The Inference Layer Cashes In: Fireworks Eyes $30 Billion as Fal Pitches $15 Billion on an $800 Million Run Rate

The AI funding cycle has a new center of gravity. On Friday, The Information’s Stephanie Palazzolo reported that two of the busiest inference providers in the market are heading back to investors with dramatically higher asks: Fal has spoken to investors about raising at a valuation of more than $15 billion, while Fireworks AI has considered a new round at a $30 billion target — up from the $17.5 billion mark it set just two months ago in a $1.505 billion Series D.

The numbers tell a coherent story about where AI money is flowing right now. The model layer gets the headlines, but the picks-and-shovels layer that actually serves tokens — image, video, audio, and text — is where revenue is compounding fastest. Palazzolo’s dispatch notes that revenue at Fal has hit an $800 million annualized pace, a figure that has doubled roughly every six months: $200 million in October 2025, $400 million by March 2026, and now double that again by September.

What these companies actually do

The category is easy to conflate with “AI cloud,” but the business models diverge in ways that matter for the valuations being discussed.

Fireworks AI, founded by former Meta engineering lead Lin Qiao, sells inference for open-weight and custom models — the infrastructure that lets enterprises run Llama-family models, fine-tuned variants, and compound AI systems without owning GPUs or negotiating hyperscaler contracts. Its July Series D was led by Atreides Management, Index Ventures, and TCV, with Nvidia among its backers. The two-step from $4 billion (October 2025 Series C) to $17.5 billion (July 2026) to a contemplated $30 billion (September 2026) is among the fastest markup ladders in the industry — if this round closes, the company will have multiplied its valuation roughly 7.5x in twelve months.

Fal operates one layer over in generative media. Its platform hosts more than a thousand image, video, 3D, and audio models — FLUX, Kling, Hailuo among them — behind a single API, and has become the default integration layer for developers building generative-media features into products. Its last confirmed round was a $140 million Series D in December 2025 at a $4.5 billion valuation, led by Sequoia with Kleiner Perkins and Nvidia participating. A $15 billion ask would be a 3.3x markup on that December mark — and a near-2x markup on the $8 billion figure Fal was discussing with investors as recently as March.

The inference gold rush, quantified

These two data points are the latest entries in what has become the defining private-market trade of 2026: the inference markup wave.

Baseten set the pace in June, closing a $1.5 billion Series F at a $13 billion valuation — triple its January mark — with the company disclosing that inference volume on its platform had grown 40x. Crusoe’s managed inference business crossed $100 million in contracted ARR less than a year after launch, on its way to a $65 million Thinking Machines Lab deal. SemiAnalysis has tracked Fireworks alone processing more than 40 trillion tokens in a single day — roughly double the entire OpenAI API’s traffic as of late March.

The demand side is doing most of the work. Every agent loop, every video generation call, every fine-tuned model deployed to production translates into sustained, compounding inference volume. Unlike training runs, which are episodic and concentrated among a handful of frontier labs, inference is diffuse and recurring — the utility-style revenue profile that private credit investors in particular have been hunting for across the AI infrastructure stack. That is precisely the cohort now bidding up the inference layer: income-oriented funds that want contracted cash flows, not venture home runs.

Why the numbers invite skepticism

A $30 billion valuation on Fireworks implies enormous conviction. Consider the arithmetic: at its last disclosed pace, the company’s annualized revenue was reported past $1 billion with daily token volume still climbing. Even granting strong growth, a $30 billion mark implies a revenue multiple somewhere north of 20x — rich for infrastructure, cheap for software, and unusual for a business whose gross margins depend on GPU economics it does not fully control.

The bear case writes itself in three parts. First, the hyperscalers are not standing still: AWS, Google Cloud, Azure, and Oracle all sell competing inference services, often at aggressive prices, and they own the silicon supply chain. Second, the open-weight model market is fluid — the marginal cost of serving a Llama-class model keeps falling, and OpenRouter-style routing means customers can shift workloads to whichever backend is cheapest this week. Third, Nvidia is both kingmaker and competitor: it backs Fireworks, Fal, and Baseten while also selling its own inference stack (NIM microservices, Launch platform) — a position that has helped every portfolio company raise, without guaranteeing any of them durable margin.

The bull case is equally straightforward. Enterprise AI adoption is still in its ramp phase, agentic workloads multiply token consumption per user by an order of magnitude over chat, and the inference specialists have proven they can win workloads the hyperscalers are too slow or too generic to serve well. Fal’s doubling revenue every six months and Fireworks’ token volumes are not projections — they are realized demand curves.

What to watch

None of these rounds is confirmed. The Information’s reporting is explicit that talks are exploratory: Fal has “spoken to investors,” Fireworks has “considered” the $30 billion target, and terms could change or collapse entirely — as Binance’s aggregation of the report noted, the financing “has not been completed, and the final details could still change.”

But the direction is unmistakable. In the span of a single quarter, the inference layer has gone from infrastructure afterthought to the most actively marked-up asset class in private AI. If Fireworks closes anywhere near $30 billion, it becomes one of the most valuable private AI infrastructure companies in the world — and the next round of marks for Together AI, Modal, DeepInfra, and the rest of the cohort will be set against it.

The inference gold rush is no longer about who found gold. It’s about who gets to price the mine.