The Geopolitical AI Race

Iman Ghanizada is a returning guest, the founder and CEO of Eclipse, and the person who pioneered Autonomic Security Operations at Google Cloud. He was born and raised in DC, worked at a lobbying firm, and spent years around the Hill before he spent years in Silicon Valley — which makes him one of the few people who can talk about both sides of the AI policy argument without guessing at how the other side thinks. Since his last appearance he has been heads-down in his lab, and he's been posting long-form on X about where model training happens, what biases it bakes in, and what the US actually sells to the rest of the world. This episode steps out of the CISO bubble. It's about the geopolitical AI race, not the SOC.

The core argument: America moved from agriculture to services in the 1970s, and those services ran on intelligence that was hard to get access to. Frontier models put that intelligence on the internet. Competitors distilled it. Now the economics are sliding down the stack — from models to compute, energy, and chips — and that's a stack China is nationalized to win. Iman puts numbers on it: 25 nuclear reactors in development, Huawei Cloud already deployed across low-income countries south of the equator, inference pitched at a fraction of US pricing, a Silicon Valley bubble "filled with token costs and semiconductors" that still needs sector-wide ROI to justify it. Stuart Mitchell co-hosts and pushes on the politics: one country treats this as a sovereign project, the other is protesting its own infrastructure. This one is for security leaders, founders, and investors who want to understand the forces that will decide which models they'll be running in five years — and for the policymakers Conor insists are listening.

RISC Point is security & compliance consulting tailored to your business. Talk to RISC Point.

Hampton North is the premier US based cybersecurity search firm. Start building your security team with Hampton North.

Sysdig is the leader in AI-powered real-time cloud defense; stop watching and start defending.

Conor opens with the frame he wanted to test: America sells intelligence. First as services, exported worldwide. Then, in the same motion, as frontier models. Now intelligence is being distributed from places that are not the United States, measured in weeks, and the argument in DC and Silicon Valley has split into two camps — go all-in on frontier models, or accept open-weight and lean in. Iman's answer is that both camps are missing the point.

His version: the original idea was that the US would invent a technology that created a wealth surplus and control it. That lasted until the models got cheaper to train and China proved it could speed up development and cut costs. But cheaper models aren't the problem. The problem is what the US economy is built on.

That intelligence used to be hard to get. It was gated behind infrastructure, compute, contracts, defense agreements, and corporations. Putting it into a frontier model and shipping it to the world took the gate down. His line is that all of that "has largely been commoditized," and the fair response — build more data centers, capture demand, Jevons' paradox — runs straight into the fact that China is doing the same thing with its own chips, its own clouds, and a nationalized energy program. He calls it something that's already happened. The question now is what comes next.

Iman doesn't hedge on open-weight itself. He calls it great, compares it to a CD anyone can plug into a computer, and says competitive markets are how the technology develops further. He also calls the whole debate irrelevant to the thing that should worry people.

What matters is where the economics land. If the value slides down the stack, how long are US contracts around the world protected? Not just professional services — Deloitte in Moldova is his example — but every multinational that does business inside another country's borders. If inference gets cheap enough, a country can run its own resources on a little inference and stop paying for American knowledge. His bet is that LLMs become utilities, and China is very good at manufacturing utilities at scale.

Conor pushes on the stack: chips and bricks at the bottom, models in the middle, applications on top. The model layer looks most at risk — enormous value captured, and then distributed almost immediately. Iman agrees inference costs coming down is inevitable, and connects it to the same playbook China used in manufacturing: learn to make the thing, make it cheaper, and use price to diminish US control. He reads the OpenAI reaction to Dean Ball's "AI Communism" comment — and Jensen's counter-announcement defending open weights — through that lens. The labs aren't afraid of competition inside America. They're afraid of what happens when the competition is between nations.

That's the tension he wants held at once: the labs vying for regulatory capture because they need infrastructure and political backing against Chinese development, and the country still needing a free and open market at home. He also flags the detail that makes the "nations" framing complicated — OpenAI employs more Chinese engineers than the leading Chinese labs. This is a battle of competition, not of people.

Conor brings the electricity number: by his memory, China adds roughly four times the electrons the US does every year, compounding. Iman stacks the rest of the supply chain on top. Twenty-five nuclear reactors in development. Domestic versions of EUV lithography machines that aren't superior yet but are on a clear ramp. And a reverse-engineering culture that AI is about to make far more effective — not just distilling model outputs, but distilling patents and intellectual property at national scale.

Conor asks whether distillation is the mechanism behind the national security claim. Iman says no — the race itself is good for the world. The failure is one of alignment. The US walked into a war with political parties that weren't aligned for it and a population protesting data centers. Capital formation is strong; venture is funding everything it should. The question is whether value gets captured fast enough.

He notes he's starting to see enterprise earnings reports show AI investment paying off, but the timing matters. If prices crash — from open models, but also from infrastructure and energy getting cheap — the rest of the world buys whatever it wants, and the US is left with export controls as its main tool to force purchases of American services. For a security leader this is the backdrop to every AI procurement conversation over the next several years: the vendor landscape you're evaluating sits on top of an infrastructure race the US hasn't decided to fight as one country.

Conor raises the thing security people don't usually question: safety training. Chinese open-weight models read as raw horsepower with baked-in bias. US frontier models arrive wrapped in guardrails that refuse meaningful work. Is there a version where safety becomes a structural disadvantage?

Iman's answer is the straw analogy — one country banning straws for the planet while everyone else keeps using them isn't a competitive strategy. Then he gets practical about what refusals mean for anyone shipping product.

Open weights fix that in a way closed models can't: you can fine-tune, inspect the model, and make it work inside your organization. He's careful to separate the model from the service provider — you can host a Chinese model anywhere, but if you run it on Kimi's cloud, you get their safety layer on top. And the bias question isn't about closed source; Chinese models carry Chinese propaganda because they're trained on a ton of Chinese text.

On whether AI safety is the concern the posts make it out to be, he's blunt: "We haven't seen the proof, but we've seen the posts." If the US slows adoption while its fiercest competitor goes full steam, the rest of the world isn't going to wait for a verdict. "We're just going to use what's best."

Conor floats intelligence sovereignty as a third path: not the US, not China, but Luxembourg or Moldova taking the best model available and wrapping their own standard of safety around it. Iman agrees that's the pitch China's president has been making — every country in control of its own sovereignty — and then asks what the pitch is for.

Two possibilities. One is that China genuinely sells sovereign infrastructure countries own and host; Huawei and Alibaba Cloud already do this and it's a growing segment. The other is a longer play.

Cheap open models are the on-ramp. Chinese compute is the next step. Then the models get good enough to close.

Conor points out this is standard playbook for anyone who isn't leading a race — open-source until competitive, then close, as Meta did with Llama. Iman's counter is that the US has its own lock-in: CUDA, deployed infrastructure everywhere, economic frameworks and compliance regimes that force the rules of the West. Unwinding that takes time. Which brings him back to where he started: the bigger risk isn't China. It's whether the US is ready to win — and winning means infrastructure, data centers, and distributing American intelligence wider, not just continuing to pay OpenAI and Anthropic for tokens.

Conor challenges the word "war." Iman loosens it — LeBron and Jordan are fierce competitors who don't hate each other — but points at the export-control conversations, the talk of banning Mercedes-Benz over 20% Chinese ownership, and a leaked Department of War request for creative ways to pressure Iran. The tone in DC is not peacetime.

Stuart names the asymmetry directly: China is selling this as a sovereign project the whole country benefits from, and the US narrative is data center protests, job loss, and kill-switch legislation. Iman's diagnosis is that three games are running with three different incentive models. Institutions creating value. The federal government as older sibling, applying controls so those institutions can compete globally — "the president, who is our CEO." And local and state politics that have to connect the federal layer down to the institutions.

His prescription, when Conor asks for one, is two-part. More engagement between DC and Silicon Valley — because the Valley is largely kids out of Stanford working on consequential technology, and DC needs a grip on how policy accelerates or decelerates the mechanics of industries. And a messaging fix so the country stops carrying "a big giant weight on our back trying to climb up Mount Everest." Winning requires alignment at the rare earths layer, the chips layer, the energy layer, and every level of government. Capital is already moving — Prometheus raised $12 billion, Fei-Fei Li raised $1 billion — but capital doesn't fix a fragmented base.

The reason the message matters, in his telling, is that the technology can actually solve the problems both sides of the political debate are worried about.

Asked what he's been building at Eclipse, Iman calls it "a bit of science fiction" and then talks about the mission instead of the product. The question he's working on is what the world's next intelligence advantage looks like — and why it matters that the answer comes from a heterogeneous culture where people can see themselves in it. Democracy, in his framing, isn't just America's narrative; it's a narrative for human life and dignity that can travel. The risk for heterogeneous cultures, Rome included, is losing connectivity to shared values until the fighting turns inward.

On the market, he is direct about how early it is. If the addressable market for distributing intelligence had been scratched, you wouldn't walk outside in San Francisco and see homeless people everywhere. Enterprises still aren't seeing ROI. There are security problems and reliability problems. All of it blocks adoption.

The close is the optimistic version of the whole conversation. Thirty years of sitting behind screens, and a technology — paired with robotics, embodiment, and manufacturing — that could let people put the screens away. He ends on John Adams: study politics and war so the next generation can study mathematics and philosophy, so the one after that can study arts and music. Conor lands there on purpose. The hard work is what's in front of us. The other side of it is worth the argument.

Guests — Iman Ghanizada, founder and CEO of Eclipse; pioneered Autonomic Security Operations at Google Cloud; born and raised in DC, worked at a lobbying firm and around the Hill; returning Zero Signal guest.

Books mentioned — None named in the conversation.

Frameworks / models / tools named — Autonomic Security Operations; DeepSeek; Kimi (K3); Llama; CUDA; Google Cloud Platform (GCP); Huawei Cloud; Alibaba Cloud; EUV lithography; Jevons' paradox; model distillation; open-weight vs. frontier model stack (chips and bricks, model layer, application layer).

Other people / shows / resources referenced — Stuart Mitchell (co-host); Dean Ball ("AI Communism" comment); Jensen Huang; Sam Altman; Dario Amodei; Dr. Fei-Fei Li ($1B raise); Prometheus ($12B raise); OpenAI; Anthropic; Meta; Intel; Deloitte; Mercedes-Benz (20% Chinese ownership); Department of War (leaked Iran memo); Xi Jinping (sovereignty narrative); LeBron James and Michael Jordan; John Adams; Iman's long-form posts on X; Iman's prior Zero Signal appearance.