The Power Allocation

Meta just became the largest corporate buyer of nuclear power in tech history. In this episode of The Power Allocation, we break down Meta's 6.6 gigawatt nuclear energy deal with Vistra Corp, TerraPower, and Oklo — and what it means for AI infrastructure investment. Key topics covered: Why Meta chose nuclear over renewables for AI data center power How nuclear's 90%+ capacity factor solves AI's ba...

What if the solution to grid constraints is to skip the grid entirely? GridFree AI, a Houston-based startup, announced its first grid-independent site in South Dallas — part of a planned three-site campus with nearly 5 gigawatts of combined capacity. Five gigawatts. That's more than many countries consume. And none of it connected to the public grid. The model: on-site power generation using natur...

Oracle and OpenAI aren't just building a data center. They're building a power plant that happens to run GPUs. Vantage Data Centers broke ground on the Lighthouse project in Port Washington, Wisconsin — a $15 billion campus comprising four data centers delivering 902 megawatts of IT capacity. This is part of the Stargate initiative, the joint venture between Oracle, OpenAI, and SoftBank. To put 90...

Google didn't just buy a clean energy company. It bought a shortcut to power. Alphabet acquired Intersect Power for $4.75 billion in cash, assuming existing debt. Intersect isn't a tech company — it's a developer of utility-scale solar, storage, and data center-ready power infrastructure. Google already had one of the largest corporate renewable portfolios in the world. This acquisition isn't abou...

Capital can be deployed in months. Power infrastructure takes years. This mismatch is the defining tension of the AI buildout. A 200-megawatt data center requires 12-24 months for land and permitting, 18-24 months for construction, and 24-48 months for grid interconnection. Money raised in 2024 may not translate into running compute until 2028 or later. The AI infrastructure winners will be determ...

Every hyperscaler has announced ambitious renewable energy targets. But renewables alone cannot power the AI buildout. Here's why. AI workloads run continuously. Solar generates power roughly 25% of the time. Wind achieves 35-45% capacity factors. Neither provides the baseload reliability that AI facilities require. Battery storage helps but remains expensive at scale. The AI buildout is driving i...

The AI buildout depends on decisions made by regulated utilities — entities most tech investors have never studied. Utilities are the quiet kingmakers of this cycle. Utilities operate under regulatory frameworks that prioritize reliability and rate stability. Adding 200 megawatts of load for a single customer requires new substations, transmission upgrades, and generation procurement — all approve...

Data centers don't just need power — they need cheap power. And the geography of electricity pricing is reshaping where AI infrastructure gets built. For a 100-megawatt facility running 24/7, each cent per kilowatt-hour represents roughly $8.7 million in annual operating cost. A 5-cent differential means $43 million per year — nearly a billion dollars over a 20-year horizon. Power pricing is creat...

You can buy land. You can order GPUs. You can raise capital. But you cannot buy a faster grid connection. Interconnection is the binding constraint on AI infrastructure. In PJM — the grid operator for the Mid-Atlantic — the interconnection queue now exceeds 2,500 projects totaling over 250 gigawatts. The average wait time has stretched beyond 4 years. Developers who secured interconnection agreeme...

Hyperscalers and AI startups operate in the same industry but play entirely different games. The difference is infrastructure. Hyperscalers sign 10-15 year power purchase agreements, coordinate grid interconnection at utility scale, deploy liquid cooling systems, and plan 5-10 years ahead. Microsoft has nuclear agreements. Google invests in geothermal. Amazon backs utility-scale renewables. Startu...

For a decade, software ate the world. Now AI is making land matter again. Meta's 2-gigawatt facility requires hundreds of acres. Microsoft's expansion spans millions of square feet globally. These aren't software deployments — they're industrial developments with 30-year time horizons requiring transmission adjacency, water rights, and favorable zoning. Developers with optioned land near transmiss...

Data centers are relocating. Not for tax incentives. Not for talent. For power. Virginia's data center corridor consumed 26% of state electricity in 2023. Dominion Energy is struggling to keep pace. Grid stress events and connection delays stretching to years are pushing capital to Texas, Ohio, Indiana, Southeast Asia, and the Nordics. Capital is following power geography, not network geography. T...

The GPU shortage narrative is fading. The real constraint on AI compute is electricity — and it's not even close. Chip fabrication is scaling — TSMC, Intel, and Samsung are investing billions. But grid capacity doesn't scale like silicon. A single NVIDIA H100 draws 700 watts. AI training clusters now exceed 100 megawatts. Goldman Sachs projects 165% increase in global data center power demand by 2...

The AI boom is not constrained by chips, algorithms, or talent. It's constrained by electricity. Capital is moving accordingly. In this episode, we examine why AI facilities require 50-150 kilowatts per rack versus 10-15 kilowatts for traditional computing, why developers face 5-7 year grid interconnection delays, and how Microsoft and Meta are deploying $145 billion combined in infrastructure bet...