MCP becomes AI backbone—but token costs spark debate

Venture Beat ↗

The gist

MCP has become the backbone for AI integration across enterprises, but soaring token costs are fueling a fierce debate about its efficiency and risk.

What to know

  • Anthropic's Model Context Protocol (MCP), launched in November 2024, turned messy AI integrations into a standardized, governed layer now backed by industry giants like Google Cloud and Bloomberg.
  • By late 2026, MCP powered over 17,000 public servers and became the control plane for authentication, discovery, and policy enforcement in major enterprise AI fleets from Broadcom, Citrix, and F5.
  • Skeptics warn that large tool sets can burn tens of thousands of tokens before a question is asked, though new on-demand loading techniques have slashed token use by 85%.

From Chaos to Governed AI

Gateways and registries transformed MCP from a standard into core infrastructure, letting enterprises unify authentication, discovery, and policy enforcement across sprawling AI systems.

MCP mattered because it turned a messy integration pattern into something enterprises could actually govern. As Strategize Your Career put it, “Before November 2024, the industry suffered from an integration problem: every model required connectors for every enterprise data source,” while ByteByteGo notes that “function calling” became widely available in mid-2023 but remained provider-specific; MCP, announced by Anthropic in November 2024 and open-sourced from day one, instead established a common layer whose standardized tools, resources, and prompts could be discovered and controlled consistently across systems.

What transformed that common layer into infrastructure was the rise of gateways, registries, and enterprise auth patterns around it. Business Analytics Review described the registry as standardizing tool schemas and authentication, while HackerNoon highlighted built-in OAuth 2.1 and remote servers that let firms “stay in control”; in practice, NeoTeo’s guidance to check the “/.well-known/oauth-authorization-server” route if sign-in discovery fails—and its note that the documented sign-in flow expects an HTTP 200 response there—shows sign-in discovery becoming operational plumbing, and NVIDIA’s October 2026 Auto Ontology showed dynamic, governed access when an MCP-connected agent ran “check_answerable,” found only “coverage of 0.5” for “ad spend,” and constrained the next step accordingly.

Sources
PPBusiness Analytics ReviewHackerNoonByteByteGo NewsletterStrategize Your CareerMedium

Vendor Buy-In Drives Ubiquity

MCP's open standard and cross-vendor adoption turned it from an Anthropic project into the default backbone for enterprise AI data integration, with major players like Google Cloud and Bloomberg scaling its reach.

MCP won because it offered vendors a simple economic bargain: standardize once, integrate everywhere. Designing with AI described MCP as “by Anthropic in November 2024 as an open standard for connecting AI assistants to external data systems,” built around the promise to “write a tool once as an MCP server, and any client that supports the protocol can use it,” and that portability became more credible as the standard moved beyond Anthropic into broader institutional backing and cross-vendor uptake rather than remaining a single-company format.

By late 2026, that abstract promise had turned into visible market adoption: Google Cloud backed a Daloopa connector, with Satish Thomas saying the partnership would equip financial professionals with verified data in workflow, while Daloopa said its MCP was “LLM-agnostic and supports AI platforms that adopt the MCP standard” and that its platform covers more than 6,000 public companies globally. Bloomberg made the same point at larger scale when The Stack reported, “Bloomberg Unveils Enterprise MCP, Letting AI Agents Query 100 Million Securities,” as the public ecosystem also ballooned to “more than 17,000 MCP servers.”

Sources

Control Layer Becomes Central

By shifting governance, audit, and access controls into centralized MCP gateways, enterprises cut onboarding time, reduced latency, and gained real-time policy enforcement across massive AI fleets.

What changed in enterprise MCP deployments was not merely connectivity but control. CASA Software said Broadcom Automic Automation, paired with MCP, let companies expose only approved workflows as agent-callable skills while retaining RBAC, audit, and execution controls; Greg Harrowsmith added that “in December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation,” giving enterprises a vendor-neutral base for governance. In practice, that made orchestration platforms, gateways, and registries the place where tool access scope, authentication, authorization, discovery, and audit requirements were explicitly managed rather than left inside each agent.

The control-layer role became clearest once fleets got large. Business Analytics Review described a team that had “spent six months wiring custom function calls into three different LLM hosts” before “Discovery moved to a single tools/list call at connection time” after adopting an MCP server; before that, “context windows regularly exceeded 80 percent utilization because entire query result sets were inlined,” and debugging meant “grepping logs across four services.” After eight weeks, latency fell 40 percent, context tokens by roughly a third, and onboarding dropped from two weeks to two days—evidence that standard discovery and observable primitives shifted governance into one surface. By early fall, vendors were selling that surface as infrastructure. Snowflake wrote that “Three enterprise earnings calls in August 2026 made the adoption curve concrete,” then defined a complete MCP gateway architecture as “a centralized control plane between AI clients and enterprise systems”; Citrix had launched NetScaler AI Gateway on April 9 and on July 9 “expanded this capability by adding Model Context Protocol (MCP) gateway functionality,” making it a unified control point, while F5 positioned its AI Gateway as a centralized runtime control layer that unified governance across models and MCP tool access. As F5 SVP of Product and Marketing Nirav Shah wrote, “F5 AI Gateway is purpose-built to close the governance gap,” and it “becomes the runtime enforcement point” for policy on every request, turning platform-level governance into real-time enforcement.

Sources

Token Costs Spark Tradeoffs

Skeptics highlight that MCP’s upfront token overhead can be steep, but new loading techniques have forced a rethink—trading universal access for smarter, leaner payloads and faster response times.

The sharpest critique of MCP was never that it could not work, but that it could become expensive before work even began. In January, an AI Engineer warned of a “meme… that the GitHub server ships like 200,000 tokens when you handshake with it,” adding that if “an LLM connects to an NCP server, it typically does download all the descriptions in one go,” then large fleets could overflow context windows; DevOps & AI Toolkit made the same point more bluntly: connect to “40 or 50 tools and you've burned tens of thousands of tokens before asking a single question.”

That did not make the skeptic case absolute, but it did make MCP a design tradeoff rather than an automatic upgrade over CLI or custom integrations. VentureBeat reported Anthropic’s “Claude Code Unveils MCP Tool Search, Slashing Token Use by 85% and Boosting Accuracy,” while an AI Engineer said on-demand loading cut MCP payloads from 150k tokens to 2k; similarly, progressive disclosure loads only “name and description” first, not “500 lines of instructions,” and one builder cut exposure “instead of 21 tools… I expect to see less,” yet Arise AI still found MCP took 71 round trips and 8 minutes versus seven CLI turns in under one minute.

Sources
DevOps & AI ToolkitVenture BeatAI EngineerFragmented - AI Developer PodcastAI EngineerAI Engineer

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.