For more than 15 years, Fastly has helped thousands of companies power experiences that are fast, safe, and engaging. But increasingly, those experiences aren’t just for humans anymore. Agents and bots are taking over the internet faster than anyone predicted. From January through May of this year, we watched AI requests on our own network climb roughly 30%, or about 6.5x faster than human traffic grew over the same time period. It’s as if we now have a second internet running alongside the one originally built for people.
So today, we’re introducing AI Runtime Control (ARC), a control plane that sits between your applications and your LLM and other model providers. Every model call you make is routed through one place to govern access, track spend, and help you stay resilient as AI traffic grows.
The Gaps Enterprises are Living with Today
Ask most engineering or platform leaders how their organization talks to LLM providers, and you'll hear some version of the same story: teams are calling OpenAI, Anthropic, Google, and other providers directly, with a scattering of raw API keys passed around between developers and projects. Where there's an attempt to get control of AI traffic, teams often end up stitching together a patchwork of point solutions where one tool is used for cost tracking, another for logging, another for rate limiting. Each point solution only handles one piece of the problem, and none of the solutions know how to talk to each other. There's no unified view of what any of it costs, no reliable way to attribute spend to a specific app or team, and no consistent logging when compliance or security asks how a prompt was handled or where a piece of customer data ended up. These are gaps that get more expensive and risky every month AI adoption grows.
Inside Fastly AI Runtime Control
AI Runtime Control closes that gap. Rather than trying to manage a series of point solutions, you benefit from a holistic, integrated approach that can scale up as you transform your business. With ARC, every model call your apps and agents make routes through one Fastly-managed endpoint, where you can issue per-app virtual keys, attribute token usage and spend back to whoever's actually generating it, and enforce policy on which models get used.
But ARC is much more than a gateway. Having one endpoint that reaches a bunch of models is useful, but it could already be considered table stakes at this point, not a true, unified control plane.
ARC picks up where the value of having a gateway runs out with model-specific virtual keys that can be easily created, rotated, refreshed, or revoked; spend that gets attributed to specific applications, sessions, individual developers, or teams; proprietary detection and guardrails to mitigate LLM-specific threats; automatic failover when a provider errors out; and full request/response logging for every AI interaction.
ARC is also bring-your-own-key (BYOK). You keep your existing provider relationships and your negotiated rates. We're not trying to get between you and your model; we're building the control layer around it - providing the governance, visibility, security and resilience layer that a single shared endpoint was never designed to provide. Here are a few of the capabilities that make that real today:
Govern access with virtual keys. Instead of one shared provider key passed around a Slack channel, every app, team, or developer gets its own virtual key. Admins can create, rotate, or revoke access instantly, with no code changes on the client side. This also gives you visibility into shadow AI and vibe-coding sprawl.
Protect against LLM-specific attacks. AI Firewall operates within ARC providing an optional layer of security against natural-language attacks targeting models. It leverages an updated version of Fastly’s SmartParse detection engine and custom guardrails to detect attacks like prompt injection and other LLM specific attacks.
Know your spend in real time. Token and request usage is visible in real time, mapped down to the individual key, session, and provider. Transition from reactive spending to proactive financial guardrails by applying enforceable monthly dollar-budget limits directly to individual virtual keys.
Build in resilience by default. A virtual key can point at an ordered list of targets: your primary model, then a fallback, then another. If the primary provider errors out, ARC automatically retries and fails over to the next target, transparent to the caller. Combined with per-key rate limiting, that means a provider outage or a runaway agent doesn't take your production AI feature down with it.
Audit everything. Full request and response logging for both prompts and completions searchable by user, model, provider, and date range. When legal or compliance asks how a piece of customer data was used, or whether it touched a public model, there's an actual answer instead of a shrug.
AI doesn't remove the need for performant, hardened, enterprise-grade infrastructure - it raises the bar for every request, whether the thing on the other end is a person or an agent acting on their behalf. Identifying that traffic, governing it, and making it resilient is where the work starts.
Get Started with ARC Today! The Best is Yet to Come
We’ve spent years becoming the control layer for how people reach your applications. ARC is us doing the same thing for how your applications reach AI. If you're already sending real AI traffic and feeling the absence of a single place to see and control it, that's exactly the gap ARC is built to close. And because ARC provides a single integration path, you can easily unlock new features and capabilities without having to reconfigure or rearchitect.
What we are announcing today is the foundation that we will continue to build on. AI Runtime Control gives you the ability to innovate with confidence and run AI like you would run your production infrastructure. Visit our documentation for more information on how to get started.