OpenAI set out to build its own infrastructure backbone. What it has built instead is a deeper dependency on the company that funds it – and the gap between those two outcomes is starting to matter.
The Infrastructure Problem OpenAI Cannot Outrun
Running large language models at scale requires an extraordinary amount of compute – server farms, cooling systems, networking hardware, and the engineering teams to manage all of it. OpenAI has none of that at the scale it needs. Microsoft does. The arrangement made sense when ChatGPT was a research experiment. Now that it is a commercial product generating serious revenue and competing against Google, Amazon, and a growing list of well-capitalized startups, that dependency looks increasingly like a structural liability.
OpenAI’s ambitions to operate its own data centers have been discussed internally and reported on by various outlets, but the company has not moved at the speed its public positioning suggests. Building proprietary cloud infrastructure requires billions in capital expenditure, long lead times for hardware procurement, and relationships with energy providers that take years to establish. Microsoft has spent over a decade building exactly that. OpenAI is starting from close to zero.
The partnership terms give Microsoft significant control over how and where OpenAI’s technology runs. Microsoft is OpenAI’s exclusive cloud provider for most commercial workloads, and that exclusivity has teeth. When OpenAI wants to expand capacity quickly – as it frequently does in response to surges in ChatGPT usage – it is not placing orders with its own operations team. It is negotiating with its biggest investor and most powerful partner simultaneously. Those two roles create an obvious tension.
That tension has not broken anything publicly yet. But the structure creates a situation where OpenAI’s growth is contingent on Microsoft’s willingness and capacity to accommodate it. Microsoft has every reason to keep ChatGPT running well – it has integrated OpenAI’s models across Azure, Copilot, and a dozen enterprise products. The incentives are mostly aligned. But aligned incentives are not the same as independent control, and OpenAI’s leadership knows that distinction matters at the scale the company is targeting.
Why the Cloud Independence Plan Has Not Come Together
OpenAI has reportedly explored building or leasing its own data center capacity, and its partnership with SoftBank on the Stargate infrastructure project was announced with the kind of fanfare that suggested a real break from Microsoft dependence. The reality is more complicated. Stargate is a long-term buildout, not an immediate solution. The compute it promises is years away from materializing at the scale OpenAI needs today. In the meantime, the company runs on Azure.
The cost structure of AI infrastructure makes this problem harder, not easier, as revenue grows. Training runs for new model generations consume compute at rates that make even well-funded companies wince. Inference – running the model for every ChatGPT query, every API call, every enterprise integration – is a continuous cost that scales directly with usage. More users means more spend, and right now that spend flows through Microsoft’s billing systems. OpenAI does not capture the margin on the infrastructure its own products require.
This is where the business model starts to look strained. OpenAI earns revenue from subscriptions and API access. It pays Microsoft for the compute that makes those products possible. The margin left over after that payment is the actual business – and that margin is constrained by a dependency the company did not fully anticipate when it was primarily a research organization. Building toward profitability while renting your own product’s engine from your largest investor is a genuinely difficult position.
Microsoft, for its part, benefits from the arrangement in ways that go beyond the infrastructure fees. Every OpenAI model that runs on Azure is a demonstration of Azure’s capability. Every enterprise customer that uses GPT-4 through Azure OpenAI Service is an Azure customer. The partnership gives Microsoft an AI credibility story it did not have before ChatGPT arrived. Unwinding it would be costly for both sides, but the costs would not fall symmetrically. Microsoft has other AI bets. OpenAI has one cloud provider.
The competitive dynamic makes the dependency more uncomfortable over time. Google runs its own models on its own infrastructure. Amazon develops its own chips and runs its models on AWS. Meta builds and trains models on its own hardware fleet. Every major AI competitor OpenAI faces is vertically integrated in ways that give them cost and speed advantages OpenAI currently cannot match. That is not a gap that closes by negotiating better terms with Microsoft – it closes by owning the stack, and OpenAI is nowhere near that.
What This Means Going Forward
OpenAI’s valuation has climbed to levels that price in an assumption of future independence – the idea that the company will eventually control its own infrastructure, set its own terms, and capture the full margin of the AI products it builds. That assumption may prove correct. But the timeline keeps slipping, and the Stargate announcement, while significant in the abstract, has not produced a single server rack that OpenAI controls today. Microsoft’s own AI expansion across its product suite adds another wrinkle – as Microsoft deepens its own AI capabilities, the two companies’ interests as partners and competitors will continue to blur.
The more immediate question is whether OpenAI can negotiate governance of this relationship on terms that protect its strategic interests as it scales – or whether each new product launch simply deepens a dependency that becomes harder to exit the longer it continues. The Stargate project is often cited as the answer to this problem. But Stargate is financed partly through SoftBank, built on land and power agreements that are still being assembled, and designed to deliver capacity in phases over years. OpenAI needs infrastructure leverage now, while the AI market is still being defined and while the margins between dominance and irrelevance are still thin enough to close.
