Fireworks AI, the inference startup Lin Qiao founded after leaving Meta's PyTorch team, raised $1.505 billion on July 16, 2026, at a valuation of 17.5 billion dollars. The company has crossed a billion dollars in annualized revenue and now serves 40 trillion tokens a day, nearly triple last year's total, a wager that businesses will keep paying to customize AI instead of renting it from the labs that build it.

In 2022, Lin Qiao left one of the most consequential engineering jobs in artificial intelligence. For seven years, as Meta's Senior Director of Engineering, she led more than 300 engineers who built and scaled PyTorch, the open-source framework that trains and runs a large share of the world's AI models, including systems at OpenAI and Google.

Qiao helped build the plumbing the entire industry now runs on. Then she left to compete with the assumption behind it: that a handful of labs should own the intelligence, and everyone else should rent it.

The startup she built to test that bet, Fireworks AI, announced a $1.505 billion Series D on July 16, 2026.

The round values the company at $17.5 billion.

Fireworks does not train a frontier model of its own. It leases graphics processing units, then helps other companies fine-tune open-weight models, such as GLM-5.2, Llama, or DeepSeek, on their own private data before serving the results in production. Fine-tuning means adjusting a pre-trained model's parameters on a narrower, proprietary dataset so it performs better on one company's specific tasks than a generic version would.

That business has just crossed a billion dollars in annualized revenue, and it now handles more than 40 trillion tokens a day, up from roughly 15 trillion a day about a year earlier. A token is the small chunk of text, often a few characters or part of a word, that a language model reads and generates one piece at a time. The more tokens a platform processes, the more AI work is actually running through it.

Ninety-five percent of that volume now comes from models customers have customized on their own data, not off-the-shelf frontier systems, according to the company.

The raise is bigger than one company's cap table. Some of the most active firms in venture capital are betting that running other companies' AI models, not building new ones, is now a business as large as the labs themselves. That bet lands at an uncomfortable moment. OpenAI, Anthropic, and Google are all expanding their own hosted inference products, and Microsoft and Amazon are folding managed open-model hosting directly into their cloud platforms. Whether an independent inference layer can survive being squeezed from both directions is the question hanging over the entire round.

Three Firms Led the Round. Fourteen More Followed.

Atreides Management, Index Ventures, and TCV led the Series D. NVIDIA, an existing investor, returned for another round.

The rest of the syndicate reads like a cross-section of venture capital's biggest names right now:

20VC
Bessemer Venture Partners
Evantic Capital
Insight Partners
Lightspeed Venture Partners
Lone Pine Capital
Menlo Ventures
Operator Collective
Ontario Teachers' Pension Plan
Original Capital
Prysm Capital
Quantum Capital
TIME Ventures

Fireworks says the new capital will expand its global compute capacity, grow its engineering team from around 200 people toward 600 by the end of 2026, and deepen partnerships with cloud providers including Microsoft and NVIDIA.

"Fireworks has assembled one of the most elite and technical teams in AI, paired with technology that consistently sets the pace for the industry and commercial momentum that very few companies have ever achieved at this scale," said Gavin Baker, chief investment officer and managing partner at Atreides Management, in the funding announcement. "We believe both frontier and open models will increasingly be used together."

Qiao framed the bet in stark terms of her own.

"There are two paths forward for AI. In one, intelligence belongs to a few big labs, and everyone else rents it. In the other, every company in the world builds specialized intelligence of its own, shaped by the domain only it understands. We are building towards the second." — Lin Qiao, co-founder and CEO, Fireworks AI

Renting Intelligence Versus Owning It

Fireworks' pitch is that a bank, a hospital system, or a law firm holds knowledge no general-purpose chatbot has: years of internal documents, customer records, and house style. Feed that data into an open-weight model and run it on Fireworks' infrastructure, the company argues, and the result beats a generic frontier model on the tasks that matter to that one business, at a lower cost per query.

Fireworks' enterprise customers include Uber, Shopify, and Doximity, according to the company. Its most concentrated relationship, however, has been with Cursor, the AI coding tool. As recently as last year, Cursor supplied roughly half of Fireworks' revenue.

That concentration became a live risk in April 2026. Cursor's board gave Elon Musk's SpaceX the option to acquire the company outright for $60 billion. Let's Data Science covered that deal in detail at the time.

Alongside the acquisition option, SpaceX struck an immediate compute partnership worth $10 billion, under which Cursor began training its Composer model on xAI's Colossus supercomputer rather than buying all of its inference from outside vendors.

SpaceX did not let the option lapse. In June, days after its own initial public offering, the company converted it into a definitive agreement to acquire Cursor outright for $60 billion in stock, a deal both sides expect to close in the third quarter of 2026.

Tech Funding News reported in July that Cursor's shift toward xAI's compute is already reducing its reliance on providers like Fireworks. Qiao has said the company's customer base is now more diversified, and its broader enterprise roster supports that claim, though a future Fireworks public listing would settle the question with real numbers.

The Token Math That Turned Heads

The number that made Wall Street pay attention was not the valuation. It was the traffic.

CNBC compared Fireworks' disclosed volume against figures Google and OpenAI have published for their own developer platforms. Google said in May that its models were processing about 19 billion tokens a minute for developers, which implies more than 27 trillion tokens a day. OpenAI said in March that its developer tools were handling roughly 15 billion tokens a minute, implying about 22 trillion a day.

Fireworks' 40 trillion tokens a day is larger than either figure, according to CNBC's analysis, even though the company is a fraction of Google's or OpenAI's size by revenue. The comparison is specific to developer and API traffic that Google and OpenAI have chosen to disclose, not the far larger volume of consumer chat traffic those companies also handle. Still, for a company most consumers have never used directly, it is a striking share of the AI industry's actual computation passing through one vendor's infrastructure.

A Crowded, Increasingly Expensive Neighborhood

Fireworks is not the only inference company commanding a double-digit-billion valuation this year. Together AI and Baseten, its two closest rivals, have both raised recently at prices that assume similar growth will continue.

Company	Last Round	Amount Raised	Valuation
Fireworks AI	Series D, July 2026	$1.505 billion	$17.5 billion
Together AI	Series C, July 2026	$800 million	$8.3 billion
Baseten	Series F, June 2026	$1.5 billion	$13 billion

At more than a billion dollars in trailing annualized revenue, Fireworks' valuation works out to roughly 17.5 times revenue, a multiple the research firm Sacra calculates using the same figures the company disclosed.

Groq, which runs inference on its own custom chips rather than renting NVIDIA GPUs, raised $650 million in June 2026 to rebuild after NVIDIA paid to license its chip architecture and hired several of its executives. Meanwhile, AWS and Microsoft Azure have each built managed hosting for customized open-weight models directly into their cloud consoles, the same service Fireworks sells as a standalone product.

Fireworks itself has written previously about serving open models like GLM-5.2 within hours of release, and it is one of the providers that made DeepSeek V4's steep price cuts available to enterprise customers earlier this year. The chip side of that same buildout has drawn its own money: Etched, which makes chips built only for inference, booked a billion dollars in orders in a similar bet that serving models cheaply is worth as much as training them.

Not Everyone Is Convinced

The skepticism predates this specific round. In May, when Fireworks was reportedly in talks to raise at a $15 billion valuation, the newsletter Newcomer reported that investors were already debating whether inference companies deserved their prices.

"It seems like VCs are just doing a revenue multiple and are assuming the margin doesn't matter," one investor told Newcomer.

The concern is structural. Fireworks, like Together AI and Baseten, leases its computing capacity rather than owning it. Hyperscalers and the AI labs themselves do not pay to rent the chips they run on, which gives them a cost advantage inference resellers cannot easily match. Newcomer's reporting also noted that Fireworks competes directly with OpenAI and Anthropic for scarce GPU allocation, even as it depends on those same companies' commercial success to keep demand for open-model alternatives growing.

Sacra's research raises a related concern it calls hyperscaler capture: as cloud providers absorb model hosting, fine-tuning, and agent deployment into single platforms such as AWS Bedrock and Microsoft's Azure AI Foundry, a specialized vendor like Fireworks risks being squeezed into an optional add-on rather than remaining the primary place enterprises run AI. Tech Funding News made a similar point after the round closed, noting that NVIDIA and Microsoft are simultaneously Fireworks' investors and, in different ways, its future competitors.

Fireworks' answer is that customization depth, not raw speed, is the durable advantage. Any customer can rent a faster chip, but building the tooling that turns a company's own data into a working specialized model is harder to copy. Whether that holds once hyperscalers bundle similar tools for free is, by Tech Funding News's own account, the question the next year of enterprise AI spending will answer.

The Bottom Line

Strip away the venture math, and the round says something simple: serving other companies' AI models is now a business large enough to mint its own decacorns, separate from the labs that build the models in the first place. Fireworks crossed a billion dollars in revenue, nearly tripled its daily traffic in about a year, and got investors to pay more than seventeen times that revenue for a piece of it.

None of that resolves the harder question sitting underneath the deal. Fireworks does not own the chips it runs on, and both its cloud partners and its largest customer have shown they are capable of building around it rather than through it. The company is betting that enterprises will keep needing a specialist to turn open models into something proprietary, even as the biggest labs and the biggest clouds each try to sell that same service themselves.

Qiao framed the choice in two sentences when she announced the round: intelligence belongs to a few big labs and everyone else rents it, or every company builds its own. Investors just backed the second path with 1.505 billion dollars. Whether enough customers agree to keep Fireworks independent is a bet nobody, including Qiao, has fully settled yet.