A thesis on where data is going
Raw data was the last generation's infrastructure problem. The companies that define the verified data layer will power the coming agentic wave and become indispensable partners to both the warehouse layer and the inference layer.
01
Enterprise data infrastructure was built for humans: deliberate, business-hours, one query at a time. Business users file a request. Analysts build (or repair) a complicated dashboard. The pace of the system has always been the pace of the people using it.
That assumption is about to break. Agents don't work in business hours, don't ask one question at a time, and don't wait their turn. The same task a human used to run once becomes a fan-out of dozens of parallel queries, executed continuously, against systems that were never sized for it.
The shift isn't theoretical and it isn't years away. It's already happening inside every enterprise running an AI deployment — and the data layer is the first thing to feel the strain.
Projected demand
More data demand from a partially agentified enterprise vs. today's human baseline.
02
The bottleneck has never been access to data. It's been the distance between a question and a trustworthy answer. That distance is what Supper is closing.
The "cost to serve" data needs in a human analyst world was high; so incumbent dashboard tools proliferated. In an agentic world, inference cost is now low enough for a new form factor to emerge: streaming answers.
03
The infrastructure shift towards agentic business processes is inevitable and will be a force multiplier for adopters.
Agents are error multipliers. A human analyst acts on a wrong number and makes one bad call. An agent fires 8 to 12 chained queries in a single session, each answer shaping the next. Feed it bad data at step one and every downstream decision is contaminated before anyone sees the output.
Many enterprise AI deployments are stalling — not because the models aren't capable, but because the data isn't trusted enough to let an agent act without a human in the loop. It works (sometimes), but it doesn't scale — and it gives back most of the value agents were supposed to create.
The underlying problem is that nobody has built a data layer designed for agents: verified, cross-system, context-driven and auditable.
04
The model providers need the data ecosystem to be open. The warehouse vendors need it to stay closed. That tension is where Supper operates.
05
The Virtual Data Warehouse is the governed data infrastructure layer for the agentic enterprise. It sits above every data source — internal databases, SaaS tools, cloud warehouses, streaming systems — and below every AI agent, BI tool, or analytical application.
It is not a database. Not a connector. Not a model. Not a bolt-on to an existing warehouse. It is a net new layer in the enterprise stack — the same way the data warehouse was a net new layer when it emerged above transactional databases in the early 1990s.
The continuous, machine-speed, multi-source, semantically precise data demand the VDW serves did not exist before agents. The existing stack was not built for it.
The VDW maintains canonical definitions of every metric, entity, and business concept across sources. "Revenue" means the same thing whether the agent is querying Salesforce, Snowflake, or a Postgres billing table. CAC, churn, pipeline, and margin are defined once, enforced everywhere, and versioned as business definitions evolve. An agent querying through the VDW never encounters ambiguous schema or conflicting metric definitions — it operates on a pre-built understanding of what the data means, not a fresh inference from raw column names.
The VDW enforces a unified access control model that treats AI agents as first-class principals alongside human users. Row-level security, column masking, data residency policies, and audit logging are enforced at the VDW layer — consistently, across every source the agent might touch in a single query. A sales rep's agent and a CFO's agent asking the same question receive answers scoped to their respective permissions.
The VDW sits between every agent query and every data source. It intercepts requests before queries are built, enforces result-set limits, routes large results to out-of-band delivery rather than the agent's context window, and applies budget controls. It eliminates the schema injection tax by serving pre-compiled, compressed semantic context rather than raw table definitions — reducing token overhead by roughly 100× and making that context cacheable at up to 90% savings. The cost per data operation is able to keep up with the new agentic demand scale.
The VDW handles joins and aggregations that span multiple source systems at the data layer, rather than requiring complex data pipeline management and bloated, byzantine warehouse structures. A question that spans your CRM, your data warehouse, and your internal product database resolves at the VDW layer, queries across multiple sources of record, and surfaces as a single governed, semantically coherent answer.
Every query — whether from a human analyst, a BI dashboard, or an autonomous agent operating at 3am — carries full provenance: what data was touched, what permissions were applied, what version of a metric definition was used, and what the result was. Compliance, legal, and finance teams have a complete, auditable record of every data access event across the enterprise, in a single place, regardless of which source system the underlying data lives in.
The bottleneck in enterprise AI isn't the models. It isn't the warehouses. It's in between — the context layer that knows what the data means, who can see it, and whether the answer is right.