Competing AI labs don’t usually adopt each other’s infrastructure. That’s what made March 2025 the real inflection point for the Model Context Protocol: OpenAI, Anthropic’s most direct competitor, officially adopted MCP across its Agents SDK, Responses API, and ChatGPT desktop app — with Sam Altman’s response to the moment reduced to a single line: people love MCP, and OpenAI was adding support across its products. Google DeepMind followed in April 2025. Microsoft and GitHub joined the steering committee in May. By the time Anthropic donated MCP to a newly formed, Linux Foundation-backed governance body in December 2025, the protocol had already stopped being “Anthropic’s thing” in any meaningful sense — it had become shared infrastructure that direct competitors were building on together.

That’s rare enough in enterprise software history to be worth understanding on its own terms — what MCP actually does, how fast it’s really moving, and where the reported numbers hold up against scrutiny and where they don’t.

What MCP Actually Does

Anthropic open-sourced the Model Context Protocol on November 25, 2024, as a standard way for AI models to connect to external tools, data sources, and services. The common shorthand is “USB-C for AI” — a single connector that lets any AI application plug into any tool, instead of every model needing a custom-built integration for every system it talks to.

The problem it solves has a name: the N×M problem. Before MCP, connecting 10 AI applications to 100 enterprise tools could require up to 1,000 separate point-to-point integrations, each one bespoke and each one needing separate maintenance. MCP collapses that into a linear problem — build one MCP server per tool, and every MCP-compatible AI client can use it immediately, with no additional custom wiring. That’s the entire value proposition in one sentence, and it’s why the protocol spread as fast as it did: it removes a cost that scaled quadratically and replaces it with one that scales linearly.

The Growth Numbers — And Which Ones Actually Hold Up

MCP’s download and adoption numbers get cited constantly, and not all of them are equally solid. The most consistently verified figure: combined Python and TypeScript SDK downloads grew from roughly 100,000 in the protocol’s first month to approximately 97 million monthly downloads by March 2026 — a jump of roughly 970x in under a year and a half.

The server ecosystem tells a similar growth story, with numbers that vary depending on when and how they’re counted. Server downloads reportedly passed 8 million by late 2025, up from that same 100,000 baseline. A May 24, 2026 pull directly from the official MCP Registry API counted 9,652 current server records and nearly 29,000 total server-and-version records, while a GitHub search for repositories tagged with the MCP-server topic returned close to 16,000 results the same day. The flagship modelcontextprotocol/servers repository itself had passed 86,000 GitHub stars.

Here’s where it’s worth slowing down: a widely repeated claim that “78% of enterprise AI teams use MCP in production” turns out not to be traceable to a real source. The most defensible enterprise figure comes from Stacklok’s 2026 software industry survey, which found 41% of surveyed software organizations in limited or broad production use of MCP servers — a meaningfully large number, just not the inflated one that’s been circulating. Separately, Fortune 500-focused research found that 28% of Fortune 500 companies had implemented MCP servers as of early 2026, against a much larger 80% that had active AI agents in production generally — a gap that says something important on its own: agent deployment is outpacing standardized tool-connection infrastructure, not the other way around.

Why Enterprises Are Actually Adopting It

Three forces explain why MCP moved from protocol to default this fast:

The integration math got impossible to ignore. Once an enterprise is running more than a handful of AI applications against more than a handful of internal systems, custom point-to-point integration becomes an engineering tax that scales worse every quarter. MCP’s linear-integration model is a direct, measurable cost reduction, not just an architectural preference.

Cross-vendor governance removed the lock-in risk. When Anthropic donated MCP to the newly formed Agentic AI Foundation under the Linux Foundation on December 9, 2025, OpenAI and Block joined as co-founders, with AWS, Google, Microsoft, Cloudflare, and Bloomberg joining as platinum members. That’s the same governance pattern that made Linux, Kubernetes, and OpenTelemetry safe long-term bets for enterprise infrastructure — no single vendor controls the roadmap, which lowers the risk of building on it.

Production deployments are already showing up at scale. Documented enterprise MCP usage includes deployments at Block, Bloomberg, and Amazon, alongside hundreds of Fortune 500 companies experimenting with or running MCP servers in production workflows — real usage, not just protocol-spec enthusiasm.

What’s Still Genuinely Unresolved

The growth story doesn’t mean MCP is a finished, risk-free technology, and treating it that way would be its own kind of overclaiming.

Security scrutiny caught up fast. Security researchers filed more than 30 CVEs against MCP implementations in January and February 2026 alone — a direct consequence of what MCP fundamentally does. Giving an AI agent standardized access to real tools means every connected tool becomes a real attack surface, and that tradeoff doesn’t disappear just because the protocol is well-governed.

Scaling has hit real technical walls. Enterprise deployments using stateful Streamable HTTP connections have run into scaling limits at production volume — the kind of problem that shows up specifically once a protocol moves from prototype traffic to genuine enterprise load, and one the ecosystem is still actively working through.

Governance is still maturing, not finished. The Agentic AI Foundation is still formalizing its working groups, interest groups, and succession procedures. The first MCP Dev Summit North America, held April 2–3, 2026 in New York with more than 95 sessions from Anthropic, OpenAI, Microsoft, Docker, and Bloomberg, marked the shift from informal multi-company collaboration toward structured governance — but “marked the shift” is different from “completed the shift.”

What This Actually Means for Enterprise AI Strategy

The practical takeaway isn’t “adopt MCP because everyone else is.” It’s that the specific problem MCP solves — the N×M integration tax — is a real, measurable cost that most enterprises running more than a few AI applications will eventually hit regardless of which model vendor they’re using. The fact that OpenAI, Anthropic, Google, and Microsoft all support the same protocol means an enterprise’s MCP investment isn’t tied to a single model vendor’s roadmap, which is precisely the kind of infrastructure bet that tends to age well.

At the same time, the gap between the 80% of Fortune 500 companies running AI agents and the 28% running MCP servers is the honest opportunity signal here: most enterprises have already committed to agentic AI, but haven’t yet standardized how those agents connect to tools. That gap is where the next 18 months of enterprise AI infrastructure spending is most likely to concentrate — and where the security and scaling issues above will need real answers, not just governance announcements, before it closes.