For most of the last decade, "AI company" meant a model lab or a software vendor. In 2026, the companies actually setting the pace of the AI buildout are the ones pouring concrete, stringing transmission lines, and signing power purchase agreements. Better models still matter. But a frontier model is worthless if there's no megawatt to run it on — and increasingly, the megawatt is the harder problem to solve. At Launch Capital, this is the shift our whole thesis is built around: AI infrastructure isn't a side bet on the AI boom, it is the AI boom.
The capital numbers make the point on their own. J.P. Morgan estimates hyperscaler capex will hit roughly $697 billion in 2026, and BloombergNEF puts the combined capex of the 14 largest data center operators globally at close to $750 billion this year — nearly double what they spent in 2025. PwC goes further out, projecting $31.6 trillion in cumulative global AI infrastructure investment through 2050, with data center capex alone climbing from about $800 billion a year today to $1.8 trillion a year by mid-century. Chip upgrades will eventually account for most of that long-run spend, but PwC is explicit about what determines where the money actually flows in the meantime: power.
And power is why the financing itself is changing shape. A decade ago, a data center was financed like an office building — one loan, one lease, one tenant profile. That model doesn't fit a $4-5 billion AI campus with a 15-year lease to a hyperscaler and equipment that depreciates on a completely different timeline than the shell around it. J.P. Morgan's June 2026 financing of Hut 8's Beacon Point facility in Texas — a $4.25 billion senior secured note offering, 100% pre-leased, at a 95% loan-to-cost ratio — is a preview of where the rest of the sector is headed: sponsors stacking development equity, project bonds, private credit, and long-term takeout capital, each layer priced against a different piece of the risk. Goldman Sachs expects private infrastructure funds, which raised a record $221 billion last year, to keep growing toward $3 trillion in assets by 2030 largely on the back of this demand. Capital discipline, not chip access, is becoming the gating factor on who gets to build.
Meanwhile, the power itself is moving behind the meter. Grid interconnection queues of five-plus years don't match a hardware refresh cycle measured in months, so developers are increasingly building their own generation rather than waiting for the utility. Roughly 2 GW of behind-the-meter capacity was already operating by mid-2026 — led by xAI's nearly 1,500 MW of gas turbines outside Memphis — with another 36% of identified projects permitted and the rest in earlier stages. In the Foley 2026 Data Center Survey, 56% of developers said they're now actively exploring co-located or on-site generation, and McKinsey projects 25–33% of all incremental data center demand through 2030 will be met behind the meter rather than through a traditional grid connection. On-site generation used to be backup power. It's becoming a primary power strategy.
Power, not compute, is the scarce resource. The sites that win are the ones where energy planning happened before ground was broken — not after. That means treating generation capacity, grid connectivity, and permitting timelines as core underwriting criteria, on equal footing with tenant quality and land value, rather than downstream logistics to solve once a lease is signed. It's also why we look as closely at a project's power plan and financing structure as we do at its anchor tenant.
The winners in this cycle won't only be the labs building the models. They'll also be the developers who can secure and finance power at scale, the utilities and independent power producers racing to keep up with demand, the construction and engineering firms executing multi-billion-dollar builds on compressed timelines, and the capital providers structuring the debt and equity that makes all of it possible. Real estate used to be won on location. AI infrastructure is won on megawatts, financing structure, and speed to power — and increasingly, those three things are the same conversation. AI runs on compute. Compute runs on power. And right now, power runs on whoever can finance and build it fastest.
Launch Capital is a Toronto-based venture capital firm and family office focused on AI infrastructure — data centers, GPU compute, and the power that runs them