Sail Research Raises $80 Million to Lower Inference Costs for Long-Horizon AI Agents

Sail Research emerged from stealth on June 25, 2026, with $80 million in combined seed and Series A funding at a $450 million valuation. Founded by former Apple and Together AI engineers, the company is building infrastructure specifically optimized to run long-running AI agents. Kleiner Perkins led the Series A round and Sequoia Capital led the seed round, with participation from Alphabet chairman John Hennessy.

Why Long-Horizon AI Agents Demand 50x More Tokens Than Chatbots

In standard chatbot setups, the primary technical goal is low latency. When a human types a prompt, the system must generate a response within seconds, requiring infrastructure optimized to spit out the first tokens as quickly as possible. For autonomous AI agents that run independently in the background, this instantaneous speed is less critical.

Instead, these long-horizon agents run complex, multi-step workflows like code reviews, compliance audits, or extensive web research that can take hours or weeks to complete. Because these processes run asynchronously without a human waiting at a keyboard, they require massive, continuous token processing rather than immediate response speed.

These deep, branching workflows consume tokens at rates 50 to 500 times higher than simple chat interactions. Even as the cost per token has declined across the industry, the sheer volume of tokens consumed by autonomous agents has driven enterprise AI bills to surge. This economic shift creates a demand for inference engines built exclusively to optimize cost and throughput.

To address this demand, Sail Research optimizes its inference stack to prioritize throughput over low latency. By processing larger batches of tokens simultaneously, the company claims it can reduce the cost of running long-horizon agents by 3x to 10x compared to traditional low-latency endpoints.

“We only care about efficiency,” CEO Neil Movva said in an interview with Fortune. “It’s quite difficult to build an inference engine for both throughput and latency at the same time.”

Customizing vLLM and PagedAttention for Multi-Agent Workloads

The platform achieves its efficiency advantages through proprietary infrastructure optimizations and deep customization of open-source inference engines. One such engine is vLLM, which utilizes an algorithm called PagedAttention to manage graphics card memory.

In traditional inference setups, graphics processing unit (GPU) memory is allocated in static, continuous blocks to hold the conversational context. This process often wastes valuable memory because the system must reserve space for the maximum possible length of the conversation, even if the actual dialogue is short.

PagedAttention resolves this by dividing the context memory into small, dynamic pages that can be scattered across non-contiguous physical memory locations. This approach allows the system to utilize GPU memory more efficiently, freeing up hardware resources to process multiple agent steps concurrently.

The company’s API is compatible with existing OpenAI-based workflows and runs optimized open-weight models. These include popular open-source models such as DeepSeek, Gemma, GLM, Kimi, and Nemotron.

How Sailboxes Suspend Runtime Billing During External API Waits

Alongside inference optimization, the platform uses virtual machines called Sailboxes to run autonomous workloads. These Linux-based sandbox environments are stateful, allowing them to remain open for days or weeks at a time while agents perform deep-branching research or automated programming tasks.

A common financial drain in agentic systems occurs when an agent pauses to wait for an external action. For example, an agent might need to wait for a third-party API response, a slow database query, or a user approval step before it can proceed.

Traditional cloud sandboxes continue to charge for hosting as long as the virtual machine is running, regardless of whether any processing is occurring. Sailboxes solve this by saving the state of the agent and shutting down the virtual machine during waiting periods. The platform only bills customers for active compute time.

Developers can customize each virtual machine by installing a software image, configuration modules, and specific settings. The platform also makes it possible to link multiple sandbox environments together to form an agent ensemble, allowing separate specialized models to collaborate.

The Stanford and Apple Engineering Roots Behind the Hardware Stack

The company’s technical direction is shaped by its founders’ backgrounds in silicon architecture and low-level software. CEO Neil Movva, 28, previously worked at Nvidia and Apple, focusing on AI accelerator design.

Movva was also the first kernels engineer hired at Together AI, where he worked on developing accelerated inference systems. That platform was designed for interactive, low-latency applications, which led Movva to identify a need for infrastructure built with different priorities.

CTO Samir Menon, who was Movva’s classmate at Stanford, previously worked on Apple’s security engineering team. Together, the founders aimed to build a vertically integrated stack that spans hardware selection, custom kernels, and global fleet orchestration.

Scoring 90.72% on the BrowseComp-Plus Deep Research Benchmark

Before exiting stealth, the startup evaluated its platform using BrowseComp-Plus, a benchmark designed to test AI agents performing complex online research tasks over extended periods. Under this evaluation, the company says its platform achieved a 90.72% accuracy score while operating at up to ten times lower cost than competing services.

Several enterprise software startups are already running active workloads on the platform. These include Parallel Web Systems, Jack and Jill, and Detail.dev.

At Detail.dev, developers use the platform to power autonomous code-review agents that analyze entire codebases over three to four hours. Without a high-throughput, low-cost platform, the company says the sheer volume of tokens consumed during such tasks would be financially prohibitive.

Sequoia, Kleiner Perkins, and Semiconductor Angels Back the Launch

The startup closed its combined seed and Series A funding rounds with several prominent venture capital firms and semiconductor executives. Sequoia Capital led the seed round, while Kleiner Perkins led the Series A. Partner Aditya Naganath represented Kleiner Perkins in the deal.

Additional venture participation came from Redpoint Ventures, Theory Ventures, Vine Ventures, CRV, A*, and Abstract Ventures.

The funding rounds also secured backing from several angel investors. These include Intel chief executive Lip-Bu Tan and Together AI chief scientist Tri Dao, alongside John Hennessy.