LLM Inference Economics — Full Breakdown + Free Tier + API Business (Sept 2025)

$ per 1M tokens

$4.22–$11.32

Optimized floor → Baseline ceiling (H100; DeepSeek‑V3.1‑class). ^{NVIDIA, NextPlatform}

H100 throughput (t/s)

45 → 90

fp16 bound → fp8 + FlashAttention‑2 + vLLM batching ^{FA‑2, vLLM, bench}

ChatGPT usage anchors

190.6M DAU · 2.5B prompts/day

^{SQ Magazine, TechCrunch}

Free token share (OpenAI)

84.8%

Derived from 1k tokens/prompt and paid usage model (20/80).

A. Cost Model

Per‑GPU monthly cost and $/1M tokens (component breakdown)

Infographic: H100 monthly TCO (mid): $1,031.85 = amort $833.33 + power $50.40 + cooling $15.12 + colocation $133.00 ^{Brightlio, CBRE}. $/M spans baseline $7.45–$11.32 and optimized $4.22–$5.13.

Per‑GPU Monthly Cost (low / mid / high)

Component	Low	Mid	High
Amortization	$694.44	$833.33	$1,111.11
Power	$50.40	$50.40	$50.40
Cooling (+30%)	$15.12	$15.12	$15.12
Colocation	$114.41	$133.00	$152.11
Total	$874.37	$1,031.85	$1,328.74

$ per 1M tokens — ranges

	Min	Mid	Max
Baseline (~45 t/s)	$7.45	$8.79	$11.32
Optimized (~90 t/s)	$4.22	$4.42	$5.13

Per‑million component contributions (mid): Baseline amort $7.10, power $0.43, cooling $0.13, colo $1.13. Optimized amort $3.57, power $0.22, cooling $0.06, colo $0.57.

B. OpenAI Chat P&L

Paid vs free usage — costs and revenue

Infographic: DAU ≈ 190.6M (source), prompts/day ≈ 2.5B (source), 10% paid ⇒ 19.1M paid, 171.5M free. Tokens/prompt = 1,000. Paid usage: 20% heavy (80/day) & 80% avg (5/day).

Token volumes

	Tokens / month
Paid users	11.4T
Free users	63.6T
Total	75.0T

Costs — full range

	Baseline min	Baseline mid	Baseline max	Opt min	Opt mid	Opt max
Paid cost	$85,216,271	$100,564,346	$129,499,006	$48,222,045	$50,584,168	$58,624,550
Free cost	$473,652,242	$558,960,484	$719,786,185	$268,029,562	$281,158,802	$325,849,151
Total cost	$558,868,512	$659,524,830	$849,285,191	$316,251,608	$331,742,970	$384,473,701

Revenue & gross profit

	Value
Revenue ($20 × paid users)	$381,200,000
Gross profit (baseline mid)	$-278,324,830
Gross profit (optimized mid)	$49,457,030
Break‑even sub price (baseline mid)	$34.60/mo
Break‑even sub price (optimized mid)	$17.41/mo

C. Anthropic Chat P&L

Paid vs free usage — scaled costs and revenue

Infographic: Claude MAU mid ≈ 24.4M (StatsUp, SQ). 10% paid ⇒ 2.4M paid, 22.0M free. Free:paid token ratio borrowed from ChatGPT derivation = 5.56:1.

Token volumes (est.)

	Tokens / month
Paid users	1.5T
Free users (est.)	8.2T
Total	9.6T

Costs & revenue (optimized mid)

	Value
Paid cost	$6,488,892
Free cost (est.)	$36,066,803
Total cost	$42,555,696
Revenue ($20 × paid users)	$48,900,000
Gross profit (opt‑mid)	$6,344,304

D. API Business

Users, traffic, and per‑million token margins

OpenAI: "Three million developers build with OpenAI" (official); reported ~2.2B API requests/day (SQ Magazine).
Anthropic: reported ~25B API calls/month (≈ 833,333,333/day) (SQ Magazine). Pricing: official. OpenAI pricing: official.

API “DAU” (active orgs/keys per day) — scenarios

Scenario	Per‑org calls/day	OpenAI API DAU (orgs)	Anthropic API DAU (orgs)
Small apps	1,000	2,200,000	833,333
Mid apps	10,000	220,000	83,333
Large apps	100,000	22,000	8,333

These are org‑level DAU estimates, not end users. A mid‑case of ~10k calls/org/day ⇒ ~220k active orgs/day at OpenAI, consistent with a subset of ~3M total developers grouped into teams/apps.

Per‑million pricing → margins

Blended price per 1M total tokens: P = p_in·(1−s_out) + p_out·s_out for output share s_out. We show 50/50 and output‑heavy 30/70 mixes. Margins subtract our serving COGS $/M (H100 TCO): best‑case (optimized‑min) → worst‑case (baseline‑max).

Model & mix	Blended price / 1M	Margin band ($/M)	Margin band (%)
OpenAI GPT‑5 — 50/50 in:out	$5.62	$1.41 → $-5.70	25% → -101%
OpenAI GPT‑5 — 30/70 in:out	$7.38	$3.16 → $-3.95	43% → -54%
OpenAI GPT‑4.1 — 50/50 in:out	$7.50	$3.28 → $-3.82	44% → -51%
OpenAI GPT‑4.1 — 30/70 in:out	$9.30	$5.08 → $-2.02	55% → -22%
Anthropic Claude Sonnet 4 (≤200k) — 50/50 in:out	$9.00	$4.78 → $-2.32	53% → -26%
Anthropic Claude Sonnet 4 (≤200k) — 30/70 in:out	$11.40	$7.18 → $0.08	63% → 1%

E. Verdict

Can $20 paid subs subsidize free chat usage?

OpenAI (ChatGPT): Gross profit (opt‑mid) = $49,457,030 ⇒ profit after free is included. Break‑even ≈ $17.41/mo.
Anthropic (Claude): Gross profit (opt‑mid) = $6,344,304 ⇒ profit with scaled free costs.
API: Per‑million unit margins are generally positive at list prices vs our cost range, suggesting API can cross‑subsidize chat if needed.

Figures are serving COGS only (GPUs, power, cooling, colocation); exclude staff, R&D, networking/egress, training amortization, idle capacity.

References