Context
Eric Schmidt, former CEO of Google, adviser to the Pentagon, and current CEO of Relativity Space, lays out his latest view of the AI race between the US and China. His argument is not just about who has the best model. It is about who can turn model progress into deployment through energy, data centers, talent, safety controls, and global distribution. That matters to product managers because many AI product bets will be shaped less by pure model quality and more by cost, infrastructure, governance, and how quickly teams learn to use the technology well.
The Big Idea
The real competition is deployment economics, not just model quality. The US still leads at the frontier, but China's gains in efficiency, energy buildout, and open distribution have cut the lead to about six months.
Schmidt's deeper point is that AI advantage now comes from a stack of choices: model architecture, compute access, openness, talent policy, and real-world adoption. For PMs, that is a reminder that the technically best product does not automatically become the winning product.
Key Insights
Software cost collapse reshapes competition
Schmidt's starting point is that the cost of software has fallen from very expensive to almost nothing. When the cost of creating software drops that sharply, the amount of software created rises just as fast. He expects the next few years to be defined by huge amounts of software, especially agents working together.
For PMs, that changes the scarce resource. The hard question is no longer only "can we build this?" The harder question becomes which workflows deserve automation, which ones need human review, and how multiple AI systems should work together without creating chaos.
China caught up differently
Schmidt says he helped design chip restrictions on China and expected those restrictions to matter for about 10 years. Instead, he now sees the gap at roughly six months.
His explanation is useful:
- China did not erase the gap by matching US chips.
- China adapted by finding algorithmic methods that use about one-tenth of the energy with some performance trade-offs.
- Once Chinese teams publish a useful open technique, US teams copy it quickly.
The PM lesson is that constraints often force a different innovation path. If a competitor loses access to one key input, do not assume the competitor is blocked. The competitor may redesign the system around a different advantage.
Power is now strategy
Schmidt thinks the US still has strong frontier labs, deep capital markets, and top talent. The weak point is infrastructure, especially energy and data centers.
He gives unusually concrete numbers:
- The US may need 60 to 80 gigawatts of power through 2030 or 2031.
- One gigawatt is roughly one nuclear power plant.
- He puts one gigawatt at about $50 billion in cost.
For PMs, this is a strategic lesson about bottlenecks. A roadmap can look strong at the feature layer while being constrained by an upstream dependency the team does not control. In national AI competition that dependency is power. In a product organization it may be data access, procurement, security review, or model inference cost.
Open weights change geopolitics
Schmidt distinguishes classic open source from open weights. In his framing, China has leaned into open source and open weights, while leading US frontier labs have mostly stayed closed source and closed weights.
Why the split matters:
- A frontier model can cost $100 million to $500 million to train, so US labs have a clear reason to protect the asset.
- China chose openness partly because it was behind and needed users fast.
- Schmidt thinks the majority of the world is likely to use Chinese models over time because many markets will choose free or cheap models that are good enough.
That is a useful product lesson. Distribution, pricing, and ecosystem strategy can matter more for adoption than raw technical quality. A premium product can lead the benchmark tables while a cheaper and more accessible product wins usage at global scale.
Safety needs coordination, not pause
Schmidt is skeptical that a meaningful slowdown is realistic. He describes a world with roughly five to seven frontier labs in the US and a similar number in China. Even if the US could organize voluntary restraint, China already believes it is behind and is unlikely to stop.
He is more concerned with what happens after release:
- Open-weight models can ship with guardrails.
- Those guardrails can be removed through a technique he calls obliteration.
- Once the raw model is widely available, cyber misuse becomes much harder to control.
For PMs, the practical takeaway is that governance cannot rely on perfect central control. If a capability will spread, the safer approach is to design operating rules, response plans, and monitoring before the system is broadly adopted.
AI adoption starts with use
Schmidt ends with a very direct leadership rule: "Start by using it." He says his own view changed when he began using long chain-of-thought reasoning and saw the depth of the system's work.
He points business leaders toward structured functions first:
- finance reports
- manufacturing inventory
- security reviews
- cash forecasts
The PM lesson is that firsthand use changes strategy quality. Leaders who have not used modern reasoning systems tend to talk about AI at the level of headlines. Leaders who use the systems on real work start to see workflow redesign opportunities that are invisible from a distance.
Mental Models & Frameworks
The AI race scorecard
Schmidt's scorecard is a helpful way to avoid vague statements like "the US is ahead" or "China is winning." He breaks the race into layers, each with a different leader.
- US advantage: model performance, talent depth, and safety coordination.
- China advantage: model cost, energy and compute buildout, deployment speed, and likely long-run global reach.
- Mixed area: military and defense AI, where Schmidt sees less visibility and only a slight US edge at most.
Use this model when evaluating any AI market. Ask which layer actually matters for your product: research quality, deployment cost, ecosystem spread, regulatory control, or speed of execution.
Frontier versus utility models
Schmidt implicitly separates AI use cases into two buckets:
- Frontier workloads: math, chemistry, and other high-end reasoning tasks where the best models matter.
- Utility workloads: accounting, reporting, and other structured business tasks where a cheaper model may be good enough.
He notes that self-hosted open-weight usage can be about three to five times cheaper on token cost, even if it is less capable. Use this split when choosing a model strategy. Not every workflow deserves frontier-model pricing.
No surprise rule
Schmidt proposes a "no surprise rule" as one example of how rival nations could reduce risk without ending competition. He compares it to missile-launch notification systems, where countries alert others so satellite systems can verify the event and avoid misreading it as an attack.
For PMs, the reusable idea is simple: when a capability is powerful enough to trigger panic or misuse, create advance notice and verification mechanisms. In product terms, that can mean pre-agreed incident reporting, release protocols, or escalation rules for high-risk AI features.
Open platform logic
Schmidt frames the open versus closed debate as an old strategic choice:
- Open platform logic: give away part of the stack so others build on top of it and spread your standard.
- Closed product logic: protect the asset, focus on the end user, and capture more direct value.
He does not present either approach as universally right. He treats them as different ways to compete. Use this model when deciding whether your company should expose capabilities, partner broadly, or keep an AI advantage proprietary.
Decision Principles
Principle: Watch bottlenecks, not headlines
- When: a team is debating AI strategy based mainly on benchmark rankings or news momentum.
- Why: Schmidt's scorecard shows that the real constraint may be energy, cost, distribution, or talent rather than model quality. A product plan that ignores the actual bottleneck will overestimate what the team can ship.
Principle: Separate innovation from diffusion
- When: a team is choosing between the most capable model and a cheaper, more accessible model.
- Why: Schmidt expects frontier models to drive the biggest scientific leaps, but cheaper models to spread faster across the world. The model that advances the state of the art is not always the model that wins the market.
Principle: Learn by using AI
- When: leaders are setting an AI roadmap, budget, or operating model.
- Why: Schmidt argues that direct use changes judgment. Teams spot better workflow opportunities once they have seen reasoning systems handle real tasks, not just demos.
Trade-offs & Nuance
Openness weakens control
Open weights can speed up adoption, improve ecosystem learning, and reduce costs. Open weights also make it easier for others to remove guardrails and repurpose the raw model for harmful use.
Schmidt does not offer an easy answer here. His point is that openness is not only a product-distribution decision. It is also a security and governance decision.
Cheap models widen influence
Schmidt thinks the long-run global user base may tilt toward Chinese models because many countries will choose products that are free, cheap, and good enough. That is a distribution advantage.
The trade-off is that the most advanced breakthroughs may still come from more expensive frontier systems. The same company or country does not need to win both layers of the market.
Security can repel talent
Schmidt is blunt that US immigration and visa policy can become self-defeating. He argues that training top researchers, then forcing them to leave, is one of the country's worst own goals.
His example is memorable: he says the person leading China's quantum program wanted to stay in the US after earning a PhD here and was forced out. The broader lesson is that defensive policy can protect short-term security while damaging long-term talent advantage.
Decentralization slows execution
Schmidt clearly prefers the US system overall, but he also admits the US pays a speed penalty for pluralism, debate, and fragmented infrastructure decisions. China can align policy, energy buildout, and deployment faster.
For PMs, the lesson is not that centralized systems are better. It is that decentralized systems need stronger coordination to move quickly on shared priorities.
Practical Application
Audit your AI dependencies
Before approving the next major AI initiative, write down which layer actually determines success:
- model quality
- token cost
- proprietary data access
- security approval
- compute availability
- integration complexity
This prevents a team from treating every AI problem like a model-selection problem.
Map internal AI bottlenecks
Schmidt's national-level bottleneck is energy. Your organization's equivalent bottleneck may be much smaller but just as decisive.
Create a one-page map for one important AI initiative:
- what slows experimentation today
- which dependency the product team does not control
- what would remove that dependency fastest
This makes it easier to invest in the real constraint instead of polishing the visible surface.
Pilot AI in finance
Start where the workflow is structured and the stakes are understandable. Schmidt specifically calls out finance as relatively easy for AI systems.
A practical first step:
- have AI draft a weekly financial or operational report
- ask the finance lead or CFO to review and correct it
- track where the AI saves time and where it introduces risk
The point is not only efficiency. It is building organizational trust through a workflow where errors are visible and review is natural.
Track open model exposure
If teams are already using open-weight or low-cost external models, do not treat that as an informal tooling choice. Inventory it.
For each meaningful use case, capture:
- which model the team uses
- where the model is hosted
- what data is sent to it
- what guardrails or filters exist
- what would happen if the model provider became a legal or geopolitical risk
That turns an invisible dependency into an explicit product decision.
Build firsthand leadership fluency
Set up a short weekly session where product leaders use a reasoning model on live work, not generic prompts.
Good starting tasks:
- drafting a product review memo
- comparing roadmap scenarios
- generating risk questions for a launch
- summarizing customer feedback and highlighting contradictions
Schmidt's point is that real use changes strategy. A leadership team that only discusses AI conceptually will move slower than one that works with it directly.
Questions to Consider
- Where is your team assuming that better model performance will win, even though lower cost, easier distribution, or faster deployment may matter more for adoption?
- Which dependency in your product organization plays the same role that electricity and data centers play in the national AI race: essential, expensive, and easy to ignore until it becomes the bottleneck?
- For which workflows does your company truly need frontier-model performance, and for which workflows would a cheaper good-enough model create more value faster?
- If your product or internal workflow depends on open-weight models, what is your plan if the model's guardrails are removed, the hosting environment changes, or legal exposure suddenly rises?
- How many leaders making AI strategy decisions in your company have actually used modern reasoning models on real work in the last month?
Bottom Line
The US still has a frontier lead, but Schmidt's warning is that leadership in AI will be decided by far more than model benchmarks. Product leaders should pay close attention to infrastructure, openness, talent policy, safety design, and direct usage habits, because those are often the real variables that determine who captures value.
Notable Quotes
"The China-U.S. competition is the defining competition of your generation, my generation, of our age." (Eric Schmidt)
"AI is the single most important thing that's going to happen for the rest of our lives." (Eric Schmidt)
"Start by using it." (Eric Schmidt)
