跳到主要内容

07.10 AI Hardware Cloud and Token Costs Enter the P&L ​

Governance metadata | Owner: Head of Finance/Legal | Last reviewed: 2026-10-22 | Next review: 2027-01-31

Inference cost is the second BOM. For the awareness side, see 08.09 / 08.10; the line items are merged into the main P&L.

📌 Statement on the separation of bases for the main case: the inference Token estimates in this document apply specifically to AI hardware with cloud large-model multimodal interaction (such as AI voice recorders, AI companion robots, and embodied hardware). The knowledge base main case, the AstroNext USD 399 smart astronomical telescope, uses an on-device embedded star-map matching algorithm and NPU local inference, with no high-frequency cloud large-model call cost, so no separate cloud Token line item is itemized in the 07.06 Single-SKU Full-Lifecycle P&L Financial Model; the two bases are fully separated and do not conflict.


1. Line Items and Formulas ​

Line itemFormulaNote
Monthly calls per unitActive users × calls per userMeasured pilot data takes priority
Cost per callToken/inference × unit priceEmbedding may be shown as a separate line
Monthly cloud cost per unitMonthly calls × cost per callFor edge inference, book electricity / chip amortization
Subscription offsetSubscription revenue − cloud − channel feesDo not look at subscription gross margin alone
AI-adjusted net margin(EBIT − annualized cloud cost) / revenueFeeds into GR1/GR4

Default redline: free-tier annualized cloud cost > 20% of hardware gross margin → change the plan / reduce frequency / move to the edge; sustained >50% → the model is unvalidated, and GR1 must not be passed on the strength of "hardware sold at a small loss to buy subscriptions" alone (unless finance grants a written waiver).

Worked Example Template ​

ItemValue
Monthly active users____
Calls per user per month____
Cost per call, USD____
Monthly cloud cost per unit=____
Annualized=×12
Compared with per-unit gross marginCompare against 07.06

2. Pricing Structure Selection ​

StructureBest fitRisk
Basic AI bundled into the hardwareImpulse / low-to-mid ticketCost explodes as active users grow
Subscription unlocks premium featuresService / wearablesChurn and experience
Pay-per-use Token packsProfessional toolsPayment psychology barrier

3. Cost-Reduction Levers ​

  1. Cache high-frequency queries
  2. Short context / structured output
  3. Small local model + cloud tiering
  4. Rate limiting and fair use
  5. Back-fill the P&L with measured pilot data

4. Cross-Check with the Chapter 08 Projections ​

  • Wearables / robot vacuum subscriptions: cloud cost must be deducted before LTV is discussed
  • Over-promising "unlimited AI" → expectation management per 08.09

08.09 / 08.10 / 07.06 / 07.07


7. Pilot Cost Dashboard Fields ​

FieldSource
Daily Token / callAPI billing
Monthly cost per unitAggregation
Free-tier shareProduct logs
Anomalous spikesAlerts

Spike > 3× baseline → check whether it is being abused.

8. Subscription LTV ​

If the Chapter 08 projections quote an LTV, they must include:
Subscription gross margin = Subscription revenue − cloud − channel − incremental customer service
It is prohibited to treat "subscription list price × renewal rate" as gross margin.

9. GR Statement Template ​

"AI cost is already in the 07.10/07.06 scenarios; free-tier annualized cloud cost / hardware gross margin = __% (≤20% is the default watch line)"


10. Scenario Playbook: The Free Tier Runs Out of Control ​

Symptom: 30 days after launch, free-tier calls surge and the monthly cloud cost > that month's gross margin.
Actions (default):

  1. T+0: rate-limit the free tier (daily call cap)
  2. T+1: cache high-frequency questions + short answers
  3. T+7: migrate premium features to subscription; evaluate edge models
  4. T+14: update the 07.06 scenarios; note it in the GR4 appendix

11. Number Alignment Rules with the Chapter 08 Projections ​

Whenever an LTV / gross margin figure involves "subscription" and the product has AI:
it must carry a "cloud assumption" footnote; otherwise it counts as an unaudited number and must not enter the fundraising package.

12. Field-Level Checklist (Finance) ​

FieldPresent?Note
Monthly calls per unit□
Cost per call□
Annualized cloud cost□
% of hardware gross margin□Redline 20%
Does the subscription deduct cloud cost□

13. Suppliers and SLA (Cloud) ​

ItemDefault requirement
Availability99.5%+ annualized (set by the business)
Price changes30 days' written notice
DataNot used for training (contractual)
Single point of failureEvaluate a dual-API or degraded-mode path

14. Cost Attribution: R&D vs. Operations ​

CostBooking treatment
First-year ramp overrunMay go into a dedicated marketing/growth line, but must be disclosed
Steady-state cloudProduct COGS or subscription COGS
One-off fine-tuningR&D expense or capitalization policy (set by finance)

15. Alignment with the 08.08 / 08.04 Projections ​

Whenever an "annual subscription fee" appears:
add the note cloud cost deducted or cloud cost not deducted — gross margin overstated.
Fundraising materials require the former by default.

MetricYellowRed
Monthly cloud cost per unit> 120% of budget>150% or > 20% of gross margin
Free-tier share>70%>85% with no conversion
API error rate>0.5%>2%

17. Cost-Reduction Roadmap Template ​

PhaseMeansTarget cost reduction
0–30 daysRate limiting + caching20–40%
1–3 monthsShort context + tiered modelsanother 30%
3–6 monthsMove high-frequency calls to the edgeanother 20–50%

19. Edge vs. Cloud Routing Table ​

TaskDefault placementReason
Voice wake-upEdgeLow latency + cloud cost savings
Complex Q&ACloudQuality
Image recognitionDepends on computeCost-sensitive
Report generationCloud + cachingFrequency is controllable

20. Subscription Redesign Communication Template (Product / Customer Service) ​

  • Explain the "fair use" policy and the reason for rate limiting
  • Provide the benefits table for the upgraded tier
  • One-off compensation options for existing users (set by operations)

21. Example of the Three Finance Scenarios Stacked ​

Tariff scenarioHardware EBITAnnual cloud costAI-adjusted
A 7.5%73.692449.69→12.5% illustrative
B 25%52.692428.69→7.2% illustrative
C 35%40.692416.69→4.2% illustrative

(Indicative; it must be recalculated with actual calls. When B/C can no longer cover the 10% post-reserve threshold, AI cost becomes the veto factor.)

22. With GR2 Tooling ​

Add to the AI hardware GR2 review: "is the cloud cost ceiling written into the product requirements".

24. Complete Worked Example (Recomputable) ​

Assumptions (fill in the numbers): 20,000 monthly active users; 40 calls per user per month; cost per call USD 0.002; per-unit hardware gross margin USD 80 (illustrative, on the 07.06 basis).

Monthly cloud cost per unit = 40 × 0.002 = USD 0.80
Annualized = 9.60
% of hardware gross margin = 9.60/80 = 12% → inside the watch zone (<20%)
If calls rise to 120: annualized 28.8 → 36% → triggers a plan change / rate limiting

Sensitivity:

Monthly callsAnnual cloud% of the 80 gross margin
204.86%
409.612%
8019.224% red
12028.836% red

25. GR2/GR4 Checkboxes (AI Products) ​

GR2 ​

  • [ ] Free-tier annualized cloud cost ≤ 20% of hardware gross margin, or a waiver has been approved

  • [ ] If subscription unlocks premium features: cloud cost has been deducted from subscription gross margin

  • [ ] The API supplier and its price-increase clauses have been signed

GR4 (T+14) ​

  • [ ] Actual calls vs. model error <30%

  • [ ] The 20% redline has not been touched

  • [ ] The rate-limiting / caching switches can be executed remotely

26. Prohibited Statements ​

  • "Subscription revenue is high, so cloud cost can be ignored"
  • "Burn scale first and do the math later" (unless there is a written stage budget from finance)
  • "Consistent with the Chapter 08 projected LTV" without a cloud assumption attached

27. Checklist (Final) ​

  • [ ] The worked example has been substituted with this product

  • [ ] The GR2/GR4 AI items are in the review package

  • [ ] Product and sales have been trained on the prohibited statements

huajun.wang Smart Hardware GTM Playbook · Methodology