On August 5, 2026, Sapiom announced a $35 million Series A to build the infrastructure layer that sits beneath enterprise AI agents. Led by Dragonfly with participation from Accel, Anthropic, Menlo Ventures, and seven other firms, the round brings the San Francisco startup’s total funding to $50 million, arriving less than a year after the company was founded.
The company’s pitch is concrete and backed by operating metrics: building an AI agent demo is the easy part. Running one reliably, at scale, without destroying the economics of the business, is where most enterprise teams struggle. Sapiom says it has processed more than 270 million transactions since launching six months ago and now handles over 100,000 agent runs per day.
The Problem: Agents Break Down at Production Economics
Sapiom’s founding premise is that the real constraint on enterprise AI agent adoption is not capability. It is cost and reliability.
Ilan Zerbib, Sapiom’s founder and CEO, made this concrete in a case shared at the time of the raise. Polsia is an AI-native company that employs no full-time staff and runs swarms of agents to operate its business. When Polsia’s revenue grew sharply this year, so did its inference spend, reaching $1.2 million per month billed to Anthropic. That level of token consumption made the company economically unsustainable even as its product worked. After engaging Sapiom to run evaluations and redesign its routing strategy, Polsia’s monthly Anthropic bill dropped to approximately $100,000: a 90 percent reduction with no reported change in product quality.
Zerbib told Semafor that for startups building agentic products at frontier-model prices, “it’s just unsustainable” when those products involve high-volume, automated tasks. His framing of the market size puts the urgency in perspective: “There’s maybe 10 million developers, maybe 50 million. If you think about agents, we’re talking about trillions of agents that will operate in the economy in the next three years.”
The math breaks down quickly for any team running agents at scale. A single developer querying Claude or ChatGPT is a small inference cost. An agent processing documents, routing customer requests, or running code reviews around the clock is an infrastructure cost that compounds with every additional workload. Sapiom’s wager is that the business controlling the routing and cost economics of that infrastructure captures a durable position in the AI stack, regardless of which underlying model wins.
Three Products, One Infrastructure Layer
Alongside the Series A, Sapiom launched three products that form what it calls an end-to-end agent operations layer.
Sapiom Router routes each model call to the most appropriate model rather than defaulting to whichever frontier model a team already uses. The routing is dynamic, evaluating available models, compute resources, tools, and services based on cost, quality, latency, reliability, task requirements, and company policy. Budgets and permissions are enforced before an action executes, and each transaction is metered and recorded in an audit trail.
Unlike OpenRouter, which aggregates access to model providers, Sapiom runs open-weight models on its own server racks in a San Jose data center. This lets it serve those models at direct compute cost rather than passing a markup through. For proprietary models from OpenAI, Anthropic, or Google, the routing layer calls the originating provider, but substitutes a hosted open-weight model when quality thresholds allow.
Sapiom Agent Studio gives engineering teams a local development environment for building, testing, inspecting, and deploying agents in the context of their existing codebases. This is the build-and-validate surface, designed to surface routing and cost issues before they reach production.
Sapiom Runtime provides managed production infrastructure with access controls, routing, automatic failover, step-level visibility, and high-throughput execution. This is the layer that converts a working agent prototype into a continuously running, auditable production service.
| Product | What It Does | Key Capability |
|---|---|---|
| Sapiom Router | Routes model calls dynamically | Cost, quality, latency optimization per request |
| Sapiom Agent Studio | Build and test environment | Inspect routing decisions before production |
| Sapiom Runtime | Managed production infrastructure | Failover, access controls, audit trail, scale |
One customer reduced its inference bill by 75 percent after moving its agents to Sapiom Runtime, according to the company. A second customer, Polsia, cut its monthly Anthropic bill by 90 percent. Both figures were shared at the time of the announcement but have not been independently verified.
Why Anthropic Is Backing a Platform That Routes Traffic Away From It
The most unexpected fact in this funding round is that Anthropic participated as an investor in a platform explicitly designed to route enterprise traffic away from Anthropic’s API when cheaper alternatives suffice.
Zerbib addressed this in Semafor’s coverage of the raise. His argument: lower costs accelerate agent adoption broadly, and growth in the number of agents running in the economy generates more demand for frontier models, including Claude, for the tasks that require top-tier capability. The routing layer grows the market for all providers by making agent deployment economically viable for a much larger base of customers.
Dragonfly Managing Partner Haseeb Qureshi, who joins Sapiom’s board as part of the round, framed it differently in the company’s announcement: “Sapiom solves the common problem in which agents operate with no budget and CTOs allocate money with no visibility into where it goes. That’s not a governance problem you solve with another dashboard. It’s an infrastructure problem, and it needs to be solved at the point where agents act and money moves.”
The timing places Sapiom alongside a cluster of enterprise AI governance platforms that became generally available or raised significant capital this week. Databricks Unity AI Gateway hit general availability with multi-model routing, spend controls, and MCP governance on August 5, the same day as Sapiom’s announcement. Snowflake Cortex AI Gateway launched in July with task-scoped permission controls. The pattern is consistent: enterprise teams adopting AI agents need a control plane that enforces economics and governance at the infrastructure level, not in the application layer.
What This Means for Enterprise AI Builders
The Sapiom raise carries three practical implications for teams building or procuring AI agent infrastructure.
Token cost is now a line-item risk. As agent workloads scale from dozens of interactions to millions, inference spend becomes a material operating cost. Teams that treat model selection as a one-time architecture decision are exposed to price changes, provider outages, and quality drift from any single vendor. Infrastructure that routes dynamically and fails over automatically reduces that exposure without requiring manual re-architecture when a provider changes its pricing.
Audit trails matter before regulators require them. Sapiom records each transaction in an audit trail and enforces budgets before actions execute. For regulated industries deploying agents against customer records, source code, or financial data, that sequence matters. A governance layer that logs only after the fact is weaker than one that enforces rules at execution time. This is the same principle that Cloudflare’s AI agent payment infrastructure applies to spending controls: constrain at the point of action, not in a reconciliation report.
Production infrastructure is not a build-yourself problem at scale. The Polsia case study suggests that even a technically sophisticated AI-native company can spend 12 times more than necessary on inference before engaging a routing specialist. For enterprise teams building agent products on tighter timelines, the cost of building and maintaining that routing logic in-house compounds quickly. Sapiom is betting that a vendor-neutral infrastructure layer, one that treats models and providers as interchangeable inputs to be optimized in real time, is more defensible over time than tools tied to a single ecosystem.
Sapiom’s Position in the Agent Infrastructure Market
Sapiom is not alone in targeting the gap between agent prototypes and production. Larger platform vendors (Databricks, Snowflake, Cloudflare) and a wave of specialist startups are all addressing the same enterprise problem from different angles. What distinguishes Sapiom’s approach is the combination of owning physical compute infrastructure for open-weight models, enforcing budgets at the point of execution rather than reporting on them after the fact, and offering an integrated build-to-run path through Studio and Runtime alongside the routing layer.
Whether that combination is defensible against platforms backed by cloud providers with far larger infrastructure footprints remains an open question. The $35 million gives Sapiom roughly 18 to 24 months to prove the model at scale before needing to answer it.
The Trillion-Agent Horizon
Sapiom’s positioning around “the next trillion agents” is deliberately outsized, but it points toward a real transition. Analysts have projected the AI agent market could exceed $180 billion by the early 2030s. Most of that growth would require agents running unattended, at high volume, for sustained periods: the exact scenario where production infrastructure, not demo capability, determines what survives contact with real workloads.
The $35 million Series A gives Sapiom runway to extend its platform to larger enterprise deployments. At 270 million transactions processed and 100,000 agent runs per day, six months after launch, the operational scale is already meaningful enough to attract a board seat from one of the more active investor-operators in AI infrastructure.
For enterprise teams evaluating where AI agent deployment costs will come from as adoption scales in 2026 and beyond, Sapiom’s raise is a useful calibration: the market has decided that production infrastructure is a real and distinct problem, separate from model capability, and the capital is beginning to allocate accordingly.