--- title: "Why Falling Costs Don't Mean Better Margins" description: "Inference costs are dropping 5-10x per year. But seat-based pricing can't handle cost variance, and usage-based pricing deflates your revenue. The real problem is that most teams can't see what each customer costs them." date: 2026-02-19 author: Kat Laszlo canonical: https://tansohq.com/blog/falling-costs-margins --- # Why Falling Costs Don't Mean Better Margins By [Kat Laszlo](https://www.linkedin.com/in/katrinalaszlo/) · February 19, 2026 AI inference costs have dropped roughly 280x since late 2022. Every quarter, a new model comes out that's cheaper and better than the last one. If you follow the headlines, you'd think margins should be improving across the board. They're not. Median AI gross margins were 41% in 2024, 45% in 2025, with a projected 52% in 2026. Compare that to traditional SaaS at 80-90%. The gap is closing, but slowly, and not because of cost deflation alone. Falling inference costs don't automatically become better margins. Your pricing model determines whether you capture the savings or give them away. And the two most common models each break in different ways. --- ## Seat-based pricing: you eat the variance Most AI companies started here because it's familiar. You charge $50 or $100 per seat per month, just like SaaS always has. The logic seems sound: as inference gets cheaper, your cost to serve each seat drops, and your margin expands. In practice, this doesn't hold. The problem is that seat-based pricing assumes roughly uniform cost per user. Traditional SaaS could get away with this because the marginal cost of serving any individual user was close to zero. It didn't matter if one person logged in twice a day and another logged in twice a month. The cost difference was negligible. AI is different. Every query burns compute. A light user might cost you $5 a month. A power user might cost you $80. GitHub Copilot learned this the hard way: at $10 per month, heavy users were costing Microsoft up to $80 each. The company averaged a $20 loss per user in the early days. Falling inference costs help with the average, but they don't fix the distribution. If your heaviest users are 10-15x more expensive than your lightest, a 50% cost reduction still leaves you underwater on power users. You've just moved from losing $40 per heavy user to losing $15. That's better, but it's not a business model. And the Jevons Paradox makes it worse. As AI gets cheaper and more capable, people use it more. The features that attract power users get better, which means power users use them even more. Inference now consumes 80-90% of all AI computing power, a complete reversal from a few years ago when training dominated. The per-unit cost goes down, but the number of units per user goes up. 84% of enterprises report significant margin erosion tied to AI workloads. 80% miss their AI infrastructure forecasts by more than 25%. Costs are falling per token, but total spend keeps climbing. --- ## Usage-based pricing: your revenue deflates too The obvious response is to charge for what customers actually consume. Per token, per API call, per compute minute. Your costs are variable, so your pricing should be too. But usage-based pricing introduces its own problems. Your revenue deflates alongside your costs. If you charge per token and token costs drop 10x in a year, either your price drops 10x (and your revenue collapses) or you maintain price and a competitor undercuts you. Jasper learned this: their per-word pricing worked until generation costs fell and ChatGPT offered a flat alternative. Revenue dropped from $120M to roughly $55-88M. There's also the predictability problem. Customers don't think in tokens. When Leena AI used consumption pricing for their employee support agents, it backfired. Users became wary of consuming the product because they couldn't predict what it would cost. After switching to an outcomes-based model, adoption improved. Enterprise buyers need to know what they're going to spend. A CIO who blows through the AI budget in Q2 has a career problem, not just a billing problem. And usage-based pricing penalizes your best customers. The people who use your product the most, who get the most value, who are the most likely to renew and expand — they get the biggest bills. Your happiest customers become your most price-sensitive. And then there's the race to the bottom. DeepSeek pushed token costs toward near-zero in early 2025, offering models at roughly $0.70 per million tokens versus $20+ for comparable Western models. If your pricing is anchored to token costs, you're competing on a metric that trends toward zero. Every efficiency improvement in the ecosystem compresses your revenue. --- ## What actually improves margins The companies reaching 60%+ gross margins aren't getting there by waiting for inference to get cheaper. **Model routing.** Using cheap models for 70-90% of simple queries and reserving expensive models for the complex ones. Companies that implement this consistently see 60-87% cost reduction. If 90% of your queries can be handled by a model that costs a fraction of a cent, your average cost per query drops dramatically even if the remaining 10% are expensive. **Per-customer cost attribution.** Knowing which customers and which features are margin-positive versus margin-negative, in real time, not at the end of the quarter. This lets you set pricing tiers, usage caps, and overage rates based on actual data instead of assumptions. **Hybrid pricing.** A base subscription covers your fixed costs and gives the customer predictability. Usage or outcome components on top capture value when customers scale. The base protects your floor. The variable component captures your upside. **Non-inference revenue streams.** Replit charges $25 per month for AI assistance but makes most of its margin on hosting, deployment, storage, and bandwidth. The AI relationship is the entry point. The high-margin services around it are the business. None of this requires waiting for the next round of cost cuts. These are operational decisions you can make now. --- ## Cheaper doesn't mean profitable Inference costs will keep falling. That's good for the ecosystem, but it doesn't fix your margins on its own. Seat-based pricing can't handle cost variance per user. Usage-based pricing deflates your revenue along with your costs. The companies that are actually improving their margins are the ones that can see what each customer costs them and price accordingly. The pricing model matters, but visibility into the data comes first. If you can't answer "which of my customers are profitable" today, falling costs won't save you. They'll just make the losses smaller. ## Key Takeaways - ✓Inference costs are dropping 5-10x per year. That's real. - ✕Seat-based pricing can't capture the benefit because cost variance per user is the problem, not average cost. - ✕Usage-based pricing deflates your revenue alongside your costs. - →The actual problem isn't what you charge. It's that you can't see what each customer costs you. *[Kat Laszlo](https://www.linkedin.com/in/katrinalaszlo/) is co-founder of Tanso, flexible pricing infrastructure for SaaS and AI.* ### Want to price with confidence? Book a quick call and we'll walk you through it.