Fireworks: the inference layer for the open model era

Every era of computing has its defining unit of production, and in the AI economy, that unit is tokens. Demand for them is exploding faster than any input in the history of technology. Hyperscalers are on pace to spend more than $800 billion on capital expenditures in 2026 alone, and analysts project trillions more through the end of the decade. 

Fireworks is at the center of the largest infrastructure build-out in history. Founded by the team that built and scaled PyTorch at Meta, Fireworks is the frontier training and inference platform for open models. Today, they process ~43 trillion tokens per day, making it one of the largest inference providers. Token consumption is expected to climb over 30x by the end of the decade. This is why we’re proudly joining Fireworks' $1.5B Series D, deepening our investment in the token economy.

The open model inflection

In the early years of the generative AI era, production AI meant expensive, closed frontier APIs. That's changing fast. Open-weight models now match frontier quality on the workloads most enterprises care about and at a fraction of the cost. With inference now a CFO-level line item, AI-native companies are optimizing for privacy, speed, quality, and cost. Increasingly, the solution is an open model post-trained on proprietary data. 

Companies don't want to “rent” their core intelligence. The vast majority of valuable data lives inside the enterprise itself, not on the public internet. These enterprises increasingly want to keep their data assets private while using their own specialized models.

The frontier platform for training and inference

Fireworks built the platform that makes that future practical: a virtual GPU cloud spanning more than a dozen cloud providers and over 20 regions, a proprietary inference engine and optimization system that serves any model at maximum speed and efficiency, and a training platform for fine-tuning and reinforcement learning. Because training and inference live in one place, every product interaction can feed the next version of the model, creating a flywheel that compounds with usage.

Fireworks just made one of the fastest-scaling stories in software history, growing from $100M to $1B+ in ARR in only 16 months. The same climb took many of our most successful investments of the prior SaaS generation, from Twilio to Shopify, four to five years. Fireworks compressed half a decade of growth into less than six quarters. 

Our diligence, among many of Fireworks's largest customers, confirmed the platform is the fastest to bring state-of-the-art open models to production, the strongest fine-tuning and RL workflows, and the most technical team customers have ever worked with. It's why Microsoft chose Fireworks as one of its few true first-party partners, shipping natively inside Microsoft Foundry on Azure. We've seen how transformative hyperscaler distribution can be. When we first invested in Anthropic in 2024, the AWS partnership gave the business immediate reach at massive scale. Fireworks' Azure partnership similarly positions it for day-one distribution across one of the deepest enterprise customer bases in the world. 

Our commitment to the AI infrastructure layer

At Bessemer, our conviction in Fireworks starts with the team. CEO Lin Qiao led the engineering organization at Meta that built the open-source framework (PyTorch) underpinning much of modern AI, and her co-founders share that pedigree, giving Fireworks the deepest bench of inference talent in the market. In April 2026, George Hu joined as President, one of the few executives anywhere to have scaled two iconic software companies. As COO, he helped scale Salesforce to billions in revenue and drove 10x growth at Twilio. As George put it when he joined, "inference is poised to be one of the biggest markets in history." Having worked with George during his Twilio years, we had first-hand experience with his GTM genius and believe it will be instrumental in enabling Fireworks' continued scaling.

Inference tomorrow will look very different from today: long-running, massively parallel agentic workloads on a substrate fragmenting across heterogeneous silicon, hyperscalers, neoclouds, and air-gapped environments. Fireworks’ vision is to become the software layer that abstracts it all away, the inference platform running on every GPU in the world, so every company can build, own, and continuously improve its own AI.

This partnership furthers Bessemer's long-standing commitment to the AI infrastructure layer. The largest AI outcomes accrue from the platforms that lead during the formative window, and in the open model era, we believe Fireworks is that platform for inference. We're thrilled to be a part of the journey with Lin, George, and the entire Fireworks team as they fuel the engine of the token economy.