Fireworks raises $1.505B as Lin Qiao scales open-model inference

Why it matters

Qiao is betting that enterprises will own customized models while paying Fireworks to operate them. The $1.505 billion round gives her scale, but margins and workload portability will decide whether inference becomes durable infrastructure or a price-driven commodity.

Lin Qiao (@lqiao) has raised $1.505 billion for Fireworks AI at a $17.5 billion valuation, giving the former PyTorch leader a large pool of capital to expand the compute and engineering behind customized AI models.

Fireworks announced the Series D on July 15th. Atreides Management, Index Ventures and TCV led the financing, with Evantic Capital, Lightspeed Venture Partners, Nvidia, 20VC, Bessemer Venture Partners and Menlo Ventures among the participants.

A Forbes analysis placed the deal alongside $1.5 billion for Baseten and $800 million for Together AI. Those three inference providers raised roughly $3.8 billion within four weeks, concentrating investor attention on the infrastructure that sits between model developers and AI applications.

Qiao previously headed PyTorch at Meta. She founded Fireworks in 2022 with six other infrastructure veterans, most of whom had worked on PyTorch or large-scale machine learning systems at Meta and Google. Fireworks' founding group includes former PyTorch core maintainer Dmytro Dzhulgakov, former PyTorch compiler engineer James Reed and former Google Vertex AI lead Chenyu Zhao.

Their founding thesis was that enterprises would eventually want intelligence adapted to their own data, workflows and customer interactions. In Fireworks' October 2025 Series C announcement, Qiao wrote that thousands of businesses would seek to own and customize their AI products rather than leave that intelligence entirely in the hands of a few model labs.

The Series D is the clearest financial test of that thesis so far. Fireworks says it has surpassed $1 billion in annualized revenue run rate, processes more than 40 trillion tokens each day and grew fivefold over the prior year. Fireworks also says over 95% of those tokens come from models specialized using customer data, a category that includes fine-tuned open weights, adapters, distillations and customer-trained models.

Those figures are self-reported and unaudited. Annualized run rate extrapolates recent performance rather than recording a completed year of revenue. At the stated run rate, the financing values Fireworks at roughly 17.5 times annualized revenue.

Fireworks sells everything around the weights

An open-weight model provides trained parameters. Running those parameters as a dependable product still requires continuous batching, caching, quantization, autoscaling, observability, security controls and billing. Fireworks sells that operating layer, allowing customers to adapt models and expose them through production endpoints without buying and managing their own GPU fleets.

That division of labor is becoming more important as AI applications move from occasional prompts to agents that call models repeatedly while reading files, planning tasks, using tools and checking their work. Gartner forecasts that inference on a 1 trillion-parameter model will cost providers over 90% less in 2030 than in 2025. Gartner also estimates that agentic systems can consume five to 30 times as many tokens per task as a standard chatbot.

That combination creates a large market for Qiao if Fireworks can keep reducing the cost of each request while usage climbs. It also creates relentless price pressure. Lower token costs make it easier for customers to move workloads or demand discounts, while cloud providers can bundle inference with compute, storage and existing enterprise contracts.

Qiao is using the round to expand Fireworks' compute infrastructure and engineering organization. Fireworks names Cursor and Harvey among customers building specialized models on its infrastructure, though Fireworks has not disclosed how much revenue comes from inference alone, how revenue is distributed across large customers or what portion is tied to minimum-spend contracts.

The race has become capital intensive

Fireworks is scaling against rivals that have raised comparable sums and are making similar ownership arguments.

Baseten's June 22nd Series F brought in $1.5 billion at a $13 billion headline valuation. Baseten said its revenue increased 20-fold and inference volume rose 40-fold over the prior year, without disclosing the underlying revenue figures. Baseten combines managed inference with post-training and engineers who work directly with customers on model optimization.

Together AI raised $800 million on July 1st at an $8.3 billion valuation, with Aramco Ventures among its investors. Together AI spans GPU capacity, training, fine-tuning and inference, giving it a wider infrastructure footprint than a serving-only provider. Together AI says its annual bookings exceeded $1.15 billion in the preceding quarter. Bookings can include contracted future work and are broader than recognized revenue.

OpenRouter raised $113 million in May for a different position in the same market. OpenRouter routes requests across models and inference providers, generally without owning the underlying compute. That brokerage model can help customers compare prices and switch providers, placing another layer between application developers and businesses such as Fireworks.

Each approach seeks control over where AI workloads run and how buyers pay for them. Fireworks wants to own the optimized runtime for specialized models. Baseten couples serving with hands-on model engineering. Together AI sells a broader compute and model-development stack. OpenRouter controls routing and provider selection.

Qiao still has to prove the margins

Fireworks does not disclose gross margin. Sacra estimates Fireworks operates at roughly 50% gross margin and has targeted 60%, reflecting the cost of GPUs and data-center capacity inside every token served. It is an outside estimate, not a Fireworks disclosure. Conventional subscription software can operate above 70%, leaving inference providers with less room for inefficient hardware utilization or aggressive price cuts.

The Series D gives Qiao capital to buy capacity, hire engineers and optimize Fireworks' stack before inference prices fall further. It also increases the performance expected from Fireworks. A $17.5 billion valuation rests on sustained token growth, improving utilization and Fireworks retaining customers even as open models and cloud services become easier to deploy.

Open weights alone do not guarantee an open production stack. A customer can own a model's parameters while becoming dependent on a provider's deployment format, optimization tools, reserved capacity and monitoring systems. Enterprise buyers will need to examine portability, hardware assumptions, minimum commitments and migration costs with the same care they apply to model quality.

RuntimeWire recently examined how agent gateways are becoming an enterprise control layer. Fireworks is making a related bid lower in the stack. Serving infrastructure determines which models can run economically, how quickly customers can swap them and how much each agent action costs.

Qiao's PyTorch background gave Fireworks credibility with engineers who already understand the burden of operating machine learning systems at scale. The next part of her bet is commercial: turning that technical position into durable customer ownership before hyperscalers, model labs and routing services compress the value of the layer Fireworks has built.