← All projects
Concept for Absurd · by Anirudh Devella

Buy the right amount of compute

Every generation on Absurd costs compute. Buy too little and users wait. Buy too much and the money sits idle. This shows a usage forecast with its range, how monitoring catches it drifting, and the trade-off behind the buffer you buy.

Fictional example. The people, numbers and events are made up to show the method. This is an outside concept, not Absurd's product or data, and not affiliated with Absurd.

Built for Absurd · the short version

Buy too little compute and users wait. Buy too much and the money sits idle.

In this sample, monitoring flags a launch-driven surge a week before usage runs past capacity.

The data is made up. To make it real, I'd need weekly usage by model family and team, plus past capacity commitments.

Who built it

Anirudh Devella, data scientist. Before this, in production:

  • 8.3% forecast error at Bluelight (YC)
  • $14M+ in invoices benchmarked at Bluelight
  • AI Coach for 10,000+ athletes at Brandlete
  • 150K leads scored at CoverRight

“I’d hire him again.” Alex Dixon, Co-founder, Bluelight (YC), managed Anirudh directly. Read it on LinkedIn

01 · A forecast with a range

Not one number, a range you can plan against

Weekly generations across all model families. The forecast for the next 10 weeks comes with a likely range and a wide range. Capacity is bought against the range, not the middle line.

Actual usageForecastLikely range (50%)Wide range (90%)Committed capacity
02 · Monitoring catches the drift

When usage leaves the range, someone hears about it

A forecast is only right until something changes, like a new video model launch. Monitoring compares every week to the range and raises an alert before capacity runs out.

03 · How much buffer to buy

Idle money vs users waiting

Buying more capacity above the forecast costs money every week. Buying less means some weeks users wait. Move the buffer and watch both sides.

Capacity above the forecast
30%
Idle capacity cost
per month, illustrative prices
Weeks users wait
expected per year

Next step

My first week with your real data

  1. Backtest a pooled forecast across model families against the last 6 months.
  2. Set the alert rule: usage outside the forecast range two weeks running.
  3. Size the next commitment from the forecast range instead of a flat buffer.