Usage-Based Billing Systems

This weekend I built an MCP (Model Context Protocol) server with usage tracking for LLMs. I was surprised to find that a "single" MCP tool request actually involves several information-gathering requests by the LLM before actual tool invocation. If you're building MCP servers, you need to think differently about billing. AI agents consume APIs fundamentally differently than humans use apps. They don't browse, they execute. They don't explore, they accomplish specific tasks with precision. This creates a unique billing challenge: What should you actually charge for? Three critical billing principles for the AI agent economy: 1. Bill for value creation, not discovery. AI agents will call tools/list and resources/list many times as they explore your capabilities. These are discovery calls like browsing your website. Don't charge for them. But when they call tools/call to actually execute a search or a calculation? That's where value is created and costs are incurred. 2. Align billing with your underlying costs. Every MCP tool call that hits your actual API should be billed. That query for "coffee shops near me" that costs you API credits? Charge for it. But returning static tool definitions or cached documentation? Keep those free to encourage integration and exploration. This aligns with FinOps Foundation's principle that business value should drive technology decisions (https://lnkd.in/ggiusZU2). Your billing model should reflect actual value delivery, not arbitrary metrics. 3. Make billing predictable and transparent. Agents need to understand costs before they act. Unlike humans who might impulse-click, agents are programmed to optimize. Give them cost estimation endpoints, clear pricing tiers, and predictable rate structures. Surprises kill automation. The FinOps Foundation's unit economics framework (https://lnkd.in/gDsyDvH3) provides excellent guidance here: focus on measuring the unit economics of the entire use case, not just individual API calls. Your MCP billing should enable cost optimization at the workflow level. Drawing inspiration from AI FinOps best practices: The FinOps Foundation's work on GenAI token pricing (https://lnkd.in/gpeGth4X) by Josh Collier and David Tepper reveals hidden costs that MCP servers should avoid: Context Window Creep: Don't penalize agents for maintaining conversation state Provider variations: Ensure consistent pricing regardless of underlying infrastructure Total Cost of Ownership: Help users understand the full cost to accomplish their business outcomes The bottom line: Apply FinOps principles to your MCP billing strategy. Make costs accessible, timely, and accurate. The companies building fair, transparent MCP billing models today will become the preferred infrastructure in tomorrow's agent-driven economy. #MCP #FinOps #APIBilling #AIAgents #UnitEconomics

We're moving away from charging for *access* to software and toward charging for the *work delivered* by software & AI agents. Don't freak out: this doesn't mean everything will become *pay-as-you-go* overnight. I can think of 7 flavors of charging for work: 1️⃣ Pay-as-you-go - No commitment, totally flexible - Enterprise procurement teams usually *hate* this! - Works best when your customers can bill-back the expense or bake it into an operating budget - Otherwise, there's a risk of customers policing their own usage (taximeter effect) 2️⃣ Subscription + pay-as-you-go - Small level of commitment helps 'lock customers in' and give them access to advanced features, support, etc. - Works well when the usage metric is getting commoditized (ex: SMS messages, compute, storage) -- you can advertise a low usage fee & make up for it with the subscription fee - Still not quite loved by enterprise procurement since their bill isn't predictable yet now includes multiple line items... 3️⃣ Three-part tariff (usage subscription + PAYG) - Similar to the above, but with a larger subscription fee that includes some level of usage "included" - Folks usually advertise the initial usage as a gift ("get your first 500 SMS messages for free!") - Including a minimum level of usage helps get the customer hooked & usually incentivizes more overall consumption 4️⃣ Usage-based subscription (high watermark) - Customers commit to a certain level of usage or tier (ex: up to 5,000 API calls per month); this is typically "use it or lose it" - Subscriptions are for a high watermark of usage -- if usage exceeds the plan in a given month, they immediate move into upgrade territory - Fear of overages + usage fluctuations encourages sales to over-sell & customers to over-buy 5️⃣ Usage-based subscription (annual drawdown) - Similar to the above, but the usage allocation can be consumed flexibly over the course of 12 months similar to a gift card - This gives the customer plenty of time to monitor adoption & plan for an early renewal/upgrade if usage is trending above their commit - Great for customers with seasonality or month-to-month usage fluctuations who still want a predictable bill 6️⃣ Roll-overs - If the customer doesn't consume their full allocation, they can "roll it over" to the next year -- typically only if they commit to a flat or increased renewal - More customer friendly, but also more painful to manage! 7️⃣ Adaptive flat rate - The customer commits to a usage-based subscription, but can use the product as much as they want with no overages/upgrades during that period - Their tier resets up/down at renewal based on their actual usage behavior - Much more predictable for customers while encouraging them to increase consumption (downside is that you could be stuck with the costs!) -- I suspect most folks will offer multiple options as they seek to balance lands, expands & tough procurement convos. The downside: complexity.

🔵 Stripe just paid $1 billion for something it could have built. That tells you everything about the complexity and urgency of usage-based billing in the AI era. The biggest shift in software monetization since SaaS is happening. Patrick Collison isn't mincing words: 𝐮𝐬𝐚𝐠𝐞-𝐛𝐚𝐬𝐞𝐝 𝐩𝐫𝐢𝐜𝐢𝐧𝐠 𝐢𝐬 "𝐭𝐡𝐞 𝐧𝐚𝐭𝐢𝐯𝐞 𝐛𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐦𝐨𝐝𝐞𝐥 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐀𝐈 𝐞𝐫𝐚," potentially as big as (or bigger than) the advent of SaaS itself. UBB - Usage Based Billing Payment processing is one layer; monetization logic is another. Stripe is focused now on both. 🔷 𝐓𝐡𝐞 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 Metronome's valuation doubled in less than a year (from $470M in February to $1B now), with 8x growth in platform volume during 2024. Their client roster speaks volumes: OpenAI, Anthropic, Databricks, Nvidia—companies where consumption-based pricing isn't optional, it's essential. The shift makes sense. AI value correlates directly with consumption: API calls, compute time, tokens processed. Traditional seat-based subscriptions simply don't capture how customers actually derive value. 🔷 𝐁𝐞𝐲𝐨𝐧𝐝 𝐌𝐞𝐭𝐫𝐨𝐧𝐨𝐦𝐞: 𝐓𝐡𝐞 𝐔𝐬𝐚𝐠𝐞-𝐁𝐚𝐬𝐞𝐝 𝐁𝐢𝐥𝐥𝐢𝐧𝐠 𝐄𝐜𝐨𝐬𝐲𝐬𝐭𝐞𝐦 This isn't a one-company phenomenon. This market has exploded with specialized players, each carving out territory. Here are four infrastructure players comparable to Metronome: ‣ Orb – Usage-based billing and pricing infra with strong adoption among modern SaaS and AI companies; Metronome itself positions Orb as its primary direct comparator. ‣ m3ter – Purpose-built usage metering and rating engine for complex B2B SaaS and hybrid models, often grouped with Metronome and Amberflo as the core UBB infra cohort. ‣ Amberflo.ai – Developer-first consumption billing that focuses on metering at scale and “AWS-style” usage pricing; regularly listed alongside Metronome and m3ter as leading UBB startups. ‣ Lago – Open‑source usage-based billing and metering, explicitly branded as a Metronome alternative and highlighted as the strongest choice when teams want control and self-hosting. Stripe chose to acquire rather than build. That signals how complex and critical this capability has become. 🔷 𝐖𝐡𝐲 𝐓𝐡𝐢𝐬 𝐌𝐚𝐭𝐭𝐞𝐫𝐬 Usage-based pricing aligns revenue with value delivery in ways subscriptions never could. It's more transparent for customers, more scalable for providers, and infinitely more adaptable to hybrid models. For financial services and fintech, this is an infrastructure-level transformation. We're not just talking about billing—we're talking about how companies capture value in real-time, optimize pricing dynamically, and build monetization as a competitive advantage. The question isn't whether to consider usage-based models. It's how quickly you can implement them before your competitors do. #Fintech #AI #monetization

🚨 Big M&A! Stripe acquired Metronome to win at AI-usage-based billing A leader in usage-based billing. Customers include OpenAI& Nvidia. Here's why this matters: - AI products break traditional billing systems. - SaaS is predictable: $50/user/month. - AI is chaos: you're charging per token, per inference, per compute minute. Usage spikes 10x overnight when a customer's product goes viral. Most billing infrastructure collapses under that load. --- Metronome built theirs to handle it. That's why OpenAI uses them. They have sophisticated "budget" based billing (set an amount and use tokens within that budget). Most Metronome customers were / are Stripe customers, so there's a natural overlap here. --- The strategic rationale works two ways. For Metronome There's only so much upside in being pure-play billing. Owning the payment — the actual unit of value — unlocks every product extension over time. For Stripe They get that killer feature that had been hurting in the AI space and can massively expand their ARPU. --- Stripe gets proven AI billing infrastructure. Metronome gets distribution to millions of businesses. Smart deal on both sides.

Hot take: people don’t actually want predictability… they want transparency. I keep hearing: “Usage-based pricing is unpredictable and buyers don’t like that” Predictability is: “Tell me my bill to the dollar a year in .” Transparency is: “Show me what’s driving it, and give me knobs.” Because we already live in a (mostly) usage world: Electricity / water: I don’t predict kWh. I just know my holiday guests took extra long extra hot showers. AWS: the whole point is you pay for what you use (compute/storage/queries), and you survive because they give you dashboards, alerts, budgets, and “hey… this spike looks suspicious.” or this S3 bucket hasn't been used in ages. Snowflake, Clay, Figma, OpenAI, Loveable, … and many more So the fix isn’t “go back to seats.” Also: seats are not coming back. Buyers ask for seats because it feels simpler—until layoffs hit, usage drops, and suddenly your “predictable” revenue is predictably… down.  Plus procurement will grind your price per seat every renewal with the classic: “We’ll open an RFP.” Congrats: you’re now funding rising LLM/inference costs by giving your innovation away for free. 🫠 The real answer is making usage feel fair: ✅ Visibility Tell customers exactly how they are using those credits. Contacts? Emails sent? Scanning inboxes? Calls? Doing research? ✅ Controls Let your customers set caps, alerts, limits - and let them know when they are about to hit them, who did it, and why! ✅ Forecasting “if you keep doing this…” and “here is a plan that will save you money …” ✅ Economizing that’s actually a win-win: “Hey, you’re trending high—upgrade to a bigger package and your unit cost drops.” Customer pays less per unit, vendor earns more, everyone stops doom-refreshing invoices. AI Agents are turning software into a usage-based economy - whether it’s outcomes or resource consumption, that economy only works if we put customers at the center: clarity, control, and trust—while staying profitable. Usage isn’t the problem. Opacity is.

How businesses charge their customers has changed more in the last twelve months than it changed in the last decade. A year ago, billing was simple. You sent a flat monthly subscription, or you sent an invoice at the end of a project. Today it looks completely different. Software companies are charging by the unit, the call, the outcome, or the workflow that just ran in the background. Service businesses are layering retainers with usage fees and performance based pricing. Agencies are running monthly fees plus per project add ons plus revenue share. AI has poured fuel on all of this. Businesses are now monetizing in tokens, API calls, and usage that didn’t exist five years ago. But the billing system on the back end of most of these businesses is still trying to send a single flat invoice for a single flat amount. I have watched this happen with a handful of my clients this year. The CEO decides to roll out a new pricing model on Monday. The finance team finds out on Tuesday that the billing tool can't actually support it. By Friday someone is in a spreadsheet trying to back into what each customer owes. That gap is exactly the problem Airwallex went after with their launch today. It's called Airwallex Billing. It's a real time billing layer that captures every transaction at the moment it happens, and lets you bill on any model you choose. You can run pay as you go, tiered subscriptions, hybrid pricing, project invoicing, outcome based fees, or whatever your business actually does. Here's why this matters more than most launches. → Real time metering means revenue isn't lost between what the business delivers and what the customer gets billed for → Subscription, usage, invoicing, and hybrid pricing all live on one platform, so you can test new pricing models without rebuilding your stack every time → It runs on the same global infrastructure that already handles payments, spend, transfers, and reconciliation, which means you aren't bolting another vendor onto an already crowded finance stack → Built in tax automation per jurisdiction, so charging customers across borders doesn't blow up your compliance → Multi entity support, so finance teams keep one source of truth even as the business splits into different revenue lines and geographies The companies winning the next five years will be the ones that can change how they charge as fast as their customers and their offers evolve. You can't do that on a billing system built for 2018 invoices. If you've been wanting to evolve how you bill but kept putting it off because the back office work felt impossible, this is worth a serious look. Dropped the launch link below for anyone curious

Imagine you buy 100 Replit Enterprise licenses for your team and 3 employees are maxing out their credits. The other 97 are barely touching it at all. That's a lot of $$$ down the drain. This is why seat-based pricing is a dying. And we're switching to fully usage-based pricing for our enterprise customers. Here's what that looks like: Instead of 100 individual buckets that expire unused, you get one shared credit pool. Your three power users burn through what they need. Occasional users dip in when they've got a project. Nobody wastes money on empty seats. At month's end, you see exactly what you consumed. No phantom charges for people who forgot the product exists. I'll be honest. We make less money this way on some accounts. A seat-based model would extract more from customers who over-buy and under-use. Usage-based flips that entirely. We only win when you actually use the product. You only pay for what you actually consume. Would you want 100% usage-based billing from Replit?

Stripe, Adyen, Airwallex and Salesforce all bought the same kind of software this year. None of them coordinated. Four of the best engineering shops in payments. Four different continents. One target: usage-based billing. The layer that meters what a customer uses, prices it, and books the revenue. Adyen is the tell. It didn't buy a single company in 20 years. Then it bought two in seven weeks. So I pulled all five deals into a single grid to see what they were really paying for. Here's the pattern. A processing fee gets cheaper as it scales. Any rival can undercut it by a few basis points. A billing system gets harder to leave as it scales. It's wired into how a company sets prices, books revenue, and closes the quarter. The rail commoditizes. The meter compounds. Run the math on a merchant doing $100M a year. As their processor, you may keep $ 200k–$400k after interchange and scheme fees. Run their billing too, and that one line is $700k. At software margins. On revenue, it's far harder to walk away from. Then there's lock-in. Swapping a processor is a weekend. Ripping out billing is a year. So why now, and not three years ago? AI killed per-seat pricing. When one agent does the work of ten people, you've sold one seat where you used to sell ten. The only way to line price up with cost is to meter what gets used. The firms with the most to lose from killing per-seat pricing are the ones killing it first. This isn't processors buying scale. It's payments companies buying a different layer. I broke down all five deals, what each buyer was really chasing, and what it means for you, in this week's edition. If you run a PSP, the per-transaction business is your floor now, not your ceiling. Full breakdown in the comments.

A flytrap doesn't catch its prey by force. It offers something sweet, waits until the insect is inside, then closes. Enterprise AI pricing works the same way. The bait was a flat monthly subscription. But the real cost depends on how much the AI gets used. Like a phone plan that looked flat but turns to be pay-as-you-go: the monthly fee connects the line, but every call adds up. With AI, each prompt burns "tokens," and the bill keeps growing. The consequences ▫️ 61% of companies cut AI projects because of unplanned cost increases ▫️ 78% of IT leaders reported unexpected consumption-based charges ▫️ Spending on AI-native tools rose 108% YoY Source: Zylo 2026 SaaS Management Index We have seen this trap before Cloud ran this play for a decade - cheap to pour data in, expensive to pull it back out. AI vendors are following the same script. The only difference AI is repricing in real time: ▫️ OpenAI: GPT-5.5 for Enterprises doubled the API input rate vs its predecessor: $2.50 → $5 per million input tokens ▫️ Microsoft: Copilot is $30/user/month on top of an M365 licence, with pay-as-you-go into Copilot Chat, SharePoint agents and Studio ▫️ Anthropic paused, not cancelled, a move from flat fee to credits ▫️ GitHub Copilot moved to usage-based billing on 1 June ▫️ Salesforce bills in Flex Credits, SAP in AI Units The biggest spenders The heaviest AI bills come from non-technical teams, turning piles of documents into slides and reports. Under usage-based pricing, a single heavy user can outspend dozens of colleagues in a day. ➡️ Only 26% of companies have real-time visibility on AI real costs (KPMG) A sticky strategy The same across vendors: ➡️ low-cost entry ➡️ embed into critical workflows ➡️ shift to usage-based billing ➡️ grow revenue. Once AI is embedded, leaving becomes infrastructure replacement. As OpenAI's Nick Turley put it: "having an unlimited plan is like having an unlimited electricity plan. It just doesn't make sense." 3 things every board should mandate now 1️⃣ Ensure AI is metered as an utility vs software. 2️⃣ Price the exit: For every AI in a core workflow, ask for the rebuild cost. 3️⃣ Request guardrails on premium models and agents. The flytrap is already open. The exit is the only number that still matters. #AI #AIGovernance #Boardroom #StratEdge

The free trial for Databricks Genie is wrapping up now that it’s officially GA, and it is switching over to a pay-as-you-go pricing model (from 6th July). Here is how data teams can maintain absolute FinOps control. It is a classic cloud playbook: offer a high-value tool for free during preview to drive adoption, then introduce metering. Because GenAI reasoning carries real infrastructure overhead, a shift to paid usage was inevitable. The good news? Databricks chose a consumption-based architecture over rigid, expensive seat licensing. This gives data engineering teams the leverage to control costs. Here is how to manage the transition and protect your budget from future price changes: • Leverage the 150 DBU Free Baseline: Every active user gets 150 free DBUs monthly. Occasional business users or executives asking light questions will still cost your organization exactly $0 on the AI layer. • Enforce Strict Genie Budgets: Alongside pricing, Databricks launched programmable budgets. Admins can set monthly thresholds per user or workspace to trigger alerts or trigger a hard stop that blocks usage the moment a limit is reached. • Isolate LLM from Execution Compute: Genie decouples semantic AI routing ($0.07/DBU) from data execution. You can keep execution costs low by optimizing your underlying Azure Serverless SQL warehouses independently. • Audit Spend via System Tables: All Genie consumption is logged directly in Unity Catalog system billing tables. You can monitor usage via SQL and instantly calculate the exact ROI of user queries. By using programmatic budgets and Unity Catalog governance, data teams can safely scale self-service data products without worrying about unexpected bill spikes. #DataEngineering #AzureDatabricks #CloudFinOps #UnityCatalog #DataArchitecture #Databricks