The index always concentrates
Twenty search engines competed to index the web in 1999. Twenty-six years later, 91% of all web searches flow through one. ARD defined the discovery layer for AI agents in June 2026. The same clock started.
The amber bar marks the chokepoints: cloud runtime is where agents execute, and agent platforms are how users and systems reach them. When ARD launched June 17, 2026, only two founding members had products in both rows: Google and Microsoft.
The other nine concentrate in the remaining layers: model serving, developer tooling, enterprise data, hardware and networking. Real infrastructure, but none of it is the runtime. An agent running on GCP does not query a neutral registry. It defaults to Google's own, which is easier to reach and harder to displace. Google held that advantage before ARD shipped.
Sources and methodology
Product assignments reflect publicly documented capabilities at ARD specification launch (June 17, 2026). ARD specification: Google Developers Blog, June 17, 2026. Founding coalition: Google, Microsoft, Hugging Face, Nvidia, Salesforce, Cisco, GitHub, Databricks, ServiceNow, Snowflake, GoDaddy. Layer categories are analytical groupings, not official ARD terminology. Search market share estimate in subtitle: StatCounter Global Stats, desktop global.