When we talk about artificial intelligence, the conversation usually hits on trillion-dollar data centers, escalating export controls, or the breathless rhetoric of a winner-takes-all U.S.-China tech war. But treating frontier AI as a zero-sum arms race erases the ground-level economic realities and shared vulnerabilities that actually govern our future. To truly understand the stakes of this transition, from the fragile economics of overbuilt infrastructure to the sudden hollowing-out of entry-level cognitive labor, you have to look past the hype and examine how technology diffuses into the real world.
In this interview, the Monitor sat down with tech guru and policy strategist Alvin Wang Graylin to dismantle Washington’s prevailing assumptions about tech competition. We dive into how best to reframe U.S. AI competition with China, the mounting crisis facing young graduates entering an automated workforce, and why treating AI as a shared public good, rather than a geopolitical weapon, is the only way to avoid mutually assured miscalculation.
Alvin Wang Graylin helps leaders understand how AI is reshaping competition, risk, and the global economy. He is a Digital Fellow at Stanford’s Institute for Human-Centered AI, a Senior Fellow at the Asia Society Policy Institute’s Center for China Analysis, and Professor of AI and Technology Policy at the University of Washington. His 35-year career spans AI, semiconductors, data centers, cybersecurity, enterprise applications and edge devices on both sides of the Pacific, including senior roles at Fortune 100 firms and he has founded four venture-backed startups. One of his startups built a ChatGPT-style conversational AI search service in China, deployed with all three Chinese mobile carriers in 2005. Today he advises governments, multilateral organizations, and Fortune 100 boards on AI strategy and governance. He is the author of Our Next Reality and co-hosts the Big Bang Tech Report, a podcast on the business impact of advanced technologies.
His Substack can be found here.
Elijah Shoaf: I’ll start with our first question here. You have reframed the U.S.-China AI dynamic from a simple race to an innovation race or a platform race, where cooperative strategies are optimal. How can the U.S. effectively signal this shift towards cooperation to Beijing when current bilateral trust is so heavily fractured?
Alvin Graylin: Right now, we’re signaling the opposite. When all of the various restrictions are getting intensified with new bills being introduced, and the Commerce Department adding new export controls, I think we’re sending a very competitive signal. In reality, particularly in the area of safety and security, that is where collaboration needs to happen.
You cannot race yourself to a safer world. You can only make the world more dangerous when you race and throw away guardrails. I would really like to see signals at the bilateral AI dialogue due by late November that something more substantive comes out of it. I would like to see both sides come in with more preparation. In the past, the Chinese probably had a little bit more preparation than the U.S. in this area. But this time, we still have another month and a half until the meetings happen. There is time for something substantive to be provided from both sides, so that we can come out of that with announcements beyond just scheduling a next meeting.
ES: Yes, one would hope it’s a good sign that we’re coming together and at least talking about the issues.
AG: Yes, after the first meeting, the fact that we have a second meeting is great. In 2024, there was only one meeting, and then nothing else happened after that. The fact that we have a second meeting and it’s got a name, it’s got people assigned to it, makes it real. Now the question is: what happens in between? Will there be more substance created? If there is, then it creates a lot more hope and the beginning of trust.
ES: In your recent Quincy brief, which is one of the reasons we reached out, you note that any U.S. frontier AI advantage would likely evaporate rapidly, making a permanent military advantage impossible without risking severe escalation. Right now, there is a “winner-takes-all” framing of AI competition, and that inadvertently increases the risk of miscalculation over Taiwan. Would you identify that fundamentally as a supply chain and alliance problem, rather than a race where once someone hits a certain threshold, they win
AG: I think there are some base assumptions right now fueling the race. One is that the world is zero-sum. Another is that there is a finish line and it’s a winner-takes-all race. None of those assumptions are true. If they aren’t true, then the way we are approaching this competition is by making very big bets with uncertain outcomes or returns.
The fact that Chinese capital expenditure is one-tenth of the U.S., yet they are within a few months of our capabilities, suggests that our investments are not paying off. By investing roughly $725 billion in AI infrastructure this year, and likely over a trillion next year, we’re also making the economy more fragile. It’s highly concentrated right now; almost half of the growth in the S&P 500 is in AI-related stocks, so if there is a correction, that is going to be relatively painful. There is over $3 trillion off-balance-sheet commitments right now from these AI and hyperscaler companies that will have to be paid off, which adds more financial fragility. The fact that there are a lot of half-built data centers right now being delayed, canceled, or lacking parts means you have a lot of committed capital in something that is not producing value for the economy. All these things suggest that a correction is coming.
The American economy will probably be affected a lot more than the Chinese economy simply because of the exposure we have. Our current strategy of trying to do everything we can to be there first is an unclear and probably unfounded strategy.
ES: Your answer brought up a point outside of my planned questions. As you noted, we invest heavily in AI across all sectors, and while the Chinese invest comparatively less, their capabilities remain very close to ours—within a few months. On the U.S. side, where do you think this disparity comes from? Is it purely speculation? Higher labor costs? Why do you think we are putting so much more money into this for a similar result?
AG: We are creating more compute capacity, so it’s not necessarily one-to-one in terms of spend versus results. But looking at compute right now, the American proportion is somewhere around 70 to 75% of the global total. China is around 10%.
One of the reasons the Chinese are not as far behind is that they have access to offshore remote compute for training. Not all labs use it, but a significant portion do. They essentially get the value of those compute cycles without having to take on the capital expenditure and financial exposure. They are getting it as needed, rather than buying the house they’re renting.
Additionally, the Chinese focus is much more on efficiency. Their lack of compute actually gives them more incentive to create greater efficiencies in memory, energy, power, and GPUs. They are essentially using hardware two or three generations behind, which means whatever they create can be much more easily diffused into the market.
That touches on your question of whether this is an innovation race versus a diffusion race. In the real world, diffusion is actually more important than innovation. It doesn’t matter who invents a technology if they can’t integrate it into their economy. The Chinese government’s focus for its labs is to make technology that can be put into education, healthcare, and manufacturing to improve the country’s productivity. It also creates greater trade with the rest of the world. Because they are making open-source models, they are being adopted globally since not everyone can afford frontier model pricing. From both of those objectives, the Chinese are much more cash-efficient in achieving their results, versus the frontier labs, which are spending a lot more money and creating more leverage and financial risk, but not necessarily generating financial returns.
ES: That connects into my next question. You highlighted that Chinese open-weight models have rapidly captured market share and diffused, holding the first six positions on OpenRouter in July of 2026 while competing with leading U.S. models at roughly one-tenth of the API price. What specific factors allow Chinese developers to achieve such high performance at a fraction of the cost, despite expanding U.S. export controls on memory and lithography?
AG: Export controls are forcing them to be more creative. Indigenous innovation is happening, and around 40 to 45 percent of all AI research papers in the world right now are coming from China.
Many U.S. labs argue that distillation is the primary reason for their success. However, based on what I am seeing, distillation alone wouldn’t allow them to reach these capability levels. They are publishing real innovations that significantly reduce requirements on memory and compute. For instance, a single paper can cut the necessary compute by 10x, and that doesn’t come from distillation. Anthropic’s own reporting found no evidence of distillation from its Mythos and Fable models, yet new models coming out of Chinese labs are near parity with them, so distillation alone can’t explain how they got there. This shows they are building those capabilities without needing to distill from U.S. models. Writing off Chinese competitiveness as mere distillation will, in the long term, be a detriment to American labs.
Interestingly, companies like xAI are embracing distillation, and Google Gemini even offers a “distill-as-a-service” option on a per-million token basis, so the U.S. labs’ stance is somewhat inconsistent. Maintaining a long-term lead is very difficult in this space because when you make something open source, you have two to three million researchers around the world fine-tuning, improving, and uploading those models back to a shared open-science repository. No matter how smart a lab’s 2,000 or 3,000 employees are, it is very difficult to compete with an army of millions of open-source researchers.
ES: To transition from the high-level state-to-state competition to the broader economy: you point to a 16% relative decline in employment for 22- to 25-year-olds in highly AI-exposed U.S. occupations, effectively breaking the structural pipeline that produces future mid-level and senior talent. So as an American who just finished my master’s degree myself preparing to enter the workforce, I’d like to know what specific sectors within the U.S. economy are driving this entry-level collapse?
AG: You are definitely not the only one feeling that pain—new grad employment is way down, and while I mentioned 16%, updated data is showing closer to 19% now, so it is getting harder. Another factor people rarely talk about is an underemployment rate of over 40%. That means you might have a college degree, but you’re working as a barista or driving for Uber; it doesn’t show up in standard unemployment numbers, but it deeply impacts personal satisfaction.
Roles in customer service, entry-level accounting, and data analysis are starting to see significant reductions in force. Much of this has historically manifested as a lack of new hiring, which is exactly what we are seeing in the continuous decline of new job postings since 2023. This is an unusual period because, historically, a rising stock market corresponds with an increase in job postings. While it may not be entirely due to AI, there is definitely a strong relation. Anyone using these technology tools regularly realizes how capable these systems are at replacing junior workers, research assistants, analytics personnel, and even simple coding jobs.
ES: With emerging technology, things can initially paint a bleak picture, particularly for those of us most affected, but you have also discussed solutions. To address the displacement of cognitive labor, which constitutes about 60 to 70 percent of the U.S. workforce, you have recommended establishing a “GI Bill for the AI age” and exploring a levy on labor-displacing automation. From a U.S. legislative perspective, how can policymakers calibrate this automation levy to adequately fund transition programs without driving American AI developers and their innovation offshore?
AG: My suggestion was actually not to tax them on profits because they are all money-losing companies, so you cannot use current systems to gain taxation revenue. You have to tax spending. They are spending roughly $725 billion this year, and likely over a trillion next year, and my suggestion is to use more of a Pigovian levy model, similar to taxes on goods with negative externalities like alcohol, tobacco, or gasoline. In addition to the cost of the goods, they create negative externalities onto society, so you charge extra taxes.
This would probably slow down investment in the industry, which is a good thing because right now we are over-investing. We actually have tax incentives encouraging overbuilding: since last year’s tax bill, companies can immediately write off the full cost of the servers and chips that fill these data centers. We are doing the opposite of what we should be doing regarding taxation and displacement issues.
I recently wrote a paper called The Last Straw discussing the $40 trillion national debt. Right now, 84% of federal tax revenue comes from individual income and payroll taxes. We tax payroll dollars at close to 30%, while capital expenditure dollars are taxed at around 5%. As soon as AI becomes capable, smart managers will immediately move from hiring people to hiring robots or virtual cognitive workers because they are taxed lower, can work longer, and don’t have healthcare or management overhead. If we keep pushing this forward as fast as we are, we will reduce total tax revenue, making it even more difficult to service the interest on our debt, let alone repay it. For those who say we’ll grow our way out of it, even if the economy grows, if the taxation system isn’t modified, we will see a net reduction in total tax revenues.
ES: Is anyone talking about this in the political sphere? People talk about AI, data centers, and water use, but it often feels like a shallow conversation. The only person I’m aware of addressing this intelligently is Andrew Yang, but he hasn’t been in a position of power for a while.
AG: You have folks like Bernie Sanders talking about essentially banning data centers, though I think he goes a bit too extreme, which causes people to stop listening. There are others talking about moratoriums for a certain period or extra taxes on data centers, which are good things. There are a number of bills being put into process, but thousands of bills are initiated every year, and fewer than a few percent pass. I don’t know if there is the political will to put together GI Bill-type initiatives right now.
Unfortunately, I don’t know how much time we have, because this transition will happen very quickly. Technology adoption follows a hockey-stick curve, meaning the impact is relatively low until it hits a threshold, which is about 80% of the work tasks of a human worker. Until AI can take over roughly 80% of a job, managers are unwilling to replace humans, so it remains augmentation. But once it crosses that threshold, the decision happens fast. On GDPval, a benchmark that tracks real-world work tasks, the best models matched or beat human experts on under half of tasks a year ago; now they are above 80%.
We are getting there. Now it’s about how quickly these are integrated into business processes, which could take anywhere from a year to two years. We likely have another couple of years before a big wave of displacement hits. The sad thing is that we won’t see much of it until it happens—it’s essentially invisible until it hits all at once.
ES: So, if I’m hearing you correctly, we likely won’t take substantive action until we actually feel the effects?
AG: Human nature, and especially governments, tend to be much more reactive. That’s why I wanted to highlight the GI Bill—it’s the only major initiative I’ve seen that got ahead of the problem. They passed it in 1944, a year before the end of the war, because they knew 15 million young men would return with no jobs, creating a major societal crisis. What most policymakers don’t realize today is that even if it’s not 60 to 70 percent of the workforce displaced, even if it’s half or a third of that, we are talking about multiples of that 15 million people. That will create both a social and economic crisis with the potential for social instability.
ES: To swing back to the macro political level of U.S.-China relations: in your recent brief, you advocated for an independent international testing institution to evaluate publicly released commercial models, explicitly keeping it separate from both Washington and Beijing’s intelligence services. What initial Track 2 diplomatic initiatives could the U.S. launch today to help build confidence between Chinese and American AI labs regarding shared safety evaluations before formal, state-level agreements are signed?
AG: We can start with simple things like shared definitions of risk and common test matrices that everyone can understand. I speak with labs on both sides, and the Chinese are very wary of U.S. testing labs because they don’t want them used as tools to unfairly exclude their models from the market. They want transparency, which is why I advocated for third-party testing facilities that are neither American nor Chinese, so both parties see them as neutral ground.
That’s a good place to start, and it serves as a global good because we want models, whether open-source or closed-source, thoroughly tested for safety and security. While you can’t guarantee 100% security, having a clear standard of what is permissible is definitely desired.
ES: Do you see any existing institution that both the U.S. and China would trust?
AG: In the paper, I mention the UK AI Security Institute (AISI). They are well-funded, staffed with informed and capable people, and have experience testing models from around the world. They would be seen as a credible place to start, though they would likely need additional funding to act as a testing arm for the world.
Just last month, Washington asked U.S. frontier labs to pause the institute’s early access to their latest models, and access was only restored after 17 days, so this isn’t just about Chinese models. Western labs are also protective of their technologies.
ES: That gets us through our questions. Is there anything else you wanted to touch on or think should be covered?
AG: I think it’s important to recognize something. Right now, AI is being positioned as a geopolitical competition between countries and a financial competition between companies. But this is a technology that will be commoditized by open source, meaning it won’t create long-term economic returns for these frontier labs.
If that’s the reality, we need to stop looking at it as a tool for financial domination or a weapon for geopolitical domination, and start treating it as a global public good that should be shared with the world. Once you view it that way, trying to hoard this technology for a single country or company is not only irresponsible, it’s immoral.
This technology can be used globally at a relatively low cost if we stop chasing the absolute frontier every week and releasing new models constantly. There’s no way for technology advancing this quickly to be absorbed into society and the economy; if it isn’t absorbed, the benefits aren’t realized. Spending trillions of dollars globally to build things that can’t be fully integrated is a misallocation of resources. We need to stop duplicating work; we currently have dozens of labs around the world training models that are within a few percent of each other, consuming scarce compute resources. If we pooled our resources, talent, and data, we could achieve more with a lot less. We could save that money and redirect it toward improving healthcare, education, and infrastructure around the world.