1 min lesson
Baseline vs. peak
Compare two rows from "Baseline vs. peak", then say when each one fits.
Step 1 of 3
Baseline vs. peakprovisioning for peak everywhere is how budgets blow up
- Component
- Steady baseline
- How you serve it
- Reserved / committed capacity
- Why
- Predictable, so capture the low per-hour rate
- Component
- Predictable peaks
- How you serve it
- Reserved sized to the band + scheduled scale
- Why
- Known shape, plan ahead of it
- Component
- Spiky burst & experiments
- How you serve it
- On-demand / serverless
- Why
- Absorb without stranding reserved spend
| Component | How you serve it | Why |
|---|---|---|
| Steady baseline | Reserved / committed capacity | Predictable, so capture the low per-hour rate |
| Predictable peaks | Reserved sized to the band + scheduled scale | Known shape, plan ahead of it |
| Spiky burst & experiments | On-demand / serverless | Absorb without stranding reserved spend |
Match the funding mode to the demand shape; reserving for peak everywhere strands capacity in the troughs.
Learn more
Full explanation
Lead time and the cost of being wrong
Lead time and the cost of being wrongGPU supply is not instant
GPUs cannot be conjured on the day you need them. Supply is constrained and lead times are real, so a forecast has to run ahead of demand by the acquisition window or the plan is fiction. That lead time is exactly why scenario planning earns its keep: you pre-commit trigger points so you are ordering capacity before the upside case arrives, not after.
The discipline that makes a forecast trustworthy is quantifying the cost of being wrong in each direction. Over-provision and you strand reserved spend on idle cards. Under-provision and a launch is capacity-starved, degrading latency or blocking growth. Those costs are rarely symmetric and naming which way you are biased - and why - is what a senior leader is listening for.
- Over-provision
- Stranded reserved spend; the discount turns into idle hours you still pay for
- Under-provision
- Capacity-starved launch; latency SLAs break or growth is throttled
- The judgment call
- Decide which error is cheaper for this program, then size the buffer toward it on purpose
If the take-home or whiteboard asks you to forecast capacity, do not hand back one number. Hand back base/upside/downside with the assumptions visible, the baseline-vs-burst split, the lead time built in and an explicit statement of which direction you biased the buffer and why. A single point estimate signals you have never owned a forecast that got tested by reality.
Learn more
Optional practice
Practice: Baseline vs. peak
QWhy must a GPU capacity forecast run ahead of demand?