Catalyst with Shayle Kann

Episode: When to Colocate Data Centers with Generation

Release Date: September 12, 2025

Episode Overview

In this episode, host Shayle Kann dives deep into the increasingly popular—and sometimes misunderstood—trend of colocating new power generation with large-scale data centers, particularly in the context of the U.S. energy landscape. Kann is joined by Brian Janis, co-founder of Cloverleaf Infrastructure and former head of energy at Microsoft, to scrutinize the real motivations and practical challenges behind these projects. The conversation examines why generation is being built behind the meter, debunks certain industry narratives, highlights alternative solutions like demand response and virtual power plants, and explores the evolving prospects for technologies like batteries, renewables, and nuclear.

Key Discussion Points and Insights

1. What Drives Onsite Power Generation at Data Centers?

• Fundamental Data Center Power Needs

  • Data centers, whether for AI or traditional cloud services, require extremely high reliability due to high capex and the critical nature of their workloads.
  • Even batch workloads for AI model training have minimal tolerance for unexpected outages.
    • “Nobody wants to build a $20 billion training model and just turn it on and off every time the electricity starts to cut out.” — Brian Janis [04:35]

• The Standard Architecture

  • Most data centers are connected to the grid, have onsite Uninterruptible Power Supplies (UPS) for short outages, and backup diesel generators for rare, longer events.
    • “That architecture: grid connection, UPS, backup gen set, that's the kind of basic, like, dominant paradigm, right?” — Shayle Kann [06:26]
  • Diesel generators are often limited by air permits and practical issues at higher scale (e.g., gigawatt-class sites). Permitting vast diesel capacity is becoming infeasible.

2. Current Hype vs. Reality: Colocated ‘Prime Power’ Generation

• Industry Trend

  • There's a surge of announcements about colocating substantial new generation (gas, nuclear, renewables) directly onsite, not just as backup but as “prime power.”
  • Kann observes this has grown from useful innovation to potential hype, where the real rationale isn't always clear.

• Arguments For Onsite Generation

  • Speed to Power/Interconnection Delays:
    • Long utility interconnection timelines (5–7+ years) have pushed operators to consider self-generation to get sites running faster.
    • “If I go to utility and they tell me it's going to be five to seven years to get the connection at the scale I want, then maybe it's faster for me to just build my own generation…” — Brian Janis [10:12]
  • Need to Match 24/7 Load with 24/7 Generation:
    • Some believe colocated baseload generation is necessary because renewables are intermittent.

• Brian’s Skepticism of Both Arguments

  • Gas Supply Not Guaranteed: Congestion in the gas grid can be as challenging as in electricity; getting unlimited gas everywhere isn't feasible [11:38].
  • Cost and Overbuild:
    • To meet data center reliability, onsite generation must be oversized, driving up capital costs and underutilizing capacity.
    • “The per kilowatt hour cost of that system is extraordinarily high.” — Brian Janis [14:36]
    • In cheap electricity markets (e.g., Texas), it’s rarely competitive: “You go off grid in Texas and you're paying... $150 to $200 a megawatt hour, 24/7. And your neighboring data center connected to the grid is paying $20 for that same power.” [14:36]
  • No Unique Cost Advantage: The supposed savings from skipping transmission and distribution are outweighed by the required redundancy buildout.

• “Bit Watt Spread”

  • The key is not the absolute cost of power, but the margin compared to the revenue from data center operations.
    • “It's not a deal killer that you're paying that much for power. It's just something you have to take into consideration...” — Brian Janis [16:48]
  • Fast, reliable startup still sometimes justifies higher onsite power costs if interconnection is impossible.

3. Is the “Grid Can’t Connect Us Fast Enough” Problem Real?

• Kann’s Characterization of Janis’ View

  • The industry may be prematurely giving up on grid connections; there’s typically more headroom than believed.
    • “[You think] there is more headroom in interconnection capacity on a reasonable timeframe on the grid, is that right?” — Shayle Kann [17:41]

• Utility Planning is About Peaks, Not 24/7 Continuity

  • Utilities solve for summer/winter peaks, not 24/7 supply. The connection bottleneck is a capacity problem, not necessarily an energy one. [18:36]
    • “The problem we're trying to solve here is not that I need 24/7 generation to match a 24/7 load, it's that I need to solve for the summer peaks and the winter system peaks in order to connect a load.” — Brian Janis [18:36]

• Role of Flexibility and Advanced Grid Solutions

  • Adding storage, advanced transmission, grid-enhancing technologies, and virtual power plants can unlock more grid capacity without new baseload plants.
    • “Transmission and generation are sort of substitutionary. And storage is the same way…” — Brian Janis [20:00]

4. Challenges and Opportunities in Orchestration

• Utility and Developer Coordination

  • The big challenge is orchestrating multi-party solutions (advanced conductors, batteries, VPPs) across over 3,000 U.S. utilities and multiple market operators.
    • “This is where my argument maybe falls apart because the orchestration of this is the real challenge.” — Brian Janis [25:12]

• Emerging Solutions: Bring-Your-Own Capacity

  • Some utilities are piloting new models, letting load customers bring part of their own capacity (including batteries or demand response) to reduce their grid profile.
    • “If you want to site a 500 MW data center and the utility says, I've got 300 MW for you... operate it such that you never pull more...” — Shayle Kann [26:51]

• Virtual Power Plants (VPPs) and Demand Response

  • Rather than onsite generation, large-scale demand response or geographically targeted VPPs can enable faster site deployment with less capital outlay. Still largely unproven at scale, but coming soon.
    • “The fastest way to do it is through a virtual demand response sort of program.” — Brian Janis [28:16]
    • “We are pretty close to doing that on a couple of projects.” — Brian Janis [29:31]

• Ideal Future: Software-Enabled, Stackable Capacity

  • In an ideal world, utilities and developers use trusted analytical tools to mix and match resources (storage, demand response, onsite generation) for cost-optimal interconnections.
    • “Ideally you would want to have a tool where you could take any point of interconnection and... spit out ‘here’s a stack of capacity, least cost to most cost...’” — Brian Janis [30:56]

5. Resource-Specific Commentary

• Onsite Wind & Solar

  • Viable in areas with abundant land and good resources (e.g., West Texas), but not scalable for the entire industry or all geographies.
    • “That's still going to be a relatively small percentage of the overall market because we can't put everything in West Texas.” — Brian Janis [33:05]

• Nuclear: Behind the Meter at Data Centers

  • Highly speculative, unlikely in the near or medium term due to build timelines, costs, and reliability uncertainties.
    • “I don't think there's a credible argument for behind the meter nuclear at a data center in the near future. And by near future, I mean the next couple of decades.” — Brian Janis [35:20]
    • “We're just saying not on site.” — Shayle Kann [36:04]

Notable Quotes & Memorable Moments

“Nobody wants to build a $20 billion training model and just turn it on and off every time the electricity starts to cut out.”
— Brian Janis [04:35]

“The per kilowatt hour cost of that system is extraordinarily high. I mean, extraordinarily high...”
— Brian Janis [14:36]

“You go off grid in Texas and you're paying somewhere between $150 to $200 a megawatt hour, 24/7. And your neighboring data center connected to the grid is paying $20 for that same power.”
— Brian Janis [14:36]

“In almost no case would that ever be true because you would overbuild that system to meet the level of reliability.”
— Brian Janis [12:03]

“The problem we're trying to solve here is not that I need 24/7 generation to match a 24/7 load, it's that I need to solve for the summer peaks and the winter system peaks in order to connect a load.”
— Brian Janis [18:36]

“Transmission and generation are sort of substitutionary. And storage is the same way.”
— Brian Janis [20:00]

“This orchestration opportunity really is, I think, the huge opportunity to end up in a world where we actually do connect a lot more of this stuff to the grid, rather than... everything is bifurcated behind the meter, which I think is a worse outcome.”
— Brian Janis [25:12]

“I don't think there's a credible argument for behind the meter nuclear at a data center in the near future.”
— Brian Janis [35:20]

Timestamps for Key Segments

• 02:22 – 04:15: Setting up the episode theme—colocated power for data centers
• 04:17 – 06:26: Core data center power requirements and architecture
• 07:06 – 09:12: Limitations and realities of existing onsite generation
• 10:12 – 16:15: Strawman arguments for onsite 'prime' generation and Brian's counterpoints
• 16:15 – 18:36: “Bit watt spread” and economic calculus
• 18:36 – 22:17: Reality of grid headroom, utility planning, and advanced solutions
• 24:03 – 26:51: Complexity of multi-party orchestration and programmatic solutions
• 26:51 – 30:32: Hybrid options to optimize interconnection and limits
• 32:04 – 36:31: Resource-specific discussion: renewables, nuclear, and the future

Tone and Style

The tone is inquisitive, candid, and occasionally skeptical—Kann and Janis often challenge industry dogmas and hype with grounded, technical arguments and industry insights. The conversation balances hard-nosed economic and engineering analysis with humor (“near, far, wherever you are, you’ll be powering my data center”), providing value both for industry insiders and listeners seeking an honest take on data center energy strategy.

Summary Takeaway

While colocating generation with data centers appeals as a shortcut around slow interconnection queues, most real-world cases are plagued by high costs, redundancy overhead, and resource challenges. The real solution, argue Kann and Janis, lies in better grid coordination, smarter use of storage and flexibility, and the hard work of orchestrating multiple technologies and stakeholders—rather than falling for the mirage of behind-the-meter baseload fixes.