AI on the Cheap vs. Stargate’s Big Splash

Welcome back to Techne! I’m a big fan of The Public Domain Review, an online journal focusing on works of art and literature now in the public domain. Hugh Aldersey-Williams, a writer and curator, just published a fascinating piece on two French architects, Étienne-Louis Boullée and Claude-Nicolas Ledoux. Both architects pioneered a revolutionary style that emphasized pure geometric forms and stripped-down classical elements. Sometimes impractical, sometimes realized, these beautiful sketches are markers in the transition toward modern architecture.  

AI’s Emerging Paradox

While the presidential transition commanded headlines this week, equally significant shifts were occurring in AI technology. Just hours before Donald Trump’s inauguration, DeepSeek released its latest model, achieving a breakthrough in AI reasoning that matches the best models of OpenAI and Anthropic but at a fraction of the cost. DeepSeek, which is backed by a Chinese financial trading firm, has been able to continually keep up with the industry leaders while navigating stringent U.S. chip export controls.

This development challenges the assumption that cutting-edge AI progress requires massive computational resources available only to better-funded Western companies. DeepSeek’s models still aren’t cheap, at the cost of $5.5 million, but the company has been able to keep up with the best companies in the U.S. even with the chip restrictions. It is helping to show that innovative engineering and efficient resource use might matter more than raw computing power in the long run. 

The AI landscape has transformed dramatically in just the first month of 2025. Things are moving fast, so let me get you up to speed.

Executive orders, past and present.

In the first couple of hours of his presidency, Trump issued 46 separate executive actions, the most consequential for AI policy being his repeal of the Biden administration’s Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. 

I’ve written about the Biden EO in a number of places because it was the organizing agenda for AI policy. It included some 150 requirements that forced various agencies to publish frameworks, studies, and undertake rulemaking related to AI. Trump’s rescission will cancel the reporting requirements attached to Biden’s EO, but the substance of the former administration’s work will remain. 

Federal agencies like the Equal Employment Opportunity Commission, the Food and Drug Administration, the Consumer Financial Protection Bureau, and National Institute of Standards and Technology have issued more than 500 AI-relevant regulations, standards, and other governance documents. Some or none of these could be on the chopping block if the various agency heads under Trump decide to unravel the regulations. However, any rollback would require administrative processes that could last months. And many of these policies, particularly those focused on consumer protection, probably won’t be repealed even after review. 

Perhaps the most consequential question is what Trump will do with the expansive authority Biden claimed under the Defense Production Act (DPA). The DPA, originally passed during the Korean War and reauthorized every five years since, has evolved from a wartime industrial policy tool into a broader instrument for addressing critical technology challenges. Biden broke new ground by using the DPA to compel AI companies to share access to their models, arguing that concentrated control of frontier AI systems poses national security risks.

While Trump’s early moves signal a pullback from Biden’s regulatory framework, he may find value in preserving these expanded executive powers over the AI industry. Ashley Mehra, a fellow at the Mercatus Center, explains more in her brief on how to lawfully apply the Defense Production Act.    

The Biden export controls. 

In the waning hours of the Biden administration, the Bureau of Industry and Security (BIS) , adopted a new rule that controls who gets the computing chips needed to train advanced AI models.

This rule put into place a three-tier system for AI chip access. Nations getting largely unfettered access to advanced chips include America’s closest allies, mainly Western nations like the U.K. and EU members, along with Asian tech powerhouses like Japan, South Korea, and Taiwan. At the other extreme, strategic competitors like China, Russia, and Iran face complete restrictions as third-tier countries. No advanced chips can be imported. But most of the world’s nations fall into a middle tier that applies complex limitations to their chip imports.

For these countries, the rules create both limits for individual companies and national caps. Any single company can import computing power equivalent to roughly 1,700 of today’s high-end Nvidia H100 chips per year without special permission. When companies need more, they must navigate a demanding approval process requiring consent from four different U.S. government agencies: the Commerce, Energy, State, and Defense departments. Even if approved, these larger orders count against their country’s annual limit of about 50,000 H100-equivalent chips. The rule also implements enhanced security standards for development facilities.

This change in policy is the latest in a series of restrictions imposed by the Biden administration on semiconductor exports. In early December 2024, the Commerce Department significantly expanded its controls on the Chinese semiconductor industry. These rules restricted sales of high-bandwidth memory to China and limited access to crucial chip design software. The Commerce Department also added 140 Chinese companies—including tool manufacturers, fabrication facilities, and investment companies involved in semiconducting manufacturing—to a list of those that can no longer sell products in the United States. Perhaps most significantly, the administration expanded the Foreign Direct Product Rule, which prohibits any product containing even a single chip designed or manufactured with American technology from being sent to blacklisted fabrication facilities. These combined measures represent the most comprehensive attempt yet to restrict China’s access to advanced semiconductor technology.

Even with all the restrictions, Chinese companies like Huawei are getting access to the most advanced chips through intermediary companies. More importantly, China’s biggest chipmaker, SMIC, has been making strides on advanced chips without the most recent technologies. As the Financial Times explained, SMIC seems to be doing the impossible but at a cost. Its advanced chips cost 50 percent more than its main competitor, Taiwan’s TSMC, and they are of lower quality. So while China can work around U.S. restrictions, it’s costing money and manufacturing efficiency.

But DeepSeek’s recent innovations underscore that chip access might not be as constraining as previously thought. 

Why DeepSeek matters.

DeepSeek, to me, is the most important firm in AI right now. 

DeepSeek was founded in May 2023 by High-Flyer, one of the largest quantitative trading funds in China. DeepSeek’s breakthrough came in December 2024 with DeepSeek-V3, a model that matched the performance of GPT-4o and Claude 3.5 Sonnet while using far fewer resources. The company achieved this through a mixture-of-experts approach that selectively activated a limited number of models parameters for each token. A token is the basic unit of text in AI models that serves as both an input and an output. It can be a word, a part of a word, or a single character like a period or a question mark. 

By demonstrating that state-of-the-art AI systems can be developed in this way, this company has lowered the barrier to entry for a range of state and nonstate actors to get into the fray. DeepSeek estimated the training cost for its model at $5.5 million, which isn’t much to be on the bleeding edge when OpenAI spent $80 million to $100 million to train ChatGPT 4.0 just a generation back. 

While previous estimates suggested only major tech companies and well-resourced nations could develop advanced AI systems, DeepSeek’s training cost represents a dramatic lowering of financial barriers. The timing is particularly significant, as it occurred just as U.S. export controls on advanced semiconductors were tightening.

Nations with substantial technical talent but limited access to advanced computing infrastructure, such as Russia, Iran, and Pakistan, might now have a viable path to developing sophisticated AI capabilities. Even individual billionaires with sufficient technical expertise could theoretically pursue similar programs, though assembling the necessary technical talent would remain a significant hurdle.

That’s the unintended consequence of U.S. export controls. While the U.S. succeeded in maintaining its choke point over high-end semiconductor access, this may have inadvertently sparked innovations that make such choke points less relevant. The proliferation of these more-efficient AI development techniques could ultimately pose greater security challenges than the concentrated development of chip-intensive systems by major state actors.

Going big with Stargate.

While DeepSeek demonstrates the potential for doing more with less, OpenAI is pursuing the opposite strategy by betting big on massive computing infrastructure. A day after repealing Biden’s executive order on AI, Trump stood alongside OpenAI CEO Sam Altman, Oracle CEO Larry Ellison, and SoftBank CEO Masayoshi Son to announce a major new investment in data centers. 

OpenAI laid out the details on X:

The Stargate Project is a new company which intends to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States. We will begin deploying $100 billion immediately. This infrastructure will secure American leadership in AI, create hundreds of thousands of American jobs, and generate massive economic benefit for the entire world. This project will not only support the re-industrialization of the United States but also provide a strategic capability to protect the national security of America and its allies.

This isn’t out of the blue. OpenAI has been talking about building a $100 billion data center that requires 2 million graphics processing units and 5 gigawatts of power to run. The plans are laid out in the company’s September 2024 “Infrastructure Is Destiny” white paper. One buildout is currently underway in Texas and others are being considered.

The key here will be getting the energy needed to run this data center and the returns to make investment worth it. 

This contrast in approaches highlights a fascinating divide in the AI industry: Some firms are finding ways to build powerful AI systems with minimal resources, while others believe that unprecedented scale is the key to future breakthroughs. 

The future is likely to lie in between these two poles. While efficiency innovations can reduce the computing needed for today’s AI capabilities, pushing beyond current frontiers might still require massive computational resources. 

Will export controls matter in the long run? 

The assumption that maintaining American technological superiority requires controlling access to cutting-edge hardware is being challenged on multiple fronts. Just as DeepSeek shows that innovation can sometimes substitute for raw computing power, China’s response to export controls demonstrates how restrictions can spark unexpected adaptations.

My views on China have changed since President Trump first placed restrictions on Huawei back in 2018. I’ve become more concerned about the leadership of Chinese President Xi Jinping and the direction he’s taking his nation. At the same time, I’ve become less certain that the export control regime can effectively slow China’s AI progress. 

I’m not the only one changing their tune. Ben Thompson, the author of the Stratechery blog, recently wrote that “you could make the case that the China Chip Ban hasn’t just failed, but actually backfired.” Dean Ball of the Mercatus Center had similar thoughts in a post on the export controls, writing that it is “either the wisest or the most disastrous decision of President Biden’s administration.”

As it considers further trade restrictions against China, President Donald Trump’s administration will have time to plan out its next steps and would do well to think carefully about its goals. 

There is time. Trump’s executive order on “America First Trade Policy” has a section on economic and trade relations with China that outlines a series of studies to be undertaken. As Bill Bishop pointed out in the Sinocism blog, this means that:

These studies will take months, giving the PRC perhaps more breathing room expected from economic pressure from the US, at least for much of 2025, as well as a window to try to negotiate some sort of a broader deal with the Trump administration. However, the language in the EO about China should be seen as a clear warning to Beijing that significant increased pressure is under consideration.

The rise of DeepSeek alongside the launch of OpenAI’s Stargate project presents a fascinating paradox: As the U.S. bets on maintaining technological superiority through massive infrastructure investments, innovative approaches from abroad are demonstrating alternative paths forward. As the Trump administration settles in and considers its approach to both AI regulation and China policy, these early developments suggest that maintaining technological advantage may require more nuanced tools than export controls alone. The challenge ahead will be crafting policies that can effectively balance national security concerns with the reality that innovation often finds a way around barriers.

Until next week,

🚀 Will

Notes and Quotes

• In his first week, Trump signed executive orders that established the Department of Government Efficiency (DOGE) by cannibalizing the U.S. Digital Service, pulled permits for offshore wind projects for power generation, and declared an energy emergency that is intended to unleash energy development. Trump also ordered the Justice Department not to enforce the TikTok law for at least 75 days. So TikTok is banned but also still alive, it seems.  
• The legal framework for space mining hinges on the 1967 Outer Space Treaty. While the treaty prevents nations and private entities from claiming ownership of celestial bodies, recent legislation like the 2015 Commercial Space Launch Competitiveness Act attempts to bypass this by treating space resources similar to “harvesting fish from the sea.” This approach reframes mining from a real property right to a permitted “use” activity, where resources become personal property once extracted. From The Space Review comes this legal primer on space resources.
• Blue Origin’s New Glenn rocket achieved orbit on its maiden flight. The reusable rocket already has contracts with the U.S. Space Force, French satellite operator Eutelsat, Southeast Asian aerospace manufacturer mu Space, Japan’s SKY Perfect JSAT, and Canadian satellite communications company Telesat, so it is expected that New Glenn will become a mainstay in the heavy-lift rocket space alongside SpaceX.
• Derek Thompson’s recent feature in The Atlantic is about how Americans are spending more time alone than ever. Here is one bit that struck me: “In 2023, 74 percent of all restaurant traffic came from ‘off premises’ customers—that is, from takeout and delivery—up from 61 percent before COVID.”
• A sign of the times, from tech policy wonk Daniel Castro: “President of the European Commission Ursula von der Leyen signals the EU is ready to move away from the United States, and towards China, in her Davos speech.”
• Constellation Energy, which operates America’s largest fleet of nuclear power plants, has announced plans to acquire power producer Calpine in a $16.4 billion deal. This acquisition would combine Constellation’s nuclear expertise with Calpine’s extensive natural gas power generation assets. While Constellation is known for its base load nuclear facilities, Calpine specializes in operating flexible gas turbine plants that can quickly ramp up or down to meet changing electricity demand. Calpine also has a major stake in geothermal energy generation.
• My former colleague Austin Vernon has a new post on drones and their impact on air power. Here is one key bit: “Drones will continue to be significant, and most effort should be on masses of low-performance aircraft optimized for new tactics. That requirement pushes heavily towards battery electric powertrains. These systems can only win by running traditional platforms into the ground with extreme sortie rate, logistical tail advantages, basing flexibility, and ease of production. Design, testing, and production capacity should be relatively inexpensive, enabling significant capability without betting the farm in the same way as a high-end program.”