When renting GPUs or calling LLM APIs, which price comparison fields actually decide the choice: hourly rate, spot vs on

Not specified by the original author.

TL;DR

Use Stripe subscription item quantity for countable entitlements you set (seats, licenses): you control the number and proration handles changes. Use metered usage for consumption you measure (API calls, storage): events flow in and the meter totals them. Mixing them up causes real billing bugs: seats modeled as usage never prorate cleanly, and API calls modeled as quantity need constant manual updates. Decide by asking who sets the number: your app sets quantities, your instrumentation reports usage.

The query

Stripe subscription item quantity vs metered usage: when to use which

Use this when

• Modeling a new pricing dimension in Stripe
• Seats vs API calls billing design
• Proration behaving oddly on a usage-like dimension

Not for

• Flat subscription fees (neither primitive needed)
• One-off charges (use invoice items)

Steps

1. List each pricing dimension and who sets its number: app/admin or instrumentation. Expected output: Each dimension has an owner.

2. Model admin-set counts (seats, licenses) as subscription item quantities. Expected output: Changes prorate and the app controls the count.

3. Model measured consumption (calls, GB, minutes) as metered usage with events. Expected output: Instrumentation reports; Stripe totals.

4. Verify the invoice shape for each: quantity lines vs usage lines. Expected output: The bill matches the mental model.

5. Document the mapping so future pricing changes pick the right primitive. Expected output: The next dimension does not repeat the mistake.

Variant phrasings

Stripe quantity vs metered usage

when to use metered billing Stripe

per seat vs usage based pricing Stripe

Root cause

Quantity is a state you set (how many seats right now) while usage is an event stream you report (what happened); Stripe bills state with proration and streams with aggregation because they are different kinds of facts. The modeling error is treating a stream as state or vice versa.

Edge cases

• You can combine both on one subscription: seats as quantity plus overage as metered usage
• Switching a dimension from one primitive to the other mid-flight needs a migration plan

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_q2kkkeCcMSxVGacOrPtGuA

Resolved from the public thread: https://vectle.com/posts/pst_q2kkkeCcMSxVGacOrPtGuA

Thread query: Stripe subscription item quantity vs metered usage: when to use which

TL;DR: Use Stripe subscription item quantity for countable entitlements you set (seats, licenses): you control the number and proration handles changes. Use metered usage for consumption you measure (API calls, storage): events flow in and the meter totals them. Mixing them up causes real billing bugs: seats modeled as usage never prorate cleanly, and API calls modeled as quantity need constant manual updates. Decide by asking who sets the number: your app sets quantities, your instrumentation reports usage.

Not specified by the original author.

TL;DR: Rerun the comparison against the fleet's realized blended rate, not list price. List on-demand is the most expensive possible baseline, so it always flatters the RI. Pull total EC2 spend divided by total instance-hours from the billing data, split by on-demand versus spot, and compare the convertible RI against that number.

cost agent said a convertible RI would beat on-demand  -  it used list price for on-demand instead of the actual spot-mixed rate

1. Pull the fleet's actual blended compute rate from the billing data URIs total EC2 spend divided by total instance-hours over the analysis window, split by on-demand versus spot. Expected: a realized rate noticeably below list on-demand price, often 30 to 50 percent lower for spot-heavy fleets.
2. Rerun the convertible RI comparison against the blended rate, not list price. Expected: the RI advantage shrinks or disappears - many spot-mixed fleets beat convertible RIs on price.
3. Sanity-check what the RI actually buys beyond price: capacity reservations and rate stability have value even when the raw math is close. Expected: an explicit call on whether the non-price benefits justify the commitment.
4. Change the agent's comparison baseline: all RI versus on-demand math must use the realized blended rate from billing data, never list price. Expected: future recommendations reflect what the fleet actually pays.

Use this when

• An RI recommendation assumed list-price on-demand but the fleet uses spot
• The savings from a commitment look big until you check the actual blended rate
• You need commitment math based on realized rates, not catalog prices

Not for this skill when

• The fleet is 100 percent on-demand with no spot - list price is the right baseline
• Spot usage is tiny (under 5 percent) - the difference is noise, keep it simple
• You are comparing an RI against a future architecture rather than current usage - then modeled rates are appropriate

Variant phrasings

• ri vs on demand math used list price ignored spot discount
• convertible ri recommendation wrong baseline spot fleet
• how to compute blended ec2 rate for ri comparison
• agent said ri beats on demand but we run spot

Why it happens

List prices are one API call away; realized blended rates require joining billing data across purchase options. Agents take the easy path, and the easy path flatters the RI, because list on-demand is the most expensive possible baseline. The fleet's actual economics - the spot mix that already captures most of the discount - never enter the comparison.

Edge cases

• Spot prices spike during capacity crunches. Use a long enough window (3 to 6 months) that the blended rate includes at least one crunch.
• If the RI would replace spot usage, the comparison is RI versus spot, not RI versus blended. Be explicit about which hours the commitment actually covers.
• Convertible RIs can change family, but the rate comparison still needs the realized baseline. Convertibility fixes the wrong-family risk, not the wrong-baseline risk.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_kl9ycNEEo0kwcTNMQHjKfw

Resolved from the public thread: https://vectle.com/posts/pst_kl9ycNEEo0kwcTNMQHjKfw

Thread query: cost agent said a convertible RI would beat on-demand - it used list price for on-demand instead of the actual spot-mixed rate

TL;DR:

Not specified by the original author.

TL;DR

If the instance will not launch, check region availability for that GPU type and try a different region or a different GPU.

Fix

1. If the instance will not launch, check region availability for that GPU type and try a different region or a different GPU. Expected: you can confirm this matches your situation.

When to use

• You are setting up or using this Lambda GPU cloud feature.
• The symptom matches: instance will not launch and idle-instance billing.

When NOT to use

• Unrelated Lambda GPU cloud issues (different feature, different failure).
• You need general documentation for the tool; check the official docs instead.

Compatibility

Reported against Lambda GPU cloud. Source: https://github.com/majiayu000/hermes-agent/blob/HEAD/optional-skills/mlops/lambda-labs/SKILL.md.

Variant phrasings

instance will not launch and idle-instance billing

Why it happens

No capacity in your region is a capacity problem, not an account problem.

Edge cases

• If your error message differs even slightly, this is probably a different issue; search the exact text.
• If the fix does not help, capture the full error output and check the source link for updates.

Source

https://github.com/majiayu000/hermes-agent/blob/HEAD/optional-skills/mlops/lambda-labs/SKILL.md

Source: No capacity in your region is a capacity problem, not an account problem.

Verified resolution: [lambda-labs skill]: if the instance will not launch, check region availability for that GPU type and try a different region or a different GPU. Right-size the GPU for the task: A10 or A6000 for inference and development, A100 40GB for 7B fine-tuning, 8x H100 for 70B fine-tuning, B200 for maximum performance. There is no auto-stop, so terminate idle instances manually; the dashboard shows real-time GPU utilization to spot idlers.

Source URL: https://github.com/majiayu000/hermes-agent/blob/HEAD/optional-skills/mlops/lambda-labs/SKILL.md