Fireworks AI Wants $40 Billion. Its Last Revenue Number Is From July.

Fireworks AI wants a valuation of $30 billion to $40 billion, people familiar with the talks told Business Insider this week. Three months ago the company closed a $1.5 billion round at $17.5 billion. The new ask roughly doubles the price before a fresh revenue number exists.

Eleven days before that report, Fireworks published a security bulletin. An intruder had used internal Fireworks credentials to reach environment variables inside a limited number of customer evaluation jobs. Containment began on the day the activity was identified.

Inference capacity is being repriced as scarce infrastructure. Trust now decides who keeps charging for it.

What the $40 billion is actually buying

Fireworks rents GPU capacity and sells three layers on top of it: serverless inference, dedicated deployments, and reserved capacity. Customers run open-weight models or their own fine-tunes through that stack. The logo list includes Cursor, Uber, Shopify, Notion, DoorDash and GitLab.

Most of the traffic is custom work. Fireworks said in July that more than 95% of the 40 trillion tokens it serves each day come from models tuned on customer data. The pitch is narrow and specific: enterprises stop renting general intelligence and start owning a specialized one.

That distinction matters for revenue quality. A general model API competes on price per token against every other general model API. A model tuned on a bank's underwriting data, or a retailer's catalogue, is harder to move once it is in production. Switching costs live in the fine-tune, and the fine-tune lives on Fireworks.

The contract mix matters just as much. Serverless inference is elastic and easy to leave. Reserved capacity is sold forward, often against a committed spend floor. Dedicated deployments carry the highest revenue per customer and the lowest churn, because moving them requires a new integration project. A vendor with a large reserved book looks safer at 40x than a vendor living on metered traffic.

The markup is a sector mark, not a company one

Modal Labs raised $355 million at $4.65 billion in May. By October it was closing $750 million at $15.75 billion, according to TechCrunch. Baseten raised $1.5 billion at $13 billion in June and was reported near $26 billion by September. Together AI reached $8.3 billion in July.

Fireworks is running the same ladder. It was worth $552 million in July 2024, $4 billion in October 2025, and $17.5 billion in July 2026. At the top of the reported range, the new mark would be ten times the October 2025 price and roughly double the July price.

Reported private marks, 2026 · Business Insider, Dealroom, TechCrunch

Why inference became the toll road

Training gets the headlines. Inference gets the invoice. Every coding assistant, every support agent, every document parser pays per token for as long as the product stays live. That recurring shape is what infrastructure investors pay a multiple for.

Open-weight models changed the supply side. An enterprise can take a strong base model, tune it on proprietary data, and serve it well below frontier API prices. Fireworks sells the rails for that maneuver and keeps the traffic on its own stack. Same-day access to new open models is part of the pitch.

The margin question stays open. Token prices fall every quarter. GPU depreciation is real, and reserved capacity carries utilization risk when demand shifts. A 40x multiple on trailing run-rate assumes gross margin holds while the price per token declines. That is a strong assumption for a business that buys the same silicon as its customers.

Competition is not only other startups. AWS, Google Cloud, Microsoft Azure and Oracle all sell managed inference, and each can bundle it into an existing enterprise agreement. Fireworks needs its custom-model layer to be meaningfully stickier than a procurement line item. The 95% figure is the company's evidence for that claim.

The cost line a 40x multiple has to underwrite

Fireworks does not own most of the metal. It leases GPU capacity and resells it as inference. That keeps the model capital-light at the income statement level and heavy at the contract level: committed capacity has to be paid whether or not customers send tokens.

Every peer is raising into the same build-out. Modal's $750 million and Baseten's $1.5 billion are rounds of comparable scale, priced within months of each other. Fireworks raised $1.505 billion in July and has spent part of the year building global orchestration for its inference stack.

The arithmetic of inference is unforgiving in one direction. Revenue scales with tokens served. Cost scales with reserved racks, power and utilization. When demand is strong, both rise together and margin looks healthy. When a customer shifts a workload to a cheaper model or brings it in-house, the revenue line moves first and the capacity commitment stays.

Hardware residuals are the other half. Residual-value trackers such as American Compute put a used H100 at 11-19% of original list price by 2027 in their conservative band. If that holds, a vendor pricing inference off a five- or six-year asset life is assuming a recovery value the secondary market has not yet demonstrated. The gap lands in gross margin, not in the headline valuation.

This is where the private markup and the public market diverge. Public neoclouds are already being repriced on credit terms. Private inference platforms are still being repriced on revenue growth. The two views cannot stay apart forever, because they describe the same underlying asset: racks of GPUs earning a spread over their cost of capital.

None of this makes the $30-40 billion ask wrong. It defines what has to be true. A 40x multiple on trailing revenue needs either durable pricing power or a cost structure that improves as volume grows. Fireworks argues the second through custom models, which carry higher switching costs and less price comparison than commodity tokens. The next disclosure is where that argument gets tested.

The October incident lands at the procurement gate

Fireworks published its bulletin on October 9. The sequence is short and specific. Identification on October 6 at 9am. Containment the same day. Direct notification to every identified affected customer. Additional API keys revoked and service-account tokens rotated on October 10, with credential review continuing.

The company says it found no evidence of unauthorized access to customer models, datasets, training data, personal information, or inference traffic inside its security perimeter.

For an infrastructure vendor, timing is everything. Security review sits before the contract. Enterprise buyers rarely accept a vendor that cannot close an incident cleanly, especially when the vendor holds model weights and evaluation data. A limited credential exposure can be survivable. A stalled renewal cycle is harder to explain to a board.

FIREWORKS AI — FUNDING AND INCIDENT TIMELINE
────────────────────────────
2024-07 · Series B at $552M valuation
2025-10 · Series C, $250M at $4B
2026-07 · Series D, $1.505B at $17.5B
2026-09 · $132M employee tender offer
2026-10-06 · Credential incident identified
2026-10-20 · Reported $30-40B talks

The tender offer is the quiet detail. Fireworks reportedly ran a $132 million employee secondary in September, with Atreides named as lead. Staff liquidity at a $17.5 billion mark is a useful signal: the company wanted to keep people before the next round, and investors wanted exposure before the next price.

What a 40x multiple has to believe

At $40 billion, Fireworks would trade near 40 times the $1 billion run-rate it disclosed in July. Dealroom puts the Series D at roughly 17.5 times. The step-up assumes either revenue grew faster than the multiple, or scarcity justifies paying ahead of it.

Three things have to hold. Inference demand keeps compounding at current rates. Gross margin survives the next round of token price cuts. Enterprise buyers treat October as a contained event.

Each one is testable. Watch the next revenue disclosure for the first. Watch the round size and the lead investor for the second. Watch renewal commentary from the design partners named in the Series D announcement for the third.

The allocator's checklist

Watch the round mechanics first. Size, lead investor, and whether Nvidia participates again matter more than the headline number. Nvidia joined the Series D alongside Atreides, Index Ventures and TCV. Its return would signal that the supply side still wants this layer financed.

Then watch disclosure discipline. Fireworks has not published a revenue figure since July. A new number separates a 40x multiple from a 25x multiple dressed as growth.

Then watch the top of the customer list. Cursor, Uber, Shopify and DoorDash are logos every inference vendor wants. Losing one is a data point. Losing two is a thesis problem.

Lin Qiao's argument is simple. Enterprises want their own models, and someone has to run them at scale with predictable latency. The reported valuation says investors believe that job is durable and scarce.

The October bulletin says trust is now part of the product. Price that correctly and the next round is defensible at almost any multiple. Ignore it and the term sheet arrives with a headline attached.