Sapiom Raises $35M Series A to Build the AI Agent Runtime Layer

Sapiom has raised a $35M Series A led by Dragonfly, bringing the San Francisco AI infrastructure company's reported total funding to $50M. The round arrived alongside the launch of Sapiom Router, Agent Studio, and Runtime, a three-part platform built to help AI agents move from convincing demonstrations into controlled production systems.

The financing matters because the expensive part of the agent economy is shifting from intelligence to execution. Models can reason, but an agent deployed inside a business still has to choose services, stay within policy, recover from failures, explain its actions, and produce a bill that somebody in finance can defend.

What Happened

Sapiom announced the Series A on August 5, 2026, eleven months after founder and CEO Ilan Zerbib started the company. Dragonfly led the round, with Accel, Gradient, Coinbase Ventures, Operator Collective, Formus Capital, and VanEck Ventures participating, while Okta Ventures, Menlo Ventures, Anthropic, and Array Ventures continued their support. Haseeb Qureshi, Managing Partner at Dragonfly, joined Sapiom's board.

The financing follows Sapiom's February 2026 seed round, which the company described as $15.75M led by Accel. Sapiom now reports $50M in total funding. No valuation was disclosed, and the company did not provide a detailed allocation of the new capital.

From Payments to a Production Platform

Sapiom first emerged around a narrow but revealing problem: AI agents can decide that they need an API, model, dataset, or service, but the commercial infrastructure around those resources still assumes that a human will create an account, hold credentials, approve payment, and review the invoice. Zerbib had seen versions of that constraint while co-founding Earny and later working in payments engineering at Shopify. The problem is financial, but it is also architectural.

That original payments thesis has widened into an execution platform. Router selects among approved models and services using requirements such as cost, quality, latency, reliability, availability, and company policy. Agent Studio gives builders a place to create, test, and inspect agents as code, while Runtime supplies sandboxes, memory, secrets, tool access, scheduling, retries, traces, and recovery for production workloads.

Why This Matters

The agent market has spent plenty of money proving that software can reason through a task. The harder question is whether that software can run continuously without turning every retry, model call, and external action into an invisible cost or uncontrolled permission. Sapiom is selling infrastructure at that exact point of discomfort, before an action happens and after operators need to reconstruct it.

Gartner forecast that more than 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. That forecast does not validate any particular Sapiom claim, but it explains why investors may see a large market in making agent behavior bounded, observable, and economically legible.

The Economics Behind the Round

Sapiom says it processed more than 270M transactions during its first six months and now supports more than 100,000 agent runs per day. The company also says one customer reduced inference costs by 75%, while another reduced monthly inference spending from approximately $1.2M to approximately $100K. Those company-provided figures are not independently audited, but they point to the operating issue Sapiom wants to own.

Inference cost is only one piece of that issue. A production agent can trigger a branching sequence of model calls, browser sessions, searches, compute jobs, data purchases, and messages, each with its own permission and price. If a team can see only a monthly total, it cannot reliably connect spending to a customer, task, outcome, or failure, which makes both ROI analysis and governance weaker.

What Dragonfly's Lead Signals

Dragonfly's lead role connects Sapiom's agent-runtime thesis to programmable payments and machine commerce. The firm describes itself as a technically focused crypto investor, while Sapiom has built around emerging payment standards including Coinbase's x402 and acquired machine-payment infrastructure company Fewsats in June 2026. The investment suggests that the boundary between AI infrastructure and financial infrastructure is getting harder to draw cleanly.

That does not mean every agent needs a wallet or every software action belongs on a blockchain. It means the next generation of autonomous systems will need payment authorization, usage controls, receipts, identity, and policy to operate at machine speed. Investors are placing capital behind companies trying to make those controls part of the execution layer instead of a patch applied after deployment.

What This Signals for Builders

For founders, the Sapiom round is a reminder that an agent product inherits every weakness in the infrastructure beneath it. A strong model cannot compensate for unknown unit economics, broad credentials, unrecoverable workflows, or logs that show activity without explaining outcomes. The same capital shift is visible in DevCuration's coverage of Baseten's AI inference infrastructure, Coralogix's AI observability platform, and AttoTude's AI connectivity stack. As agent capabilities become easier to access, reliable execution becomes a sharper source of differentiation.

Sapiom still has to prove that its platform can become a standard across a crowded and fast-moving AI infrastructure market. The Series A gives the company more room to expand Router, Agent Studio, and Runtime, hire across engineering and operations, and support customers operating agents at scale. The larger wager is that the companies that win the agent economy will not be the ones with the loudest demos, but the ones that can show what ran, what it cost, what was permitted, and what happened when the system failed.