The Signal
The agent economy can look diverse at the application layer while remaining highly concentrated underneath. An enterprise might use five different agents but still depend on the same cloud, the same model family, the same identity provider, the same protocol stack, and the same settlement asset. DN’s Agent Concentration Risk Index measures that hidden dependency.
Five layers where concentration can hide
| Layer | What can concentrate | Why it matters | Failure mode |
|---|---|---|---|
| Models | Frontier model families and inference providers | One outage, policy shift, or price change can affect many agents at once. | Correlated capability loss. |
| Cloud | Compute, storage, orchestration, networking | Agent stacks often share hyperscale infrastructure even when applications differ. | Regional or provider-wide disruption. |
| Identity | Authentication, authorization, credentials | Agent permissions can depend on one identity plane. | Single-point access failure. |
| Protocols | MCP, A2A, payment or discovery standards | Open standards reduce lock-in, but implementation monocultures can still form. | Ecosystem-wide compatibility shock. |
| Payment rails | Stablecoins, processors, settlement networks | Agent commerce can inherit concentration from settlement assets. | Liquidity, depeg, compliance or access shock. |
DN Agent Economy Herfindahl Dashboard
Use market shares when you have them. HHI is calculated as the sum of squared percentage shares. A higher score means greater concentration.
If shares do not sum to 100%, the HHI is a lower-bound estimate because smaller providers are omitted. Do not aggregate all remaining firms into one “Other” bucket if you want a true HHI.
Cloud is already a visible concentration layer
In Q2 2026, AWS, Microsoft, and Google represented about 63% of worldwide cloud infrastructure spending. That does not mean every agent deployment has the same dependency, but it shows how easily an apparently diverse agent ecosystem can converge onto a small infrastructure base.
Multi-cloud architecture can reduce provider concentration, but it also introduces operational complexity. Diversification only helps if the failover path is actually tested.
Stablecoins show how settlement concentration can emerge
DefiLlama’s 29 October 2026 data shows USDT at roughly 60.8% of total stablecoin market capitalization. That creates a clear example of agent-payment concentration: a business can support many wallets and chains while still depending on the same settlement asset.
Concentration is not automatically bad. Deep liquidity, network effects, and broad acceptance can improve utility. The risk appears when the organization mistakes convenience for resilience.
Open protocols reduce lock-in, not concentration by themselves
MCP and A2A are important because they make interoperability more plausible. MCP’s 2026 specification emphasizes scalable, stateless infrastructure and hardened authorization. A2A has moved into production use with support from more than 150 organizations.
But an open protocol can still become a monoculture if most implementations depend on the same SDK, gateway, hosted registry, or identity provider. Standards reduce switching friction. They do not guarantee operational diversity.
Model concentration is harder to measure
Frontier-model concentration changes quickly and is difficult to measure with one clean global denominator. Model usage also differs sharply between consumer chat, enterprise APIs, coding, search, and agent workflows.
DN therefore treats model concentration as a dependency question unless a defensible market-share dataset exists: how many independent model families can your workflow use, how quickly can you switch, and how much performance degrades after switching?
The seven questions every agent stack should answer
- What percentage of critical workflows depend on one model family?
- Can the same workflow run on a second cloud without major redesign?
- Is identity centralized through one provider or credential system?
- Can tools switch between MCP, A2A, direct APIs or alternative interfaces?
- What happens if the primary payment asset or processor becomes unavailable?
- Are failover paths tested, or merely documented?
- How much performance, cost, and reliability degrade during failover?
Concentration versus fragmentation
| State | Advantage | Risk |
|---|---|---|
| High concentration | Liquidity, compatibility, lower integration cost | Correlated outage, pricing power, policy shock |
| Moderate concentration | Strong ecosystem with credible alternatives | Fallback may still be weaker |
| High fragmentation | More independent options | Higher integration cost and inconsistent standards |
How to reduce concentration risk
- Maintain at least one tested model fallback for critical workflows.
- Separate application logic from cloud-specific services where feasible.
- Keep identity and authorization portable.
- Prefer open standards when they reduce switching costs.
- Track settlement-asset and processor concentration separately.
- Test failover quarterly, not only after an outage.
- Measure degradation during fallback, not just whether fallback technically works.
Methodology and limitations
Version 1.0, 29 October 2026. DN uses market-share evidence for cloud and stablecoins, plus dependency scoring for model, identity, protocol, and orchestration layers. The dashboard implements the Herfindahl-Hirschman Index when real shares are available.
Limitations: concentration changes quickly, and market share is not identical to operational dependency. A firm with strong portability may face less risk than market-wide concentration implies.
Commercial disclosure: no cloud, model, protocol, identity, or payment provider paid for inclusion.
Primary sources
- Synergy Research Group, Q2 2026 cloud market shares
- DefiLlama, stablecoin market capitalization and USDT dominance
- Model Context Protocol, 2026-07-28 specification
- Linux Foundation, A2A adoption milestone
FAQ
What is agent concentration risk?
The risk that apparently diverse AI-agent systems depend on a small number of underlying models, clouds, identity providers, protocols, or payment rails.
What is HHI?
The Herfindahl-Hirschman Index is the sum of squared market shares. Higher values indicate greater concentration.
Is concentration always bad?
No. Concentration can improve liquidity, compatibility, and efficiency. The problem is unmanaged dependency and correlated failure.
Do open protocols solve concentration?
They reduce switching friction and interoperability barriers, but concentrated implementations, gateways, or hosting can still create dependency.
