Lowy Institute AI Snapshot Puts the US Ahead of China, With Limits

The Lowy Institute AI snapshot places the United States ahead of China, despite China’s advantages in minerals and open-model distribution. Published November 8, 2026, the assessment compares 14 indicators across resources, technology, AI capabilities, and adoption. Its finding challenges a familiar shortcut: popularity among developers does not automatically establish leadership across the entire AI market.

Researchers Jack Sato and Hunter Marston distinguish American strengths in computing and consumer distribution from Chinese strengths elsewhere. Their comparison gives the US an overall lead while showing why individual indicators can support very different headlines. Neither country controls every resource, distribution channel, or source of influence.

For developers and enterprise buyers, the significance extends beyond national rankings. Choosing an AI system means choosing dependencies involving infrastructure, software access, operating requirements, and governance. The report provides a starting point for examining those dependencies. It does not establish which model belongs in every organization’s workflow.

What the Lowy Institute AI snapshot actually changes

The new assessment changes the frame of comparison more than the underlying competitive landscape.

Lowy’s AI power comparison assembles indicators that usually appear in separate debates. It places materials, infrastructure, model capability, and distribution within the same framework. Its measures receive equal weight at each level, and indicator scores compare the countries against their combined totals.

That approach makes an argument about how to understand technological power. A country’s position cannot rest entirely on a successful chatbot launch. Likewise, industrial resources alone cannot establish that its software reaches more users or performs better.

The distinction matters because national competition often becomes a sequence of isolated announcements. A model release supplies a performance headline. A chip restriction supplies a security headline. A download milestone supplies an adoption headline. Readers then encounter each development as evidence of a single overall winner.

Those developments answer different questions. Combining them requires an explicit judgment about their relative importance. Lowy makes its framework visible, allowing readers to examine that judgment instead of treating it as an unstated assumption.

An overall score should therefore function as a structured interpretation. It should not become a substitute for examining the underlying categories. An organization selecting software needs a different weighting from a government assessing industrial vulnerability.

Consider a team evaluating an assistant for confidential engineering documents. Consumer popularity would offer limited evidence about whether that system meets its requirements. Deployment control, document handling, answer quality, and maintenance obligations would carry more immediate weight.

A business introducing an assistant to customers would face another calculation. Ease of access, response reliability, integration support, and continuity could outweigh the ability to operate the model independently. Neither decision contradicts a national assessment. Each applies a narrower objective.

The publication also creates a useful baseline for subsequent reporting. Future claims of leadership can be checked against the dimension they actually affect. A distribution gain should update the distribution discussion. It should not silently become evidence of equivalent progress in manufacturing.

This is the central contribution of the Lowy Institute AI snapshot: it makes competing definitions of leadership harder to ignore. The assessment favors the United States overall, while leaving room for Chinese advantages that matter to particular users and industries.

Readers should preserve both parts of that finding. Removing the overall conclusion understates the report’s argument. Removing the unevenness turns an analytical framework into a victory announcement that the evidence cannot sustain.

US China AI competition runs through different distribution channels

Winning developer attention and reaching everyday users are related achievements, but they measure different kinds of influence.

Lowy reports that Chinese models accounted for 78% of combined US and Chinese model downloads in its comparison. American models attracted more than 90% of the combined website traffic and more than 80% of measured app use. Those figures describe distinct channels, rather than interchangeable estimates of total market share.

An open-weight model makes its learned parameters available for download, subject to its license. That availability lets another organization operate or adapt it. A hosted assistant instead provides access through infrastructure managed by a service provider.

This difference changes the relationship between the model developer and the eventual user. Someone using a hosted assistant usually encounters a recognizable product. Someone using an application built around downloaded model weights might never see the original developer’s name.

China AI adoption can therefore expand without matching growth in visits to Chinese chatbot websites. The reverse also holds. Heavy traffic to an American assistant does not establish that American downloadable models dominate independent deployments.

However, downloads need careful interpretation. Hugging Face’s download counting rules explain that its statistics count requests to designated files. They do not directly count unique people, active organizations, or completed production deployments. Counting behavior also varies with file formats and libraries.

That limitation does not make download statistics meaningless. They offer evidence of activity around model distribution. The mistake comes when a distribution event becomes a claim about sustained operational use without additional evidence.

A downloaded model might support evaluation, experimentation, adaptation, or a deployed service. Those activities have different commercial and strategic implications. The download count alone cannot reliably distinguish among them.

Website traffic also has limits. A visit does not establish that someone completed useful work, returned regularly, or selected a paid service. App activity requires similar care when comparisons use different measurement systems or access conditions.

The useful question is what each channel enables. Hosted distribution gives providers a direct relationship with users and control over product updates. Downloadable models give downstream operators more responsibility for deployment and greater scope to shape the surrounding application.

Neither arrangement guarantees better outcomes. Operating a model independently introduces maintenance and security responsibilities. Depending on a hosted service introduces exposure to changes in availability, product behavior, and access conditions.

For developers, the implication is practical. A national adoption headline should not settle a technical choice. The relevant comparison concerns whether a particular distribution arrangement supports the application’s requirements over time.

For policymakers, the implication differs. Influence can accumulate through branded consumer products or through underlying components embedded in other systems. A useful assessment needs to examine both pathways without pretending their activity measures describe the same population.

America’s computing advantage still depends on physical inputs

A lead in AI software does not eliminate dependence on the industrial systems that make the software possible.

Compute means the processing resources used to train and run AI systems. Lowy estimates that American AI data centers hold nearly 40 times China’s combined computing capacity. It also finds that China leads production in 30 minerals on the American critical-minerals list, compared with eight for the United States.

The critical minerals list covers 60 materials considered vital to the US economy and national security. Its purpose concerns supply vulnerability across the economy. It is not a ranking of materials by their individual importance to AI.

That distinction prevents another misleading shortcut. Leading production across more listed minerals does not automatically translate into an equivalent advantage in finished AI systems. Materials must pass through manufacturing, equipment, infrastructure, and operational processes before they support usable computing.

Yet the opposite shortcut is equally weak. Superior models do not make upstream dependencies disappear. A supplier can hold an important position without producing the most visible end product.

This creates a persistent tension within US China AI competition. Different participants can influence different stages of the same industrial system. Leadership at the application layer offers no automatic protection against constraints elsewhere.

Electricity adds another dependency. The International Energy Agency’s AI energy outlook projects data-center electricity consumption rising from 485 terawatt-hours in 2025 to 950 in 2030. It identifies bottlenecks in energy equipment and other infrastructure as constraints on near-term expansion.

A terawatt-hour measures electricity consumed over time. It does not measure the quality of an AI model or the usefulness of its answers. The energy outlook nevertheless explains why installed computing equipment cannot be assessed separately from the systems supplying it.

Announced infrastructure and operating infrastructure deserve separate treatment. A proposed facility still needs the physical and commercial conditions required to provide reliable service. Treating every announcement as immediately available capacity would obscure those conditions.

For an enterprise customer, this industrial discussion matters through service commitments. The operational question concerns dependable access when workloads increase or providers change their offerings. A national capacity estimate cannot answer that question for an individual contract.

For governments, the question concerns resilience across linked dependencies. Expanding one input might expose another constraint. More equipment creates limited benefit if the surrounding infrastructure cannot support it. A diversified supplier base also requires practical alternatives, rather than names on a policy document.

The Lowy Institute AI snapshot is most useful here as a reminder to follow the chain of dependence. Its categories identify different sources of strength. Readers still need evidence about how those strengths connect in actual production and deployment.

The competitive picture therefore resists a clean separation between software and industry. Each depends on the other, while improvements in one area can leave vulnerabilities elsewhere unresolved.

Capability leadership does not guarantee deployment leadership

The strongest general-purpose model does not automatically become the most suitable system for every operational task.

Anthropic’s AI leadership scenarios separate model intelligence, domestic adoption, global distribution, and societal resilience. The company argues that capability matters centrally, while acknowledging that effective deployment can offset a disadvantage in intelligence.

That perspective supports a broader reading of competition, but it requires attribution. Anthropic develops AI models and participates in the market it analyzes. Its strategic assessment offers an informed viewpoint, rather than independent confirmation of every competitive claim.

The underlying distinction remains useful. Model capability concerns what a system can do under specified conditions. Deployment concerns whether an organization can make that capability useful within its own processes.

A benchmark, a standardized evaluation used to compare systems, can illuminate a defined set of abilities. It cannot automatically establish performance across every workplace. Tasks, documents, operating constraints, and error consequences differ.

Consider a hypothetical engineering team comparing assistants against its internal documentation. One system might produce more fluent explanations but cite the wrong version of a technical specification. Another might offer less polished prose while locating the correct document consistently.

That example does not establish that a particular national model family performs better. It shows why the buyer’s evaluation must connect capability to the actual task. A general ranking leaves that connection unfinished.

Deployment control creates another consideration. A team handling sensitive material might favor an arrangement that keeps processing within an approved environment. Another team might prioritize managed operations because it lacks staff to maintain its own model infrastructure.

These choices involve costs in effort, responsibility, and flexibility, even when no subscription comparison enters the discussion. Independent operation requires someone to manage updates and investigate failures. Managed access requires someone to assess the provider’s policies and service commitments.

China AI adoption should be examined through this operational lens. Interest in downloadable models can reflect demand for flexibility and control. It does not establish that every downloaded system reaches production or replaces an existing service.

American consumer reach also needs an operational lens. Familiarity can help introduce a product into an organization. Familiarity alone cannot establish suitability for confidential work, specialized documents, or decisions requiring reliable evidence.

For knowledge workers, the connection is especially direct. An assistant’s usefulness often depends on the material it can retrieve and the evidence it preserves. A stronger general model cannot compensate for missing or outdated source documents.

Teams evaluating an AI knowledge base should therefore examine information access alongside model choice. The national competition provides context, while the organization’s own evidence determines whether an assistant delivers useful results.

This is where capability and adoption meet without becoming identical. Better models expand the range of possible applications. Successful deployment determines which possibilities become dependable work. Reporting should track progress in both areas and resist declaring one a complete substitute for the other.

Public trust cannot be inferred from technical leadership

Producing widely used AI systems does not automatically confer authority to write the rules governing them.

The governance question concerns who establishes acceptable practices, accountability, and oversight. It differs from the question of which country produces the most capable model. Users can rely on a product while distrusting the institutions associated with it.

Pew Research Center’s September 2026 AI regulation survey found median trust of 43% for China, 35% for the United States, and 34% for the European Union. These results came from 12 surveyed countries, mostly middle-income economies.

The survey’s scope matters as much as the percentages. Pew asked related questions in a different group of 25 countries in 2025. A change in aggregate results across those samples does not, by itself, establish that the same populations changed their views.

This methodological caution has consequences for the Lowy Institute AI snapshot. Its governance discussion should prompt closer examination of attitudes across countries. It should not encourage a universal claim about global public opinion.

Trust is also specific to the question asked. Confidence that a government can regulate AI does not establish a preference for its commercial products. Nor does using a foreign assistant necessarily express approval of that government’s policies.

For enterprise buyers, governance becomes concrete through responsibility. Someone must decide which information an assistant can access. Someone must investigate harmful or incorrect outputs. Someone must determine when human review is mandatory.

Those decisions cannot be delegated to a national leaderboard. An organization needs arrangements that remain understandable when something goes wrong. The relevant evidence includes operating procedures, enforceable commitments, and clearly assigned ownership.

For governments outside the United States and China, the same distinction creates room for independent choices. They can evaluate imported systems against domestic requirements without treating procurement as an unconditional endorsement of a supplier country.

That does not make the choices politically simple. Technical dependencies and policy relationships can interact. Still, treating every purchase as complete geopolitical alignment would erase the practical differences between applications and institutions.

There is also a limitation in the snapshot’s overall structure. Equal weighting makes the framework readable, but equal weights represent an analytical choice. They do not prove that each category has an equal effect on future outcomes.

A government facing supply disruption might place greater weight on industrial inputs. A software company expanding internationally might emphasize distribution. A public agency handling sensitive records might prioritize accountability and operating control.

None of these priorities invalidates the report. They show why readers should separate its overall assessment from their own decision criteria. The framework supports scrutiny when its assumptions remain visible.

The skeptical conclusion is therefore specific: the snapshot organizes evidence about leadership, while leaving important questions about outcomes unresolved. It does not establish universal product superiority, inevitable future dominance, or an internationally accepted right to govern AI.

What to watch after the Lowy Institute AI snapshot

The next useful evidence will connect competitive claims to sustained use, working infrastructure, and enforceable operating choices.

The first signal concerns adoption beyond initial interest. Over the coming months, model releases and distribution updates deserve attention when they include evidence of continued deployment. Download totals remain useful, but evidence of recurring operational use would strengthen claims about durable influence.

For Chinese downloadable models, the key distinction concerns experimentation versus continuing dependence. A model tested during development occupies a different position from one supporting an organization’s regular operations. Reporting that documents the transition would clarify the significance of distribution gains.

The same standard should apply to American services. Increased attention does not necessarily mean sustained usefulness. Evidence of continued organizational use would better support claims that consumer reach translates into lasting operational influence.

This signal would strengthen Lowy’s interpretation if American distribution advantages persist while Chinese open-model activity remains concentrated in experimentation. It would weaken a simple reading of that interpretation if Chinese models gain documented, sustained use across independently operated applications.

The second signal concerns infrastructure entering service. Company updates and public disclosures should distinguish proposed facilities from operating capacity. Evidence about electricity access, equipment availability, and commissioning would help explain whether announced expansion becomes usable computing.

That evidence matters on both sides of the competition. Additional operating infrastructure would strengthen the position of whichever ecosystem can use it effectively. Delays would weaken conclusions that rely on announced capacity arriving as expected.

The third signal concerns governance becoming operational. Watch for procurement requirements, implemented safeguards, and published assessments that change how organizations can deploy AI. Membership announcements and broad policy statements provide less evidence than rules with defined responsibilities.

This signal would challenge the assumption that technical leadership automatically determines governance influence. Institutions outside the leading model-producing countries can still shape practical requirements. Their influence becomes clearer when those requirements affect real deployments.

For readers following US China AI competition, these signals offer a disciplined alternative to reacting to every leadership claim. Each asks whether a measurable development changes the relationship between capability, distribution, and dependence.

Keep the original evidence alongside each subsequent update. Record which population an adoption statistic covers, which stage an infrastructure project has reached, and which obligations a governance announcement creates. That record makes later comparisons more useful.

Before your next AI purchase or deployment, translate the Lowy Institute AI snapshot into questions about your own exposure. Which provider controls access to the system? Which documents can it use? What happens when its behavior changes? Who handles failures, and how easily can the organization move its work?

Compare candidate systems using representative tasks and clear operating requirements. Preserve the results so a new release can face the same evaluation. Follow the national competition for context, then require evidence at the level where your organization actually works. Which dependency would be hardest to replace if the competitive landscape shifted?