is dispatching young software engineers with links to his technology companies to enter government offices and gain access to data and systems that help power the vast US federal bureaucracy.
The Office of Personnel Management, the General Services Administration, and the US Treasury Department’s Bureau of the Fiscal Service are boringly named, but constitute the institutional and financial plumbing of the US government.
They manage HR files and set policies for millions of government workers, oversee US government facilities and IT systems, and make and receive trillions of dollars of payments each year. It also manages the US federal debt.
So what? As Musk gains access to these systems, he is also gaining access to vast power. On Monday, Musk said he was “shutting down payments” by the US Department of Health and Human Services to a charity that provides services to refugees.
What’s more? DOGE also appears to have used access to the US Agency for International Development’s internet domain and email system to take down the aid agency’s website and deactivate thousands of employee emails, part of a broader attempt by Trump to shut down the agency and fold it into the State Department.
What’s next? A former Tesla engineer—who has been put in charge of technology transformation at GSA—indicated that part of DOGE’s plan is to centralize huge amounts of data across the federal government and to adopt an AI-first approach to government services.
This raises the prospect that DOGE may have its sights set on training foundational models with one of the world’s biggest sources of still-largely-untapped (and highly sensitive) information.
What are the risks? The speed and ferocity of the DOGE takeover, which has shattered norms and standard operating procedures, and which many experts believe violates the law, is raising alarm bells over cyber-risk.
When an Office of Personnel Management database containing sensitive security clearance applications and other data on millions of federal workers was hacked by Chinese cyber operators in 2015, it was one of the worst government data breaches in American history.
DOGE’s effort to put sensitive citizen and government data in a centralized repository for the purpose of training artificial intelligence could create an even more tempting target.
Other types of disruptions could emerge if payment systems are compromised or experience outages.
DOGE’s efforts to gain access to these sensitive systems led one longtime Treasury official to resign last week, after resisting the attempt.
The upshot? Abe Newman, a Georgetown professor and one of the co-authors of Underground Empire—which describes how the US leverages financial and informational dominance to wield global power—warns that seizure of these power centers could lead to manipulation of the government’s workforce, buildings and IT usage.
Ultimately it could bring about a kind of surveillance and control scheme that would install Trump loyalists across the rest of the government. That’s probably a feature, not a bug, for Trump supporters who want to “dismantle the administrative state”, but it would imply more democratic backsliding.
The pushback against DOGE has started, with Democratic Minority Leader Chuck Schumer on Tuesday calling it an “unelected shadow government” conducting a “hostile takeover of the federal government” and other lawmakers warning the United States was in a constitutional crisis.
What we’re reading:
More from CSIS on the DeepSeek moment.
Reports of China’s retaliatory trade action against escalation by Trump.
This press release on new uses for data generated by Quantinuum’s H2 quantum computer.
What we’re looking ahead to:
6 - 7 February: The Inaugural Conference of the International Association for Safe and Ethical AI, Paris, France.
10 - 11 February 2025: AI Action Summit in Paris, France.
11 - 13 February 2025: World Governments Summit 2025, Dubai, United Arab Emirates.
12 February 2025: Chief AI Officer Summit UK, London.
2 - 4 June 2025: AI+ Expo and Ash Carter Exchange in Washington, DC.
9 - 11 July 2025: AI for Good Global Summit.