Beta
EconomyReading · ~3 min · 71 words deep

Gross Margin (AI Labs)

AI lab gross margins (50-70%) sit below pure SaaS (80-90%) because inference compute is a variable cost, not amortized software.

TL;DR

AI lab gross margins (50-70%) sit below pure SaaS (80-90%) because inference compute is a variable cost, not amortized software.

Level 1

Gross margin = (revenue - cost of goods sold) / revenue. For a SaaS company, COGS is mostly hosting + support (10-20% of revenue). For an AI lab, COGS includes inference compute (GPUs, electricity, cloud fees), which can be 30-50% of revenue. Result: SaaS hits 80-90% gross margin; AI labs sit at 50-70%. This gap drives different capital dynamics and valuations.

Level 2

Anthropic reported ~60% gross margin in 2024 leaks. OpenAI's is estimated at 40-55% depending on mix (API heavy = better margins; ChatGPT consumer = worse because usage is unlimited on subscription). Training cost is a separate line (R&D) and doesn't hit COGS directly. Efficient architectures (MoE, distillation) and cheaper GPUs (Blackwell, H200, MI300) push COGS down year over year · AI lab gross margins are trending up as compute gets cheaper.

Level 3

Gross margin by product line: API margins 60-75% (high volume, optimizable). Consumer ChatGPT Plus: low-40s% due to heavy-use outliers. Enterprise tier: 70-80% via volume commits + caching. Model distillation is the single biggest margin lever · Claude Haiku serves at ~2× gross margin of Opus on same-type queries. Margins also shift with compute pricing · 2025's H100 glut cut inference costs 40%, bumping gross margins across the board.

The takeaway for you
If you are a
Researcher
  • ·AI lab GM: 50-70% · SaaS GM: 80-90%
  • ·COGS = inference compute + hosting
  • ·MoE, distillation, chip-price cuts all push GM up
If you are a
Builder
  • ·Not directly visible to end users
  • ·Bigger margins mean more room for price cuts
  • ·Trend: margins improving year over year
If you are a
Investor
  • ·GM determines how much valuation multiple a lab earns vs SaaS comps
  • ·Watch the trajectory · improving GM = operating leverage coming
  • ·OpenAI + Anthropic both trending up through 2025-26
If you are a
Curious · Normie
  • ·How much profit AI companies keep after paying for computers to run AI
  • ·Less than normal software because AI needs lots of computers
  • ·Getting better as AI gets more efficient
Gecko's take

Gross margin is the quiet valuation battle. AI labs that crack 70% get the SaaS multiple.

Inference compute is a per-query variable cost. SaaS hosting is relatively fixed. GPU + power per query adds up.