For two decades the binding question in computing was "can we get chips?" Around 2024 it became "can we get power to the chips?" That shift is this layer: energized land, power contracts, generation, and the shell around them — the slowest assets in the stack to create, and suddenly among the most contested.
The bottom of the stack owns the physical asset: land with grid access, power purchase contracts, generation capacity, the interconnect, and the building shell that compute eventually fills. The physics are plain:
"AI compute is, physically, a machine for turning electricity into intelligence."
Everything above this layer can be bought, rented, or re-architected on a timeline of months. Power cannot. Permits, physics, and multi-year lead times make energized capacity the hardest part of the stack to fake or rush — which is exactly why its scarcity is durable, and why "powered land" with grid access has become a prized asset in its own right.
The boundary matters: this layer owns the physical asset. The moment a company operates that asset as compute — renting GPUs or serving tokens — it is doing the work of the Data Centers & Compute layer. Land, contract, generation, shell: that is the whole of Layer 01, and all of it.
Goldman Sachs' 2030 forecast for global data-center capacity rose 78% in eighteen months — and every revision moved up.
Hand-curated from Goldman Sachs Research (Feb 2025) and its
July 2026 revision as reported by Yahoo Finance, with LBNL actuals for context;
committed as data/lps-capacity-forecasts.json. The 168 GW interim
estimate is the prior figure the July 2026 note raised; its publication date is
not stated, so it is shown undated between the two dated revisions.
The forecast went from roughly 122 GW in February 2025 to 168 GW, then to 217 GW by July 2026 — up 78% in about 18 months, with roughly $6 trillion of implied capex. Each revision moved up. The actuals point the same direction: US data centers grew from 58 TWh of electricity in 2014 to 176 TWh in 2023, reaching 4.4% of all US electricity, per LBNL's report to Congress. When the same forecaster keeps raising the same number, the story is not the estimate — it is the direction of every revision.
Land, Power & Shell books the smallest slice of the dollar — roughly $0.06 — but it is the layer with the fastest cash-on-cash returns, because the asset earns the moment it is energized and leased. And unlike every layer above it, this slice appreciates as pushback grows: each permit denied and each moratorium signed makes capacity already in hand worth more.
Illustrative split — assumptions in Methodology.
The map of this layer runs from generation — the gas bridge, nuclear and SMR baseload, behind-the-meter builds — through grid and transmission, to cooling, where liquid cooling becomes mandatory as chips densify. The real bottleneck sits in the middle: interconnection. Getting a new project connected to the grid takes years of queue process, which is why value accrues to whoever controls scarce, dispatchable, near-term power rather than to whoever files the next application.
A footnote on the queue itself: LBNL's national interconnection backlog peaked around end-2023 (~2,600 GW of projects seeking connection, roughly twice the installed US fleet) and has since shrunk on withdrawals. The bottleneck argument is the years-long queue process, not an ever-growing backlog.
Gavin Baker calls regulation the single biggest tail risk for AI, and the warning signs are concrete. In July 2026, New York's governor signed an executive order pausing state environmental permits for data-center projects of 50 MW and up for as long as a year — an executive-order moratorium in the state that hosts the financial capital of the country. (The legislature had passed a broader 20 MW moratorium bill in June; the governor did not sign it.)
Some of the fuel for that pushback is misinformation. The most cited example: Karen Hao's bestseller Empire of AI overstated data-center water usage by a factor of roughly 4,500. The error was publicly corrected — but the corrected impression travels far more slowly than the original claim. Baker points at the counter-evidence: new data-center builds that lowered local power prices and sustained blue-collar jobs. The gap between the pushback and the reality is itself the LPS opportunity — opposition suppresses new supply while demand compounds, which is precisely the condition under which energized land explodes in value.
"As data centers get more pushback, energized land can explode in value."
That is Chamath Palihapitiya's framing, and he is positioned behind it: he claims roughly 6 GW of grid and behind-the-meter power coming online between now and 2029 with developer Anita Verma-Lallian — of which roughly 1–1.5 GW is publicly verifiable today (the Hassayampa Ranch acquisition in Arizona, May 2025). He calls LPS the most obvious and fastest path to cash-on-cash returns in the stack. Meanwhile the demand hitting this layer keeps scaling: SpaceX is reported to be targeting 8 GW of compute on a $250B-plus plan, with more than 2 GW by end-2026 — a figure Baker relays with open skepticism as near-implausible, which is itself a measure of how large the asks have become.
Dwarkesh Patel's August 2026 analysis argues that compute growth of roughly 3x per year may be unsustainable: fabs are bottlenecked and wafer reallocation is hitting a wall. Power and land form the other wall. The compounding logic for this layer is that every constraint upstream — silicon, packaging, turbines, transformers — raises the value of energized capacity already in hand, because capacity in hand is the one thing a constrained system cannot quickly mint more of.
The bull case has three named failure modes. Stranded power contracts, if AI demand disappoints and multi-year commitments outlive the workloads that justified them. Regulatory and local opposition, which cuts both ways — it makes existing capacity scarcer, but it can also strand a specific project. And the intermittency problem: renewables' variable output against a workload that wants 24/7 baseload, which is why the gas bridge and nuclear restarts carry so much of the near-term buildout.
Builds power-first: sites campuses on stranded and behind-the-meter energy, bringing generation and shell together before the grid queue can say no.
Owns already-energized sites with secured interconnects — capacity built for bitcoin mining, now converting to AI campuses years ahead of any new queue position.
The institutional owner: generation, transmission, and land at global scale, deploying tens of billions into AI infrastructure as an asset class.
Every other bottleneck in this stack clears with enough capital. This one clears with permits, grid interconnections, and generation coming online, and all three run on multi-year clocks. That makes power scarcity structural, and it's why forecasters keep revising the same 2030 number upward instead of the boom cooling off. I'd rather own the ground this whole thing has to stand on than pick winners four layers up.