---
title: "Fireworks AI wants $40 billion for tokens nobody can publicly count"
description: "The open-model inference cloud is in talks to roughly double its valuation three months after a $17.5 billion Series D, on a $1 billion ARR whose claimed token volume dwarfs what public trackers show. The gap has structural explanations, but no public reconciliation, and the late-round buyer carries the risk."
publisher: "The Inference"
section: "Deals"
published: 2026-10-28T15:05:14.277Z
modified: 2026-10-28T15:05:14.277Z
canonical: https://theinference.org/article/fireworks-ai-wants-40-billion-for-tokens-nobody-can-publicly-count
language: en
keywords: "Private Markets, AI Infrastructure, Inference, Valuation, Venture Capital"
---

# Fireworks AI wants $40 billion for tokens nobody can publicly count

> The open-model inference cloud is in talks to roughly double its valuation three months after a $17.5 billion Series D, on a $1 billion ARR whose claimed token volume dwarfs what public trackers show. The gap has structural explanations, but no public reconciliation, and the late-round buyer carries the risk.

[Fireworks AI](https://theinference.org/markets/companies/fireworks-ai) is [in early talks to raise at a valuation of $30 billion to $40 billion](https://www.businessinsider.com/fireworks-ai-is-seeking-new-funding-at-around-valuation-2026-10), roughly double the price it fetched in July [1]. The step-up rests on a claim the company put on the record this week: [more than $1 billion of annualized revenue](https://aimagazine.com/news/fireworks-ai-inference-open-models-and-us-1bn-arr) [2]. Its other headline number is 40 trillion tokens served each day [2].

What anyone can count of that volume on OpenRouter, the public routing tracker, is 40 to 80 billion tokens a day [3]. About a thousandth of the claim is publicly visible, and nobody has reconciled the rest. [The last time we covered Fireworks, in September, the story was cheaper tokens](https://theinference.org/article/fireworks-ai-cuts-the-token-meter-with-ember-1); now it is the price attached to them.

## The meter is dark for structural reasons. Dark is not counted.

[OpenRouter's methodology states that requests a user keeps private are excluded](https://openrouter.ai/rankings) before its rankings are computed [4]. That matters because 95% of Fireworks' tokens come from models customers specialized on their own data, served on private enterprise endpoints rather than through a public aggregator [2]. Some volume bypasses routers entirely: Forum Edge AI, a joint venture formed this month, handed Fireworks dedicated [Nvidia](https://theinference.org/markets/companies/nvidia) B300 capacity at its Dallas data center under a [$94.4 million, three-year contract](https://cryptobriefing.com/forum-edge-ai-fireworks-nvidia-b300-deal/) [5].

[Together AI](https://theinference.org/markets/companies/together-ai) shows the same pattern. [It claimed about 13 trillion tokens a day in August](https://x.com/ai_canaries/status/2114391248047841441) [3]. OpenRouter showed 65 billion [3]. So the gap has a mechanical explanation. What it lacks is a receipt: no contract list, no audited revenue, no third-party count matching the claim.

*Chart: **Reported daily token volume runs about a thousand times above what public trackers show** Bars chart comparing Fireworks and Together AI reported daily token volumes with what OpenRouter's public rankings show, in trillions of tokens per day*

*Daily token volume as reported by each company against the volume visible in OpenRouter's public rankings, in trillions of tokens per day; Fireworks' visible range of 40 to 80 billion is drawn at its midpoint. Dashed bars are company-reported figures, not independently verified. Sources: Fireworks and Together AI statements, via AI Magazine and X; OpenRouter rankings. [2][3]*

## The buyer underwrites a number; everyone else gets a counter

The late-round investor is being asked to pay roughly 22 times projected year-end revenue [6]. In July it priced the same company at $17.5 billion, on a Series C of $4 billion the year before [1]. Buyers are weighing that step-up weeks after Fireworks disclosed that an intruder used internal credentials to reach some customer accounts [1].

*Chart: **Three months doubled the private mark, to a range that tops Nebius' market value** Range chart of Fireworks AI's Series C price in 2025, its Series D price in July 2026, the reported $30 billion to $40 billion range in current funding talks, and Nebius' market value near $40 billion*

*Fireworks AI's Series C (2025) and Series D (July 2026) prices, the $30 billion to $40 billion range reportedly under early discussion, and listed peer Nebius' market value. The talks range is unagreed and drawn from a single report; the Nebius figure is market commentary, not a filing. Sources: Business Insider; X (AndersInvests). [1][6]*

Listed counterpart [Nebius](https://theinference.org/markets/companies/nebius), which owns much of the infrastructure Fireworks rents, is [marked near $40 billion of market value](https://x.com/AndersInvests/status/2113914977132097749) [6]; if the private mark holds, earlier holders are rewarded and comparables like Nebius would reprice with it. The load-bearing hop in Fireworks' chain is real and observable: enterprises fine-tune open models and serve them in production, [Notion](https://theinference.org/markets/companies/notion) cut latency on its AI features to 350 milliseconds on the platform, and [Uber](https://theinference.org/markets/companies/uber), [Shopify](https://theinference.org/markets/companies/shopify) and Cursor are named customers [2]. Where observation ends is the revenue line.

The inference layer deserves its premium, and the open-model path is the one worth building on. But a buyer paying double July's price should underwrite a receipt, not a token counter: contracted, audited revenue, checked before the wire moves.

## Takeaway

Fireworks AI is shopping a $30 billion to $40 billion round on a $1 billion ARR and 40 trillion claimed daily tokens, but public trackers show about a thousandth of that volume. The gap has structural explanations, yet no contract list, audit or third-party count stands behind the claim, and the late-round buyer is being asked to pay 22 times projected year-end revenue on faith.

## Sources

1. [Business Insider, Infrastructure startup Fireworks AI is in early talks for new funding at a valuation between $30 billion and $40 billion, 20 October 2026](https://www.businessinsider.com/fireworks-ai-is-seeking-new-funding-at-around-valuation-2026-10)
2. [AI Magazine, Fireworks AI: Inference, Open Models and US$1bn ARR, October 2026](https://aimagazine.com/news/fireworks-ai-inference-open-models-and-us-1bn-arr)
3. [X (Ai Canaries), post on Fireworks and Together AI token volume versus OpenRouter visibility, October 2026](https://x.com/ai_canaries/status/2114391248047841441)
4. [OpenRouter, LLM Rankings, methodology section, accessed 28 October 2026](https://openrouter.ai/rankings)
5. [Crypto Briefing, Forum Edge AI signs $94.4 million Fireworks deal for Nvidia B300 GPU capacity, 26 October 2026](https://cryptobriefing.com/forum-edge-ai-fireworks-nvidia-b300-deal/)
6. [X (AndersInvests), post on the Fireworks and Nebius valuation gap, 24 October 2026](https://x.com/AndersInvests/status/2113914977132097749)


---

Fireworks AI wants $40 billion for tokens nobody can publicly count — The Inference. Canonical: https://theinference.org/article/fireworks-ai-wants-40-billion-for-tokens-nobody-can-publicly-count