An inference provider runs AI models on its own or rented chips and sells the output by the token through an API, so an app can use a model without building its own serving stack. On Fireworks AI’s pricing page, the highlighted standard serverless rates for a million output tokens run from $0.20 for NVIDIA Nemotron 3.5 Lightning 30B A3B to $15.00 for Kimi K3 and Ember-1: a 75x spread on one API, by our arithmetic.
Two Fireworks customers have just shown what else an app hands such a provider. Cursor says it moved all of its inference traffic off Fireworks after a breach it traces to a key it had issued to the company, and Notion says it is no longer serving open-weight models through Fireworks. As we read their accounts, an app gives a provider two things: its traffic and access to some of its data.
Business Pill · Usage-Based Pricing
An evergreen Business Engineer explainer on usage-based pricing, paying for what you use. It teaches the general idea only and does not discuss Fireworks, Cursor or Notion.
The key insight: As we read Fireworks’ pricing page, an inference provider sells one unit, the token, at very different prices: a 75x spread between the cheapest and dearest output rates in its highlighted table, by our arithmetic, with discounts for waiting (Batch) or for accepting interruptions (Spot).
CNBC described Fireworks’ business in July: it competes with Amazon and Google “to host models that developers can weave into applications”, and by managing computing infrastructure for models it does business in the inference cloud market, “alongside startups such as Baseten and Together AI”. It has also started providing GPUs for training, “like neoclouds CoreWeave, Lambda and Nebius”. CNBC reported that Fireworks “relies on computing power from more than 20 suppliers, including Microsoft”.
Fireworks’ case study on Cursor gives a buyer’s reason. It says Fireworks let Cursor “run large-scale reinforcement learning without building and operating a dedicated inference stack internally”, and quotes Cursor’s research lead saying the team would prefer to have engineers improve training “rather than spin up an inference effort”.
Fireworks AI serverless Standard rate per million output tokens for selected models, from its pricing page as read on 15 November 2026. The 75x spread ($15.00 / $0.20) is our arithmetic.
How the Pricing Works
Fireworks’ pricing page offers three ways to buy. Serverless is “Pay per token · From $0.05 / 1M input tokens”, with no commitment and usage drawn from prepaid credits. Reserved Throughput reserves capacity “by the minute” with a throughput SLA. Fireworks Sovereign is a contract under which Fireworks manages inference on GPUs in the customer’s own cloud account or data center.
“Every serverless request bills on input, cached input and output tokens”, the page says. On Kimi K3, a million output tokens cost $15.00 against $3.00 for input, five times as much, by our arithmetic. Cached input is “billed at half price or less”. Batch jobs get “50% off input and output”, and Spot capacity is 30% off Standard, though “Throughput can drop to zero”.
How Big the Volumes Are
Fireworks said in its Series D post in July that it serves more than 40 trillion tokens every day. That is about 463 million tokens a second on average, by our arithmetic. CNBC set the figure against Google’s disclosure of about 19 billion tokens per minute for developers, which CNBC said implies more than 27 trillion a day.
Fireworks put its annualized revenue run rate above $1 billion at the time, and CNBC reported that about half of its revenue came from Cursor as of last year. Our piece on the 121 days covers that concentration; our Inference Explained covers the cost paid on every AI answer.
Traffic, Keys and Data
Cursor’s update shows access beyond live requests. Cursor says it had issued Fireworks an access key and shared a read-only storage bucket of coding agent traces, “which we were permitted to use for model training”. Cursor says that key was exposed by Fireworks; the Fireworks bulletin we read on 15 November does not mention Cursor. Our piece on Cursor’s update covers its account.
Fireworks says its Inference, Training SDK and Nexus products are zero data retention by default and that it has not detected any impact to them. Notion’s advisory says it used Fireworks for classification jobs and “as a general inference provider to serve open-weight models inside our product”.
How a Customer Switches
The two customer accounts list similar steps. Cursor says it moved all inference traffic off Fireworks, deactivated the key it had issued, cut off access to the data sets it shared, rotated every exposed credential and expanded network access restrictions “so compromised credentials can’t be used outside their intended location”. Notion says it disconnected its systems from Fireworks by the morning of 3 November, and that on 4 November it “deleted all its data in the Fireworks system”.
As we read the pricing page, the serverless tier carries no commitment, while Reserved Throughput and Sovereign do. Our piece on Fireworks’ case studies counts how many of them still feature Cursor and Notion.
The Structural Read
As we read it, the business runs on volume: more than 40 trillion tokens a day at Fireworks, about 463 million a second by our arithmetic, sold at prices quoted per million tokens.
As we read the Cursor and Notion accounts, leaving a provider is mostly a matter of access: both list keys, data and connections to cut, and Fireworks’ serverless tier carries no commitment.
“Every serverless request bills on input, cached input and output tokens”
Three Implications
FOR APP BUILDERS Fireworks lists Batch at 50% off input and output and Spot at 30% off Standard, and says Spot throughput can drop to zero.
FOR SECURITY TEAMS Cursor says it expanded network access restrictions across its service providers and added secret scanning across its code repositories.
The framework says: “Training builds the brain. Inference is the brain working.”
As we read it, an inference provider rents out the working part, billed per token, which is why its price list and its customers’ traffic matter more than any single model. The framework does not discuss Fireworks, Cursor or Notion.
What Is Not Established
These are list prices from the highlighted standard serverless table as read on 15 November, and the page says models not listed are priced by size; pricing pages change, and enterprise contract terms are not on the page. The documents we read give no customer count from Fireworks itself; Notion’s advisory describes Fireworks as powering “tens of thousands of customers”. We did not contact Fireworks, Cursor or Notion.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
An inference provider sells model output by the token, and in Fireworks’ highlighted standard serverless rates a million output tokens cost from $0.20 to $15.00 depending on the model, a 75x spread, by our arithmetic. Fireworks says it serves more than 40 trillion tokens a day, and Cursor and Notion say they have moved traffic off its platform.
A note on sourcing. Prices and tiers come from Fireworks AI’s pricing page as read on 15 November 2026; the token volume and run rate from its Series D post of 15 July; the market description and peers from CNBC’s report of 16 July; the buyer’s rationale from Fireworks’ Cursor case study of 26 June. Cursor’s and Notion’s steps come from Cursor’s update dated 13 November and Notion’s advisory updated 12 November, and Fireworks’ data-retention statement from its security bulletin. The 75x, the five times and the tokens a second are our arithmetic. We did not contact Fireworks, Cursor or Notion.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
