Small AI Data Centers Are Booming for the Wrong Reason

The most telling detail in this year's AI infrastructure race is not another gigawatt megaproject. It is that OpenAI and Anthropic, the two companies signing the biggest compute deals in history, are now also hunting for tiny data centers, and the reason is not clever strategy so much as a bottleneck neither company can build its way around.

Key takeaways

OpenAI and Anthropic are pursuing data center deals as small as 20 to 30 megawatts, a fraction of the gigawatt-scale campuses they usually chase, according to CNBC.
Local opposition blocked or delayed 45 U.S. data center projects worth $68 billion in the second quarter of 2026 alone, per research group Data Center Watch.
Nearly $200 billion in data center investment was disrupted across the first half of 2026, with 843 opposition groups now active in 49 states.
Anthropic's own $45 billion, 460 megawatt deal with Nscale in West Virginia shows how far behind schedule the giant projects already are.

Small AI data centers are suddenly the hot new deal size, and that alone should tell you something is off. According to CNBC, both OpenAI and Anthropic have been quietly sounding out compute providers for facilities in the 20 to 30 megawatt range, tiny compared to the multi-hundred-megawatt and gigawatt sites both companies have spent the past two years locking up. Anthropic has been looking at the U.K. and the Nordics; OpenAI has focused on the Nordics as well, according to the same reporting, corroborated separately by Tom's Hardware, Invezz and Finimize.

The Megaprojects Are Stuck, So the Deals Got Smaller

It is worth remembering how big "big" has gotten. In August 2026, Anthropic agreed to pay Nscale $45 billion over six years for roughly 460 megawatts of computing capacity at a data center campus in West Virginia, a deal confirmed by Bloomberg, CNBC, Forbes, Yahoo Finance and Nscale's own regional partners. That is the kind of commitment that defines the current AI boom, the same one behind the roughly $700 billion Big Tech is expected to spend on AI infrastructure this year: enormous checks written years in advance for power and chips that do not exist yet.

Those mega-deals take years to deliver, because they require new transmission lines, new substations and, in many cases, entirely new power generation. A 20 to 30 megawatt site, by contrast, can often plug into an existing building or a colocation facility that already has grid approval. Structure Research analyst Jabez Tan told CNBC that "speed to usable capacity is the draw," adding that stitching together several smaller deployments can still add up to substantial capacity. An OpenAI spokesperson framed it as portfolio management, telling CNBC the company is "building a diversified compute portfolio to meet growing demand for AI around the world."

That framing is not wrong, but it undersells the pressure behind it. Companies do not usually diversify into slower, less efficient infrastructure unless the fast option has stopped being fast. As techdrifting has covered before, the AI data center buildout hit real supply constraints well before this year, and the small-site hunt looks like the next stage of that same story rather than a new one.

Local Opposition Is Now a Bigger Drag Than Chip Supply

The clearest evidence that the gigawatt pipeline is jammed does not come from either AI lab. It comes from Data Center Watch, a research project run by 10a Labs that tracks community pushback against data center construction across the United States. Its count for the second quarter of 2026 found that local opposition blocked or delayed 45 projects worth a combined $68 billion, a figure reported consistently by Bloomberg, Gizmodo, Tom's Hardware and Insurance Journal. In the first quarter, the same tracker logged about 75 disrupted projects worth roughly $130 billion. Add the two quarters together and nearly $200 billion in planned data center investment ran into a wall in just six months.

High voltage transmission lines carrying electricity across the countryside

The opposition is not a handful of loud town meetings. Data Center Watch counted 843 active opposition groups spread across 49 states, missing only Hawaii, and found that 30 state legislatures have introduced or adopted rules covering where data centers can be built and how much electricity or water they can draw. A Gallup survey from March 2026, cited in Gizmodo's reporting, found that seven in 10 Americans oppose building an AI data center in their own area, including 48% who are strongly opposed. Some states, including New York and Minnesota, have already moved toward moratoriums or paused approvals on new projects, according to that same reporting.

Put plainly: the constraint on AI compute in 2026 is not primarily Nvidia's production line. It is zoning boards, ratepayer advocates and residents who do not want a 24-hour industrial facility with its own substation next door. That is a much slower problem to solve than building more chips, because it runs through hundreds of separate local governments rather than one company's factory schedule.

Why Inference Workloads Make Small Sites Viable at All

The small-site strategy only works because of a technical shift in what the compute is actually for. Training a frontier model still benefits from one enormous cluster of tightly networked chips in a single location, which is exactly what gigawatt campuses are built for. Running inference, the day-to-day work of answering user prompts once a model is trained, does not have the same requirement. It can be split across many smaller, geographically scattered sites without a meaningful performance penalty, which is why 20 to 30 megawatt facilities are suddenly a usable building block instead of an afterthought.

✦ Free Newsletter ✦

Never miss a story

Tools, tutorials and AI deep-dives - straight to your inbox, every week.

That distinction matters for anyone trying to judge whether this is a sign of AI weakness or AI maturity. It is not evidence that demand for AI is shrinking; if anything, both companies are adding capacity through every channel they can find. It is evidence that the easy, single-site version of scaling is over, and both companies are now paying an efficiency cost, smaller sites, more contracts, more logistics, in exchange for capacity they can actually get online this year rather than in 2028.

Will This Push AI Subscription and API Prices Higher?

Not directly, and not soon. Smaller sites cost more per megawatt than gigawatt megaprojects, since ready-to-use capacity carries a premium. That premium is small next to a company's total infrastructure spend, so it likely will not appear as a line item on a ChatGPT Plus or Claude bill. The bigger risk is slower feature rollouts, not higher prices.

A small modular industrial building exterior in a rural area

Who Should Actually Track This Story

This matters most to enterprise buyers negotiating multi-year AI compute or API contracts, cloud and colocation investors, and anyone in local government dealing with a proposed data center. For that audience, a few things are worth watching:

If a vendor's promised capacity timeline slips more than two quarters, ask specifically whether it depends on a still-pending grid interconnection.
Favor providers with a mix of committed gigawatt-scale and smaller, already-permitted sites; a single-source mega-deal carries more schedule risk right now than it did in 2024.
Watch state-level legislation in your operating region, not just federal AI policy, since siting and water rules are now being written county by county.
Treat quarterly Data Center Watch-style opposition tracking as a leading indicator for AI compute availability, the same way supply-chain analysts watch chip lead times.

If you are simply a day-to-day user of ChatGPT, Claude or another AI assistant, this is safe to ignore for now. Nothing here points to a near-term outage or price shock for consumer plans, only a slower, lumpier path to the next round of capacity increases.

Macro close-up of illuminated fiber optic cables

The Steelman: Maybe This Is Just Smart Diversification, Not a Crisis

The strongest counterargument is that none of this is actually alarming. Large companies routinely run a mixed portfolio of big, slow assets and small, fast ones, and calling that a "bottleneck" gives ordinary risk management a scarier name than it deserves. OpenAI's own explanation, a diversified compute portfolio, is a completely normal thing for any infrastructure-heavy business to say, and it would be happening to some degree even without local opposition, simply because distributed inference is technically sound.

That view is fair, but it does not explain the scale of what Data Center Watch is tracking. Portfolio diversification does not usually involve $198 billion in projects getting blocked or delayed in six months, or 843 organized opposition groups, or 30 state legislatures actively writing new siting rules. Companies diversify for efficiency; they do not usually diversify because their preferred option is under legal and political siege in dozens of jurisdictions at once. The honest read is that this is both: a technically sensible move that is also, in large part, a response to a real and growing constraint that neither OpenAI, Anthropic nor Nvidia controls.

What to Watch Next

The number to track is not the next headline gigawatt announcement, it is Data Center Watch's next quarterly figure. If the value of blocked or delayed projects keeps climbing past the $68 billion logged for the second quarter, expect more AI labs to follow OpenAI and Anthropic into smaller, faster-to-permit sites, and expect that to show up eventually as slower feature rollouts rather than sudden price increases. If the figure falls instead, that is a real signal that permitting reform or new power deals, like the kind covered in techdrifting's look at AI's nuclear power push, are starting to unstick the pipeline.