Startup Bets on Tokenization to Forge Next Computing Giant

"Give me tokens. Just give me tokens. I need them fast, cheap, and I need them now."

That's the relentless demand Parasail CEO Mike Henry hears from developers building on generative AI models. Parasail offers a cloud computing service tailored for AI model inference, and Henry revealed to TechCrunch that the platform generates a staggering 500 billion tokens daily. Talk about maximizing token output.

Henry previously served as an executive at Groq, the LLM-focused chipmaker, where he developed the company's cloud service—an early insight that AI developers would seek specialized cloud processing. Now, after a year out of stealth, Parasail has secured a $32 million Series A funding round to scale this vision.

While Henry has a background in physical chip design, Parasail isn't tied to owning its own hardware. The company operates a hybrid model: it owns some GPUs but primarily rents processing time across 40 data centers in 15 countries, supplemented by purchases from liquidity markets. This orchestrated approach aims to drastically reduce the cost of inference requests.

By intelligently allocating workloads and avoiding demand peaks, Parasail positions itself to compete with firms that own their own silicon but may be constrained by existing customer commitments and fixed workloads.

The company's growth potential hinges on the continued rise of open-source models and AI agents developed outside major frontier labs. Parasail's leadership and investors point to the increasing cost and complexity of using proprietary APIs from companies like Anthropic and OpenAI as a key driver.

A hybrid architecture is indeed emerging, according to Andreas Stuhlmüller, CEO of Elicit. His startup, which raised a $22 million Series A to build an AI research assistant for scientific literature, serves top pharmaceutical companies that use the tool to analyze data from tens of thousands of papers.

"We've shifted more towards open models because sending hundreds of thousands of requests to a single API endpoint is challenging," Stuhlmüller told TechCrunch. This is especially true as Elicit employs AI agents that break tasks into strategic, longer-horizon operations. Open models handle initial screening to lower costs, while a more capable frontier model delivers the final answer.

The surge in model queries, fueled by the integration of AI agents into software development, is fueling investment in infrastructure providers like Parasail that enable low-cost inference. Samir Kumar, a partner at Touring Capital who co-led the funding round, told TechCrunch he anticipates inference will constitute at least 20% of future software development costs.

So, how large a share of this market can Parasail capture? In the crowded cloud computing sector, CEO Mike Henry argues that his company's exclusive focus on inference (it does not handle training) and its flexibility in serving startup clients without long-term contracts differentiate it from larger enterprise-focused cloud providers and even well-funded inference specialists like Fireworks AI and Baseten.

Of course, focusing primarily on seed and Series B startups in the volatile AI sector carries its own set of risks.

Steve Jang, a partner at Kindred Ventures and the other co-leader in this round, believes the economics of model deployment will necessitate the kind of compute brokerage Parasail provides. And this demand is only set to grow with the broader adoption of AI for content generation and robotics.

"Everyone thought there was an AI bubble. There is no AI bubble," he told TechCrunch. "Inference demand is far outstripping supply."