Eighteen months ago, the Model Context Protocol was a 47-page technical specification that Anthropic published on a Tuesday afternoon with minimal fanfare. The document described a JSON-RPC 2.0-based framework for connecting AI language models to external tools and data sources through a standardised client-server architecture. The problem it addressed was real but appeared narrow: the N times M integration problem, where connecting ten AI applications to one hundred external tools required one thousand custom integrations, each written specifically for a single combination of model and tool. MCP proposed to collapse that complexity by providing a universal interface that any MCP-compatible AI client could use to connect to any MCP-compatible server without custom integration work.
The efficiency gain was genuine. What was not obvious at the time of release was that solving the integration problem at the AI layer would position MCP as the defining infrastructure standard of the agentic AI era, controlling the interface between AI systems and the entire digital economy.
The adoption trajectory confirms that positioning. OpenAI adopted MCP in March 2025, integrating it across the ChatGPT desktop application and later the Apps SDK. Google integrated MCP into Gemini via Google AI Studio and Vertex AI Agent Builder in April 2025. Microsoft shipped MCP servers for GitHub, Azure, and Microsoft 365 through Copilot Studio and Azure OpenAI Service during 2025. By April 2026, the MCP Python SDK had crossed 164 million monthly downloads on PyPI, and over 10,000 public MCP servers were active across GitHub, npm, and dedicated registries. Seventy-eight percent of enterprise AI teams report at least one MCP-backed agent in production as of Q1 2026, up from 31% a year earlier. No developer tooling standard achieves that adoption velocity. Infrastructure standards do.
The Technical Foundation That Makes MCP Consequential
Understanding why MCP has become consequential rather than merely convenient requires understanding what the protocol actually controls. MCP defines the interface layer between AI agents and the rest of the digital economy. Every enterprise system, data source, and tool that an AI agent needs to access to perform useful work must either implement MCP natively or be wrapped in an MCP server. The MCP server is the gatekeeper that determines what the agent can see, what actions the agent can take, what data flows back to the AI system, and what audit trail is created for those interactions. The protocol defines not just the communication format but the trust model, the capability discovery mechanism, and the permission structure that governs how AI agents relate to the systems they interact with.
Before MCP, the fragmentation of AI integration created a natural limit on how much consequential work AI agents could autonomously perform. If every integration required custom engineering work, the practical ceiling for AI agent capability was determined by developer capacity to build and maintain connectors. MCP removes that ceiling by making integration a commodity rather than a custom engineering task. The average time to integrate a new tool with an AI system dropped from 18 hours to 4.2 hours as MCP adoption matured. That compression does not just make integration faster. It changes the economics of AI agent deployment in ways that make autonomous agent operation at scale commercially viable for a much wider range of organisations and use cases than custom integration economics would have permitted. The infrastructure implication is that MCP has directly enabled the AI agent deployment surge that is now visible across enterprise software markets.
The Governance Decision That Sealed MCP’s Infrastructure Status
The December 2025 decision by Anthropic to donate MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, OpenAI, and Block, was the governance event that formally elevated MCP from a product to infrastructure. The Linux Foundation’s track record of stewarding foundational technology standards, including Kubernetes, PyTorch, and Node.js, gave MCP the governance credibility that enterprise procurement teams and CIO organisations require before committing to any foundational technology standard. AWS, Google, Microsoft, Cloudflare, and Bloomberg joined as founding members alongside Anthropic, OpenAI, and Block, creating a governance structure whose member list reads as a complete mapping of the enterprise AI supply chain.
The Linux Foundation donation is not primarily a technical event. It is a strategic signal. When Anthropic chose to donate a protocol it created and controlled to a neutral foundation governed by its largest competitors, it made a deliberate calculation that MCP’s value as ubiquitous infrastructure exceeds its value as a proprietary Anthropic advantage. That calculation is the same one that has historically preceded the emergence of foundational technology standards. IBM donated key networking protocols to open standards bodies. Sun Microsystems open-sourced Java. Google open-sourced TensorFlow. Each of these decisions reflected the recognition that the technology in question would generate more value as infrastructure than as a proprietary product, and that the donating organisation’s competitive position depended more on the ecosystem built on top of the standard than on control of the standard itself. Anthropic’s MCP donation reflects the same logic applied to the agent connectivity layer.
The Enterprise Software Vendor Reckoning
The most immediate and commercially significant consequence of MCP’s emergence as a standard is the pressure it creates on enterprise software vendors to implement MCP servers for their products. The dynamic is similar to what happened when major platforms converged on REST APIs as the standard for web service integration in the early 2010s. Software vendors who did not provide REST APIs became progressively less accessible to the ecosystem of applications building on web services infrastructure, creating competitive disadvantage that was not primarily about API quality but about ecosystem compatibility. MCP is creating the same dynamic at the AI agent integration layer. An enterprise software product without an MCP server is becoming less accessible to AI agents than competitors who have shipped MCP support, regardless of the underlying software’s functional quality.
HubSpot shipped its official remote MCP server in general availability on April 13, 2026, joining Salesforce, GitHub, Slack, Stripe, and Figma among the enterprise software vendors who have shipped MCP servers in response to enterprise AI buyer demand. The enterprise vendor MCP server is becoming a standard feature rather than a differentiator, with the speed of adoption reflecting the commercial reality that enterprise AI buyers are now evaluating software procurement decisions partly on the basis of MCP compatibility. That selection pressure accelerates MCP adoption independently of Anthropic’s or OpenAI’s direct influence. The protocol’s vendor-neutral governance, now that it sits under the Linux Foundation, removes the objection that implementing MCP creates dependency on a competitor’s technology, making MCP adoption a straightforward infrastructure investment rather than a competitive positioning decision.
The A2A Layer That Completes the Agentic Stack
MCP’s emergence as the standard for connecting AI agents to tools and data is being complemented by the Agent-to-Agent protocol, released by Google in April 2025 and merged with IBM’s Agent Communication Protocol under the AAIF in August 2025. A2A addresses the horizontal dimension of agentic AI, the communication between AI agents that need to coordinate on complex tasks, while MCP addresses the vertical dimension, the connection between individual AI agents and the external tools and data they need to act. The two protocols operate at different layers and are explicitly designed to be complementary rather than competitive. Together they are defining the full protocol stack for agentic AI in a way that no single organisation’s proprietary architecture could achieve.
The combination of MCP for tool connectivity and A2A for agent coordination creates a foundation for multi-agent AI workflows that was not practically achievable before standardised protocols existed at both layers. An AI agent coordinating with other agents through A2A and accessing enterprise data and tools through MCP can execute complex, multi-step workflows across enterprise systems with a degree of autonomy that requires no custom integration engineering beyond initial MCP server deployment. That capability is what the AI infrastructure market is calling agentic AI, and its commercial deployment depends on MCP and A2A providing the protocol foundation just as HTTP and TCP/IP provided the protocol foundation for the web economy.
As covered in our analysis of the export controls splitting the global AI infrastructure market, the standards and protocols that define how AI systems interact with the digital economy are as strategically significant as the hardware that runs those systems. MCP and A2A together are those standards.
The Competitive Implications for AI Model Providers
MCP’s position as the universal agent connectivity standard creates an asymmetric competitive dynamic among AI model providers that has not yet been fully incorporated into how the market evaluates their relative positions. The model provider whose client is most compatible with the widest range of MCP servers, whose MCP implementation is most reliable and performant, and whose developer experience for building MCP-connected agents is most polished will capture a disproportionate share of enterprise AI agent deployment regardless of underlying model quality differences.
Anthropic’s position in this dynamic is structurally advantaged because MCP is its protocol. Claude’s MCP client implementation is the most mature, the most thoroughly tested against the widest range of servers, and the most deeply integrated into Anthropic’s developer tooling and documentation. As the protocol’s creator, Anthropic has had the longest development runway and the most direct access to the protocol specification’s evolution. However, the donation of MCP governance to the AAIF limits Anthropic’s ability to maintain that advantage through protocol control. Any unilateral attempt by Anthropic to fork the protocol or create Claude-specific extensions that provide compatibility advantages would undermine the vendor-neutral governance that makes MCP valuable as a standard. Anthropic’s competitive advantage in the MCP ecosystem must therefore come from implementation quality rather than specification control, which is a more sustainable competitive position but also a more contestable one.
OpenAI’s Strategic Position
OpenAI’s adoption of MCP in March 2025 was initially framed as a pragmatic compatibility decision that gave ChatGPT access to the growing MCP server ecosystem without requiring OpenAI to build or maintain those integrations. That framing understated the strategic significance of the decision. By adopting MCP, OpenAI implicitly validated Anthropic’s protocol as the industry standard at a moment when it could have promoted its own GPT Actions framework as an alternative. That validation, from AI’s largest commercial operator, was the single most consequential adoption event in MCP’s history. It created the cross-vendor adoption dynamic that makes MCP a standard rather than a preferred option.
OpenAI has since shipped its own AGENTS.md standard for agent project instructions, which complements rather than competes with MCP, and has participated in the AAIF governance structure as a co-founder. The combination of MCP for tool connectivity, A2A for agent coordination, and AGENTS.md for project context is emerging as the complete protocol stack for agentic AI. OpenAI’s contributions to all three layers of that stack position it as one of the defining architects of the agentic AI era’s infrastructure, despite the fact that none of these protocols originated within OpenAI. The company’s strategic pivot toward agentic AI as a commercial priority, visible in its Operator and Deep Research products and its $40 billion fundraise partly earmarked for agentic infrastructure, reflects its assessment that the protocol layer and the applications built on it are where the most durable commercial value in AI will accumulate.
The Data Center Infrastructure Dimension
The MCP adoption surge has a physical infrastructure dimension that the protocol discussion typically omits. MCP-connected AI agents running at enterprise scale generate significantly different compute and network traffic patterns than the chatbot and copilot AI applications that preceded them. A chatbot processes discrete conversational turns with predictable latency requirements. An AI agent executing a complex multi-step workflow through MCP connections to ten or fifteen enterprise systems generates a different pattern of compute demands, network calls, data retrieval operations, and action executions that are less predictable in duration and more variable in resource intensity than conversational AI workloads.
The infrastructure implication is that the shift from conversational AI to agentic AI, enabled by MCP standardisation, creates a new category of AI infrastructure demand that data center operators, cloud providers, and enterprise IT teams are still learning to plan for. Agentic workloads that use MCP to access real-time data, execute API calls, and coordinate with other agents through A2A generate compute patterns that are neither the short, bounded inference calls of chatbot responses nor the long, sustained GPU cluster utilisation of model training runs. They are dynamic, event-driven, and potentially long-running, with latency requirements that depend on the responsiveness of multiple external systems rather than a single model inference endpoint.
As covered in our analysis of data center leases that were not written for AI workloads, the physical infrastructure frameworks that govern how AI compute is provisioned and priced were not designed for the workload characteristics that agentic AI is producing.
The Edge and Hybrid Infrastructure Requirement
The latency requirements of MCP-connected agentic AI create a specific infrastructure pressure toward edge deployment and hybrid cloud architectures that centralised hyperscale data center models cannot fully satisfy. An AI agent that needs to access a corporate ERP system, retrieve data from a local database, and execute actions in an on-premise workflow management platform faces round-trip latency constraints that are incompatible with routing all MCP traffic through a remote cloud data center. Enterprise deployments of MCP-connected agents therefore require either edge inference infrastructure that keeps model execution physically close to the MCP servers being accessed, or hybrid architectures that intelligently route agent traffic between cloud and on-premise resources based on latency and data sovereignty requirements.
This hybrid infrastructure requirement is creating demand for a new category of enterprise AI infrastructure deployment that does not map cleanly onto existing cloud, on-premise, or colocation infrastructure models. The managed MCP endpoint services that AWS, Google Cloud, and Azure are developing are partly a response to this demand, providing a managed infrastructure layer that handles MCP server deployment, scaling, and connectivity management within the cloud provider’s own network fabric.
However, enterprises with significant on-premise systems and data sovereignty requirements need MCP infrastructure that extends beyond the cloud provider’s network boundary into the enterprise’s own data center environment. That requirement is creating market demand for enterprise MCP gateways, on-premise MCP server hosting platforms, and hybrid MCP routing infrastructure that does not yet have well-established vendors or established architectural patterns. The protocol that is becoming AI’s most consequential connectivity standard is also creating infrastructure requirements that the data center industry has not yet fully identified, much less built the capacity to serve.
The Security Surface Area That Governance Must Address
The rapid adoption of MCP has created a security surface area that the protocol’s governance structure is still working to address adequately. The combination of autonomous agent action, broad data access, and a public server ecosystem that has grown faster than security review processes can keep pace with has produced a set of vulnerabilities that are real, documented, and present in production systems. Security researchers identified that 43% of public MCP servers have at least one vulnerability, and 5.5% already contain poisoned tool descriptions, where malicious instructions hidden in a tool’s metadata can cause an AI agent to take harmful actions without the user’s awareness. In controlled testing, tool poisoning attacks succeed 84% of the time when agents run with auto-approval enabled.
The security dimension matters for the infrastructure argument because infrastructure that is insecure at scale creates regulatory risk that can reshape the governance and deployment landscape faster than commercial adoption can respond. The EU AI Act’s requirements for AI systems in high-risk applications, the evolving FERC frameworks for AI systems interacting with energy infrastructure, and the US government’s growing attention to AI agent access to critical systems all create regulatory trajectories that will eventually intersect with MCP’s role as the universal agent connectivity protocol. The AAIF’s governance of MCP will determine whether the security standards embedded in the protocol are sufficient to satisfy those regulatory requirements before regulation is imposed externally.
OAuth 2.1 authorisation support was added to the MCP specification in April 2026, improving the authentication and permission framework. However the gap between specification requirements and actual implementation across the 10,000-plus public MCP server ecosystem remains a vulnerability that the protocol’s governance structure must resolve before MCP’s security model is adequate to its role as foundational AI infrastructure.
The Infrastructure Investment Implications
The emergence of MCP as a foundational AI infrastructure standard creates investment implications that extend across the entire AI infrastructure stack. At the cloud layer, the hyperscalers who have shipped the most complete and best-supported MCP server implementations are capturing developer mindshare and enterprise adoption in ways that compound into platform lock-in at the agent layer. AWS, Google Cloud, and Azure are all building managed MCP endpoint services that allow enterprise customers to deploy MCP servers without managing the underlying infrastructure, creating a new category of cloud managed service revenue that attaches to existing enterprise relationships.
At the enterprise software layer, the MCP server has become a mandatory product feature for vendors seeking to maintain competitive access to enterprise AI buyers who are evaluating software procurement on MCP compatibility. The vendors who ship well-designed, well-maintained MCP servers with comprehensive tool coverage and robust security implementations will capture a market position in the agentic AI era that those who ship minimal or poorly maintained servers will not. At the security layer, the vulnerabilities in the public MCP server ecosystem create a market opportunity for MCP security tooling, server vetting services, and enterprise MCP gateway platforms that manage and secure MCP server access within corporate environments. Qualys, CrowdStrike, and dedicated MCP security startups are all developing products in this category.
As covered in our analysis of the AI infrastructure workforce crisis nobody is planning for, the skills required to deploy, secure, and operate AI infrastructure at enterprise scale are in critically short supply. MCP adds a new dimension to that skills gap — the ability to design, implement, and secure MCP servers that meet enterprise security and reliability requirements is a skill set that most enterprise technology organisations do not currently have at adequate scale.
What MCP’s Trajectory Means for the Next Five Years
The trajectory of MCP over the next five years will depend on three variables whose outcomes remain unsettled. The first is whether protocol-level improvements, governance-mandated standards, or regulatory requirements can resolve the security vulnerabilities in the public server ecosystem before a significant security incident creates a political environment that slows enterprise adoption or triggers restrictive regulation. The second is whether A2A achieves the same cross-vendor adoption that MCP has achieved, completing the protocol stack for multi-agent AI and enabling the complex autonomous workflows that represent the ceiling of agentic AI’s commercial potential. The third is whether alternative connectivity standards, particularly those developed within the Chinese AI ecosystem that is diverging from US-aligned protocols under export control pressure, fragment the global MCP ecosystem in ways that limit the network effects that drive MCP’s value.
Of these three variables, the security question is the most immediately consequential. A protocol that controls how AI agents access enterprise data and execute actions across enterprise systems is, by definition, critical infrastructure. Critical infrastructure that is demonstrably insecure at scale in 43% of its public implementations does not remain unregulated indefinitely, regardless of how valuable the underlying standard is. The AAIF’s governance of MCP’s security roadmap and the pace of security improvement across the public server ecosystem will determine whether MCP’s infrastructure trajectory continues at its current pace or encounters the regulatory friction that insecure critical infrastructure eventually attracts.
Governance and Infrastructure Control Remain Unresolved
The protocol Anthropic open-sourced on a Tuesday afternoon has become something considerably more consequential than its authors could have predicted. Whether it achieves the full infrastructure status its adoption trajectory implies depends on whether its governance keeps pace with its growth. The question of where MCP-connected agent workloads run, who manages the MCP server infrastructure, and how enterprises meet the latency and data sovereignty requirements of agentic AI is as consequential for the data center industry as the question of who controls the protocol specification. These questions remain open. The industry will answer them over the next five years. Those answers will determine how much of the AI era’s value creation flows through the infrastructure layer rather than the application layer above it.
