GapLimit

R02Collision Lab

The power market does not care what the computer computes

Supported
To the grid, a model-training cluster and a mining hall first appear as load. That resemblance invites a clean story: AI pays more, miners leave. The real collision is messier. Power quality, uptime, latency, cooling, chips, land and interconnection can matter as much as the price of a megawatt-hour.

The short answer

The two industries can compete where grid capacity and interconnection are scarce, but they are not interchangeable tenants. AI often values dense, reliable, connected compute. Mining can value cheap, interruptible and geographically awkward power. The strongest competition is for prepared sites and firm capacity; the weakest is around stranded or frequently curtailed energy.

Original GapLimit object

The load collision matrix

Competition grows where requirements overlap. This is a comparison framework, not a market-share estimate.
  1. 01
    Firm powerUsually valued more by continuous compute
  2. 02
    Interruptible powerMining can stop when revenue falls below power cost
  3. 03
    Prepared siteSubstation, cooling, fiber and permits create overlap
  4. 04
    Stranded energyLow overlap when connectivity or uptime is poor
01

The shared bottleneck

IEA and U.S. energy research point to rapid data-center demand growth. EIA also identifies data centers and cryptocurrency mining as large-load contributors in places such as Texas. A substation, transmission line, transformer queue or water permit can make two otherwise different businesses meet in the same application stack.

02

Uptime changes the bid

A miner can stop hashing and forgo expected revenue. Some AI workloads can shift time or place, but customer-facing inference and expensive training runs may value continuity. This does not make every AI load inflexible or every mine responsive. It makes interruption tolerance a variable that site comparisons must expose.

03

Chips are not fungible

Mining uses specialized hashing hardware. AI relies on accelerators, memory and network fabrics suited to different calculations. A mining building cannot become an AI campus by changing the logo. Electrical distribution, cooling, fiber, redundancy, security and construction quality determine how much of the site is reusable.

04

Follow the site, not the slogan

The useful unit of analysis is a specific grid node and project: available megawatts, energization date, firmness, curtailment contract, network access and cooling. Global totals describe pressure. Local constraints decide who actually collides.

Thesis audit

Pressure the bridge

Causal bridge
Compute demand → shared need for powered sites → interconnection scarcity → different willingness to pay and curtail → geographic reallocation.
Counterforce
Mining can seek power and interruption profiles that high-availability AI facilities reject; direct displacement is not inevitable.
What would prove it wrong?
The broad displacement thesis fails if new generation and transmission grow with both loads, or if their site requirements keep overlap small.
Largest uncertainty
Project-level power contracts and curtailment economics are often private. Public capacity announcements are not measured consumption.

Source ledger

Evidence carrying this piece

Sources accessed 2026-09-06. Links point to the originating institution where available.
  1. IEA · Energy and AIData-center electricity demand and concentration
  2. U.S. EIA · Tracking cryptocurrency mining electricityMining load, mobility and demand-response evidence
  3. U.S. DOE/LBNL · Data center energy useU.S. data-center energy baseline and projections