AI wrapper economics

An AI app that calls a frontier model API and marks up tokens is economically a reseller until it adds durable value beyond the raw model. "Our moat is the model" is usually false unless you train frontier systems few others can afford. For most application companies, assume the model layer commoditizes.

The risks

• Margin compression. Providers cut prices; competitors race to the bottom; your markup shrinks unless value-add grows.
• Commoditization. If the experience is a thin chat UI, the provider or an open-source stack can replace you.
• COGS surprise. Free tiers and generous limits light money on fire when a prompt goes viral.
• Concentration. One upstream outage or ToS change is an existential product event.

Durable layers: proprietary data and evals, workflow integration into a job-to-be-done, multi-model routing, memory and retrieval tuned to a domain, compliance, latency SLOs, and distribution into an existing customer base.

Price from customer value and alternative cost, then check that COGS leave acceptable gross margin. Cost-plus ("tokens × 1.5") is easy to undercut.

Walking through the arithmetic

Assumptions (illustrative, not a vendor quote):

• You charge $20 / 1M output tokens.
• Upstream model API $8 / 1M.
• Other variable COGS $2 / 1M.

Revenue per 1M tokens     = $20
Upstream + other COGS     = $8 + $2 = $10
Contribution              = $10
Gross margin              = $10 / $20 = 50%

Heavy user on a $20/month unlimited plan who consumes 5M tokens:

Upstream = 5 × $8 = $40
Other    = 5 × $2 = $10
COGS     = $50 against $20 revenue → negative 150% margin

Credit pack: $20 includes 1M tokens; overage at $25 / 1M. Power user uses 4M:

Revenue = $20 + 3 × $25 = $95
COGS    = 4 × $10 = $40
Margin  = ($95 − $40) / $95 ≈ 58%

Monthly book: 10,000 paying users, ARPU $24, COGS 45% of revenue.

Revenue      = 10,000 × $24 = $240,000
COGS         = 0.45 × $240,000 = $108,000
Gross profit = $132,000

If you spend $90,000 on S&M and $100,000 on R&D/G&A, you still burn ~$58k that month. Usage gross margin does not excuse an oversized opex base — see burn and runway.

Provider price cut: upstream $8 → $4. Keep list at $20 and margin becomes 70% — until competitors force $12 list. Plan for pass-through races.

Common mistakes

• Negative-margin whales on uncapped plans.
• AI wrappers with no eval or data moat — price wars with the provider's own app.
• Free-tier abuse with no caps or alerts. A viral prompt is a COGS spiral.

Check yourself

$20 unlimited, 5M tokens, $10 COGS per 1M. What is margin on that user?

Negative 150%. $50 COGS against $20 revenue.

What stays after the model commoditizes?

Data, workflow embed, routing, evals/trust and distribution — not the markup on tokens.

Key takeaways

• A token reseller is not a platform until value sits above the model.
• Unlimited plus variable COGS loses money on power users.
• Credit packs and overage protect margin. Model the whale first.
• Provider price cuts are temporary margin until a pass-through race.

Further learning

The New Business of AI (a16z)At a technical level, artificial intelligence seems to be the future of software. AI is showing remarkable progress on a range of difficult computer science problems, and the job of software developers – who now…Andreessen Horowitz

Next, commits and credits.