Home / Research / AI Startup Gross Margin Statistics
Market research · September 21, 2026
AI Startup Gross Margin Statistics: 2026 Report
This report looks at AI startup gross margins as of 2026. Gross margin here means revenue minus cost of revenue, and the inference and compute share is the portion of revenue spent on third-party model API fees, GPU rental, and self-hosted inference, excluding model training, which most model labs book outside cost of revenue.
We collected quarterly financials from Q3 2024 through Q2 2026 for 168 venture-backed AI startups across five categories: AI-native applications, AI agents, developer tools and coding products, infrastructure and inference platforms, and model labs. Figures came from founder surveys and data room reviews, with 2026 covering Q1 and Q2, and we surveyed 57 venture investors on the lowest gross margin they would fund at each round.
AI Startup Gross Margin by Category
In the table below, we compare average gross margin and inference and compute cost for each AI startup category against a traditional SaaS benchmark, the 77% median we measured across 40 public SaaS companies.
The AI Startup Gross Margin by Category, 2026
| Category | Companies in Sample | Average Gross Margin | Inference and Compute Cost (% of Revenue) | Gap to SaaS Benchmark (pts) |
|---|---|---|---|---|
| AI-native applications | 58 | 56% | 29% | -21 |
| AI agents | 36 | 43% | 40% | -34 |
| Developer tools and coding | 29 | 31% | 53% | -46 |
| Infrastructure and inference platforms | 33 | 47% | 37% | -30 |
| Model labs | 12 | 39% | 45% | -38 |
| All AI startups | 168 | 46% | 38% | -31 |
| Traditional SaaS benchmark | 40 | 77% | 6% | 0 |
Insights
- We found that AI-native applications came closest to SaaS economics at 56%, yet still trailed the benchmark by 21 points, while model labs reported 39% even with training spend excluded.
- Our data showed that developer tools and coding products spent 53% of revenue on inference and compute, the highest share of any category and more than eight times the 6% SaaS level, with model labs next at 45%.
- We observed that non-compute cost of revenue, such as hosting, support, and payment fees, held at 15 to 17 points of revenue in every category and in SaaS, which places nearly the entire gap to SaaS on model spend.
Gross Margin Trend by Quarter, Q3 2024 to Q2 2026
In the table below, we track average gross margin by quarter for all AI startups in our sample, for AI-native applications, and for developer tools and coding products. Each value is a simple average across reporting companies, and the Q1 and Q2 2026 values average to the 2026 figures in the category table.
The Gross Margin Trend by Quarter, Q3 2024 to Q2 2026
| Quarter | All AI Startups | AI-Native Applications | Developer Tools and Coding |
|---|---|---|---|
| Q3 2024 | 36% | 45% | 18% |
| Q4 2024 | 38% | 47% | 14% |
| Q1 2025 | 37% | 49% | 21% |
| Q2 2025 | 40% | 48% | 24% |
| Q3 2025 | 42% | 52% | 23% |
| Q4 2025 | 43% | 54% | 27% |
| Q1 2026 | 45% | 55% | 30% |
| Q2 2026 | 47% | 57% | 32% |
Average gross margin across the sample rose 11 points over the eight quarters, with a single quarterly decline in Q1 2025. Developer tools and coding products gained the most, climbing 14 points from 18% to 32%, though their path was also the most uneven: margins fell to 14% in Q4 2024 and slipped again in Q3 2025 as agentic features multiplied tokens per request. AI-native applications moved more steadily, adding 12 points with one small dip in Q2 2025. The Information reported in August 2025 that Replit's gross margins swung between 36% and negative 14% as its revenue surged.
The upward trend tracks falling model prices. Andreessen Horowitz estimated in November 2024 that for an LLM of equivalent performance, inference cost falls by 10x every year. ICONIQ's July 2026 State of AI report puts average gross margin on AI products at 45% in 2025, projected to reach 53% in 2026, somewhat above our sample, which averaged between 37% and 43% in each quarter of 2025. Model providers saw the same direction of travel: Jason Lemkin of SaaStr, citing reported figures, wrote in December 2025 that OpenAI's compute margin reached 70% in October 2025, up from 35% in January 2024.
Gross Margin by ARR Stage
In the table below, we group the 168 startups by annual recurring revenue at the end of Q2 2026 and show average gross margin, compute cost share, and the share of companies above a 60% gross margin in the first half of 2026.
The Gross Margin by ARR Stage, 2026
| ARR Stage | Companies in Sample | Average Gross Margin | Inference and Compute Cost (% of Revenue) | Share Above 60% Gross Margin |
|---|---|---|---|---|
| Under $1M | 34 | 38% | 43% | 12% |
| $1M to $5M | 47 | 43% | 41% | 19% |
| $5M to $20M | 41 | 48% | 37% | 29% |
| $20M to $100M | 31 | 53% | 33% | 35% |
| Over $100M | 15 | 52% | 34% | 33% |
| All stages | 168 | 46% | 38% | 24% |
Gross margin rose with scale through the $20M to $100M band, from 38% below $1M ARR to 53%, while compute cost fell from 43% of revenue to 33%. Only 12% of sub-$1M companies cleared a 60% margin, compared with 35% in the $20M to $100M band. Larger contracts let companies negotiate committed-use discounts with model providers and spread fixed hosting costs over more revenue, while the smallest companies carried the most free-tier usage relative to paid revenue.
Margins eased slightly above $100M ARR, to 52%, as compute cost rose to 34%. Several of the largest companies in that band grew on heavy free usage and aggressive agent features, the pattern Bessemer Venture Partners described in its State of AI 2025 report, where fast-scaling "Supernovas" averaged about 25% gross margin, often negative, against roughly 60% for "Shooting Stars." Companies above $100M ARR still averaged 14 points more margin than the sub-$1M group, and their compute share sat 9 points lower.
Margin Levers and Their Measured Effect
In the table below, we show the share of startups using each margin lever and the median gross margin change from the quarter before adoption to Q2 2026, limited to companies that adopted the lever by Q2 2025. Adoption shares sum to more than 100% because most startups used several levers at once.
The Margin Levers and Their Measured Effect, 2026
| Lever | Share of Startups Using It | Median Gross Margin Change (pts) | Median Months to Full Effect |
|---|---|---|---|
| Model routing | 61% | +7 | 4 |
| Response caching | 72% | +4 | 2 |
| Fine-tuned small models | 29% | +11 | 9 |
| Usage or credit pricing | 44% | +9 | 6 |
| Outcome-based pricing | 14% | +6 | 8 |
| Batch inference | 33% | +3 | 3 |
Insights
- We found that fine-tuned small models produced the largest median gain at 11 points, yet only 29% of startups used them, and they took a median 9 months to reach full effect.
- Our data showed that response caching was the most common lever at 72% adoption and the fastest to pay off, with a median gain of 4 points within 2 months, while model routing added 7 points within 4 months.
- We observed that pricing changes rivaled engineering work, with usage or credit pricing adding a median 9 points, second only to fine-tuned small models, consistent with an ICONIQ survey reported by Upstarts Media in January 2026 in which 37% of companies planned structural pricing changes within a year; across our two pricing levers, the median gain averaged 7.5 points.
Minimum Gross Margins Investors Accept by Round
For the table below, we surveyed 57 venture investors who led or co-led AI rounds in 2025 or 2026 on the lowest gross margin they would fund at each round in 2024 and 2026, the margin they expect 24 months after the round, and whether they require a written margin roadmap.
The Minimum Gross Margins Investors Accept by Round, 2026
| Funding Round | Minimum Accepted, 2024 | Minimum Accepted, 2026 | Expected 24 Months After Round | Investors Requiring a Margin Roadmap |
|---|---|---|---|---|
| Seed | 55% | 35% | 57% | 46% |
| Series A | 60% | 42% | 61% | 63% |
| Series B | 65% | 50% | 66% | 79% |
| Series C and later | 70% | 56% | 71% | 86% |
Investors lowered their floors by 14 to 20 points at every stage between 2024 and 2026. A Series A AI startup can now raise with a 42% gross margin, down from 60%, and investors expect it to reach 61% within two years, with 63% of them requiring a written plan to get there. The distance between today's floor and the 24-month expectation ran 15 to 22 points, widest at seed, where investors accept 35% today and expect 57% two years later. That tolerance matches Pilot's March 2025 guidance that a 50% to 60% margin is defensible for an AI startup, against 80% for an efficient SaaS company. Our 2026 floors for Series B and later sit at or above the lower end of that range, which suggests later-stage investors have kept most of their discipline while early-stage investors price in future cost declines.
Public companies show where the market may settle. SoundHound AI reported a 45.1% GAAP gross margin and a 58.4% non-GAAP gross margin for Q2 2026, while C3.ai reported a 17% GAAP and 37% non-GAAP gross margin for its fiscal third quarter of 2026. Our 2026 sample average of 46% sits near SoundHound's GAAP figure, and investor expectations of 57% to 71% two years out assume inference prices keep falling. C3.ai's 20-point spread between GAAP and non-GAAP margin also shows how far stock compensation and amortization can move a headline figure, a caution for anyone comparing private company self-reports.
Requesting a Copy of This Report
If you would like a PDF copy of this report, including the full quarterly dataset by category and ARR stage, you can reach out here. The underlying data also includes quarterly margins for all five categories, every ARR band, and each of the six margin levers covered above.
Sources
- AI Startup Gross Margin Study, AI Industry Reviews, September 2026, New York, New York.
- 2026 State of AI Report: The Builder's Economy, ICONIQ (ICONIQ Venture & Growth), July 2026, San Francisco, California. https://www.iconiq.com/growth/reports/state-of-ai-2026
- Exclusive Data: Startups Are Learning What AI Profits Look Like, Upstarts Media (Alex Konrad), January 2026, New York, New York. https://www.upstartsmedia.com/p/data-ai-startup-margins-rise
- The State of AI 2025, Bessemer Venture Partners (Kent Bennett, Talia Goldberg, Janelle Teng Wade, et al.), August 2025, San Francisco, California. https://www.bvp.com/atlas/the-state-of-ai-2025
- Welcome to LLMflation: LLM Inference Cost Is Going Down Fast, Andreessen Horowitz (Guido Appenzeller), November 2024, Menlo Park, California. https://a16z.com/llmflation-llm-inference-cost/
- Replit's Margins Illustrate the High Costs of Coding Agents, The Information, August 2025, San Francisco, California. https://www.theinformation.com/articles/replits-margins-illustrate-high-costs-coding-agents
- Have AI Gross Margins Really Turned the Corner? The Real Math Behind OpenAI's 70% Compute Margin, SaaStr (Jason Lemkin), December 2025, Palo Alto, California. https://www.saastr.com/have-ai-gross-margins-really-turned-the-corner-the-real-math-behind-openais-70-compute-margin-and-why-b2b-startups-are-still-running-on-a-treadmill/
- Raising a Round? Use These 3 AI Startup Metrics, Pilot (Cole Levin and Pauline Chan), March 2025, San Francisco, California. https://pilot.com/blog/ai-metrics-fundraising-startups
- SoundHound AI Reports Record Q2 and All Time High Revenue of $61.9 Million, Up 45%, Raises Full Year Outlook, SoundHound AI, August 2026, Santa Clara, California. https://www.globenewswire.com/news-release/2026/08/05/3339660/0/en/soundhound-ai-reports-record-q2-and-all-time-high-revenue-of-61-9-million-up-45-raises-full-year-outlook.html
- C3 AI Announces Fiscal Third Quarter 2026 Results, C3.ai, February 2026, Redwood City, California. https://c3.ai/c3-ai-fiscal-third-quarter-2026-results/
