Menlo’s Investment in Fireworks: The Runtime for Specialized Intelligence

AI inference is quickly becoming one of the largest markets in the world, and Fireworks is leading the expansion with its platform for specialized intelligence and low-cost token production. Today, Menlo is proud to be partnering with Fireworks in its $1.5 billion Series D.

As the leading AI labs push the capability frontier forward, they are opening another massive market in parallel: a separate frontier for specialized intelligence, where speed, cost, and control matter as much as raw capability.

This second front came into sharp focus in 2026, when models gained the ability to sustain long-running tasks without constant human attention. That shift unlocked entirely new classes of economically valuable work—and pushed one of computing’s fastest-growing markets onto an even steeper trajectory.

Inference demand didn’t just grow; it broadened. Open-source usage on OpenRouter has expanded more than 10x since the start of the year. And while frontier models remain essential, the workloads they unlocked also created demand for a much wider range of inference shapes—models that are faster, more economical, or fine-tuned to a company’s own data, workflows, and performance requirements.

Fireworks is building the runtime for this world, where custom and open-source models increasingly work alongside frontier models in production. The platform’s growth reflects the scale of this emerging production layer: Daily token volume has nearly tripled since late last year, from 15 trillion to 43 trillion, while annualized revenue recently reached $1 billion.

The Leading Platform for Open Models

Fireworks begins before the first production token is served. The platform brings continued pretraining, supervised fine-tuning, and reinforcement learning onto the same stack as inference. Teams can start with an open model, adapt it to their proprietary data and product feedback, deploy it immediately, and continue improving it from real-world usage. Training and serving become one continuous loop.

Once a model reaches production, inference is more than just weights in a box. Running a model in production requires specialized engineering across hundreds of thousands of possible combinations of hardware, quantization, sharding, speculative decoding, batching, and kernels. The right configuration changes with the workload—and with each customer’s priorities across price, latency, and quality. Fireworks’ proprietary FireAttention stack automates that search, extracting more performance and better economics than any other provider.

The result is a single system for turning proprietary data into custom weights, and custom weights into production intelligence with the most efficient token delivery. Many of the AI applications with the largest, most sophisticated needs in the world choose Fireworks: Cursor trained Composer 2, its frontier-level coding model, on the platform. Vercel delivered 40x improvements in latency to v0 users by partnering with Fireworks on reinforcement fine-tuning and speculative decoding. And Factory is using Fireworks to give customers up to 15x more work for the same spend with open-source options.

A World-Class Infrastructure Team

Few teams are better suited to build this layer. CEO Lin Qiao previously led PyTorch at Meta, overseeing the development and productionization of the open-source framework that became foundational to modern AI. CTO Dmytro Dzhulgakov was one of PyTorch’s core maintainers and a senior leader within Meta’s AI organization.

The rest of Fireworks’ seven-person founding team is similarly formidable: former leaders of Meta’s ads infrastructure, News Feed machine learning, PyTorch ranking systems and compiler development, alongside the former AI lead of Google Vertex. Collectively, they helped build the infrastructure supporting some of the world’s largest AI systems, from the kernel layer up.

More recently, Fireworks added to its leadership president George Hu, who previously helped scale Salesforce 50x to $5 billion, before leading Twilio’s 10x growth journey. His arrival gives Fireworks the operating muscle to match the size of its ambitions. 

Building in the Token Path

The AI infrastructure stack is reorganizing around the token path: the compute, data, and orchestration that turn intelligence into a product. At Menlo, we believe AI’s most enduring infrastructure companies will be built along this value chain.

Our portfolio reflects that conviction: from Anthropic at the capability frontier, to OpenRouter in routing, Gimlet and Modal in compute, Neon and Pinecone in databases, and Unstructured in data infrastructure—with more to be announced soon, as we are actively partnering with founders building the defining building blocks of the emerging AI stack.

Fireworks sits at the center of our thesis. We could not be more excited to partner with Lin, Dmytro, George, and the entire Fireworks team as they expand and accelerate this market for the next generation of AI applications and developers.