Data centers are moving farther from established hubs, but cheap land is not the prize. Investors are paying for credible access to power, transmission, fiber and permits.
AI data centers are moving farther from Europe’s hub cities just as powered land is getting more expensive.
JLL says hyperscale campuses in the 2026–28 pipeline average 175 km from a hub city, compared with 46 km in the earlier period. At the same time, prime powered land in the FLAP-D markets reached about €2.26 million per MW in 2026, up 82% from 2021.
For investors, those numbers expose the real trade. JLL says primary markets command 2.3 times the price of secondary locations and four times that of tertiary ones. On a hypothetical 100 MW project, the implied difference is about €128 million versus secondary land and €170 million versus tertiary land.
The opportunity is moving outward. The risk is assuming that distance alone creates value.
175 Kilometers From the Hub
The UK government’s Compute Roadmap distinguishes between AI training, which is generally less dependent on location, and inference, which can benefit from proximity to users or data. That distinction gives large training projects more geographic freedom without making latency irrelevant to every workload.
Kajaani, Finland, shows what that freedom can look like when infrastructure already exists. CSC moved data-center operations there in 2012, using a former paper-mill site with industrial infrastructure and renewable power. XTX Markets has announced plans to invest more than €1 billion in a five-data-center complex on a 478-acre site. On August 31, 2026, EuroHPC signed a €387.8 million contract for LUMI-AI, expected to be available in 2027.
Northern Sweden offers an older version of the same pattern. Facebook’s Luleå data center went live in 2013. In Narvik, Norway, Nscale says its campus is designed for 230 MW, while an expanded Microsoft agreement will add more than 30,000 NVIDIA Rubin GPUs. Nscale has also announced financing tied to the project and a potential further 115 MW expansion.
The likely effect is not a blanket repricing of remote land. It is a wider search for places where remoteness and infrastructure happen to overlap.
North Wales makes the timing problem harder to ignore. The UK’s North Wales AI Growth Zone spans Prosperity Parc on Anglesey and Trawsfynydd in Gwynedd, and the government’s criteria require sites to demonstrate access to at least 500 MW by 2030. Separately, the Wylfa small-modular-reactor program is planned around three reactors producing at least 1.4 GW, with grid deployment targeted for the mid-2030s.
Those are large numbers attached to different clocks. A site whose economics depend on future generation cannot be valued as though that generation were already deliverable.
Requested Megawatts Are Not Delivered Megawatts
Alberta offers the clearest numerical warning.
The provincial government says about 19,565 MW of new data-center load was seeking grid connections as of July 30, 2026. Alberta’s record peak load was 12,785 MW. Its first-stage allocation for new large-load connections was 1,200 MW, and that allocation had already been fully consumed.
The queue was therefore more than 15 times the first-stage allocation.
That suggests a basic distinction for investors: demand for power is abundant; credible delivery of power is not.
Meta’s Sturgeon County project sits on the other side of that divide. Alberta describes it as a C$13 billion, 1 GW facility backed by new private generation.
Saskatchewan is drawing the distinction into policy. Bell and the provincial government announced a 300 MW data center in the Rural Municipality of Sherwood near Regina. Saskatchewan says SaskPower has begun transmission-interconnection work and SaskTel will connect the facility into Bell’s national fiber backbone.
The province’s August 27, 2026 framework requires new data-center proponents to supply their own power. Saskatchewan says more than 30 applications are in its queue.
Interest can fill a queue quickly. Financing generation is harder.
Tennessee presents a different version of the same constraint. CBRE says data centers represented 18% of the state’s industrial load in 2025 and cites more than 6 GW of new generation capacity from the Tennessee Valley Authority. Yet CBRE also identifies local opposition and infrastructure delivery as constraints. Statewide supply does not eliminate site-level interconnection risk.
And in West Texas, CBRE says projects in Armstrong, Wilbarger, Haskell, Gray, Roberts, Ector and Reeves counties reached energization milestones in the first half of 2026. Based on projects already approved in the construction pipeline, it expects West Texas to become one of North America’s five largest colocation markets by 2028.
That forecast matters because West Texas is not winning on proximity to a traditional data-center hub. It is winning, if CBRE’s forecast holds, on the ability to turn energy geography into computing geography.
The Industrial Inheritance
Some of the most useful sites were built for another era of heavy industry.
Kajaani’s CSC facilities occupy a former UPM paper-mill complex. In Harjavalta, Finland, Fortum says it worked with the city on zoning and grid-connection planning for the Sievari industrial area and in April 2026 signed a site-development agreement supporting Nscale’s proposed data center.
The attraction is not industrial nostalgia. Existing sites can bring together pieces that otherwise have to be assembled separately: land, grid planning, zoning and infrastructure.
AEP shows the same constraint from the utility side. In July, the company said it had 69 GW of contracted load growth through 2030 backed by signed agreements, alongside a US$78 billion five-year capital plan.
One juxtaposition matters here. JLL’s European data puts prime powered land at €2.26 million per MW. JLL also forecasts average global shell-and-core data-center construction costs of about US$11.3 million per MW in 2026, excluding land and active IT hardware, while AI technology fit-out can cost as much as US$25 million per MW.
For a hypothetical 100 MW project, that means about US$1.13 billion of shell-and-core construction and as much as US$2.5 billion of AI technology fit-out.
Against those sums, the value of shaving land cost has to be judged against the cost of waiting for power.
That does not make land irrelevant. JLL’s own multiples imply about €0.98 million per MW for secondary powered land and about €0.57 million for tertiary locations. On 100 MW, the arithmetic difference remains substantial.
But the relevant unit is no longer acreage alone: it is acreage attached to infrastructure that can support the project on schedule.
Buildings Have to Catch Up, Too
The shift is not only geographic.
CBRE notes that older CPU deployments commonly used roughly 3–10 kW per rack, while 100 kW-class GPU racks require liquid cooling. NVIDIA reference material for GB300 NVL72 systems describes full-rack power of up to 142 kW.
That does not make every lower-density colocation building obsolete. It does make electrical distribution and cooling architecture part of the investment case rather than an engineering detail.
Supply in the established North American markets is tight at the same time. CBRE reported 1.4% primary-market vacancy in the first half of 2026, with 80.4% of capacity under construction already preleased. Less than 1,500 MW remained available for preleasing, which CBRE said was roughly six months of demand at the then-current absorption rate.
The result is pressure from both directions. Existing markets have little slack, while AI hardware raises the technical bar for whatever gets built next.
Power Risk Becomes Financial Risk
Hardware efficiency complicates long-range demand assumptions. NVIDIA says its Vera Rubin NVL72 can deliver up to 10 times more tokens per megawatt than GB200 NVL72 on a specified Kimi-K2 Thinking inference benchmark. That is a vendor benchmark under defined conditions, not a general law of compute efficiency.
Still, the implication is worth underwriting: power demand depends not only on workload growth but also on how much computing future hardware can extract from each megawatt.
CoreWeave’s filings show another way infrastructure risk reaches the financial statements. In the second quarter of 2026, three customers represented 36%, 26% and 10% of quarterly revenue. The company reported US$35.6 billion of total indebtedness at June 30, and several debt facilities carried effective interest rates between 9% and 15%.
Its filings also disclose a past case in which insufficient power affected a customer project and led to service credits.
Power delivery, in other words, can become revenue risk.
Large-load tariffs can move the risk again. AEP Ohio’s data-center tariff requires qualifying customers to pay minimum demand charges that can reach 85% of contracted capacity, includes exit-fee provisions and requires financial viability. AEP says signed large-load agreements and rate structures could produce up to US$16 billion of cost offsets for existing customers.
Saskatchewan’s model pushes the obligation farther upstream by requiring new proponents to supply their own power.
Manitoba goes further in another direction. Its 2026 legislation imposes a levy on electricity supplied to large-scale data centers and certain other intensive users. If no specific rate is prescribed, the statutory default is 100% of the applicable monthly energy and demand charges. In June, Premier Wab Kinew explicitly framed the province’s priorities as saying no to AI hyperscale data centers in order to prioritize other large projects.
The physical site can work and the policy still can fail.
What Investors Are Actually Buying
A public project announcement proves that a project has been announced. It does not prove that surrounding land has repriced.
Cold weather can help cooling economics. It cannot substitute for deliverable power, transmission capacity, fiber or permits.
An electrical connection request is not a financed megawatt.
Those distinctions change how sites should be compared. Options or conditional purchase agreements can tie acquisition to defined power, planning and fiber milestones. The relevant principle is conditionality rather than any universal timetable.
CBRE says community resistance and zoning delays have also become serious obstacles, making local approval as important to site selection as power and fiber in some markets.
So the map can move only where several systems move together.
JLL’s pipeline says the map already is moving. The average hyperscale campus in its 2026–28 comparison sits 175 km from a hub city.
The investment question is which of those distant megawatts can arrive on time.