The loudest signal of 2026 didn't come from a new language model or a silicon breakthrough. It came from Wall Street. On June 9, Apollo spearheaded a $35 billion capital solution for Broadcom's new AI XPV Platform, flanked by Blackstone and a syndicate of global banks. The venture targets more than 20 gigawatts (GW) of compute capacity for frontier AI labs through 2028.
This is an architectural shift, not just a financial one. The deal bundles chips, networking, power infrastructure, and long-duration customer commitments into a single, asset-backed mechanism. By doing so, it transforms AI infrastructure into an industrial product Wall Street can finance at scale. Consequently, AI capital expenditure is no longer capped by hyperscaler balance sheets. It has mutated into a private-credit engine, carrying leverage, aggressive utilization assumptions, and counterparty risks the broader market has barely begun to price in.
Simultaneously, Crusoe—a vertically integrated developer building for clients like OpenAI and Microsoft—hit the brakes on Project Jade. This 1.8 GW campus near Cheyenne, Wyoming, designed to scale to 10 GW alongside adjacent natural gas generation, had secured unanimous local approval in January 2026. Construction was slated for the first quarter. Its sudden pause exposes the defining paradox of this buildout: capital is limitless, but physical execution is suffocating. With up to 70 data-center proposals swarming Cheyenne, local authorities are already weighing a 12-month moratorium.
The Real Bottleneck Was Never Silicon
For two years, the AI investment thesis fixated on GPU scarcity. That narrative is dead. The binding constraint today is energized capacity.
In 2023, U.S. data centers drew roughly 176 terawatt-hours (TWh), or 4.4% of national electricity. The Department of Energy now projects that figure could surge to between 325 and 580 TWh by 2028—gobbling up to 12% of total U.S. demand. Globally, the sector's power hunger could eclipse 945 TWh by 2030.
These aren't incremental bumps. AI clusters operate less like corporate office parks and more like aluminum smelters—concentrated, non-cyclical, 24/7 industrial behemoths. The aging grid was never designed to absorb this shock. Interconnection queues stretch for years. Transformer lead times have blown out to four years, with prices tripling. Turbine backlogs are equally gridlocked. Equipment shortages and permitting hurdles now threaten nearly half of all U.S. capacity slated for 2026.
The previous "green hyperscaler" era relied on an accounting fiction. Renewable power purchase agreements (PPAs) let tech giants claim annual clean-energy parity. But an AI cluster demands hourly, firm baseload. A solar PPA in Texas does nothing to solve a 2 a.m. interconnection bottleneck in Virginia. As workloads grow denser, the chasm between carbon accounting and physical deliverability is impossible to hide.
Natural Gas Wins the Near Term (and It Isn't a Moral Statement)
Energy Secretary Chris Wright isn't mincing words, calling AI a "Manhattan Project" priority requiring "massively more energy." His administration has positioned natural gas as the unavoidable near-term bridge, alongside nuclear, accelerating permits for behind-the-meter gas generation that bypasses the grid entirely.
RBC Capital Markets estimates U.S. natural gas demand from data centers will hit 6.1 billion cubic feet per day by 2030. The Project Jade blueprint—pairing a massive data center with a dedicated on-site gas hub—is rapidly becoming the industry standard precisely because it cuts the utility out of the equation.
What Smart Capital Is Actually Buying
The market is drastically overvaluing press releases and ignoring the mundane industrial choke points that dictate reality. The true scarce assets aren't generic "data centers." The real leverage lies in energized land, gas interconnects, transformers, turbines, liquid cooling systems, water rights, air permits, and the sheer operational grit to assemble them on time.
An announced gigawatt is a phantom; an energized gigawatt is a fortress. The conversion funnel is merciless. Every element—site control, fuel supply, high-voltage gear, local politics, construction labor—must align perfectly. Many projects won't survive the friction.
For investors, the proprietary play is clear: own the bottleneck. Buy the suppliers whose order books are pulled forward by every sprawling pipeline announcement, regardless of whether those campuses ever break ground. Be highly skeptical of heavily levered private-credit structures banking on flawless execution and zero chip depreciation.
AI hasn't escaped the physical economy; it has crashed headlong into it. The next phase of this trade won't belong to the company pitching the best AGI narrative. It will belong to whoever secures the turbines, the gas molecules, the permits, and the political tolerance first.
NOT INVESTMENT ADVICE
Sources: https://www.crusoe.ai/resources/newsroom/crusoe-and-tallgrass-announce-ai-data-center-in-wyoming