Why the AI Race Is Not Just a Race for Better Models
The most visible contest is between systems that answer questions, write code and generate images. But a country cannot turn a clever model into durable power without the physical, commercial and institutional machinery beneath it.
AI capability sits on a stack. Control, access and resilience can differ at every layer.
An AI company releases a new model. It leads a benchmark, attracts users and briefly appears to be winning. Then a rival offers a cheaper model, another integrates AI into software used by millions, and a government restricts access to advanced chips. Which event changed the race?
All of them. “The AI race” sounds like one finish line, but it combines several contests: scientific progress, industrial capacity, commercial adoption, military and intelligence capability, standard-setting and control over critical dependencies. A benchmark measures only a slice.
The 2026 International AI Safety Report describes compute, algorithms and data as core drivers of progress, while also distinguishing the resources used to train a model from those used every time it answers.1 That already gives the race more than one track.
A model is the visible tip of a much larger stack
At the top are models and the applications people see. Beneath them sit specialised talent, training and operational data, software tools, cloud platforms, networks, processors, data centres and electricity. Rules and institutions cross the whole stack because they determine what may be built, bought, exported and trusted.
This is not a simple ladder in which ownership of every layer is required. Firms rent compute; countries import chips; developers adapt open-weight models; utilities serve foreign-owned data centres. But each dependency changes who can set prices, interrupt access or decide the terms of use.
The United States’ 2025 AI Action Plan makes this full-stack logic explicit by pairing innovation with infrastructure and international strategy, including proposed export packages spanning hardware, models, software, applications and standards.2 The strategic product is not merely a chatbot. It is an ecosystem.
Open the stack beneath the model
Distribution, workflow integration and user trust determine whether a powerful model changes actual decisions or productivity.
Conceptual stack. The layers interact rather than operating as a one-way pipeline.
A bottleneck can matter more than an average
Imagine a system with excellent researchers and a capable model but no dependable access to accelerators. Or abundant chips without enough grid capacity to run them. The scarce complement can set the pace for the whole system.
The OECD proposes analysing national AI compute through three dimensions: capacity, effectiveness and resilience. Hardware availability alone is insufficient if access, skills, policy or security prevent useful work.3 Europe’s AI programme follows similar logic by combining computing infrastructure with data, skills, adoption and rules rather than treating compute as a stand-alone trophy.4
Change one constraint in a fictional AI system
In this simplified example, usable output is limited by the lowest of the three indices.
What the model means
Each index starts at an arbitrary 100. One button lowers a single index to 35. “Usable output” is the minimum, not a forecast or a formula for national AI power. It illustrates complementarity: surplus strength elsewhere cannot instantly replace a missing input.
Training the model is not the end of the compute race
Training adjusts a model’s parameters using data and large computational runs. Inference happens when the trained model processes a user’s request. A country or company can gain access to an existing model yet remain dependent on someone else’s infrastructure each time it is used.
Reasoning techniques can also spend more compute after training to improve an answer. The International AI Safety Report notes that this inference-time scaling has become an important source of capability gains.1 So attention focused only on the largest training run misses the recurring cost and strategic importance of deployment.
Move from invention to everyday operation
A large training run can concentrate computation before release. Its cost does not reveal the later resources needed to operate the service.
Process map only. “Train once” is shorthand: real developers run experiments, post-training and updates.
Energy and construction speed have become technology policy
Advanced processors cannot work without buildings, substations, cooling and reliable electricity. The IEA’s 2026 analysis treats electricity supply, grids and local bottlenecks as central constraints on data-centre expansion.5 That is why national AI plans now discuss permits, power generation and technical trades alongside research.
Europe reported 19 AI factories in operation by April 2026 and was preparing larger gigafactories.6 The United Kingdom’s 2026 compute roadmap links sovereign capability to reserved compute access and a long-term effort across chips, systems and software.7 These initiatives are not proof of future leadership. They show what governments believe must be assembled to compete.
Why a software race becomes an infrastructure race
Servers also need networking, buildings, cooling, power conversion, maintenance and a suitable electricity connection.
There is no single national scoreboard
A frontier model can expand scientific and military options. Widespread adoption can raise the productivity of firms that did not invent it. Resilient infrastructure can keep essential services running. Trusted governance can make organisations more willing to deploy systems where mistakes matter.
These outcomes can diverge. A country may lead frontier development while another captures value by embedding imported models into manufacturing, healthcare or public services. A third may specialise in chips, data centres or safety evaluation. Calling only one of them the winner hides the mechanism producing national benefit.
Change what “winning” means
It emphasises the strongest general-purpose systems and the research ability to improve them. That is important, but it does not automatically measure diffusion or resilience.
Qualitative lens, not a ranking of countries. The buttons change the definition of success, not measured national scores.
Sovereignty is a gradient, not a sealed border
“Sovereign AI” can mean several things: authority over sensitive data, guaranteed compute for national priorities, the ability to run a model locally, domestic firms at key layers, or independence across the whole stack. Those are not equivalent.
Even an entirely domestic model may rely on imported fabrication equipment, software or energy technology. Conversely, a country using foreign models can preserve meaningful control through local hosting, procurement rights, interoperability, evaluation and the ability to switch suppliers. Absolute autarky is neither the only form of agency nor an easy endpoint.
China’s 2025 Global AI Governance Action Plan combines respect for national sovereignty with calls for infrastructure development, worldwide adoption, safety and cooperation.8 The EU’s cloud policy frames sovereignty around reducing strategic dependencies while maintaining an open market.9 The shared concern is control; the proposed institutional answers differ.
Climb the sovereignty ladder
Fastest route to capability, but the provider controls important terms, infrastructure and model changes. Procurement and exit rights matter.
A spectrum of control, not a quality ranking. Moving right can increase autonomy, cost and capability requirements.
Rules and trust are productive capacity too
When an AI system enters a hospital, bank, power network or government office, technical accuracy is not the only requirement. Security, accountability, privacy, reliability and a process for handling failure affect whether it can be used at all.
NIST’s AI Risk Management Framework treats governance, mapping, measurement and management as continuing functions across the AI lifecycle.10 ASEAN’s guidance similarly aims to align responsible deployment across different legal and commercial settings.11 Standards can impose costs, but they can also lower uncertainty and make adoption possible. Governance therefore sits inside the race, not outside it.
Trust can accelerate—or block—adoption
Whether those uses become durable deployment depends on evidence, integration, rules and the consequences of error.
A middle power does not have to imitate a superpower
Most countries will not fabricate every advanced chip or train the largest general-purpose model. That does not make them spectators. They can build leverage in specialised research, local-language systems, domain data, regional standards, energy-efficient infrastructure, public procurement and rapid adoption in sectors where they already have expertise.
Malaysia’s National AI Action Plan 2026–2030 describes an applied, human-centred programme spanning sectors rather than a promise to recreate every layer of the global stack.12 That is a different strategy: connect imported and domestic capabilities to national problems while strengthening the layers that make dependence manageable.
The sensible portfolio may combine foreign services for speed, locally hosted systems for sensitive work, domestic adaptation for language and context, and regional cooperation for governance and scale. The question is not “Can we own everything?” It is “Which dependencies are acceptable, which capabilities must remain available, and where can we create value others need?”
If one country has the best model, has it won the AI race?
Consider adoption, infrastructure, dependencies, security, standards and who captures value from widespread use.
The model is the headline. The stack beneath it determines who can build, deploy, govern and keep using it.
Sources & further reading
Research checked September 5, 2026. Forward-looking plans are identified as plans, not achieved outcomes. Interactive values are explanatory and do not rank countries.