The artificial intelligence industry is usually analyzed as if it were the Olympics. Every few weeks another model runs onto the track, posts a slightly better benchmark score and declares itself the fastest machine alive. For a few days there’s a new champion. Then another model arrives.
That framing misses the real contest. The strategic battle in AI is no longer only about who builds the smartest model — it’s about who controls the assets everyone else needs. OpenAI and Anthropic are trying to preempt the scarce resources required to build and distribute frontier intelligence. Chinese labs are responding with open-weight models and cost leadership. One side is buying the wells before the desert gets crowded. The other is trying to make water free.
From Best-In-Class to Strategic Control
In my book, Decoding the Software Landscape, I outline the strategies companies use to attack a technology market: becoming best-in-class, creating a narrow niche, establishing yourself as the industry’s core, targeting noncustomers, pursuing cost leadership, preempting strategic assets, or using freemium and open-source distribution. As a market matures, the battle typically shifts toward platform strategies, bundling and attempts to tip the ecosystem toward a preferred standard.
Generative AI’s first phase was a best-in-class contest — OpenAI, Anthropic, Google and others racing to build models that were visibly better at reasoning, writing, coding and image generation. But the advantage doesn’t hold. Researchers move between labs, techniques spread through papers and distillation, and customers increasingly run several models at once. A company can lead a benchmark on Monday and be explaining its limitations by Friday. Once performance converges, the question stops being “who has the best model” and becomes “who controls the assets everyone else needs.”
OpenAI Is Buying the Battlefield
OpenAI’s infrastructure strategy is the clearest example of asset preemption. Through Stargate and a growing set of partnerships with cloud providers, chipmakers and infrastructure operators, OpenAI is securing enormous quantities of compute, energy, chips and data-center capacity — not just to meet its own needs, but to make that capacity scarcer for everyone else.
This is preemption in its classic form: acquiring a strategic resource before the market fully prices it, so that once it becomes scarce, competitors pay more, wait longer or settle for less.
The risk is real. If models become dramatically more efficient, if demand grows more slowly than expected, or if new architectures reduce the value of brute-force scaling, today’s infrastructure position could turn into an expensive overhang. OpenAI appears to be betting that the cost of owning too much infrastructure is smaller than the cost of showing up to the decisive fight without enough compute.
Anthropic Is Fortifying the Gates
Anthropic is preempting a different mix of assets. Its relationship with Amazon provides capital, AWS distribution and specialized infrastructure, and the company has built a strong position among enterprise, security-sensitive and regulated customers.
In regulated industries, trust functions as a strategic asset rather than a marketing line. Banks, hospitals, government agencies and critical-infrastructure operators move slowly, run extensive security reviews and don’t switch AI vendors casually. That friction is also defensibility: once a provider is trusted and integrated, replacing it is costly.
Anthropic’s more interesting move is the Model Context Protocol (MCP), an open standard that lets AI applications connect to tools, databases and external workflows. Strategically, this is a coring move — identifying an industry-wide coordination problem and positioning your technology as the standard the rest of the market organizes around. Anthropic doesn’t need to own every MCP connector; it benefits whenever developers and enterprises build their AI infrastructure around a protocol it helped establish. The fact that competitors have also adopted MCP arguably strengthens the strategy: it reduces Anthropic’s control over the protocol but increases the odds that it becomes the industry standard. Coring isn’t about ownership — it’s about defining the language the market uses. Models change; standards can survive several generations of them.
China Is Trying to Flood the Market
Chinese AI labs occupy a different position. They generally can’t match the combined U.S. advantage in advanced semiconductors, hyperscale infrastructure and global enterprise distribution by running the same playbook, so they’re not building a smaller version of the same castle — they’re undercutting its foundations.
DeepSeek, Alibaba’s Qwen, Moonshot AI and others have released increasingly capable open-weight models that developers can download, modify and run on their own infrastructure. These are more accurately described as open-weight than fully open-source — publishing weights doesn’t necessarily reveal training data or methods — but the economic effect can still be significant. Open weights let companies and governments deploy locally, customize freely, keep sensitive data in-house and reduce dependence on a proprietary U.S. API.
This combines several strategies at once: cost leadership, since developers avoid paying a premium per query; targeting noncustomers, since organizations barred from sending data to U.S. clouds can now deploy locally; freemium-style distribution, since broad access drives adoption while monetization shows up in hosting and support; and a tipping strategy, since each developer who builds on an open model strengthens its ecosystem.
This is strategic retaliation. If one competitor is trying to create power by controlling scarcity, the logical response is to destroy the scarcity. Chinese labs don’t need the highest margin in the foundation-model market — they can win strategically by compressing proprietary pricing globally and making Chinese models the default foundation in price-sensitive and sovereignty-sensitive markets. The goal may not be to own the toll road. It may be to make the toll road unnecessary.
Open Models Don’t Need to Win Outright
U.S. companies have already responded. Meta has long championed an open-model ecosystem, Google participates through its own open-model releases, and even OpenAI has released open-weight models — evidence that no closed-model provider can fully ignore proliferation as a strategy.
The market looks likely to settle into a barbell: large, controlled frontier systems monetized through premium applications and enterprise contracts at one end, and open-weight models built to attract developers and prevent a rival ecosystem from becoming default at the other.
Open models don’t need to capture the whole market to change its economics — they just need to be good enough that customers can credibly threaten to switch. Large enterprises will often still prefer proprietary providers for support, security and accountability, and self-hosting an open model is rarely as cheap as the zero-dollar license implies. But capable open models still put a ceiling on what proprietary providers can charge.
Google, Microsoft and Amazon Own the Roads
Other players are running different strategies entirely. Google combines vertical integration with distribution: custom chips, cloud infrastructure, foundation models and reach through Search, Workspace, Chrome and Android. It doesn’t need to win every benchmark if Gemini is already embedded in products billions of people use daily.
Microsoft’s edge is different — it can host models from multiple providers while retaining control of enterprise identity, workplace data and the customer relationship. Its likely objective is to commoditize the model layer while preserving the value of the workflow and distribution layers it owns.
Amazon, through AWS and Bedrock, offers access to multiple proprietary and open models while monetizing the compute, storage and networking underneath all of them — it doesn’t need to build the winning model if every model still runs on its infrastructure. Nvidia benefits from both camps: frontier labs need massive compute, and the spread of open models pushes thousands of additional companies and governments to deploy AI on their own hardware. Centralized or fragmented, someone still has to sell the shovels.
The same technology can be a strategic core for one company and a commodity for another. Open models threaten the economics of companies selling model access while strengthening the economics of companies selling the hardware to run them.
The Application Layer Should Expect Pressure
This has real implications for AI application companies. Foundation-model providers aren’t passive suppliers — they’re platform companies with large research teams, deep balance sheets and a clear appetite for entering adjacent markets. They can cut prices, add features, bundle products, release open models or move directly into profitable workflows. That makes generic AI wrappers vulnerable: an app whose main advantage is a clever prompt or temporary access to a better model may find its edge disappears with the next product announcement.
But best-of-breed isn’t dead — generic best-of-breed is. An application company can still build a defensible position around something a model provider can’t quickly reproduce: proprietary data, regulatory approval, customer trust, deep integrations, distribution or an end-to-end workflow. A model provider can copy a feature; it can’t instantly reproduce years of clinical integrations or a regulator’s trust. That’s where durable AI businesses are likely to emerge — in healthcare, cybersecurity, financial infrastructure, government, legal services and industrial operations, where the value isn’t a prettier summary but embedding into how decisions actually get made.
Watch for Retaliation
A market-entry strategy is incomplete without accounting for how competitors will react, and that’s especially true in AI. A startup may find genuine white space, but if that space sits next to a large platform company’s core business, it should assume the incumbent has noticed. The real question isn’t whether the entrant has a good product — it’s whether the incumbent has both the capability and the motive to respond. A niche application may be ignored; one that starts controlling valuable enterprise workflows is far more likely to draw a competitive response.
Follow the Strategy, Not the Benchmark
When evaluating an AI company, four questions matter more than its latest benchmark score: What scarce asset is it preempting? What part of the market is it trying to commoditize? Who controls the distribution through which it reaches customers? And what will the incumbent do if it succeeds?
The debate between open and closed AI isn’t mainly philosophical — it’s strategic. The industry isn’t just deciding whether models should be open. It’s deciding where scarcity will remain, where value gets commoditized, and which layer of the stack ultimately collects the profits.
The best model might lead a benchmark for a few weeks. The best strategy determines who owns the market after the parade is over.
