Anthropic runs at roughly 40% gross margins on $30 billion in annualized revenue. OpenAI posted 33% gross margins in 2025, burning through $8.4 billion in inference costs that year. GitHub Copilot lost $20 per user per month for two years before restructuring to five pricing tiers. Even Snowflake, at 66.5% gross margins, is considered strong for consumption-based cloud and mediocre by SaaS standards.
Traditional SaaS built an industry on 80-90% gross margins. That buffer funded 40% sales-and-marketing spend, 20% R&D, and the growth-at-all-costs playbook that defined the 2010s. AI companies operate at roughly half that margin, and the playbook does not survive the math.
Ten Points Left Over
Start with the arithmetic nobody puts on a pitch deck. SaaS companies historically spend 40% of revenue on sales and marketing. At 80% gross margins, that leaves 40 points of gross profit after S&M to cover R&D, G&A, and still generate operating income. The model works because the buffer is enormous.
At 50% gross margins, the same 40% of revenue on sales and marketing leaves 10 points. Ten points for engineering, operations, legal, office space, and profit. That isn't a business. That's a countdown.
So AI companies face a binary choice: either generate twice the revenue per sales dollar, or spend half as much on selling. Most are doing neither. They're running an 80%-margin playbook on a 50%-margin product and calling the resulting burn rate "investment in growth."
The Margin Data Converges
Multiple independent sources land in the same range. ICONIQ Growth tracked AI company gross margins improving from 41% in 2024 to 45% in 2025 to 52% in 2026. Bessemer Venture Partners identified two AI startup archetypes: "Supernovas" hitting $40M ARR in year one at roughly 25% gross margins, and "Shooting Stars" growing more like SaaS companies at 60%. Neither archetype touches 80%.
The spread depends on product type, and the differences are structural rather than a phase companies grow out of.
| Product Type | Typical Gross Margin | Why |
|---|---|---|
| Pure inference (wrappers, chatbots) | 30-50% | Every user action triggers real compute |
| Hybrid (AI features in SaaS shell) | 50-70% | Non-AI features carry traditional margins; AI features drag the average |
| Outcome-priced (per-resolution, per-result) | 60-67% | Higher pricing power offsets higher cost variance |
Datadog runs at 80.8% gross margins selling monitoring software. Anthropic runs at 40% selling the models Datadog's customers use. Same industry, half the margin, and the gap isn't a startup phase. It's the cost of running inference on every user interaction. In traditional SaaS, the marginal cost of the 10,000th user was effectively zero. In AI products, the 10,000th user costs the same to serve as the first, and sometimes more if they run complex multi-step workflows.
What Breaks at 50% Margins
Sales compensation
AJ Bruno's QuotaPath team studied hundreds of comp plans at AI-native companies and found them modifying compensation at 3.5x the rate of average SaaS companies. Some have abandoned deal-size commissions entirely, paying flat $5,000 bonuses per new logo regardless of contract value.
The arithmetic explains why. A 15% commission on an 80% margin deal consumes 18.75% of gross profit. The same commission on a 50% margin deal consumes 30%, and at some point the commission on an incremental deal exceeds the margin that deal generates. Multi-year deal bonuses have nearly disappeared too, because locking in a price for three years is a liability when inference costs shift quarterly.
The companies adapting fastest pay per-logo bonuses where the customer relationship is the asset, speed bonuses that reward skipping lengthy proof-of-concept cycles, and lower base commission rates paired with expansion incentives.
Fundraising math
The traditional SaaS fundraising formula assumed 80% gross margins: $1 of ARR valued at $8-12 of enterprise value, with Rule of 40 as the quality benchmark.
At 50% margins, the same revenue converts to less profit. Kyle Poyar at Growth Unhinged estimates AI companies need 2-3x more revenue to reach the same profitability milestones as traditional SaaS. A $10M ARR AI company at 50% margins has $5M of gross profit, while a $10M ARR SaaS company at 80% has $8M. That $3M gap is not rounding error. It can be the difference between a Series B and a bridge round.
Public markets have already repriced, with SaaS names shedding $285-300 billion in market value over 48 hours as investors reassessed how AI affects revenue models. Private markets are slower, which means founders using SaaS benchmarks to model their AI company's economics are presenting numbers the cost structure underneath does not support.
The power user inversion
Traditional SaaS loves power users: high engagement, strong retention, near-zero marginal cost. More usage is pure upside.
AI products invert this. Power users consume the most inference, generate the most tokens, and run the most complex workflows, so their marginal cost scales with every interaction. One customer makes 200 queries per month, another makes 20,000, and both pay the same subscription. In traditional SaaS the second customer is your best account. In an AI product, that account may be quietly costing you everything.
GitHub Copilot is the canonical example. Some individual users cost Microsoft $80 per month while paying $10, and the fix wasn't efficiency. It was pricing: enterprise tiers at $39 per user, reaching 40% margins on $300 million in enterprise licensing.
Without per-customer cost data, the variance between your most and least profitable accounts is invisible. At 80% margins the buffer absorbs it. At 50%, it determines whether you have a business.
ARR Is Now a Vanity Metric
In traditional SaaS, ARR was a reliable proxy for business health because nearly all revenue converted to gross profit. In AI products, every customer carries real compute cost that doesn't disappear at scale, so two companies can report $10M ARR with wildly different gross profit depending on customer mix, usage patterns, and inference spend. OpenAI projected $25 billion in cash burn for 2026 despite $12.7 billion in annualized revenue, with revenue growth and loss growth moving in the same direction.
Todd Gagne at Ibbaka frames it precisely: "For the first time in twenty years, software companies have to care about marginal cost again." The corollary: if cost scales with usage, price must scale with usage.
The downstream effects are everywhere once you look. CAC payback periods calculated on revenue overstate the actual payback when a meaningful slice of each dollar goes to compute. LTV models built on revenue rather than gross profit produce numbers that feel good and mean less. A 10x revenue multiple on an 80% margin business implies a very different valuation than the same multiple at 50%. None of this makes ARR meaningless. It just means ARR alone no longer tells you whether you have a viable business. Gross profit does.
Pricing Becomes a Weekly Decision
Krzysztof Szyszkiewicz at Monetizely describes a pattern common among his clients: waking up to a monthly bill from Anthropic or OpenAI that's 2x higher than the previous month. When your largest cost input changes monthly, annual pricing reviews are too slow.
Fynn Glover at Schematic puts it more precisely: "A credit cost. A usage limit. A model tier. An overage threshold. These are not annual pricing reviews. They are weekly operational decisions."
DeepSeek dropped API prices 50% overnight when it released V3.2 in September 2025, bringing input token costs to $0.028 per million, roughly one-tenth the cost of GPT-5 at $1.25 per million. Any AI company building on those models saw its cost-to-serve change in a single day.
Traditional SaaS could set prices annually because the cost side was stable. AI companies operate in an environment where a provider drops prices 40%, a cheaper model launches for your use case, or your largest customer doubles inference volume, and each event reshapes your margin structure. These things happen quarterly, not annually, and every one of them is effectively a pricing decision whether or not anyone priced it deliberately. The companies that iterate pricing fastest capture the most margin. The ones reviewing annually operate on stale data for eleven months of the year.
The Model Cost Deflation Trap
There's a common objection here: model costs are dropping fast, so margins will fix themselves.
The data says otherwise. DeepSeek cut inference costs 50% in 2025 through sparse attention, OpenAI dropped GPT-4o pricing 80% between launch and early 2026, and Anthropic's per-token costs fell as Haiku and Sonnet got cheaper. And yet Anthropic's gross margins came in 10 percentage points below their own internal projections, and OpenAI's inference costs surged to $8.4 billion in 2025 despite cheaper models.
The mechanism is straightforward: cheaper models unlock new use cases, which drive more usage, which fills the cost gap. When inference gets 50% cheaper, customers don't pocket the savings. They run 3x more queries. Jevons paradox, the same dynamic that played out in cloud computing, bandwidth, and storage, applies to AI inference with equal force.
Companies waiting for model costs to solve their margin problem are making the same mistake as the SaaS companies in 2015 that assumed cloud costs would make on-premise competitive. The cost curve helps, but it doesn't change the fundamental economics. AI products will always carry meaningful marginal cost per interaction, and the operating model has to account for it.
What an AI-Native Operating Model Requires
Measurement comes first. Lead with gross profit, not ARR. Segment customers by profitability, not just size or growth rate, and gate growth spending on real unit economics rather than CAC-to-LTV ratios calculated from revenue. Most dashboards and board decks still lead with ARR; gross profit appears three slides deep, if at all.
Visibility is what makes the measurement possible. At 80% margins you could afford not to know your cost-to-serve at the customer level, because the buffer absorbed the variance. At 50%, the gap between your most and least profitable customers can be the difference between sustainability and scaled losses, so per-customer, per-feature cost attribution stops being a reporting nicety and becomes the foundation.
Compensation is the next domino. Commission structures have to reflect actual margin on deals rather than revenue, which in practice means per-logo bonuses aligned to land-and-expand motions, speed incentives that shorten sales cycles, and no multi-year premiums that lock in prices the business may need to change within months.
Then there's cadence. Weekly pricing iteration isn't a luxury for AI companies; when your largest cost input changes quarterly, pricing has to follow, and that requires cost monitoring, billing flexible enough to adjust without an engineering sprint, and pricing authority sitting with the product team instead of locked inside annual contracts.
And finally, accept the band. AI company margins will likely settle around 55-65% at maturity, not 80%. The 80% era funded "grow now, optimize later." At 50% there is no buffer, every month of incorrect pricing compounds, and the discipline of understanding costs starts on day one rather than at Series B.
Where This Lands
AI-native margins are roughly half of traditional SaaS margins, and truthfully that's not a problem to solve. It's a cost structure to operate within. The companies that restructure around it compound advantages in sales efficiency, pricing accuracy, and capital allocation. The ones that keep running the old playbook burn runway wondering why the growth math never works.
Either way, the adjustment starts with seeing the margin you're trying to restructure. Bear Lumen gives you that view per customer and per feature, continuously, with no invoice exports or month-end spreadsheet reconciliation to maintain.