The race is no longer about intelligence alone
For much of the past three years, the public story of artificial intelligence was simple: better models would beat worse models, and the company with the best benchmark numbers would inherit the future. That story has become much less useful. The real competition now runs along three fronts at once: model quality, product distribution, and control over the infrastructure that makes AI cheap enough, fast enough, and safe enough to use at scale. In that contest, OpenAI, Anthropic, Google, Apple, Microsoft, and Meta are not merely competitors. They are also customers, suppliers, partners, and, increasingly, rivals to their own users. The result is an industry that looks less like a software market than a power struggle over the operating system of the digital economy.
The market itself is narrowing at the top. One analysis cited by Brookings says the general-purpose large language model market is effectively an oligopoly dominated by Google, OpenAI, and Anthropic, which together controlled almost 90% of the $37 billion enterprise market by the end of 2025. That kind of concentration matters because it means the direction of AI will be shaped by a small number of firms making choices not just about engineering, but about pricing, access, and policy. When so much of the frontier is controlled by so few, competition and regulation start to blur into the same conversation.
Frontier labs are becoming platforms
OpenAI and Anthropic were once sold as model companies, pure and simple: laboratories producing increasingly capable systems and licensing them out. That distinction is now breaking down. As AI systems become embedded in office software, coding tools, search, customer service, and device assistants, the model layer has started to behave like a platform layer. The company that owns the model often wants the user, the workflow, the data, and the relationship with the enterprise customer. That is why the industry’s business model has become so contested. The same company that supplies the brain increasingly wants the hands, eyes, and memory too.
OpenAI is the clearest example. It remains the emblematic frontier lab, but its ambition now extends far beyond model releases. Its partnership with Microsoft gives it distribution through Copilot and access to cloud infrastructure at enormous scale. Yet the partnership is also a warning sign for the rest of the industry: when a model provider is deeply embedded in a customer’s workflow, it may end up competing with that customer’s own products and services. Brookings notes that AI companies increasingly compete with their customers, especially in enterprise markets where software vendors, startups, and internal IT teams once depended on neutral model providers. Neutrality is hard to preserve when the most valuable asset is the customer relationship itself.
Anthropic has taken a different path, emphasizing safety, governance, and controlled enterprise deployment. That strategy has made it attractive to firms that want powerful models without the reputational and compliance risks that can come with more freewheeling deployment. But Anthropic’s recent refusal, according to Axios, to sign an industry letter backing the open-weight ecosystem also hints at the strategic fault line inside AI itself. Some companies see openness as a way to broaden adoption and reduce dependence on a few closed providers; others see it as a threat to quality control, monetization, and safety. The split is not merely technical. It is ideological and commercial.
“The AI industry is increasingly dividing along economic lines, with infrastructure companies embracing open-weight models while frontier labs remain more skeptical.”
That divide helps explain why the debate over open-weight AI has become so heated. Open-weight models are not the same as fully open-source systems, but they are far easier to run, customize, and distribute than tightly controlled proprietary products. For chipmakers, cloud firms, and infrastructure players, openness can be a feature: it drives usage, creates demand for compute, and lowers the barrier to experimentation. For frontier labs, it can threaten pricing power and increase the risk of misuse. The question is not whether openness is good or bad. It is which parts of the stack benefit from it, and who loses leverage when it spreads.
Google’s advantage is patience and integration
Google enters this phase of the race with advantages that are easy to underestimate. It has one of the strongest model families, vast infrastructure, a dominant search and ads business, and a product ecosystem that spans cloud, Android, productivity tools, and consumer devices. In a market where distribution increasingly matters as much as raw capability, that breadth is invaluable. Google does not need to win a single spectacular release to shape the outcome. It can embed AI across products until the distinction between search, assistance, and workflow disappears.
That is why Google’s challenge is not simply technical. It is strategic self-cannibalization. The company’s core search business still matters enormously, and AI changes the economics of answering questions, recommending products, and mediating information. A company can either defend the old model or use AI to build the new one before somebody else does. Google appears determined to do the latter, even if that means transforming its own revenue engine. The logic is familiar from earlier platform transitions, but the stakes are higher now because the interface itself is changing. In the AI era, whoever controls the answer controls much of the user’s path through the internet.
Enterprise demand strengthens Google’s hand. Brookings highlights that Google is one of the three dominant general-purpose model providers in the enterprise market. That matters because business customers are less interested in spectacle than in reliability, compliance, and integration. Google can offer all three, or at least the credible promise of them, through its cloud platform and its long experience serving large organizations. If AI becomes a core layer of enterprise software, the firms best positioned to sell it are those already trusted with data, identity, and infrastructure.
Microsoft is the broker of the new era
If Google is trying to preserve and rebuild its own platform, Microsoft is trying to mediate between platforms. Its alliance with OpenAI has given it a powerful position in enterprise AI, especially through Copilot and Azure. The company is not just selling models; it is selling access, workflow, and procurement simplicity. For corporate buyers, that is often what matters most. They do not want to assemble an AI stack from scratch. They want a vendor that can sit inside existing procurement, security, and identity systems.
That makes Microsoft perhaps the most important broker in the industry. It benefits when enterprises adopt AI through its cloud and productivity layers, whether the underlying model comes from OpenAI, another provider, or increasingly from a mix of systems. The company’s leverage comes from being the default organizer of the AI workplace. Yet it also faces a structural risk: if frontier model providers become too strong, they may capture more of the value than Microsoft does; if open-weight models become too dominant, pricing pressure may erode margins across the stack. Microsoft wants AI to be everywhere, but not so cheap that it becomes invisible.
That tension is reflected in the scale of spending now coursing through the sector. Campaign US reports that Microsoft, Alphabet, Meta, and Amazon are on track to spend more than $650 billion on AI investments in 2026, with money flowing toward data centers, specialized chips, and cooling systems. The numbers are staggering, but they make economic sense if AI is becoming core infrastructure rather than a discretionary software feature. A serious bid for AI leadership is no longer a matter of software talent alone. It is a capital-intensive industrial project.
Meta is betting on openness and scale
Meta occupies a peculiar position. It is one of the largest AI infrastructure spenders, a major proponent of open-weight models, and a company whose core business is still advertising. Its support for open-weight AI is therefore not just philosophical. It is strategic. Open models can accelerate adoption, attract developers, reduce dependence on rivals’ closed systems, and help Meta position itself as the champion of a more distributed AI ecosystem. They can also lower barriers for the next generation of apps built on top of Meta’s own platforms.
Meta’s advantage is distribution at scale. Billions of users already live inside its products. If it can turn AI into an ambient layer across messaging, creation, advertising, and commerce, it can make AI feel less like a separate tool and more like a permanent feature of digital life. That may be more powerful in the long run than winning a temporary benchmark contest. In consumer technology, ubiquity often beats elegance.
But Meta’s openness has limits. A company that supports open-weight models also needs to protect its own competitive edge, especially in advertising and recommendation systems. That means Meta’s version of openness is likely to be selective and strategic rather than ideological. The company wants a world in which AI is widely used, but not one in which the most valuable advantages are easily copied. In that sense, Meta is not fundamentally different from its peers. It is simply more willing to dress its commercial interests in the language of openness.
Apple is turning AI into a device feature, not a cloud product
Apple’s role in the AI race is distinct because its strategic unit is not the model but the device. While OpenAI, Anthropic, Google, Microsoft, and Meta compete to define the cloud intelligence layer, Apple is trying to make AI feel local, private, and useful on the device itself. That is a different proposition. It emphasizes on-device processing, user control, and tighter integration with hardware rather than the spectacle of giant frontier models. In a market increasingly worried about privacy, that may prove persuasive.
Apple has long understood that trust can be a feature. Its AI strategy appears designed to preserve the company’s reputation for privacy while still incorporating the capabilities users now expect. Rather than asking consumers to accept that every request must travel to a remote server, Apple wants to place at least part of the intelligence closer to the user. That is not only a technical choice. It is a political one. In the age of AI, privacy is becoming a competitive differentiator.
Apple’s method may also be the most durable. Not every AI interaction needs the largest model. Many tasks can be handled by smaller, faster systems tuned to the device and the context. If Apple succeeds, it could redefine consumer expectations again, just as it once did with the smartphone. The company does not need to lead the frontier lab race to remain central. It only needs to make AI feel indispensable in the place people already spend their lives.
Regulation will not settle the race, but it will shape the winners
For all the industry’s technical drama, regulation may be the decisive variable. Governments are no longer asking whether AI should be governed; they are asking how much control is enough, which rules should apply to which systems, and who is responsible when AI causes harm. That question matters differently for each company. Frontier labs worry about liability, model evaluations, and deployment constraints. Platform firms worry about privacy, competition, and data usage. Device makers worry about how much processing happens on the phone versus the cloud. The regulatory burden will not fall evenly.
Data privacy is the most immediate fault line. The more AI systems rely on user data to personalize responses, the more they resemble surveillance infrastructure. The more they are used across workplaces, the more sensitive the question of retention, training, and access becomes. Enterprises want models that are powerful without being porous. Consumers want convenience without surrendering too much control. Regulators, meanwhile, are trying to define what counts as personal data in systems that learn from patterns rather than just records.
That is why AI regulation will likely favor firms with the largest compliance budgets and the strongest existing trust relationships. Big Tech has been regulated before, and often survives by turning regulation into a moat. The same may happen again. Smaller firms can innovate quickly, but they have less room for error when the costs of mistakes, fines, or product recalls are high. In that sense, regulation may not slow the oligopoly so much as harden it.
The new AI order is about control, not just capability
The most important change in the AI market is that success is no longer measured by a single model release. It is measured by whether a company can control the entire arc from compute to interface, from data to distribution, from safety policy to enterprise procurement. That favors firms that already possess scale, capital, and customer relationships. It also helps explain why the industry’s fiercest debates are happening not over whether AI will matter, but over who gets to define the terms on which it matters.
OpenAI and Anthropic may still own the prestige of the frontier. Google may have the deepest integration. Microsoft may have the best enterprise brokerage. Meta may have the broadest distribution. Apple may have the strongest privacy story. But none of them can dominate the future alone, because the future is being built as an ecosystem of competing dependencies. The AI race is therefore less a sprint to a single finish line than a struggle to set the rules of a permanent infrastructure layer.
That is what makes the moment so consequential. The companies now shaping AI are not just making software; they are building the architecture through which information, work, and identity will increasingly flow. The decisive question is no longer which model is smartest on a benchmark. It is which company can persuade users, enterprises, and governments that its version of intelligence is the safest, cheapest, most convenient, and least dangerous way to live and work. In the end, that may be the hardest form of intelligence to build.
