At 100,000 daily active users, cloud AI costs your enterprise $1.095 million per year for a standard text feature. At 500,000 DAU, that is $5.475 million per year. At 1 million DAU, $10.95 million per year. The on-device build premium is $60,000-$80,000. The break-even is measured in weeks, not years.
Key findings
At 1M DAU with 5 text AI queries per user per day, cloud AI costs $1.095 million per month. On-device build premium is $80,000 one-time. Break-even is less than 3 days of cloud spend.
Voice AI at 500K DAU (5 minutes of dictation per user per day) costs $547,500-$2.19 million per month via cloud APIs. On-device Whisper costs $0 per minute after the $30,000-$60,000 build investment.
Image AI has the highest per-interaction cloud cost at $0.04-$0.12 per image. At 100K DAU generating 1 image per day, cloud image generation costs $1.2 million to $3.6 million per year.
Wednesday built all three AI types — text (llama.cpp), voice (Whisper), and image (MNN/QNN/Core ML) — in Off Grid at production scale, validating the on-device cost model with real-world usage data.
Why scale changes the math
Cloud AI cost scales linearly. Double your users, double your monthly AI bill. At low scale, this is manageable. At enterprise scale, it is the largest line item in your infrastructure budget.
On-device AI cost does not scale. The build investment is fixed. Whether you have 10,000 users or 10 million, the per-query cost is zero. The model runs on the user's device. Your infrastructure cost does not change.
The break-even point — where cumulative cloud costs equal the on-device build premium — shifts dramatically with scale. At 5,000 DAU, break-even might be 18 months out. At 500,000 DAU, break-even is measured in weeks. At 1 million DAU, it is measured in days.
This analysis covers three scale points — 100K, 500K, and 1M daily active users — and three AI feature types: text, voice, and image. Each has a different cost profile and a different break-even calculation.
Text AI cost model
Text AI pricing is based on tokens — roughly 750 words per 1,000 tokens. Cloud AI at enterprise pricing: $0.003 per query (assuming an average query involves approximately 500 input tokens and 200 output tokens at GPT-4o class quality).
At 5 queries per user per day — a conservative estimate for a productivity or documentation AI feature:
| DAU | Cloud monthly | Cloud annual | On-device build | Break-even |
|---|---|---|---|---|
| 100K | $45,000 | $540,000 | $70,000 | 1.6 months |
| 500K | $225,000 | $2,700,000 | $70,000 | 10 days |
| 1M | $450,000 | $5,400,000 | $80,000 | 5 days |
At 10 queries per user per day — an active use case:
| DAU | Cloud monthly | Cloud annual | On-device build | Break-even |
|---|---|---|---|---|
| 100K | $90,000 | $1,080,000 | $70,000 | 24 days |
| 500K | $450,000 | $5,400,000 | $70,000 | 5 days |
| 1M | $900,000 | $10,800,000 | $80,000 | 3 days |
Voice AI cost model
Voice AI is priced per audio duration. Cloud speech APIs: $0.004-$0.016 per 15 seconds. For a field service or clinical documentation use case, assume 5 minutes of dictation per active user per day.
5 minutes = 20 fifteen-second segments = $0.08-$0.32 per user per day.
| DAU | Cloud monthly (low) | Cloud monthly (high) | On-device build | Break-even (midpoint) |
|---|---|---|---|---|
| 100K | $240,000 | $960,000 | $45,000 | Under 1 week |
| 500K | $1,200,000 | $4,800,000 | $45,000 | Under 2 days |
| 1M | $2,400,000 | $9,600,000 | $60,000 | Under 1 day |
Voice AI has the fastest break-even of the three types because cloud voice APIs are expensive relative to on-device alternatives. Wednesday built on-device Whisper in Off Grid. The build cost is real and bounded.
Image AI cost model
Image generation is priced per image. Cloud image APIs: $0.04-$0.12 per image depending on resolution and model. For an enterprise app where users generate one image per day on average — a realistic estimate for creative, inspection, or documentation use cases:
| DAU | Cloud monthly (low) | Cloud monthly (high) | On-device build | Break-even (midpoint) |
|---|---|---|---|---|
| 100K | $120,000 | $360,000 | $120,000 | 30 days |
| 500K | $600,000 | $1,800,000 | $140,000 | Under 1 week |
| 1M | $1,200,000 | $3,600,000 | $160,000 | Under 4 days |
Image AI build cost is highest ($80,000-$160,000) because shipping on-device image generation requires three separate hardware backend implementations: MNN for ARM64 Android, QNN for Snapdragon 8 Gen 1+ NPU, and Core ML for iOS. Wednesday built all three in Off Grid. That is the reference implementation.
100K DAU: full analysis
At 100,000 daily active users, cloud AI is already expensive but the break-even is still measured in months for text and weeks for voice and image.
For a text AI feature at 10 queries per user per day:
- Cloud AI Year 1 cost: $1,080,000
- Cloud AI Year 2 cost: more if user base grows; same if stable
- On-device build: $70,000 one-time
- On-device Year 1 total: $70,000 + $15,000 model maintenance = $85,000
- Year 1 savings from on-device: $995,000
For a voice AI feature at 5 minutes of dictation per user per day (at midpoint cloud pricing):
- Cloud AI Year 1 cost: $7,200,000
- On-device build: $45,000 one-time
- Year 1 savings from on-device: $7,155,000
At 100K DAU, the on-device build premium is recouped within the first month of operation for both text and voice features. The 3-year savings are in the millions regardless of feature type.
500K DAU: full analysis
At 500,000 DAU, cloud AI costs are not manageable. They are the dominant infrastructure cost of the product.
For a text AI feature at 5 queries per user per day:
- Cloud AI Year 1: $2,700,000
- Cloud AI 3-year total (assuming 10% annual user growth): $8,910,000
- On-device 3-year total: $70,000 + $45,000 maintenance = $115,000
- 3-year savings: $8,795,000
For a voice AI feature at 5 minutes per user per day:
- Cloud AI Year 1: $7,200,000 (at midpoint pricing)
- On-device 3-year total: $45,000 + $30,000 maintenance = $75,000
- 3-year savings: over $22,000,000
At 500K DAU, the question is not "can we afford on-device AI?" It is "how quickly can we get it shipped?"
1M DAU: full analysis
At 1 million daily active users, cloud AI inference for a standard text feature costs more in one month than the entire 3-year on-device total.
For a combined text, voice, and image AI suite at 1M DAU:
- Cloud AI monthly: $900,000 (text at 10 queries/day) + $6,000,000 (voice at 5 min/day) + $2,400,000 (image at 1/day) = $9,300,000 per month
- On-device build for all three: $80,000 + $60,000 + $160,000 = $300,000
- On-device 3-year total: $300,000 build + $135,000 maintenance = $435,000
- Cloud AI 3-year total: $334,800,000
The 3-year savings from on-device AI at 1M DAU is $334 million. Even accounting for estimation ranges and more conservative usage assumptions, the order of magnitude is conclusive.
Read more case studies at mobile.wednesday.is/work
How Wednesday builds cost models for enterprise boards
The numbers in this article are the starting point for a board-level AI investment conversation, not the ending point. Every enterprise deployment has specific variables: the actual query frequency for the specific use case, the growth trajectory of the user base, the cloud vendor's specific enterprise pricing, and the data sensitivity classification that determines whether compliance costs apply to the cloud option.
Wednesday builds custom cost models for enterprise clients before any scoping or architecture decision. The model captures those specific variables and produces a break-even analysis, a 3-year total cost of ownership comparison, and a sensitivity analysis showing how the numbers change under different growth assumptions.
Off Grid is the reference implementation for all three AI types at production scale. 50,000 users using text, voice, and image AI with zero per-query cloud infrastructure cost. The cost model is not theoretical — it has been validated in production.
For enterprise teams whose CFO needs to approve an AI budget, the cost model comes before the technical specification. The numbers at scale almost always favour on-device. The question is whether you want to build that case before or after you have already committed to a cloud AI architecture.
Originally published at https://mobile.wednesday.is/writing/ai-feature-cost-per-user-scale-on-device-vs-cloud-2026
Top comments (0)