"Inference, without the hardware bet."
The AI inference cloud that owns almost no GPUs - and turns that into the whole point. It rents capacity across the planet, routes your model to whatever chip is cheapest right now, and moves roughly 500 billion tokens a day.
There is a strange thing happening in artificial intelligence, which is that everybody wants to buy the same picks and shovels at the same time, and the picks and shovels are graphics cards, and there are not enough of them, and the ones that exist are expensive and occasionally obsolete before you finish paying for them. This is a bad thing to build a company around, and Parasail's founders know it, because they built companies around it before.
Parasail's answer is to not buy the shovels. Instead it rents them - from data centers, from GPU liquidity markets, from wherever spare compute happens to be sitting idle - and stitches the whole scattered supply into one thing it calls an AI Supercloud. You send it a request. It finds the hardware. You are billed, more or less, by the token. You never had to sign a multi-year lease on a rack of chips you hoped would still be useful in 2029.
The company describes this in five words that are refreshingly free of jargon: "Inference, without the hardware bet." The bet in question is the one everyone else is making. If you own the GPUs, you win big when demand and prices cooperate, and you eat a very large fixed cost when they don't. Parasail's wager is that flexibility - the ability to route a workload to the cheapest available chip at any given second - beats ownership, at least for the customers it wants.
Those customers are, deliberately, the ones the big clouds treat as an afterthought: seed-stage and Series-B AI startups that can't tell you what their inference bill will look like next month, let alone next year. Parasail sells them optionality. No long-term commitment, open-source and custom models supported, pay for what you use. It is, in a sense, the anti-enterprise-contract.
And the volume is real. The company says it processes on the order of 500 billion tokens a day. Do the arithmetic against a headcount of roughly 32 people and you get a business where the leverage is unmistakably in the software, not the org chart - which is exactly what an orchestration company is supposed to look like.
Give me tokens - fast and cheap.
Strip away the branding and Parasail is doing something close to arbitrage. It treats global GPU capacity the way a trading desk treats a commodity: as supply to be sourced cheaply, priced intelligently, and matched to demand. The internal machinery it uses to decide where a given workload should run is described, evocatively, as a "GPU permutation engine."
Pool GPU supply from ~40 data centers, 15 countries, liquidity markets, and Parasail's own clusters.
Automatically tune each model endpoint for speed, performance, and cost - then route to the best-fit hardware.
Deliver serverless or dedicated inference on pay-per-token economics - no hardware to own.
Because it isn't locked into a single fleet of owned chips, Parasail can undercut providers that are - and it can flex up and down with a customer's demand instead of charging them for idle silicon. That is the entire strategic idea, stated plainly.
A fabric of global compute that automatically optimizes model endpoints for speed, performance, and cost across ~40 data centers and Parasail's own clusters.
On-demand, pay-per-token API access to open-source and custom models. Spin up, ship, scale down - no commitment.
Reserved capacity for real-time and batch workloads that need consistent, predictable throughput.
Large-scale batch jobs - RAG pipelines, data processing, model evaluation - run cost-effectively across pooled supply.
Parasail plays in the crowded AI inference cloud space, where the pitch to developers is always some version of "cheaper, faster tokens." The differentiator it leans on is supply-side: not owning the hardware, and therefore not being trapped by it. The bar below is illustrative - a rough sketch of positioning, not audited market share.
Competitors named in press coverage: Fireworks AI, Baseten, Together AI, plus hyperscaler GPU offerings.
Former Chief Product Officer at Groq, where he helped build the cloud offering, at a company whose AI-chip ambitions later drew a $20B valuation orbit. Before that he founded and led Mythic, the analog AI-chip startup, raising $165M. His new company deliberately sells software leverage instead of silicon.
Previously founded and led Swift Navigation, a precise-positioning company that generated hundreds of millions in sales and raised roughly $250M in venture funding over his tenure. Brings a track record of scaling deep-tech infrastructure businesses.
Around them sits a small, senior bench: engineers and researchers drawn from NVIDIA, Amazon, Stripe, Uber, Groq, and Blue Origin, with PhDs in machine learning and computer science and expertise spanning distributed systems, GPU inference, compiler design, and Kubernetes.
| Round | Amount | Date | Lead / Notable Investors |
|---|---|---|---|
| Seed | ~part of $42M total | 2025 | Kindred Ventures, Basis Set, Threshold, Black Opal, Buckley |
| Series A | $32M | Apr 2026 | Touring Capital & Kindred Ventures (co-lead), Samsung NEXT, Flume, Banyan |
The Series A capital is earmarked to deepen orchestration and inference optimization, accelerate go-to-market, and strengthen partnerships across the GPU and data-center ecosystem. Total funding to date: roughly $42 million.
Mike Henry and Tim Harris begin building an inference cloud that rents GPUs instead of owning them.
The company publicly launches as an AI deployment network with access to a large supply of on-demand GPUs.
Touring Capital and Kindred Ventures co-lead a round bringing total funding to ~$42M to expand the AI Supercloud.
It runs AI model inference in the cloud, orchestrating rented GPUs across ~40 data centers in 15 countries so developers can deploy open-source and custom models without buying hardware, paying largely per token.
Mike Henry (Co-Founder & CEO, former Groq CPO and Mythic founder) and Tim Harris (Co-Founder, former Swift Navigation founder & CEO).
About $42M total, including a $32M Series A in April 2026 co-led by Touring Capital and Kindred Ventures.
It owns almost no GPUs. Instead it treats global compute like a liquidity market and routes each workload to whichever hardware is cheapest and fastest - avoiding the long-term hardware bet.
Primarily AI-native startups from seed through Series B, plus developers and enterprises running large-scale LLM inference. It processes on the order of 500 billion tokens a day.
Company file compiled from public sources including parasail.io, TechCrunch, PR Newswire, Data Center Dynamics, and BusinessWire. Figures such as token volume, data-center count, and funding are as reported publicly and are approximate. Market-positioning graphic is illustrative, not audited.