Meta has erected six tent style structures next to its New Albany, Ohio campus and is filling them with AI training hardware, according to satellite imagery and permit filings surfaced by Cleanview founder Michael Thomas. Each tent runs about 125,000 square feet, which puts the cluster in the same physical footprint as a midsized traditional data hall but assembled in roughly a quarter of the time. Construction ran from April into June 2026, and the company is powering the site with roughly 200 megawatts of modular gas turbines rather than waiting for utility interconnect.
The mechanics matter for anyone planning compute capacity in the next eighteen months. A conventional hyperscale build needs deep concrete, redundant utility feeds, multistage commissioning, and 18 to 30 months on the calendar before a single GPU draws power. The tent approach borrows two ideas from outside the cloud world. The first is the Fremont parking lot tent that Tesla erected during the Model 3 ramp in 2018. The second is the trailer mounted gas turbine pattern that xAI deployed in Memphis to bring its Colossus cluster online while the grid caught up. Stitched together, the result is what Thomas called the "Mad Max phase" of the AI race: temporary buildings, behind the meter generation, and accelerators worth billions sitting inside fabric walls.
The capex math behind the canvas
The financial pressure pushing Meta toward fabric and turbines is real. Meta has guided to roughly 145 billion dollars of data center and capex spend across the program, and the stock is down about 5 percent year to date as Wall Street pushes back on the burn rate. Mark Zuckerberg telegraphed the tent strategy to The Information last year when he talked about multi gigawatt clusters built under weatherproof shells. What we are seeing now is the strategy moving from slide to satellite photo.
It is also happening as the Wall Street Journal reports that Meta's Muse Spark model is complete internally but that the developer APIs to access it have slipped repeatedly. In other words, the pressure to deliver inference and training capacity is colliding with the pressure to ship product, and tents are how the capex side closes the gap. Brick and mortar would have meant another two earnings cycles of waiting.
How we are rebudgeting GPU and colo for 2026
For us, the operator implications are concrete and time sensitive. We are doing three things this quarter for our retail and digital commerce clients. First, we are pulling forward our reserved capacity decisions by one quarter, because if three or four hyperscalers copy this tent and turbine pattern, H200 and B200 class capacity in secondary US markets will loosen faster than the official roadmaps suggested in Q1. That changes the hedging math on reserved versus on demand pricing and on multi region failover for training jobs by roughly fifteen to twenty percent on our model.
Second, we are writing a new procurement clause for any net new colo agreement signed after August. The clause requires the vendor to disclose whether the contracted megawatts are utility sourced or behind the meter turbine, and to provide quarterly attestation on both. Buyers who skip this will end up with carbon accounting gaps that finance and ESG cannot reconcile at year end. We have already heard pushback from two Tier 2 colo operators saying the disclosure is "competitively sensitive," which is itself a signal worth pricing in.
Third, we are scoping a build versus buy refresh for clients spending more than ten million annually on training compute. The break even for renting GPU hours from a tent cluster operator versus owning depreciated H200 inventory in a leased cage has tightened sharply in the last sixty days. If Meta and xAI keep dropping cost per token at the current slope, owning silicon past 2027 looks weaker than it did at the start of the year for everyone outside the foundation model labs.
Where the risks land
Fabric structures are weather sensitive, and a tornado or ice load event in central Ohio could take a billion dollar cluster offline in minutes. Off grid gas turbines also push emissions disclosure into messy territory: scope 2 reporting assumes utility power, and scope 1 may not capture leased turbine output cleanly. Insurance underwriters are still pricing this risk, and we have heard from at least one broker that AI data center coverage is being repriced quarterly rather than annually. That is a procurement cost line most of our clients have not budgeted.
The regulatory pressure is gathering as well. Ohio legislators have already flagged concerns about local air permits for the turbines, and a Federal Trade Commission inquiry into AI infrastructure competition could pull tent style sites into scope. Behind the meter generation at this scale also tests state utility commission rules that were written for industrial cogeneration in the 1990s, not 200 megawatt AI clusters with no grid hookup.
The dates and thresholds we are watching
Three signals will decide whether the tent pattern hardens into industry standard or stays a Meta one off. The first is Meta's Q2 2026 earnings release, expected in late July, where any revision to the 145 billion capex figure or a new line item on temporary structure depreciation will tell us whether the financial controllers have blessed the tactic for fiscal 2027 and beyond. The second is the general availability date for Muse Spark developer APIs, currently unscheduled per the WSJ reporting: if that date lands before October it suggests the tent capacity is already producing usable inference, and if it slips into 2027 the gap between capex and product becomes a Wall Street story. The third is whether the Public Utilities Commission of Ohio opens a formal docket on behind the meter gas generation for hyperscale AI loads, which a state commissioner hinted at in May; a docket would set the template that Texas, Virginia, and Iowa regulators are likely to follow within twelve months, and would cap how aggressively any of us can plan turbine backed compute into 2027 budgets.
