The US can lose the modern tech arms race

... in the dumbest, most un-American way

As countries invest in and develop new technologies, these technologies in turn shape the economic landscape of the world, the military capabilities of countries, and the diplomatic clout countries have as they pursue their national strategies. The most geopolitically consequential tech advancement of the 20th century was the atom bomb, decisively establishing America as a superpower at the end of World War II and shaping the global landscape of nuclear deterrence, Cold War rivalries, and more.

Semiconductor advancements in the middle to late 20th century reshaped geopolitics as well. Computers delivered to homes didn’t just advance the productivity of workers and serve as the infrastructure for the internet & the modern, lucrative world of e-commerce, they also advanced the capabilities of sensors and information-processing of data from censors, created cyber threats and opportunities, and improved the capabilities of military hardware (cars, tanks, boats, planes, drones, etc.).

Nuclear weapons and semiconductors had radically different patterns of technology development. Nuclear weapons were first developed in the Manhattan Project, on secure facilities of secretive groups within Los Alamos National Labs financed exclusively by federal funding for national defense. Semiconductors, meanwhile, were developed in the private sector, powered by a mix of venture capital funds and enduring revenues from successful products leading to future rounds of venture capital funds, yesterday’s founders becoming tomorrow’s investors, and more.

In a previous life, I worked on disruptive technologies, AI, and complex systems for national security at Sandia National Labs. Nuclear weapons were disruptive technologies that changed the global landscape of diplomacy, deterrence, and the costs of conflicts. Semiconductors were disruptive technologies that changed the processing of information and paved the way for the internet, a disruptive technology improving economic productivity by changing the way commerce, communication, and collaboration occur. Perhaps not ironically, my work at the National Labs was cut short by a mix of DOGE cutting contracts based on keywords (e.g. grants involving the “diversity” of technologies could very well be cut based on the keyword “diversity”) and greener pasture opportunities to advise the NIH Office of the Director, in collaboration with the White House Office of Science and Technology Policy and National Security Council, on the dangerous Gain of Function Executive Order.

Even farther back in time, I received a PhD from Princeton University studying the mathematics of competition in biological and economic systems. This interdisciplinary background paved the way for work in finance, epidemiology and bioattribution, and the development of AI tools for national security.

All that’s to say: hear me out as I talk about AI technological competition du jour.

The Unproven AGI Hypothesis

AI is disruptive technology, this much is clear, but exactly how disruptive, in which applications, and to what effect remains to be seen. It’s important to separate fact from hypothesis and hypothesis from theory when making consequential decisions.

The fact is: the internet generated a lot of data and neural networks are very good at approximating those data, guessing blank or masked words or estimating a picture that comes with corresponding text.

The hypotheses in the AI world are many, but chief among them is the hypothesis that neural networks can achieve what we call Artificial General Intelligence or AGI, a hypothetical level of intelligence where computers beat human brain power at every task. The hypothesis (if we’re kind, and not calling it speculation) then extrapolates that AGI, by beating human brain power in every task, will itself generate new technologies and thereby lead to a positive feedback loop of technological development that leaves the losers of this tech arms race so far in the dust they look like Australopithecus holding a rock compared to futuristic humans and technologies powered by AGI.

AGI does not exist today. True to their mathematical nature, neural networks are what we call universal function approximators, or mathematical models that can approximate any function with enough data. Trained on the internet, these mathematical models thus approximate the internet with soft-ish step-functions (SoftMax, Relu, …, functions that look like a step on a staircase with round edges). The weird, fuzzy curvature of data on the internet is approximated with an interesting, granular set of approximately, locally-constant steps defined by a very complicated set of functions fed into functions that, at the end of the day, are soft-ish step-functions.

If we detach ourselves from the AGI hypothesis and focus on the mathematics of neural networks as universal function approximators, it’s perhaps no surprise that these mathematical models haven’t gotten much better than, well, the internet, the data on which they were trained. Below is a plot of such model performance across a variety of tasks. I intentionally remove “AI” from “AI model” in the prior sentence because I aim to separate fact from hypothesis, and model is a fact but AI is a poorly-defined hype buzzword masquerading as a hypothesis. The plot below shows model performance over time by several tasks relative to “human performance”, where “human performance” is typically some measure of central tendency of humans - the average or median etc. Over time and across tasks, models initially get better quite quickly, but slowly these models trained on the internet seem to plateau around 0-20% better than the average/median human.

The fact that models trained on the internet don’t get that much better than the midpoint human is consistent with the fact that these models are universal function approximators approximating the internet, and a weighted draw of words from the internet may be better than the midpoint human depending on how we weight the draw. The hypothesis of AGI remains a hypothesis in light of this data - AGI has clearly not been achieved yet, but that doesn’t mean it can’t happen. It seems very difficult to disprove the hypothesis of AGI - that would require some study of human brains in all their possible configurations compared to computer models, along with more agreeable definitions of “intelligence” that help us compare the two, but we note that this difficult hypothesis to disprove makes it a bit worrisome as a scientific hypothesis. At a minimum, the hypothesis that billions of dollars of compute would generate AGI has been disproven.

The AI Focus of US Tech Strategy

… but what if we put hundreds of billions more? Will we get AGI then? If not hundreds of billions, what if we try trillions?

This is the current path the US approach to AI is on. Silicon Valley VCs, many of whom became famous as founders in the Dot Com era or subsequent SaaS boom and App craze, has been investing heavily in this next silicon-based tech craze. These VCs are closely connected with some of the largest founders and leaders of the AGI pursuit such as Sam Altman, former YCombinator President and current OpenAI CEO. The Trump Administration also has close ties to this AI community, including dominant players in the tech VC world like Marc Andreesen, former founder of Netscape Navigator and founder of Andreesen Horowitz, one of the largest venture capital firms in the world with ~$90 billion assets under management. Mr. Andreesen is an outspoken believer in AI and AGI who plays a major role shaping the Trump Administration’s AI policy, appointed to the President’s Council of Advisors on Science and Technology alongside NVIDIA CEO Jensen Huang, Meta CEO Mark Zuckerbeg, and others.

The PCAST membership reveals the technological myopia of the Trump Administration. Oracle, Dell, NVIDIA, Meta, former internet founders, and other influential individuals in the Silicon Valley orbit and orbit of silicon-based technologies abound. There is one professor in quantum physics from UC Santa Barbara and one CEO of a fusion company, but the rest are all classical silicon-based economic organisms. With investments in finance as well as investments in technological development, putting most of your eggs in one basket is a good thing if you can pick the right basket. It’s a bad thing if the core assumptions behind that basket end up not holding.

The AI basket is premised on AGI. The small network of individuals driving significant venture capital investment and tech policy on AI, all patriots, strongly believe AI will be the most important disruptive technology of the 21st century, that the time to invest in AI is now to ensure we become the futuristic society powered by AGI and our adversaries are left as Australopithecus with their silicon rocks. They may be right. However, in an administration that prizes loyalty in a way that can at times discourage dissent, they would be wise to analyze the alternative in which their core assumptions don’t hold.

For example: what if the data on the internet, and data collected by platforms aiming to fine-tune AI, is not hyperintelligent and thus models trained on such data will not yield breakthroughs? What if innovations aren’t discoverable by interpolation between existing words with SoftMax functions of functions, but instead come about by radical new syntheses of ideas, new words that don’t exist in training data, new concepts that must then be canonized before adopted by humans? Atomic Theory was first posed in the early 1800s yet wasn’t widely adopted among humans until the early 1900s. Heliocentric theories of the solar system were proposed a long time before adopted, and even mathematical proofs like the existence of irrational numbers took the death of the old guard to usher in the era of adoption. The core hypotheses of AGI aren’t just that AGI can exist and be usefully hyperintelligent, but that humans would recognize and adopt such intelligence, that we’d form policy and invest funds based on the recommendations of these models.

What if we’re wrong?

In 2025, Sam Altman secured $40 billion in venture investments to pursue AGI. This investment reportedly came from Japan-based SoftBank, Microsoft, and others. The 2025 megadeal, combined with similar large venture-growth deals of later-stage AI foundation model labs, created an unprecedented surge of venture investment concentrated in a few small deals. On top of these deals, there exists significant venture debt in the AI world, and, most notably, a web of circular dealmaking between chip manufacturers (think: NVIDIA), hyperscalers who host chips on which AI models are trained (think: Microsoft, Google, AWS), and AI model developers (think: OpenAI, Anthropic et al.).

The recent figure below by Bloomberg shows the nature of circular deals in the AI community. NVIDIA can invest billions in model companies, either as debt or equity but often as equity with the promise that these billions go back into buying NVIDIA chips or running models on NVIDIA GPUs hosted by hyperscalers who buy NVIDIA GPUs.

These deals aren’t localized within the private sector, but go side-by-side with federal science policy and even international sovereign wealth moves as well. The Trump Administration proposed a Stargate Strategy, a $500 billion public-private endeavor to build AI infrastructure in the US. Formed Stargate LLC, this joint venture is owned by OpenAI, SoftBank, Oracle, and Emirati state-owned investment firm MGX. Never mind for now the conflicts of interest on the public side of this endeavor (e.g. crypto co-owned by Trump and his advisor Steve Witkoff’s son being used by Saudis to fund this endeavor in ways that directly increase Trump and Witkoff’s’ wealth). Instead, we note that there are a small number of players - OpenAI, Microsoft, NVIDIA, Oracle - showing up again and again in both each others’ private investments, the President’s Council of Advisors on Science and Technology, and AI infrastructure projects announced by the White House.

Focusing on the people and entities involved isn’t an effort to demonize them, but to understand the social processes and potential blind spots of their assessments. I took the same approach when analyzing epidemiological forecasts and public health policy during COVID, arriving at a method of analysis I coined “socioscientific inefficiency analysis”, i.e. finding the blind spots of social groups driving market-relevant science and science policy.

Circular deals are combining with federal policy advice driven by people with the most exposure to these circular deals in AI, the most to gain and the most to lose. Again, perhaps in Ayn Randian fashion these dealmakers are the Atlases now empowered by American Dynamism at the federal level to usher in a golden era of Artificial General Intelligence raising the United States beyond the heights of Atlantis as a marvelous, fantastic, futuristic new civilization.

… or perhaps these internet platform and data companies and their people are wrong about the economic and military value of models approximating data from the internet.

The economic hypothesis can be undercut if open-source models, such as those frequently released by Chinese competitors, prove as good as proprietary models such as OpenAI’s models behind their $40B 2025 fundraising round. While it’s possible that complex systems of models of models, mixtures of experts and other ways of having models interact could lead to more intelligent models, it’s also possible the core problem of universal function approximation of data on the internet, and the questionable intelligence of interpolation between such data as the path to innovation, proves insurmountable and undermines the AGI thesis.

The military value of AI is something I won’t discuss much of here except to say: autonomous vehicles are interesting capabilities and I won’t argue against investment in these capabilities, but the decisions these vehicles make are very narrow, they are not AGI computer-brains innovating new technologies but rather they are smart bullets and clever wingmen designed for the purpose of neutralizing adversary capabilities.

If the small set of people and companies driving these imbalanced AI investments and science policies are wrong, then the long-term technological competitiveness of the US can be undermined by this technological and financial misadventure. The combination of debt, circular deals, and lofty forecasts of GPU demand based on promised revenues from AI models all point to a massive bubble, undermined further by recent development such as Chinese open-source models, Chinese acquisition of lithography capabilities to manufacture their own GPUs, and compounding economic crises such as the cavitation of oil and other supplies in the Straight of Hormuz that can have unintended consequences on the cost of energy and therefore the cost of compute and profitability of AI businesses at a time when the entire network of Atlases sweating in expensive suits in the DC summer is depending on their profitability to pay off debts and meet revenue forecasts.

If these advisors failed to anticipate the sensitivity of their economic thesis to open-source models from China, that would be a massive failure in analytic foresight, a failure to contend with potential adverse consequences of their course of action. Of course the adversaries and competitors of the US would try to undermine these efforts. If all of our eggs are in one technological basket then that reduces the number of baskets our adversaries need to disrupt in order to frustrate our national strategy. Someone clever and analytical like me is probably sitting in Beijing, looking at the US fixation with AI and considering low-cost ways to press their needles against our bubble. I won’t go so far as to brainstorm specific ideas for them, but I’m also no longer in the business of brainstorming ideas for the US government following DOGE’s cuts, my departure from national labs, and my return to life as a private citizen trying to balance my own portfolio.

Beyond the world of adversarial economic strategy, there is a world of missed technological opportunities caused by overinvestment in AI. As venture firms grow larger to support the institutional giant LPs, as venture deals grow larger and preferentially allocated for later-stage AI companies, funds available for every other kind of technology become harder to find. Biology, quantum, fusion, chemistry, psychology and human services (including e.g. vocational or training services to accelerate workforce development), and so many other important areas of science and technology remain underinvested. While AI accelerates, these other fields of science lag, dragging along desperately applying for grants without the winds of hype to support their worthy endeavors.

The Un-American Nature of Such a Failure

American economic competitiveness is not made by Atlases, but by disruptors. As disruptors become Atlases, incumbents with control over venture capital allocation and federal policy, we lose the widespread, rebellious entrepreneurial spirit that drives American innovation and dynamism.

We’re also seeing a critical gap in the human resources side of American tech development. Scientists are primarily trained in academic labs, requiring years of training to become proficient in advanced technological areas like biology, quantum physics, computer science, and more. Scientists following the path of scientific excellence rarely encounter wealth along the way, and yet wealth or access to capital is becoming a driving force behind whether or not people found companies, let alone succeed. Capital allocators are missing many scientists with great ideas for truly disruptive technologies because these scientists are outside the social networks of investment circles increasingly focused on a smaller and smaller range of technological advancements, further and further distanced from the front lines of science and technology across fields. While investors gloat about the availability of capital for tech ventures, scientists view a desert of capital scarcity driven by oases of capital consumed by AI.

America was the underdog from the time of its founding, and underdogs fight harder. We were underdogs against the Brits in the Revolutionary war, requiring innovative new guerilla military strategies and utilization of our home-field advantage to overcome a militarily superior adversary. We were underdogs in World War II, yet we demonstrated an economic and logistical nimbleness in our industrial capacity, and we funded a rag-tag bunch of scientists studying atoms to develop nuclear weapons. We were not the underdogs in the later half of the 20th century, but we funded underdogs developing computers, developing mobile phones, developing apps. At one point in time, AI too was an underdog, it was the disruptor, but as of today AI is the Atlas that isn’t shrugging but rather using its incumbent gravity to pull capital and federal policy in its favor.

This fixation with AI can cost us dearly. With trillions of dollars locked up in circular deals among the same AI players advising the president on federal policy, we run the risk of capital misallocation. By putting all our technological eggs in one basket, we provide a basket for adversaries to target. By watering the hungry trees of unprofitable AI labs, we neglect the garden of crops that can sustain us and bear fruits in seasons when AI is not doing well. The Atlas is acting like a grown child still sucking on the teat of the US government, unable to sustain its own R&D or infrastructure for a technological venture of historic scale whose core hypothesis - AGI - could very well turn out to have been a fantasy all along. The consequences of such a technological and economic misadventure could leave us feeling like Australopithecus as our adversaries race ahead with well-balanced tech portfolios.