A Case for American AI Leadership
A retired veterinary internist on competition, caution, and keeping our eyes on the evidence
Where I stand
I’ll be upfront about my views. I believe the United States should keep building artificial intelligence, and keep leading in it. I also believe we are still in the toddler phase of this technology. Toddlers grow fast, stumble often, and need watching. Both things are true at once.
I use several AI systems in my own work, and I’m open to all of them: Anthropic’s Claude, OpenAI’s ChatGPT, Google’s Gemini, xAI’s Grok, and open-source models anyone can download. For software work, I think Claude is currently the best I’ve used. But that is exactly why competition matters. Today’s leader is not guaranteed to be tomorrow’s, and no single company should be locked in by rules written in its favor.
In my last two articles, I tried to follow a few principles, and I’ll hold myself to them here:
• Truth and honesty come first, even when the evidence is inconvenient for my own side.
• History over hype. Judge a technology by what it has actually done, not by utopian or doomsday predictions.
• Separate the messenger from the message. Ask who benefits from a claim, then judge the claim on its merits anyway.
• Dream, plan, reassess. Be optimistic, build a step-by-step plan, check progress honestly, and change course when needed.
This is a political topic, and people I respect disagree with me. I’ve included their strongest arguments near the end.
Still in the toddler phase: history over hype
Most of today’s AI debate is about the future: either an age of abundance or the end of humanity. Both are predictions. The record of what AI has actually done is something we can evaluate.
That record is genuinely impressive:
• AlphaFold predicted the structure of virtually all 200 million known proteins and earned its creators a share of the 2024 Nobel Prize in Chemistry. A prediction is not an experiment. It still removed a 50-year bottleneck.
• Baricitinib, an arthritis drug flagged by AI as a possible COVID treatment in early 2020, was proven in trials and received full FDA approval in May 2022 for hospitalized adults who need oxygen, a ventilator, or ECMO. The benefit was real and modest. Deaths in one key trial were 4.7 percent with the drug and 7.1 percent without it.
It also includes humbling failures:
• IBM’s Watson for Oncology was marketed as a revolution in cancer care and produced what internal documents described as unsafe and incorrect treatment recommendations. MD Anderson shelved its project after about $62 million. It never treated a patient.
• AI-designed drugs have cleared the first safety stage of trials more often than older drugs, but in phase 2 they still fail at about the same rate, roughly 40 percent, and that count is still a handful of finished trials.
The lesson from that history is neither panic nor blind enthusiasm. AI is a powerful tool that works best when its claims are tested. That is the lens I’ll use for the rest of this article.
Why competition matters
In medicine, I learned to be wary of any treatment protocol that only one supplier could provide. The same instinct applies here. A healthy AI industry has many players: large companies with closed models, and open-source models that anyone can download, inspect, and build on.
Why open models matter. Open-source models let universities, small labs, startups, and independent researchers use and study AI without paying a gatekeeper. They also let outsiders check the work. In a field this young, that independent scrutiny is worth a great deal. David Friedberg made the same point on the October 2 All-In podcast: open-weight models already run on machines in countries all over the world. You cannot uninvent them, and you cannot put them back in one company’s vault.
The lock-in risk. Regulation is not automatically bad. But every rule carries compliance costs, and large companies can absorb costs that would sink a small lab or an open-source project. Rules written with only the biggest players in mind can end up protecting them from competition, whether or not anyone intended it. Economists call this regulatory capture.
The patchwork problem. Instead of one national standard, the US now has a growing pile of state rules:
• By May 2026, state legislators had introduced more than 2,100 AI-related bills, with inconsistent definitions and requirements.
• States enacted 109 AI laws in the first half of 2026 alone, according to NYU’s Center on Technology Policy.
• Some proposals would hit open models directly. One New York bill would require every generative AI developer, including open-source developers, to publish every source of data used to train their models.
A large company can hire lawyers to track 50 sets of rules. A university lab or open-source volunteer cannot.
To be fair to the other side: many states say they are acting because Congress has not, and their laws often target real harms, such as chatbots that impersonate licensed professionals. When the Senate voted on a proposed 10-year moratorium on state AI enforcement, it was defeated 99 to 1. The answer is likely a clear federal standard that sets a floor for real harms, not simply blocking the states.
The race with China
The United States still leads in AI, but the lead is narrow, and it depends on how you measure it.
The models. Different scorekeepers reach different conclusions:
• Epoch AI, an independent research group, found that since 2023 every model at the frontier of AI capability came from the US, and Chinese models trailed by an average of about seven months.
• In April 2026, the federal government’s Center for AI Standards and Innovation evaluated DeepSeek V4 Pro and found it about eight months behind the frontier.
• Stanford’s 2026 AI Index, measuring by head-to-head user ratings, found the gap much smaller: the top US model led the top Chinese model by only a few points.
Notably, nearly all leading Chinese models are open-weight, free for anyone in the world to download. That is a powerful way to win users globally, and it is one reason I don’t want US rules to make open American models impractical.
The energy. This is where China’s advantage is clearest. In 2025, China added roughly 540 gigawatts of new power capacity, about eight times what the US added. BloombergNEF expects China to add more than six times as much generating capacity as the US over the next five years. On All-In the day after the White House summit, the hosts put the same fact in longer form: China has roughly doubled its grid every decade, while US grid capacity has been essentially flat for 25 years. The scarce thing is no longer only the chip. It is the electron.
The honest caveats. Cheap electricity makes AI cheaper to run, not automatically better. China’s grid is also fragmented, which limits moving power between regions. And the US still leads in the most advanced chips. China is a formidable competitor, not an unstoppable one.
Energy is the bottleneck
AI runs on electricity, and a lot of it. Data centers used about 4 percent of all US electricity in 2024, and that share is climbing. If we want to keep leading, we need to build power plants, transmission lines, and data centers faster than we have been.
Right now, that buildout is meeting serious resistance:
• In the first half of 2026, nearly $200 billion worth of data center projects were blocked or delayed.
• Organized opposition groups more than doubled in early 2026, to 833 groups in 49 states, and statewide moratorium proposals were introduced in 14 states.
• Here in New York, Governor Kathy Hochul announced a one-year pause on new large-scale data centers, citing electricity demand, utility costs, and community impacts.
Listening to the neighbors. I want to be fair here, because this is not mostly a story of outside agitators. The opposition is largely local and bipartisan. A Gallup survey in March found that seven in ten Americans oppose building an AI data center in their own area. Their concerns are concrete: higher power bills when utilities build new capacity for large users, water for cooling, noise, and the loss of farmland and open space. As someone who has spent years caring for a piece of land, I understand that.
The answer is to solve those problems, not dismiss them. That means:
• Site them wisely. America has plenty of land away from neighborhoods. Locate large data centers in rural and industrial areas, not next to homes.
• Pay landowners fairly, and share the benefits with host communities through tax revenue and local jobs.
• Make data centers pay their own way. The cost of new power plants and transmission lines should never land on household bills. Done right, the added tax base and grid investment could even lower local rates.
• Build the power first. Require the electrical infrastructure to be in place, with new generation such as natural gas, nuclear, or solar with storage, before a facility comes online.
• Be a good neighbor. Require low-water cooling where water is scarce, and living fences of trees and shrubs around buildings to dampen sound and screen the view.
• Speed up permitting for power plants and transmission lines without skipping real environmental review.
It’s also worth thinking critically about the loudest claims on both sides. Some warnings about data centers are overstated, and some industry promises are too rosy. The fair test is to check each specific claim against the evidence, the same way I’d evaluate a treatment.
Is there foreign influence? Some, yes. I went looking for evidence of a foreign-backed campaign behind the opposition, because I suspected one. Here is what I found.
What’s documented:
• A China-linked network. A report from the Bitcoin Policy Institute traced anti-data center activism to a network of US nonprofits, including CodePink, funded primarily by Neville Roy Singham, an American businessman based in Shanghai. In January 2026, CodePink published an article attacking specific US data center projects and framing the fight as opposition to a new Cold War with China.
• A request to investigate. In September 2026, two House committee chairmen asked the Treasury Department and IRS to investigate whether tax-exempt groups tied to Chinese interests helped fund data center opposition. Their letter alleges the Singham network moved $278 million into US advocacy groups through shell companies and an anonymous fund. That is an allegation under review, not yet a finding.
• A caught influence operation. OpenAI banned a cluster of likely Chinese accounts that had used ChatGPT to produce anti-data center content.
• Foreign billionaire money. According to the American Energy Institute, 12 US groups opposing data centers have received more than $39 million from foreign donors, mainly Swiss, British, and Danish. Swiss billionaire Hansjorg Wyss alone has given nearly $14 million to four groups, including the Sierra Club and Indivisible, that called for a national moratorium.
What the evidence doesn’t show:
• The influence operation had little reach. OpenAI said its posts got little traction, and that it exploited existing local fights rather than creating them. A Clemson University researcher who tracks foreign influence campaigns said his team hasn’t found much.
• Big donations aren’t earmarked. Groups like the Sierra Club work on many issues, so a foreign donation doesn’t mean that money paid for data center opposition specifically.
• The messengers have stakes too. The Bitcoin Policy Institute mainly advocates for the crypto industry, which also needs cheap power, and many prominent promoters of the foreign-influence theory are invested in AI.
• Much of it is ordinary politics. In Texas, for example, a Democratic-backed campaign openly hosts data center town halls with its candidates. Foreign nationals are barred by federal law from donating to US candidates, and I found no evidence of politicians being paid by foreign sources to oppose data centers. The documented channel is foreign money into nonprofits, which is legal but deserves far better disclosure.
My conclusion. Foreign actors, including a China-linked network, have tried to amplify data center opposition, and foreign billionaires fund some of the groups involved. That deserves transparency and investigation. But the evidence so far shows foreign actors exploiting a genuine local movement, not creating one. Most opponents are simply neighbors worried about their power bills, and the best answer to them is a better plan.
Chamath Palihapitiya put a sharper frame on this the day after the summit. He called the data centers themselves a national-security asset, because the country that can turn electricity into computing wins the race. I think he is right about the stakes and wrong if that line is used to bulldoze the neighbors. National security is a reason to build. It is not a reason to send the bill to the household next door.
The White House accord: who was in the room, and who signed
On September 29, 2026, President Trump hosted a lunch at the White House that David Sacks, the White House AI adviser and an All-In host, later called the Bretton Woods of superintelligence. The point of the nickname is the guest list. It was not only the model companies. It was the stack: chips, cloud, software, and the people who have to power it.
Who was in the room. Jeff Bezos of Amazon was there. So were Satya Nadella of Microsoft, Jensen Huang of Nvidia, Lisa Su of AMD, Sundar Pichai of Google, Mark Zuckerberg of Meta, Dario Amodei of Anthropic, Alex Karp of Palantir, and Elon Musk. Sam Altman was not there. OpenAI was represented by its president, Greg Brockman, who sat with the group and signed. Speaker Mike Johnson was there. Sacks was there and helped broker the paper.
Who signed. The one-page White House Accord on Super Intelligence was signed by President Trump and six frontier companies: Anthropic (Amodei), Google (Pichai), Meta (Zuckerberg), Nvidia (Huang), OpenAI (Brockman), and xAI, now part of SpaceX (Musk). Bezos was at the table and did not sign. Nadella was at the table and Microsoft did not sign. Apple was not in the signing group. Being in the room and putting a name on the page are different facts, and earlier drafts of this article blurred them.
How it came together. On the October 2 All-In episode, Sacks said a short White House statement had been circulated in advance, and that during the lunch Zuckerberg and Huang walked through the principles. The signed page had more meat than the draft. Trump later called it morally binding, and compared it to a constitution. It is not a statute. It has no fine written into it.
What they agreed to. Four layers, for the most advanced systems:
1. Internal monitoring for cyber, biological, and chemical risks.
2. An internal review team to make sure those controls actually work.
3. An independent outside auditor to test the controls.
4. A board-level committee that receives the results.
Sacks added a point I had not seen in the news write-ups. A board that ignores an auditor’s report, he said, risks losing its directors-and-officers insurance. The Federal Trade Commission and the Securities and Exchange Commission can also police a public promise a company fails to keep. That is not the same as a new AI law. It is a reason the promise is not quite free.
The same morning. Hours before the lunch, Trump signed an order launching America.gov, a government chatbot meant to be one front door for federal services, from passports to park questions. Agencies were told to connect their public services within 90 days. I mention it because it is the other half of the day: not only a promise about future models, but a working tool for ordinary people. A tool that answers from government pages will sometimes contradict a politician. That is a feature if the pages are the source, and a problem if nobody checks them.
Why I like the accord. It moves beyond pure self-policing. Outside auditors and a board committee are more than a press release. It also put direct competitors at one table, including people who are suing each other, and got six of them onto one page. That is the kind of practical safety net that fits a technology in its toddler phase.
The fair criticism. The accord is voluntary. The companies helped draft it, and they hire the auditors. Critics noted that only weeks earlier several of the same executives had said the industry should be regulated by government. CNBC’s headline the next day was that AI safety looked more unsettled, not less.
In keeping with my own approach, I see the accord as a good first step in the plan, not the finish line. The real test is the reassessment: whether the audits happen, whether the results are made public, and whether companies change course when an auditor finds a problem.
What the builders said the next day
The October 2 All-In episode is the best window I have into the ambition in that room, because one of the hosts was in it. A few points are worth keeping, with the usual caution that these men invest in and advise the industry they are describing.
• The whole stack, not just the chatbots. Sacks’s case is that this was the first time the chip makers, the cloud companies, and the model labs were in one conversation about both the opportunity and the risk. Bezos being there, without signing, fits that picture. Amazon builds the buildings and sells the computing. It is not, today, one of the six frontier model labs on the page.
• Electrons, not slogans. The repeated line was that the generation of electrons is now the constraint. Chips matter. Power matters more if the chips have nowhere to plug in. That matches the China numbers above.
• Open models are not a switch you can flip off. Friedberg said there is no turning off open-weight models, no shutting down every data center, and no single command center. His prediction was a rise in defense systems, not a pause. I share the first half. The second half is a forecast.
• Whoever wins, wins. Chamath relayed Trump’s line from the opening: whoever wins superintelligence wins. I will use the researchers’ meaning here, not the rebrand. The tools we have are not a machine that beats humans at everything. The race is still real.
The Wall Street Journal, cited on that episode, called the AI buildout the biggest economic bet in US history. The same week, second-quarter growth was revised up, and August payrolls rose by 162,000. I am not going to hang a boom on one month. I am going to say the bet is already large enough that getting the power and the rules wrong will show up in ordinary life, not only in a lab.
Taking misuse seriously
Supporting AI does not mean ignoring its dangers. A tool powerful enough to help discover a new antibiotic could, in the wrong hands, help someone do harm. Preparing for that is part of supporting AI, not a reason to oppose it.
The risks that worry me most are concrete, not science fiction:
• Biological and chemical threats. The fear is that AI could give someone without expert training a step-by-step path to a dangerous pathogen or toxin. The White House accord specifically requires monitoring for these risks.
• Cyberattacks. AI systems that can write code can also find and exploit weaknesses in software, hospitals, and power grids.
• AI agents acting on their own. Newer AI systems can carry out tasks without a person approving each step. This summer, OpenAI was still investigating how its own AI agents ended up probing the systems of Hugging Face ahead of a July security breach there. Incidents like that are exactly why outside audits matter.
• Fraud and deception. Fake voices, fake videos, and convincing scams are already here.
The right response is targeted: focus rules and testing on the specific capabilities that could cause catastrophic harm, rather than burying every developer, large or small, under blanket restrictions. A scalpel, not a sledgehammer.
The other side, honestly
I promised to give the strongest counterarguments, and here they are. Thoughtful people hold these views, and some of them may turn out to be right.
• Serious researchers are worried. This isn’t only activists. Geoffrey Hinton, often called the godfather of AI, briefed senators on the dangers this September. Anthropic’s own CEO, Dario Amodei, has called on the industry to slow down. He then sat at the White House table and signed. Both facts belong in the record. When the people building a technology urge caution, it deserves a hearing.
• “Racing” can mean cutting corners. The central worry is that companies under pressure to win may skip steps in safety oversight. The China competition argument, however valid, can also become an excuse to avoid any rules at all.
• Voluntary promises have limits. Without penalties, a commitment holds only as long as it’s convenient. Efforts in Congress to pass binding safeguards stalled this fall.
• Ordinary people bear real costs. Higher electric bills, water use, and lost farmland fall on local communities, while the profits flow elsewhere.
• Jobs and power. AI could disrupt many jobs, and a small number of companies could end up holding enormous economic and political power.
• An ounce of prevention. As we say in medicine, prevention is better than cure. Some argue that sensible rules now are better than rushed, heavy-handed rules after a disaster.
I don’t think these arguments mean we should stop or pause. But each one names a real problem that a good plan has to address.
Dream, plan, reassess: what good policy looks like
I approach big goals the same way every time: start with a bold dream, build a step-by-step plan, reassess honestly along the way, and change course when the evidence says to. Here’s how I’d apply that to AI.
The dream. America leads the world in AI that cures diseases, grows more food on less land, and makes life better for ordinary people, built here, powered by abundant American energy, and trusted by the public.
The plan.
1. One clear national standard that sets a floor for real harms, so innovators aren’t navigating 50 different rulebooks.
2. Keep the door open for open-source models, so universities, small labs, and startups can compete with the giants.
3. Build energy at scale, with data centers paying their own way and host communities sharing the benefits.
4. Targeted safety rules for the genuinely dangerous capabilities, such as biological, chemical, and cyber threats, with independent testing.
5. Strengthen the White House accord over time: publish audit results, and make the commitments enforceable if voluntary measures fall short. Invite the companies that were in the room but did not sign, Amazon and Microsoft among them, onto the same page.
The reassessment. Measure honestly and publicly. Are the audits happening? Are electric bills rising near data centers? Is the gap with China growing or shrinking? Are open models thriving or being squeezed out?
Changing course. If voluntary measures fail, be willing to make them binding. If a rule is choking competition, fix it. If the risks prove larger than expected, slow down. Admitting a plan needs adjusting is not weakness. It’s how good doctors, good farmers, and good countries get things right.
Raising a toddler
Anyone who has raised a toddler knows the balance. You don’t lock them in a room because they might fall, and you don’t leave them unattended near the stairs. You let them explore, you childproof the dangerous spots, and you keep watching.
That’s how I see AI in America right now. We should keep building, keep competing, and keep our lead, because the benefits are real and our competitors won’t wait. We should also take misuse seriously, listen to the neighbors who host our data centers, and demand that safety promises be tested and verified.
Most of all, we should hold every side of this debate, including my own, to the same standard: truth and honesty, history over hype, and evidence over fear. I’ll keep learning, and I’ll keep reassessing. I hope you will too.
Notes on sources for this revision
• All-In podcast, October 2, 2026, “Trump’s Super Intelligence Summit, AI Safety Accord, GDP Beats, Midterm Predictions.” https://www.youtube.com/watch?v=ZJKs08oU1zg
• AP, September 29, 2026: accord signed by Trump, Amodei, Pichai, Zuckerberg, Brockman, Huang, and Musk. https://apnews.com/article/trump-ai-anthropic-musk-595796511f110fc006cca0d01329733e
• CNN, September 29, 2026: Bezos, Nadella, Karp, Su, and Brockman among those present. https://www.cnn.com/2026/09/29/business/amodei-huang-karp-trump
• Forbes, September 29, 2026: Bezos attended; Brockman signed; Altman did not. https://www.forbes.com/sites/saradorn/2026/09/29/trump-says-billionaire-ai-execs-want-tremendous-self-regulation-in-white-house-meeting/
• Politico, September 29, 2026: America.gov executive order. https://www.politico.com/news/2026/09/29/trump-signs-executive-order-to-launch-ai-powered-america-gov-01096992
