Parasail

  Executive Briefing

Parasail is a San Francisco–based AI infrastructure company that operates what it calls an "AI Supercloud" — a distributed inference network that aggregates underutilized GPU capacity from more than 25 cloud partners across 40+ data centers in 15+ countries and exposes it through a single, OpenAI-compatible API. The company's founding thesis is that the GPU market mirrors the early internet's bandwidth fragmentation: enormous idle supply is scattered across competing providers, while developers face rate limits, vendor lock-in, and pricing premiums. Parasail acts as the coordination layer — a programmable, neutral orchestration fabric that matches inference demand to the best available supply in real time.

The company was co-founded in late 2023 by Mike Henry (CEO) and Tim Harris (co-founder and board member). Henry's background is unusual for an inference startup: he holds a PhD in Electrical and Computer Engineering from Virginia Tech, founded Mythic in 2012 — an analog AI-acceleration chip company that raised approximately $165 million in venture capital — and then served as Interim Chief Product Officer at Groq in 2023, where he built Groq's cloud LLM offering from the ground up. It was there that Henry observed a structural shift: most AI infrastructure spend was migrating outside the "big three" legacy hyperscalers, yet the supply of GPU capacity that sat outside those clouds was fragmented and inaccessible. Harris brings a complementary pedigree, having previously co-founded Swift Navigation in 2013, a company building centimeter-accurate positioning software for autonomous vehicles that raised approximately $250 million.¹²

Parasail ran a closed beta beginning January 2025, passed seven-figure annual recurring revenue during that period, and launched publicly on April 2, 2025.³ By April 2026, the platform was reported to be processing more than 500 billion tokens per day, with revenue growing at approximately 30% month-over-month — figures that are self-reported and not independently verified.⁴ A $32 million Series A round announced in April 2026 brought total disclosed funding to $42 million and added Touring Capital, Kindred Ventures, and Samsung NEXT as backers.⁴ The company positions itself as critical infrastructure for the AI agent era, arguing that agentic workloads generate orders of magnitude more token volume than human-in-the-loop applications and require the elastic, contract-free supply that no single cloud provider can economically offer.

  At a Glance

  Origins & Founding

The intellectual origin of Parasail lies in Mike Henry's tenure at Groq in 2023. While building Groq's consumer-facing LLM cloud product, Henry mapped the emerging structure of the AI compute market and concluded that the majority of inference workloads were moving away from AWS, Google Cloud, and Azure toward a fragmented ecosystem of GPU neoclouds, colocation providers, and specialized inference hardware vendors. Yet no neutral intermediary existed to aggregate this supply and make it accessible to developers with a consistent API, SLA, and pricing model.³

Henry's analogy was the content delivery network industry of the early 2000s: bandwidth sat stranded in carrier networks until companies like Akamai and later Cloudflare built abstraction layers that transformed fragmented supply into programmable infrastructure. Parasail was designed to play that role for GPU inference — a "supercloud" sitting above competing providers, routing workloads intelligently based on cost, latency, model availability, and hardware generation.⁵

Tim Harris joined as co-founder, bringing experience scaling a capital-intensive infrastructure business in Swift Navigation, where autonomous-vehicle positioning systems required integrating heterogeneous hardware and real-time signal processing at global scale — a structurally similar coordination problem.

The company was incorporated in late 2023. Seed funding of $10 million, led by Basis Set Ventures and Threshold Ventures with participation from Buckley Ventures and Black Opal Ventures, provided the capital to build the multi-cloud orchestration control plane and begin onboarding GPU supply partners.³

  History & Timeline

    2023 — Founding and stealth

Parasail was founded in late 2023 by Mike Henry and Tim Harris. The founding team remained in stealth while building the core multi-cloud Kubernetes orchestration layer and establishing initial partnerships with GPU cloud providers. The company closed a $10 million seed round from Basis Set Ventures, Threshold Ventures, Buckley Ventures, and Black Opal Ventures.³

    2025 — Closed beta, public launch, and early traction

In January 2025, Parasail opened a closed beta and began onboarding its first developer customers. The company reported passing seven-figure ARR during the beta period — a notable milestone before public availability.³ On April 2, 2025, Parasail launched publicly, announcing an OpenAI-compatible API, four deployment modes (serverless, dedicated serverless, dedicated, and batch), and an on-demand GPU fleet that the company claimed exceeded Oracle Cloud Infrastructure's entire GPU footprint at the time — a claim not independently verified.³⁵ Early customers included Weights & Biases, Rasa, SambaNova, Oumi, Elicit, mem0, Gravity, and Kotoba. At launch the team was reported at approximately 12 employees.

    2026 — Scale and Series A

By April 2026, Parasail reported processing more than 500 billion tokens per day across its network, with revenue growing approximately 30% month-over-month (both figures self-reported).⁴ On April 15, 2026, the company announced a $32 million Series A co-led by Touring Capital and Kindred Ventures, with participation from Samsung NEXT, Flume Ventures, Banyan Ventures, and existing investors. Samir Kumar (Touring Capital) and Steve Jang (Kindred Ventures) joined as board observers. The platform served approximately 30 open-weight models as of mid-2026.⁴

  What They Offer — Products & Platform

Parasail's product is a single inference platform accessible via an OpenAI-compatible REST API that abstracts the underlying GPU fleet. Developers can deploy any open-weight model from Hugging Face in under five minutes and approximately five lines of code, without managing infrastructure, negotiating contracts, or filing capacity requests.³

The platform offers four deployment modes:

Serverless is the default and most accessible tier: fully managed, auto-scaling inference billed per token with no minimum commitment, no rate limits, and no quotas. The platform handles load balancing, cold-start mitigation, and hardware selection automatically.

Dedicated Serverless provides guaranteed throughput with an isolated GPU pool, suitable for latency-sensitive production workloads that need predictable performance without owning specific hardware.

Dedicated reserves physical GPUs for a customer's exclusive use, enabling custom model weights, fine-tuned adapters, and maximum control over the inference environment. Pricing is per GPU-hour.

Batch is an asynchronous, offline processing mode designed for high-volume workloads that tolerate latency — document processing, dataset annotation, model evaluation pipelines. Batch pricing is offered at a 50% discount off serverless rates, with an additional 30% discount for cached-prompt tokens.⁶

Beyond raw inference, the platform supports function calling and structured JSON output across nearly all served models, image and video understanding, voice agent workloads targeting sub-500 ms end-to-end latency, and multi-model agent orchestration — routing different tasks to different models within a single agentic pipeline.

  Technology & Infrastructure

Parasail's core technical innovation is a proprietary multi-cloud Kubernetes control plane that treats competing GPU cloud providers as heterogeneous worker nodes within a single logical cluster. Standard Kubernetes does not natively bridge clusters across data centers or cloud providers; Parasail built the federation and scheduling layer on top of Kubernetes to solve this, enabling workloads to be dispatched to whichever provider has suitable hardware available at optimal cost.⁵

The GPU fleet spans more than 25 cloud supply partners — confirmed partners include CoreWeave and Lambda Labs — plus Parasail-operated clusters, aggregating across 40 data centers in 15+ countries. Available hardware includes NVIDIA H200, H100, A100, and RTX 4090 GPUs, covering the principal inference-grade silicon generations. The company claims its aggregated on-demand fleet exceeded Oracle Cloud Infrastructure's entire GPU footprint as of April 2025 — a positioning statement that has not been independently verified.³

Workload placement uses a combination of ML-driven routing models and human-in-the-loop optimization to match model architectures, transformer configuration parameters, and GPU generations. The goal is to minimize time-to-first-token and maximize throughput without requiring customers to perform their own inference performance engineering. Real-time telemetry drives continuous kernel selection and routing adjustment.

The platform supports FP8 quantization at no added cost relative to the listed serverless price, and FP16 at a 30% markup. This quantization-aware pricing model allows cost-sensitive workloads to benefit from reduced memory bandwidth without penalty, while high-precision workloads can opt up.⁶

Batch processing infrastructure achieves reported cost reductions of 80–90% compared to real-time inference for equivalent token volumes, enabled by bin-packing workloads across available capacity during off-peak periods.⁴

  Model Catalog & Performance

Parasail's catalog focuses exclusively on open-weight models, with the company emphasizing day-0 availability for newly released frontier open-weight systems. As of mid-2026, approximately 30 models were available on the platform; nearly all support function calling, approximately 29 of 30 support JSON mode, and 17 of 30 include reasoning capabilities.⁴ Model availability is fluid as the frontier moves quickly, and the specific catalog will have evolved since the data below was captured.

Notable models served include:

• DeepSeek V4 Flash (Max/High variants) — 1M context window, blended pricing reported at approximately $0.11/M tokens.⁷
• Llama 4 Maverick — Meta's flagship open-weight multimodal model, 1M context window.
• Kimi K2.6 — reported highest intelligence index score (54) on the platform as of mid-2026.⁷
• GLM-5.1 (FP8) — intelligence index 51, priced at approximately $0.90/M tokens; GLM-5 (FP8) at intelligence index 50.
• gpt-oss-120b (high/low variants) — reported throughput up to 152 tokens/second output, time-to-first-token of approximately 0.82 seconds.⁷
• MiMo-V2.5 — reported fastest output speed at 152 tokens/second and lowest serverless price at $0.09/M tokens, offering the best price-performance ratio on the platform.⁷
• Gemma 4 26B A4B — priced at approximately $0.10/M tokens.
• Qwen series (multiple variants).

All performance benchmarks cited are sourced from Artificial Analysis evaluations and are reported figures; independent verification has not been confirmed.⁷

  Pricing & Performance Position

Parasail's pricing spans roughly a 10x range across its serverless catalog, from $0.09/M tokens (MiMo-V2.5) to $0.90/M tokens (GLM-5.1). Dedicated GPU pricing is reported at $0.65–$3.25/GPU-hour depending on hardware tier.⁶

Batch pricing follows a tiered structure based on parameter count:

FP16 carries a 30% markup across all tiers. Cached-prompt tokens receive an additional 30% discount in batch mode.⁶

The company claims its serverless pricing represents a 15x–30x cost saving versus OpenAI and Anthropic's proprietary model APIs, and a 2x–5x saving versus other open-source inference providers — figures that are self-reported and depend on the comparison model and workload type.⁴ These claims reflect the structural cost advantage of open-weight models served on aggregated, underutilized GPU capacity versus proprietary model APIs with higher margin structures.

On performance, independent benchmarks from Artificial Analysis report MiMo-V2.5 at 152 tokens/second output speed, and gpt-oss-120b at 0.82 seconds time-to-first-token — competitive with leading inference specialists for the respective model sizes.⁷

  People & Leadership

Mike Henry (Co-Founder & CEO) is the operational and technical lead. His career arc spans academic research (PhD, Electrical and Computer Engineering, Virginia Tech), deep-tech hardware founding (Mythic, 2012–2023, an analog AI accelerator startup that raised approximately $165 million), and large-scale inference product leadership (Interim CPO at Groq, 2023). The combination of chip-level AI systems understanding, startup execution experience, and direct exposure to the inference market at one of its fastest-growing companies gives Henry an unusual vantage point in the GPU-cloud space.¹³

Tim Harris (Co-Founder & Board Member) contributes serial founder credibility and infrastructure-business expertise from Swift Navigation, which he co-founded in 2013 and grew to approximately $250 million in total venture capital raised before departing. Swift Navigation's core product — centimeter-accurate GPS positioning for autonomous vehicles — required integrating heterogeneous hardware sources and real-time signal arbitration at global scale, a structurally analogous coordination problem to GPU inference routing.¹

Board observers from the Series A include Samir Kumar (Partner, Touring Capital) and Steve Jang (Partner, Kindred Ventures).⁴

Team size was reported at approximately 12 employees at the April 2025 public launch; current headcount is not publicly disclosed.

  Funding, Ownership & Business

Parasail has raised a total of $42 million across two disclosed rounds as of April 2026, with no valuation publicly stated at either round.⁴

The $10 million seed round (closed 2024) was led by Basis Set Ventures and Threshold Ventures, with Buckley Ventures and Black Opal Ventures also participating. This capital funded the construction of the multi-cloud orchestration platform, supply-side partner integrations, and the January 2025 closed beta.³

The $32 million Series A (announced April 15, 2026) was co-led by Touring Capital and Kindred Ventures, with Samsung NEXT, Flume Ventures, Banyan Ventures, and existing investors participating. The round came alongside reports of 500B+ tokens/day processing volume and 30% month-over-month revenue growth.⁴ Series A capital is intended for inference optimization engineering, GPU supply partner expansion, and go-to-market acceleration.

The business model is consumption-based: serverless and batch workloads bill per token; dedicated deployments bill per GPU-hour. There are no platform fees, seat licenses, or minimum commitments for serverless usage. The company passed seven-figure ARR during its closed beta, suggesting early commercial validation before public launch. All revenue and growth figures are self-reported by the company and have not been independently verified.

  Customers & Partnerships

Parasail's stated primary target segment is AI-native startups from seed stage through Series B, and specifically teams building agentic applications where token volume scales non-linearly with product usage. The "tokenmaxxing" framing — the company's term for AI agents consuming orders-of-magnitude more tokens than human-in-the-loop workflows — is central to its customer acquisition pitch.⁴

Notable customers include Elicit (AI research assistant), mem0 (memory layer for AI agents), Gravity, Kotoba, and Venice. Early launch customers and design partners also included Weights & Biases, Rasa, SambaNova, and Oumi.³⁴

On the supply side, confirmed GPU cloud partners include CoreWeave and Lambda Labs, with the broader network reported to encompass 25+ providers across 40 data centers in 15+ countries. The platform's OpenAI-compatible API reduces switching friction and enables customers to migrate workloads from other inference providers without code changes beyond endpoint and key configuration.

  Competitive Position

Parasail competes most directly with Fireworks AI, Baseten, Together AI, and Lepton AI in the open-weight inference API market, and more broadly against the hyperscaler GPU cloud offerings from AWS, Google Cloud, and Azure for raw compute. Its differentiation rests on several structural choices that distinguish it from single-provider inference services.

The aggregated supply model is the primary differentiator: by drawing GPU capacity from 25+ competing providers rather than owning a single fleet, Parasail can offer burst capacity at a scale no individual neocloud matches, without the capital expenditure of building owned data centers. This also means the platform is provider-agnostic and can route to whichever supply source offers best price-performance for a given model and hardware generation.

The no-contracts, no-rate-limits positioning targets a real pain point: most inference providers impose per-minute or per-day token quotas that create bottlenecks for agent workloads with spiky, unpredictable volume patterns. Parasail's elimination of these constraints is a direct response to the architectural needs of LLM-powered agent systems.³⁴

Day-0 model support — the commitment to offer newly released open-weight frontier models on launch day — is another differentiator in a market where developers building with the latest models cannot wait weeks for provider support. The company's inference optimization layer (automated kernel selection, routing, quantization-aware scheduling) removes the performance engineering burden that would otherwise fall on customers deploying models across heterogeneous hardware.⁵

The CDN analogy that underpins Parasail's public positioning — GPU supply as fragmented bandwidth, Parasail as Cloudflare — is a coherent strategic frame, though whether the GPU inference market will consolidate similarly to CDN markets remains an open question.

  Outlook & Roadmap

Series A capital will be deployed toward three priorities: deepening inference optimization (kernel tuning, speculative decoding, model-specific routing), expanding the GPU supply partner ecosystem across additional regions and data centers, and accelerating go-to-market with the AI-native startup segment.⁴

The company's core strategic bet is that the AI agent era will multiply per-customer token volumes dramatically — making elastic, contract-free inference supply a structural necessity rather than a convenience — and that a neutral aggregation layer above competing GPU providers is a durable position analogous to how CDN operators sit above ISPs. As the frontier of open-weight models advances rapidly (with DeepSeek, Qwen, Llama, Kimi, and others shipping high-capability models in quick succession), Parasail's value proposition depends on maintaining day-0 coverage and price-performance leadership.

Specific capacity targets, additional partner announcements, and product roadmap details beyond the broad priorities above have not been publicly disclosed. Outlook statements should be treated as directional rather than commitments.

---

  References

1. Parasail public launch announcement — Business Wire, April 2, 2025
2. Parasail $32M Series A announcement — PR Newswire, April 15, 2026
3. Parasail GPU fleet vs Oracle — TechCrunch, April 2, 2025
4. Parasail raises $32M for tokenmaxxing developers — TechCrunch, April 15, 2026
5. Parasail compute brokerage deep-dive — Next Platform, April 3, 2025
6. Parasail provider benchmarks — Artificial Analysis
7. Parasail pricing documentation
8. Touring Capital investment thesis — Touring Capital
9. Kindred Ventures announcement
10. SiliconAngle Series A coverage

  References

1. Founder backgrounds and prior affiliations — Business Wire, April 2025. ↩ ↩² ↩³

2. Tim Harris / Swift Navigation background — PR Newswire, April 2026. ↩

3. Public launch details, closed beta, seed round, and early customers — TechCrunch, April 2, 2025 and Business Wire. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

4. Series A, token volume, and revenue growth figures — self-reported by Parasail; TechCrunch, April 15, 2026 and PR Newswire. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵

5. Infrastructure and CDN analogy — Next Platform, April 3, 2025. ↩ ↩² ↩³ ↩⁴

6. Pricing tiers and quantization details — Parasail pricing docs. ↩ ↩² ↩³ ↩⁴

7. Model performance benchmarks are reported figures from Artificial Analysis — artificialanalysis.ai/providers/parasail. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶