The infrastructure play in artificial intelligence has turned peculiar. While some companies pour billions into building data centers packed with their own chips, Parasail has taken the opposite bet: aggregate everyone else's computing power, then sell access to it cheaper than the hyperscalers can.
It's working, apparently. The San Mateo-based startup just closed a $32 million Series A, co-led by Touring Capital and Kindred Ventures. Samsung NEXT, Flume Ventures, and Banyan Ventures joined existing backers in the round, bringing Parasail's total haul to $42 million since inception.
What the company calls an "AI Supercloud" is, in essence, a brokerage. Parasail doesn't own data centers—it pools GPU capacity from dozens of providers across 40 facilities in 15 countries, then offers developers flexible access through APIs. Feed it a model inference request, and the platform routes it to wherever capacity is cheapest and available. NVIDIA H100s, H200s, A100s, even consumer-grade 4090s. AMD's MI300X chips, too.
The pitch is surgical: Parasail claims it can cut your inference bill by a factor of 15 to 30 compared to OpenAI or Anthropic, though these figures have not been independently verified. Even stacked against other open-source inference providers, Parasail claims it undercuts them by two to five times. No long-term contracts. No minimum commitments. Just pay for what you use.
"AI builders shouldn't have to become infrastructure experts to ship great products," said Mike Henry, Parasail's founder and CEO, when the funding closed. It's a line that sounds almost too convenient, but the company's usage numbers suggest developers are listening.
500 Billion Tokens a Day
Parasail now processes half a trillion tokens daily, according to the company—a tenfold increase from the 40 billion it was handling last September, based on a partnership case study from that period. Volume like that doesn't happen quietly. The startup counts customers including Elicit, mem0, and Venice, alongside earlier adopters like Weights & Biases, which needed large-scale access to DeepSeek models, and Rasa, which cited better latency serving European users.
Parasail claims 30 percent month-over-month revenue growth since the company's public launch a year ago, when it raised a $10 million seed round led by Basis Set Ventures. That kind of compounding, if it holds, is the sort of trajectory that makes VCs sit up.
"Everyone thought there was an AI bubble," Steve Jang of Kindred Ventures said. "Inference demand is far outstripping supply." Perhaps—though venture capitalists have a vested interest in believing the capacity crunch story. Still, the numbers Parasail is posting suggest genuine demand for cheaper, more flexible inference infrastructure.
The Founder's Second Act
Henry is no first-time operator. He previously founded Mythic, an analog AI chip startup that raised over $165 million before pivoting hard when the market shifted. He also served briefly as interim chief product officer at Groq, the inference chip company that's become something of a darling among AI developers chasing speed. Tim Harris, Parasail's co-founder, runs Swift Navigation, a GNSS positioning firm that's pulled in north of $250 million.
Their bet is that the inference market will fragment—geographically, across hardware types, and by cost tolerance. Not every developer needs the latest H200s humming in a Tier 1 data center. Some workloads can run on older silicon. Others need to stay in-region for latency or compliance reasons. Parasail's orchestration layer is designed to handle that complexity without forcing customers to manage it themselves.
TechCrunch has dubbed this approach "tokenmaxxing," a somewhat breathless term for what is essentially cost optimization at scale. Fireworks AI and Baseten are playing in similar territory, though each has taken slightly different architectural paths.
What the Money Buys
The fresh capital will fund expansion: more partnerships with GPU providers and data center operators, deeper investment in orchestration and optimization, and a bigger deployment footprint. Parasail is hiring developer evangelists, account executives, and senior engineers—the usual growth-stage roster.
Samir Kumar of Touring Capital framed the investment around what he called "the control layer" for inference workloads. As models proliferate and hardware options multiply, someone needs to build the software that makes sense of it all. Parasail is betting that layer is worth building—and that developers will pay for it if the price is right.
Whether the company can maintain its cost advantage as it scales, and whether hyperscalers will eventually undercut the aggregators once inference stabilizes, remains an open question. For now, though, Parasail has momentum. And in a market where inference costs still make founders wince, that might be enough.
