SaaS gross margin benchmarks at seed in 2026
Pure SaaS still clears 80%, but AI-inference businesses now ship at 50-65%. Here are the 2026 margin bands and the numbers VCs underwrite at seed.
SaaS gross margin benchmarks at seed in 2026
SaaS gross margin benchmarks at seed in 2026 are no longer the textbook 80%. Pure subscription SaaS still ships at 75-85%, AI-inference businesses cluster at 50-65%, and services-heavy products land at 40-60%. VCs now underwrite the margin band that matches your stack, not a one-size rule. Use the archetype, not the average.
The 80% rule is breaking. For a decade, the SaaS investor reflex was simple: anything under 75% gross margin is a services business in disguise. That heuristic was built for hosted software with negligible compute COGS, and it does not survive AI.
The 2026 question changed. A vector-search startup serving GPT-class queries pays real money per call. A workflow SaaS with a small inference loop pays cents. A managed-services-with-software wrapper pays humans. The seed VC question stopped being "are you above 80%?" and became "is your margin defensible at your stack, and what's the ratchet?"
The 2026 SaaS gross margin bands at seed
The cleanest read is by product archetype. The bands below are what credible seed-stage businesses report on Carta-tracked rounds, and what operator-led analyses cite as honest 2026 ranges.
| Archetype | Gross margin band | What sits in COGS |
|---|---|---|
| Pure subscription SaaS (no LLM, no GPU) | 75-85% | Hosting, third-party APIs, payment processing, support |
| AI-inference SaaS (LLM API calls per user) | 50-65% | Model API spend, vector store, hosting, support |
| AI-inference SaaS with own model (GPU rental) | 40-60% | GPU compute, model serving infra, hosting |
| Services-heavy / managed AI | 35-55% | Implementation labor, model spend, dedicated SRE |
Two anchors set the floor. First, Sequoia's 2024 AI infrastructure analysis models the end-user of GPU compute at a 50% gross margin assumption, which is now the unofficial floor most AI seed VCs tolerate before asking hard questions. Second, inference unit costs dropped roughly 1,000x in three years per a16z, going from around $60 per million tokens in 2021 to $0.06 for comparable quality by 2024. The bands above are a moving target, biased upward over time.
What counts as COGS for SaaS in 2026
The COGS line is where most seed decks lose credibility. Treat any cost that scales directly with revenue as COGS. Everything else lives in OpEx.
- Always COGS: hosting (AWS, GCP, Vercel), CDN, model API spend (OpenAI, Anthropic), vector DB, transactional email, payment processing fees, customer-facing support headcount.
- Usually COGS: dedicated SRE time on production, third-party data feeds you resell, fine-tuning compute if it ships with the product.
- Never COGS: sales reps, marketing tools, R&D salaries, founder compensation, office, general G&A.
The single most common founder mistake is parking OpenAI bills under "R&D" because the engineering team manages the integration. If usage scales with active users, it is COGS. Putting it elsewhere inflates your gross margin and an experienced seed VC will catch it in diligence.
ā Good: "COGS includes hosting, OpenAI API, Pinecone, Stripe fees, and one solutions engineer. 2026 gross margin: 58%." Specific, defensible, no surprises.
ā Bad: "Gross margin: 82%. (OpenAI costs sit in R&D for now.)" Reads as either inexperience or a massage. Both lose the round.
Why AI startup margins are structurally lower
AI products burn variable compute. A user query that costs $0.04 in inference is permanent margin drag until you renegotiate the model contract, switch to a cheaper provider, or shift to a smaller in-house model. Even with the 1,000x cost compression a16z documented, absolute spend stays material because product complexity (longer contexts, agentic chains, multi-step reasoning) absorbs the per-token savings.
Two structural pressures keep AI margins compressed:
- Limited pricing power. Sequoia notes AI infrastructure faces competitive pressure that constrains how much startups can charge above cost. Thin wrappers do not control the floor.
- GPU economics flow through. Sequoia's same analysis shows GPUs are roughly half of total data-center cost, setting the upstream price your provider passes through.
A practical implication: if you ship an AI product and your deck claims 80% gross margin, expect the lead VC to ask for a token-level breakdown. The credible answer is the 50-65% band plus a 12-month plan to lift it. For the pricing side of that plan, see how to price an AI product with token costs and margins in mind.
What margin VCs underwrite at seed in 2026
The benchmark has bifurcated. For pure SaaS, the bar is unchanged: 75%+ or your deck gets questions. For AI-inference SaaS, the modal seed VC underwrites 50-65% today, betting that future inference deflation plus pricing leverage gets you above 70% by Series B.
Selectivity is the context that matters. Carta reported the number of new seed investments fell 22% in Q3 2024 versus the prior quarter, even as median pre-money seed valuations climbed to $14.9 million. Fewer rounds, higher prices: VCs pay up only when the unit economics story is honest and the path to expansion is concrete.
The deck slide that lands shows three things: your current margin band, the levers (model switch, prompt compression, caching, tier-gating high-cost features), and one ratchet event that lifts margin by 10 points within 12 months. That is the story 2026 seed checks underwrite. Margin sits inside a broader pattern of AI startup metrics VCs want at seed; pair it with the rest of the package.
FAQ
What is a good gross margin for a seed-stage SaaS startup in 2026? Pure subscription SaaS should ship at 75-85% gross margin. AI-inference SaaS now clears at 50-65% with a credible deflation path. Anything below 40% gets read as a services business, regardless of how you label it.
How much does AI inference cost and how does it affect SaaS gross margin? Inference prices dropped roughly 1,000x in three years per a16z, from about $60 per million tokens in 2021 to $0.06 for comparable quality in 2024. Absolute spend still matters because longer contexts and agentic workflows absorb the savings, which is why AI-first SaaS gross margins cluster 20-30 points below pure SaaS.
What should I include in SaaS COGS for investor decks? Include every cost that scales with active users: hosting, model API spend, vector stores, payment fees, customer-facing support, and dedicated production SRE. Exclude R&D salaries, sales, marketing, and G&A. Parking OpenAI bills outside COGS is the most common red flag VCs catch in diligence.
What gross margin do VCs expect from AI-first startups at seed? Most seed VCs underwrite 50-65% gross margin for AI-inference businesses in 2026, with the expectation that 70%+ is reachable by Series B through model switches, caching, and pricing changes. Sequoia's GPU economics analysis uses a 50% assumption as the working floor.
How do you price SaaS when inference costs are high? Move to usage-based or outcome-based pricing so revenue scales with the cost driver. Tier the highest-cost features (long context, premium models, agent loops) behind higher plans. Add caching and prompt compression on the highest-volume queries to lift margin 5-15 points without changing the customer experience.
Related on the hub
- Traction metrics for VCs in 2026: what IC memos screen for ā Related traction metrics guide.
- Seed round valuation 2026: the benchmark report ā Related fundraising basics guide.
- The H1 2026 State of Seed Fundraising Report ā Related fundraising basics guide.
Run this playbook inside Causo.
Match to the best-fit partner at 1,000+ funds, draft a hyper-specific email, and send from your email ā in one place.