Physix FrontierConducting an Off-Grid Power Audit for AI Projects
A friend recommended an "Off-grid AI Power Health Check Sheet," so I gave it a try. A Bloomberg Tech report mentioned that Anthropic's investors led funding for an off-grid AI power startup. Here, "off-grid AI power startup" refers to companies providing self-sourced electricity, energy storage, and cooling for AI devices, independent of municipal grids. Anthropic has also noted that training frontier AI models will quickly require gigawatt-level power, and the US AI industry will need at least 50 GW of capacity in the coming years. For engineering teams, this translates to scheduling: Can compute, power, cooling, and O&M be delivered together?
I spent less than a week building a simple sheet in Excel. This sheet brings product, engineering, finance, and O&M together to judge whether an AI project is suitable for local, edge, or off-grid deployment.
Day 1: Set up the framework. Open Excel, create a new blank workbook, and rename Sheet1 to Power Health Check. Enter headers in A1 to G1: Project Name, Scenario, Daily Call Volume, Latency Requirement, Model Size, Offline Tolerance, Owner. Once the headers appear in A1:G1, the framework is ready. Fill row 2 with a real project, e.g., A2: Customer Service Summary, B2: Internal Doc Summarization, C2: 20000, D2: 2 seconds, E2: Small Model, F2: Can go offline, but must sync results back, G2: Product Team A. Select row 2, click Format under the Home tab, and choose Wrap Text so text doesn't cram into one cell.
The key here is clarifying the "Scenario." Model size is just a rough descriptor; for instance, small models are generally easier to deploy locally. Judgment also requires looking at daily call volumes, duration per call, and whether the business can wait. An easily overlooked point is asking only about model size while ignoring whether the business can tolerate latency.
Day 3: Add scoring and decision logic. Create a new worksheet named Scoring. Enter headers in A1 to C1: Dimension, Score, Notes. Define six dimensions: Power Continuity, Network Reliability, Cooling & Noise, O&M Response, Data Compliance, Total Cost. Score from 1 to 5, where 5 means conditions are mature and 1 means completely unprepared.
| Dimension | Key Question | Passing Signal |
|---|---|---|
| Power Continuity | Is there a stable power source? | Can sustain one shift during outage |
| Network Reliability | Can it connect online? | Results can be synced back after disconnect |
| Cooling & Noise | Server room temperature | Equipment can be near office areas |
| O&M Response | Who fixes it? | On-call staff available for night failures |
| Data Compliance | Where is data stored? | Data can remain local |
| Total Cost | How much for equipment? | Budget includes power, cooling, labor |
Enter Total Score in D1 and =SUM(B2:B7) in D2. Once D2 displays the total score for the six dimensions, you can use it for review meetings. In my testing, a score above 24 allows for a small-scale pilot, 18–23 requires fixing conditions first, and below 18 means don't rush to launch. This threshold is temporary and may not suit all teams.
Hold a review meeting after one week; don't just submit a sheet. I pull one person each from Product, Engineering, Finance, O&M, and Legal. Each person speaks on only two things: which column they own, and what evidence they lack. Multinational teams need to add time zones and acceptance boundaries. For example, the Singapore team says offline is okay, but syncing results depends on local networks; German colleagues ask if data leaves the country; US teams care about equipment procurement cycles. Organizationally, this sheet turns vague consensus into an acceptable checklist.
Pitfalls mainly occur in three places. Treating "Model Size" as "Power Demand" is a common misjudgment. Size is just model magnitude; inference also depends on concurrency, caching, quantization, and batching. Calculating only equipment purchase costs while ignoring cooling and O&M underestimates expenses. Off-grid sounds like plugging a battery into a mini PC, but in reality, you face temperature, fans, noise, backup networks, and on-site inspections. Turning assessment into blame-shifting distorts the health check. Team growth is important; the health check should expose process gaps. I made similar judgments when doing replaceability assessments before: Tools are mirrors.
Once the sheet works, take a small project for a two-week pilot. Record one power outage, one network jitter, and one manual restart daily, then backfill real data into the Notes column. You can also export the sheet as a README, clearly stating assumptions and owners, making it easy for other teams to reuse.
This health check sheet cannot make purchasing decisions for you, but it moves AI projects from asking "Can it run?" to asking "Can it run long-term?"
📌 This article is compiled from Bloomberg Tech. Original text: https://www.bloomberg.com/news/articles/2026-09-10/anthropic-investor-leads-funding-for-off-grid-ai-power-startup
Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.
