The race is no longer about the smartest model

For much of the past two years, the public imagination treated artificial intelligence as a horse race. Each new model arrived with a leaderboard score, a demo and a promise that the future had just been upgraded. That frame is now obsolete. The contest among OpenAI, Anthropic, Google, Microsoft, Meta and Apple is less about who can produce the cleverest chatbot than who can turn AI into a durable platform for work, search, software, consumer devices and national infrastructure.

The shift is visible in where the money is going. Big tech is pouring hundreds of billions of dollars into data centres, chips, power, cooling and cloud capacity, because the real bottleneck is no longer model architecture alone but the industrial base needed to run it at scale. Estimates cited this year put the combined AI spending of major technology firms at more than $650 billion in 2026, with Alphabet, Meta and Microsoft all committing extraordinary sums to the underlying infrastructure that makes generative AI viable. That is not the spending pattern of a passing fad; it is the footprint of a new industrial order.

It is also a sign that the market has matured into an oligopoly. Brookings notes that the general-purpose large-language-model market is dominated by three firms—Google, OpenAI and Anthropic—and that they controlled almost 90% of the enterprise market by the end of 2025. That concentration matters because it means the competitive battle is no longer just technological. It is commercial, political and regulatory. The companies are not only trying to build better models; they are fighting over who gets to mediate information, automate work and collect the data generated in the process.

OpenAI: the pioneer under pressure

OpenAI remains the symbolic centre of the AI boom. It helped make the modern generative-AI era legible to consumers, investors and policymakers, and it still occupies a uniquely visible place in the public imagination. But visibility is not the same as dominance. OpenAI faces the classic problem of the first mover: it set expectations so high that every product improvement is interpreted as proof of leadership, while every rival advance is treated as evidence that the field has caught up.

Its real challenge is structural. The company needs enormous compute to stay competitive, enormous capital to pay for that compute, and enough revenue growth to convince investors that the economics can eventually work. Yet the more OpenAI becomes a platform, the more it risks becoming a utility whose products are copied, bundled or commoditised by larger firms with deeper distribution. In enterprise markets, that means Microsoft can package similar capability inside existing workflows. In consumer markets, Google can fold AI into search, Android and Chrome. OpenAI may still define the cultural narrative, but narratives do not automatically become moats.

That tension is especially acute because OpenAI is now competing in several arenas at once: model quality, developer tools, consumer assistants, enterprise adoption and agentic systems that can perform tasks rather than merely answer questions. The company has the strongest brand in the field, but it also faces the most obvious exposure. The more famous the company becomes, the more every safety lapse, pricing decision or governance dispute becomes a referendum on the entire industry.

Anthropic: safety as strategy

Anthropic has built its identity around a different wager: that trust, reliability and alignment will matter as much as raw capability. In a crowded field, that is not a moral flourish but a commercial thesis. Brookings describes the market as an oligopoly, and within that small group Anthropic has tried to distinguish itself by becoming the preferred choice for enterprise buyers who care about governance, risk management and controlled deployment.

That positioning is especially powerful in regulated sectors. Banks, insurers, pharmaceutical firms and government contractors do not merely want a model that can write persuasive prose; they want systems that can be monitored, confined and audited. Anthropic has understood that one of the most valuable assets in AI is not novelty but permission. The company’s emphasis on safety is therefore also an attempt to become the vendor that enterprise compliance teams can bless without a lengthy internal revolt.

There is, however, a strategic irony. The more Anthropic markets itself as the safe alternative, the more it must prove that safety can scale without becoming a constraint on speed. In AI, caution can be a differentiator, but it can also become a ceiling. If the fastest-moving buyers conclude that the safety premium is worth paying, Anthropic prospers. If they decide that safety features are easy to mimic and hard to monetise, its advantage narrows quickly.

Google and Microsoft: the distribution giants

If OpenAI and Anthropic are trying to persuade the world to adopt their systems, Google and Microsoft are trying to make adoption feel inevitable. Their advantage is not just compute; it is reach. Google already sits at the front door of the internet for billions of people, while Microsoft owns the workplace software stack through Windows, Office and Azure. Those positions turn AI from a standalone product into a layer embedded in daily habits.

Google’s strategy is the most comprehensive. It can push AI through search, advertising, enterprise software, cloud services and Android. That gives it a rare ability to test, refine and distribute AI across consumer and business environments at once. It also means Google has much more to lose. If AI changes how people search for information, Google must reinvent the very product that made it powerful. If it fails, it risks becoming the incumbent disrupted by the technology it helped pioneer.

Microsoft, by contrast, is less exposed and in some ways better positioned. Its partnership with OpenAI has made it the default enterprise gateway to frontier AI, and Copilot gives it a way to bundle intelligence into the software millions of workers already use. That is the essence of platform power: not inventing every feature first, but making new capabilities feel native. The company’s strategy is to make AI a procurement decision, not a leap of faith.

Both companies understand that enterprise adoption will depend on more than benchmark performance. It will depend on integration, reliability, data controls and legal defensibility. Firms do not buy transformation; they buy workflows that save time without creating liability. Google and Microsoft are best placed to sell that proposition because they already sit inside the systems companies rely on every day.

Meta: open models, closed incentives

Meta occupies a more ambiguous position. It has become one of the industry’s most important infrastructure investors and one of the largest backers of open model development, but its core business remains advertising. That gives it a singular incentive structure. Meta is not trying to win AI by charging for premium assistants; it is trying to use AI to improve engagement, targeting and content production across its enormous social platforms.

That strategy can look like a philosophical commitment to openness, but it is also a hard-nosed business play. Open models expand influence by lowering barriers for developers and researchers, while simultaneously making Meta’s technology more ubiquitous. Yet openness has limits. If frontier AI becomes too commoditised, the profits may accrue not to the model maker but to whoever controls the interface, the data or the user relationship. In that scenario, Meta’s openness might help the ecosystem more than Meta itself.

Still, the company benefits from a crucial reality: for all the talk of AGI and assistants, much of AI’s economic value today lies in advertising, recommendations, content ranking and automation of existing digital services. Meta is exceptionally good at those things. It may not dominate the public conversation in the way OpenAI does, but it has the scale to shape the practical economics of everyday AI use.

Apple: the privacy test

Apple enters the AI race from a different angle altogether. It does not need to win the model contest in the same way because its power lies in the device layer. Apple controls the hardware, the operating system and the user experience, which means it can decide how much AI lives on the device, how much leaves it and what users are asked to permit.

This matters because AI has created a fresh privacy dilemma. The most useful systems often become so useful by absorbing more context: emails, messages, calendars, documents, photos, locations and habits. The more context a model has, the better it performs; the more context it has, the more dangerous it becomes. Apple’s pitch is that AI can be powerful without being promiscuous. On-device processing, hybrid cloud architecture and tight permissions are not just technical choices; they are the foundation of Apple’s brand.

That makes Apple perhaps the most important arbiter of the consumer AI era. Many users will not choose between competing foundation models directly. They will choose devices and apps. If Apple can make AI feel private, seamless and local, it may define the acceptable social norm for AI on personal devices. If it cannot, then the industry may drift toward a more intrusive model of ambient surveillance, in which convenience steadily erodes consent.

Regulation is becoming the real battlefield

The most consequential AI competition may take place not in product launches but in rulemaking. Governments now understand that the companies leading in AI are also accumulating unusual control over data, compute and digital infrastructure. That makes regulation both inevitable and contested. The central policy question is no longer whether AI should be governed, but how to govern it without freezing innovation or entrenching incumbents.

Big tech has a complicated relationship with this prospect. On one hand, large firms often welcome regulation because it can raise compliance costs for smaller rivals. On the other hand, the same rules can constrain product design, slow deployment and create liability for model errors, copyright disputes and privacy violations. The companies most likely to survive heavy regulation are those that can absorb its costs and convert compliance into a selling point. That again favours the giants.

Privacy is the pressure point that links all these debates. AI systems are hungry for data, but data rights are increasingly contested. Training models on public and private information raises questions about consent, ownership, retention and deletion. Enterprise customers want guarantees that their data will not leak into a shared model. Consumers want assistants that do not become surveillance tools. Regulators want transparency, but transparency can clash with trade secrets and security. The result is a policy environment in which every advantage is double-edged.

The next phase will reward systems, not slogans

The defining fact of this moment is that the AI race has expanded from model quality into systems power. The companies that matter most are not necessarily the ones that produce the most dazzling demo, but the ones that can coordinate chips, data centres, software distribution, developer ecosystems and regulatory legitimacy. In that race, the advantages of scale are enormous. Compute can be bought, but only by firms with access to capital markets and long-term infrastructure planning. Distribution can be built, but only by firms already embedded in daily life. Trust can be earned, but only by firms willing to absorb the cost of compliance and the discipline of restraint.

That is why the AI rivalry increasingly resembles the older contests that shaped the digital age. Search, mobile operating systems, cloud computing and social networks all began as product battles and ended as infrastructure battles. AI is following the same path, only faster and with more consequences. The early winners may still change, but the field is already narrowing around a small set of companies that can afford the future they are creating.

The public has been told to watch the models. The wiser move is to watch the plumbing: the chips, the data centres, the regulatory settlements, the privacy compromises and the distribution deals that determine whether AI becomes a useful assistant, a corporate utility or a new form of private infrastructure. The companies competing today are not merely selling software. They are negotiating who gets to shape the terms on which intelligence itself is delivered.

In the end, the decisive advantage may not belong to the company with the best model, but to the company that can make AI feel unavoidable, trustworthy and economically indispensable at the same time.