Six months ago, Fireworks AI was a promising but largely anonymous inference startup. Today it sits at a $2 billion valuation, its CEO is fielding calls from the largest enterprises on the planet, and its revenue has surged more than tenfold in a single year. The question confronting Silicon Valley isn’t whether AI inference — the process of actually running trained models to produce outputs — matters. It’s whether Fireworks can hold its ground as every major cloud provider scrambles to own the same turf.
Lin Qiao doesn’t look like someone who just closed one of the most talked-about funding rounds in enterprise AI. In a recent interview with Business Insider, the Fireworks AI cofounder and CEO described a company that has grown from relative obscurity into a critical piece of infrastructure for businesses deploying large language models at scale. Fireworks raised $552 million in a Series D round led by Greenoaks Capital Partners in early 2026, catapulting it to a $2 billion valuation and placing it among the most richly funded pure-play inference companies in the world.
The numbers are striking. Fireworks reported that its annualized revenue run rate has grown more than 10x year over year, though the company has not disclosed a precise dollar figure. Its customer base now includes major enterprises across financial services, healthcare, and technology. And its platform processes billions of inference requests daily — a volume that would have been unthinkable for a startup of its size just 18 months ago.
But here’s what makes Fireworks genuinely interesting, and what separates it from the dozens of AI infrastructure companies chasing venture capital: it doesn’t train models. Not a single one. Fireworks is entirely focused on the deployment side — taking models built by OpenAI, Meta, Mistral, Google, and others, and running them faster and cheaper than the companies that created them. It’s a bold positioning choice. And it’s working.
Lin Qiao’s path to this moment is anything but typical. A former engineering leader at Meta, she spent years working on PyTorch, the open-source machine learning framework that became the backbone of modern AI research. She saw firsthand how much engineering effort went into training models, and how comparatively little attention was paid to running them efficiently in production. “Training gets all the headlines,” she told Business Insider. “But inference is where the actual value gets delivered to users.”
She’s right, at least by the numbers. Industry analysts estimate that inference now accounts for roughly 60% to 80% of total AI compute spending across major enterprises, a figure that continues to climb as companies move from experimentation to production deployment. Training a frontier model is a one-time — if extraordinarily expensive — event. Running that model millions of times a day, for millions of users, is a permanent cost center. And it’s the cost center where Fireworks has planted its flag.
The company’s technical approach centers on what it calls a “compound AI” architecture, which allows customers to chain together multiple models, retrieval systems, and custom logic into a single inference pipeline. This matters because most real-world AI applications don’t rely on a single model call. A customer-service chatbot might need to retrieve relevant documents, run a classification model, generate a response with an LLM, and then check that response against safety filters — all within a few hundred milliseconds. Fireworks optimizes this entire chain, not just the individual model calls.
Speed and cost. Those are the two metrics that enterprise buyers care about most when selecting an inference provider, and Fireworks has engineered its stack to win on both. The company claims latency figures that are 2x to 4x faster than running equivalent models through major cloud providers’ default inference endpoints. On cost, Fireworks says it can deliver inference at a fraction of the price of hyperscaler alternatives, largely because it has built custom kernel-level optimizations and uses a mix of GPU types — including Nvidia’s H100s and AMD’s MI300X chips — to match workloads to the most cost-effective hardware.
That AMD detail is worth pausing on. Nvidia dominates the AI chip market so thoroughly that any company willing to run production workloads on non-Nvidia silicon is making a meaningful technical and strategic statement. Fireworks has been one of the most vocal proponents of AMD’s data center GPUs, and the relationship appears to be mutually beneficial: AMD gets a credible reference customer for its AI accelerators, and Fireworks gets access to chips that are both cheaper and more readily available than Nvidia’s perpetually supply-constrained offerings.
The competitive dynamics around inference are intensifying. Amazon Web Services, Google Cloud, and Microsoft Azure have all made significant investments in their own inference infrastructure, recognizing that whoever controls the inference layer controls the recurring revenue stream from AI deployment. AWS launched its custom Inferentia chips specifically to capture this market. Google has its TPUs. Microsoft has tightly integrated OpenAI’s models into Azure with optimized inference endpoints.
So why would an enterprise choose Fireworks over its existing cloud provider? Lin Qiao’s answer, as relayed to Business Insider, is model flexibility and vendor neutrality. Enterprises increasingly want to run different models for different tasks — a smaller, faster model for simple classification, a larger model for complex reasoning, an open-source model for tasks where they need full control over the weights. The hyperscalers tend to steer customers toward their own models or their closest partner’s models. Fireworks is agnostic. It runs over 100 models on its platform, and customers can switch between them without re-architecting their applications.
This model-agnostic approach has resonated particularly well with regulated industries. Banks, insurers, and healthcare companies often can’t afford to be locked into a single model provider, both for risk management reasons and because regulatory requirements may dictate that they maintain alternatives. Fireworks has signed several large financial-services contracts in recent months, though the company declined to name specific clients.
The fundraising trajectory tells its own story. Fireworks raised a $25 million Series A in 2023, a $150 million Series B in mid-2024, and then the massive $552 million Series D. The company appears to have either skipped or quietly closed a Series C — a pattern increasingly common among fast-growing AI startups where investor demand outpaces the company’s original fundraising plans. Greenoaks, the lead investor in the latest round, has a reputation for backing infrastructure companies with strong unit economics, which suggests Fireworks’ margins are healthier than those of many AI startups burning through cash to subsidize usage.
Not everyone is convinced the standalone inference market will persist. Some investors and technologists argue that inference will inevitably be commoditized — that as hardware improves and open-source serving frameworks mature, the margins in running someone else’s model will compress to near zero. It’s the classic infrastructure bear case: what looks like a high-value chokepoint today becomes a commodity tomorrow.
Lin Qiao has heard this argument. She doesn’t buy it. Her counterpoint is that inference optimization is not a static problem. As models get larger, as multimodal capabilities expand, and as enterprises demand lower latency at higher throughput, the engineering challenges don’t get easier. They get harder. “The models keep changing,” she said. “The workloads keep changing. The hardware keeps changing. If you’re not continuously optimizing across all three, you fall behind.”
There’s evidence to support her view. The shift toward mixture-of-experts architectures, longer context windows, and real-time multimodal processing has created new inference bottlenecks that didn’t exist a year ago. Serving a model with a 128,000-token context window is a fundamentally different engineering problem than serving the same model with an 8,000-token window. Fireworks has invested heavily in techniques like speculative decoding, continuous batching, and dynamic routing to address these challenges.
The company has also moved aggressively into what it calls “function calling” and “structured output” capabilities — features that allow AI models to interact with external APIs, databases, and tools in a reliable, programmatic way. This is the plumbing that makes AI agents possible, and it represents an enormous market opportunity. If AI agents become the dominant interface for enterprise software — a possibility that virtually every major tech company is now betting on — then the inference provider that can reliably orchestrate those agent workflows will capture a disproportionate share of value.
Fireworks isn’t alone in chasing this opportunity. Together AI, Anyscale, Baseten, Modal, and Replicate are all competing in various segments of the inference and model-serving market. Together AI, in particular, has emerged as a formidable competitor, having raised substantial funding and built a similarly model-agnostic platform. But Fireworks’ 10x revenue growth and $2 billion valuation suggest it’s pulling ahead of the pack, at least for now.
The broader market context matters here. Enterprise AI spending is projected to exceed $300 billion globally by 2027, according to estimates from IDC and Gartner. But much of that spending is shifting from experimentation budgets to production infrastructure budgets — a transition that directly benefits companies like Fireworks. Proof-of-concept projects run on whatever’s convenient. Production systems run on whatever’s fastest, cheapest, and most reliable. That’s the shift Fireworks is riding.
And then there’s the open-source factor. Meta’s Llama models, Mistral’s offerings, and a growing roster of high-quality open-weight models have created a world where enterprises don’t need to pay OpenAI or Anthropic for every inference call. But running open-source models efficiently at scale requires serious infrastructure expertise — exactly the kind of expertise Fireworks sells. In a sense, the open-source model movement is Fireworks’ best friend. Every new open-weight model released is another potential workload on its platform.
Lin Qiao has been deliberate about building Fireworks as an engineering-first company. The team is disproportionately composed of systems engineers and compiler specialists, not machine learning researchers. This reflects the company’s thesis that the hard problems in inference are systems problems — memory management, kernel optimization, hardware utilization, network throughput — rather than model architecture problems. It’s an unglamorous bet. But it’s a defensible one.
The $552 million war chest gives Fireworks significant runway to invest in hardware, hire engineers, and potentially acquire smaller competitors. Lin Qiao told Business Insider that the company plans to use the funding to expand its global infrastructure footprint and deepen its enterprise sales organization. International expansion, particularly into Europe and Asia, is a priority — regions where data sovereignty concerns make cloud-agnostic inference providers especially attractive.
The risk factors are real. Nvidia could optimize its own inference stack to the point where third-party optimization becomes less valuable. The hyperscalers could slash inference pricing to loss-leader levels to keep customers on their platforms. A new hardware architecture — custom ASICs, neuromorphic chips, or something not yet invented — could render current optimization techniques obsolete. And there’s always the possibility that the AI spending boom itself could slow, though few serious observers expect that in the near term.
Still, Lin Qiao’s timing looks impeccable. She identified inference as the bottleneck before most of the industry did. She built a company around solving it before the hyperscalers mobilized. And she raised enough capital to compete at scale before the window closed. Whether Fireworks ultimately becomes a generational company or an acquisition target for a larger player, it has already proven something important: in the AI gold rush, the picks-and-shovels metaphor isn’t quite right. The real money isn’t in making the tools. It’s in running them.
WebProNews is an iEntry Publication