Most people do not know Fastly exists until it does not. At 9:47 UTC on June 8, 2021, the San Francisco company began returning errors across 85 percent of its network. Newsrooms, retailers, government sites and software services blinked out together. The problem was not a cyberattack or a reckless customer. It was a valid configuration change that woke a software bug Fastly had quietly deployed 27 days earlier.
One minute later, monitoring caught the disruption. Forty minutes after onset, engineers had identified the triggering configuration. Within 49 minutes, 95 percent of the network was operating normally. It was a brutal public demonstration of Fastly's product: one programmable layer sitting between much of the internet and the people asking it for a page.
The apology mattered because Fastly resisted an easy excuse. The customer did nothing wrong. The failure belonged to the platform, its tests and its blast radius. Fastly disabled the offending configuration, shipped a permanent fix that day and promised to examine why quality assurance missed the bug. It also pointed toward stronger isolation using WebAssembly, the technology underneath its edge-compute product.
The company that sells distance
Fastly's business begins with an irritating fact: geography makes software slow. A request that travels from a phone in Sydney to a server in Virginia and back pays a latency tax before the application does useful work. A content delivery network stores or computes the answer closer to that phone. The shorter trip makes pages load faster, video behave better and origin servers work less.
Traditional CDNs were often treated as delivery pipes. Founder Artur Bergman wanted a control surface. As CTO of Wikia, he watched a constantly edited community site struggle with the old trade-off between cached speed and fresh information. Airline seats, hotel rooms, product inventories and breaking-news pages had the same problem. Fastly, launched in 2011 with seven employees, made its cache configurable through code and APIs. “Fastly was built by developers, to empower developers,” Bergman later wrote.
The short version
That programmability remains the difference Fastly emphasizes against Akamai, Cloudflare and the CDN products bundled into Amazon, Google and Microsoft clouds. Customers can purge old content quickly, manipulate requests and responses, stream logs as traffic happens and deploy serverless code near the user. The platform now spans network delivery, a next-generation web application firewall, DDoS and bot protection, API security, observability, object storage and edge compute.
A front door for GitHub, newsrooms and shops
Fastly's customers tend to be organizations for which the website is not a brochure but the business: software platforms, publishers, retailers, streaming companies and financial services. Its public case studies include GitHub, Shopify, GOV.UK, the Guardian, USA TODAY Co. and JD Sports. In late 2025, Fastly counted 628 enterprise customers, defined as customers producing more than $100,000 in annualized revenue during the quarter.
GitHub supplies the cleanest example of what a buyer can actually do. It first used Fastly to build a feature. When its previous CDN provider suffered an outage, GitHub moved much of the remaining content over to Fastly overnight, while under attack, without waiting for a sales representative. Fastly eventually sat in front of GitHub.com, Pages and raw content. Repeated requests could be collapsed into one trip to GitHub's origin, reducing strain exactly when traffic was hostile.
That story answers the question that infrastructure marketing usually dodges: what changed the customer's mind? An incumbent failed, urgency removed committee theater, and a self-service product was ready. The copyable idea is not “buy Fastly.” It is to make adoption possible during the moment of pain: documented APIs, reversible configuration and a useful path that does not begin with a calendar link.
Usage is the meter, security is the expansion
Fastly mostly charges for consumption: data transferred, requests processed, regions served and products switched on. Its public pricing includes 100 GB of monthly bandwidth in the free tier. Beyond that, rates vary by region and volume; enterprise contracts are negotiated. This is convenient when a product is small and naturally expands with demand. It is less charming when a viral event or bot storm turns a traffic graph into an invoice.
The published rate card makes the trade concrete. After the free allowance, the first 10 TB of monthly full-site-delivery bandwidth was listed at 12 cents per GB in North America and Europe, 19 cents in much of Asia and Australia, and 28 cents in Africa and India, before lower-volume tiers and separate request charges. A North American service moving one extra terabyte therefore starts around $120 for bandwidth alone; security, compute, support and other traffic add their own meters. Large buyers negotiate. Small teams should do something less glamorous: replay a normal month, a launch month and a bot-attack month through the calculator. Usage pricing works beautifully when cost follows valuable activity. It works poorly when the activity is unwanted, margins are thin or nobody owns the bill.
The company spent heavily to become more than a delivery network. In October 2020 it acquired Signal Sciences, a developer-friendly web and API security company, for $775 million - about $200 million in cash and the rest in Fastly shares. The purchase brought protection that could run at the edge, in a cloud, in containers or on premises. In the first quarter of 2026, Fastly said security revenue grew 47 percent year over year, helping total quarterly revenue reach a record $173 million, up 20 percent.
The broader numbers are less tidy. Fastly lost $121.7 million in 2025 even as its loss narrowed from $158.1 million a year earlier. Its ten largest customers supplied 32 percent of annual revenue. Usage grows with customer success, but traffic optimization can also reduce the bill; large customers have leverage, and enterprise deals can take months or more than a year to ramp. This is infrastructure economics, not a frictionless subscription fairy tale.
AI gives the edge a new job
AI traffic is turning the CDN from delivery van into border guard. In June 2026, Fastly reported that AI traffic in its dataset had grown 30 percent from January through May, 6.5 times faster than human traffic. Some crawlers are useful, some scrape content without permission, and some autonomous agents need to prove who they are before buying anything. Fastly now sells AI bot controls, semantic caching for repeated model prompts and an MCP server for managing services. Its Skyfire partnership aims to verify agent identity and payments at the edge.
The logic is consistent with Bergman's original idea. Put the decision close to the request. A publisher can allow one crawler, meter another and block a third. An AI application can reuse a semantically similar response rather than pay a model to generate it again. A retailer can inspect an agent before it reaches checkout. The edge is valuable because it sees the request before the expensive systems behind it do.
What to copy - and when not to
Builders can borrow five moves. Treat configuration changes like code, complete with testing and staged rollout. Stream telemetry while an event is happening instead of waiting for batch reports. Make purges and rollbacks fast. Design customer actions as potentially global events. Finally, publish a postmortem that separates trigger from cause: a customer can trigger a latent defect without causing it.
Multi-CDN is the obvious answer to concentration risk, but not a free one. Duplicate configuration drifts. Features differ. Observability fragments. Teams must routinely test failover or discover during an incident that their insurance policy is decorative. It makes sense when the expected cost of downtime exceeds the engineering and traffic cost of maintaining two working paths. For everyone else, a well-tested single provider may be the more honest system.
Fastly's story is not a clean march from speed to victory. It is a study in useful tension. The network became valuable because one control plane could change behavior everywhere. The network became dangerous for exactly the same reason. Fifteen years after launch, Fastly is pushing that control plane into security and AI, where every early decision matters more. The company still sells milliseconds. Now it also sells judgment about what those milliseconds are allowed to do.