A veteran of both US and Chinese AI industries argues the US-China AI arms race is largely a myth used to justify bad policy and overspending.
Is the US-China AI arms race real? #
Not in the way it gets described in Washington and in corporate earnings calls. There is real competition between American and Chinese AI labs, but the “winner take all” framing, the idea that one country will cross a finish line and lock in permanent dominance, doesn’t hold up. Intelligence is turning into a commodity fast: open-weight models are shrinking, running on cheaper hardware, and closing the gap with closed frontier systems. That trend undercuts the entire premise of an arms race with a single victor, and it’s a point made forcefully by Alvin Graylin, a technologist who has worked in AI on both sides of the Pacific since the early 1990s, in a recent conversation about US-China AI policy.
TL;DR #
- The arms race framing treats AI as a zero-sum contest with one winner, but intelligence is becoming commoditized software that spreads regardless of who “wins.”
- Open-weight models are shrinking fast enough to run on consumer hardware like a Mac Studio, while closed frontier models still require racks of GPUs costing millions of dollars.
- Much of the urgency behind the arms race narrative functions as industry lobbying that channels public money and policy attention toward a small number of companies with high margins to protect.
- A drug discovery example from China shows companies using smaller, specialized models (around 10 billion parameters) rather than massive general-purpose systems, undercutting the idea that only the biggest model wins.
- Heavy data center investment carries financial risk: reporting cited in the conversation put hyperscaler off-balance-sheet debt and obligations near $3 trillion, compared to roughly $1.6 trillion in subprime debt at the peak of the 2008 crisis.
- The bigger near-term danger isn’t a foreign state “winning” AI, it’s non-state actors and economic disruption , including labor market effects already showing up among young workers.
- Young people entering the workforce are advised to build broad, hands-on experience across full project lifecycles rather than specializing narrowly, since narrow specialization is exactly where AI adds the least differentiated value.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
What does “commoditization of intelligence” actually mean? #
It means the cost of accessing high-quality AI capability keeps falling, and the capability itself keeps spreading beyond the handful of companies that first built it. Open-weight models are the clearest evidence: newer, smaller architectures can now run on a Mac Studio, while comparable closed-source models still need multiple server racks worth millions of dollars. That gap is closing, not widening.
The comparison used in the discussion is electricity. Nobody “won” the electricity race. Once the technology matured, it became infrastructure that everyone eventually got access to. Software behaves the same way, arguably faster, because it can be copied, distributed, and run locally once someone releases the weights. A model that requires a data center today often runs on a laptop within a year or two, once it’s been quantized and optimized. That trajectory makes hoarding intelligence at the national level a losing strategy, not because of politics, but because of how software physically propagates.
Why does the arms race narrative persist if it isn’t accurate? #
Because it serves specific interests. A small number of companies benefit from high profit margins tied to being seen as the sole providers of frontier-level intelligence. Framing AI development as a national security race justifies enormous capital expenditure, favorable regulation, and public funding directed at those same companies. If intelligence is actually commoditizing, that story becomes harder to sustain, so there’s a structural incentive to keep the “arms race” framing alive even as the underlying technology undermines it.
This isn’t a claim that competition doesn’t exist. Labs and countries clearly compete. The distinction is between healthy competition, which drives useful progress, and a zero-sum “winner takes all” narrative that treats the field as a single race with a finish line. That narrative is what drives misallocated national security spending and discourages companies from prioritizing safety over speed, since safety work looks like a competitive handicap if you believe someone else is about to cross the finish line first.
Is bigger always better when it comes to AI models? #
No, and this is one of the clearer counterpoints to the arms race framing. A leading drug discovery company in China, cited as an example, doesn’t use a massive general-purpose model with trillions of parameters for its work. It uses a roughly 10-billion-parameter model specialized for the specific disease it’s targeting. Smaller, targeted models can outperform giant general models for well-defined problems, and they cost far less to run.
That has a direct policy implication: pouring national resources into building ever-larger frontier models isn’t obviously the best use of capital if narrower, cheaper, specialized models already solve real problems. Money spent chasing the biggest model could instead go toward infrastructure that has more visible payoff (hospitals, power plants, roads) rather than data centers built on the assumption that scale alone wins the race.
What’s the financial risk of over-investing in AI infrastructure? #
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The buildout of AI data centers is being financed partly through debt and off-balance-sheet obligations at a scale that invites comparison to prior credit bubbles. Reporting referenced in the conversation, attributed to the Wall Street Journal, put roughly $3 trillion in off-the-books debt and obligations among the top American hyperscalers. That’s roughly double the $1.6 trillion in subprime debt outstanding at the height of the 2008 financial crisis.
That doesn’t mean AI infrastructure spending is worthless, but it does mean the industry is making a large, concentrated bet on a specific narrative: that demand for frontier-scale compute will keep growing at current rates indefinitely. If that assumption breaks, either because open-weight models satisfy most real-world demand more cheaply, or because returns on AI investment underwhelm expectations, the financial exposure is significant and concentrated in a few firms.
What’s the real risk if it isn’t a state-vs-state arms race? #
The more immediate risks are economic disruption and misuse by non-state actors, not a foreign government “winning” AI. On the labor side, there’s already measurable strain. A Stanford study, referenced in the discussion and reportedly updated with continuing findings, found roughly a 19% gap in payroll outcomes for workers aged 20 to 25 compared to what would be expected relative to other age groups, concentrated in industries exposed to AI adoption. Total US payrolls have reportedly been declining since 2023 even as stock market valuations climbed, a divergence described as unusual since the two measures have historically moved together.
The explanation offered isn’t that AI has replaced most jobs outright. Adoption is still uneven, with an estimate that only around 20% of companies have meaningfully adopted AI and perhaps 5% have done so well. But companies that build and sell the tools (“the ones selling shovels”) are seeing early gains, while broader hiring, especially of younger, less experienced workers, is softening. Employers appear to be substituting AI for entry-level tasks rather than laying off experienced staff, which concentrates the pain on people just entering the workforce, with little social safety net to cushion the transition.
How should people prepare, if not by picking a specialized AI-proof job? #
The advice is to build broad, “T-shaped” experience: know a lot of things reasonably well, and go deep on at least one thing by actually building, deploying, and eventually sunsetting something end to end. The reasoning is straightforward. If your value comes from narrow specialization, that’s precisely the kind of work AI systems already do competently, and worse, a specialist who can’t independently judge or critique an AI’s output becomes a rubber stamp rather than a check on it. That’s a fragile position professionally.
Broader experience, paired with genuine hands-on project ownership, builds the pattern-recognition and judgment that’s harder to commoditize. Reading history, philosophy, psychology, and management alongside technical skills isn’t a nostalgic add-on, it’s what lets someone connect disparate pieces of information in ways a narrow specialist, human or AI, tends to miss.
Frequently Asked Questions #
Is there really no competition between the US and China in AI?
There is real competition, but the article’s argument is against the “winner take all” framing specifically, not against competition existing at all. Labs in both countries are racing to build capable models, but no single country is positioned to permanently lock out the other, because open-weight models keep closing the gap and spreading capability regardless of who develops it first.
Other agents ship a demo. Remy ships an app. #
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Why do open-weight models undercut the arms race narrative?
Because they commoditize intelligence quickly. A capability that once required a data center with racks of GPUs costing millions of dollars can, within months, be quantized and run on consumer hardware like a Mac Studio or eventually a laptop. That makes it very difficult for any single country or company to maintain exclusive control over frontier-level AI capability for long.
Who benefits from the arms race narrative if it isn’t accurate?
A small number of companies with high profit margins tied to being seen as the primary providers of frontier AI benefit most, since the narrative justifies favorable policy treatment, public funding, and continued high valuations. It also benefits parties who want to justify large defense and infrastructure spending framed around AI competitiveness.
What’s the actual evidence that young workers are being affected by AI right now?
A Stanford study cited in the discussion found roughly a 19% payroll gap for workers aged 20 to 25 in AI-exposed industries relative to expected levels, and total US payrolls have reportedly declined since 2023 even as stock markets rose, a divergence not typically seen historically.
What should someone do differently given this environment?
Build broad experience across many domains while also going deep enough on at least one project to see it through an entire life cycle, from design to deployment to retirement. Narrow specialization without hands-on ownership is the profile most easily replaced or devalued by AI tools.