Anyone paying attention to the AI tech space has heard of Model Context Protocol (MCP), the lightweight protocol that lets data sources, APIs and other tools talk directly with the AI. It’s the shift from providing the AI static data to it being able to query the latest data in real time.
But what happens when MCP is connected to petabyte-scale production data, not sampled datasets but real, live time series logs data? The answer is less simple than the vision suggests.
MCP Adoption Today
In practice, it’s an on-call engineer using MCP to diagnose what has changed in the last 6 hours. Or a CDN operations team investigating cache performance anomalies in plain English instead of having to write SQL. Security analysts querying billions of logs lines instead of relying on the accuracy of sampled datasets.
This is happening all across the industry in high volume data environments such as observability platforms, CDN analytics, security log stores, IoT telemetry, anywhere teams manage massive time-series data and need fast answers. Companies like Hydrolix, which operate in this space, have built MCP servers to connect AI tools directly to petabyte-scale data stores.
The teams that have been fastest to adopt are the ones lost in high-volume, time-sensitive data, because for them, the gap between “data exists” and “data is accessible” has always been the real bottleneck. These aren’t AI-first companies. They’re the teams that were drowning in data long before anyone called it an AI problem.
The problems showing up at scale
Traditional data infrastructure was built for human use. MCP uses them completely differently and the problems are showing up quickly.
Query Safety and Cost control: When MCP in conjunction with an LLM is generating SQL for a database that has petabytes of data, a slow query isn’t just a slow query. It becomes a resource drain that can rack up a huge bill accidentally. The model doesn’t know that scanning a month of data is much more expensive than scanning just a day, it just asks the next logical question.
Authentication and Identity: Most early MCP implementations are local set ups where the user is likely using their own credentials. While the MCP spec has made fairly fast progress on things like OAuth, IdP integration, proper role-based access, real world implementation is fragmented. The standard exists but getting it consistently implemented across every client and server in production is where the risk lives.
The Observability Blind Spot: When an AI tool queries a data source, for the most part it is anonymous, unattributed traffic that can be hard to distinguish from human traffic. How much of your query volume is coming from AI tools and MCP? You don’t know. Is MCP driving faster incident resolution for your on-call engineers? You can’t measure it. Without this visibility, it becomes harder to justify increased resources to support scaling adoption.
Data Volume Meets Token Limits: This is an interesting one because in traditional BI, a human understands the schema and knows roughly what they’re looking for. In the world of MCP, when someone asks the LLM “what has changed in the last 6 hours” the model then has to make a choice about how to query that data to fit within the context window. And that choice, aggregation or sampling, is reasoning that is happening before actual data reasoning begins. And by the end, the user has no idea what data has been considered or left out. In traditional analytics, you can always go look at the raw data returned. With AI as the intermediary, the raw data was never in the room.
Evolving Data Infrastructure
Data infrastructure has not typically been the most exciting topic when it comes to innovation but it really determines whether AI becomes transformative or just expensive. Here’s what early adopters are learning.
Understand the cost model before connecting anything. Even better, have an idea of what a runaway query might cost. Putting query guardrails in place at the core of the data, like query timeouts, row limits, scope restriction becomes a non-negotiable.
While the pilot phase may start with using personal API keys, that is fine for experimentation but it’s essential to have a plan and timeline for proper authentication before “temporary” becomes the default, we’ve all been there.
Before even starting with MCP, or as close to the start as possible, instrument MCP traffic so you can distinguish between human traffic, measure adoption, attribute costs and/or build a business case to expand.
Start with the teams feeling the most pain, usually the operators sitting closest to the data and systems that need answers preferably in minutes and not days or hours. They will let you know what works, what breaks and what doesn’t scale and these can be used as proof points for a broader rollout.
The model wars and the latest AI headlines will keep running. But MCP is where AI actually connects to real data in production, and the infrastructure underneath is what decides whether it works. The organizations paying attention to that layer now will be the ones that deliver.
Ashley Vassell, Senior Product Manager, Hydrolix