Your AI Training Cluster Thirsty? Let's Talk Water.
We ran the numbers: A 10k H100 cluster can consume 2 million gallons of water a month. Here is the math and the engineering fix.
There's one number that matters for energy investing: LCOE.
Levelized Cost of Energy. $/MWh over the project lifetime.
It's how you compare:
Everything else is noise.
LCOE = Total Lifecycle Cost (NPV) / Total Lifetime Energy (NPV)That's it. Total cost divided by total energy, both discounted to present value.
| Parameter | Value |
|---|---|
| Technology | Solar PV |
| Capacity | 100 MW |
| CAPEX | $80,000,000 |
| Annual O&M | $800,000/year |
| Annual Generation | 200,000 MWh |
| Project Lifetime | 25 years |
| Discount Rate | 8% |
| Degradation | 0.5%/year |
Result: LCOE = ~$42/MWh
That's your target to beat. If another project has a higher LCOE, you need a good reason to choose it.
IRR tells you the return, but not whether the project is the best use of capital. LCOE tells you if you're producing energy efficiently.
NPV is scale-dependent. A 1GW project will have higher NPV than a 100MW project, but that doesn't mean it's more efficient. LCOE normalizes for scale.
Payback period ignores what happens after payback. A project that pays back in 5 years but produces expensive energy for 20 more years is worse than one that pays back in 7 years with cheap energy.
LCOE is sensitive to a few key inputs:
Understanding these sensitivities helps you assess project risk.
We built an open-source LCOE calculator that runs entirely in Python. No dependencies. Takes JSON in, gives analysis out.
It also does:
Try Our Free LCOE Calculator
Calculate the levelized cost of energy for any project. No signup required.
Open CalculatorIf you're evaluating energy projects and not leading with LCOE, you're doing it wrong.
Everything else is context. LCOE is the decision.
Use it, fork it, improve it. Our LCOE calculator is open source and free to use.
We ran the numbers: A 10k H100 cluster can consume 2 million gallons of water a month. Here is the math and the engineering fix.
Traditional SaaS is too slow for energy markets. We pivoted to 'Autonomous Organization as a Service'—software that works while you sleep.
Giving an agent 30 tools costs $0.45 per run. We implemented a 'Code-First Skills' pattern to drop that to $0.003.
Grid interconnection is the #1 bottleneck for AI. Google X's Tapestry project is trying to virtualize the grid to fix it.
News tells you what happened yesterday. Markets tell you what will happen tomorrow. We built an agent to trade on the difference.
Starting August 2025, mandatory environmental reporting kicks in for AI models. Most CTOs are completely unprepared.
We forced our AI agents to fight. The 'Bull' vs. The 'Bear'. The result was better decisions than any single model could produce.
Grid carbon intensity varies by 3x throughout the day. We built a scheduler that pauses AI training when the grid is dirty.
We didn't want to pay for a Bloomberg terminal, so we wrote a 950-line TypeScript scraper that builds our own intelligence feed.