Enterprise AI has quietly crossed a line. A year ago, most organizations were still asking whether generative AI belonged in production. Today, the question has shifted to whether their data can keep up with agents that plan, call tools, execute multi-step work, and touch core systems on their own. At Dell Technologies World 2026, Dell Vice Chairman and COO Jeff Clarke put it bluntly: reasoning-model token use is up 320x year over year even as unit token costs have fallen roughly 80%, and in most enterprises 80 to 90% of data is still unstructured and scattered across dozens of systems. Agents are getting smarter and cheaper to run per token. The data beneath them is not getting easier to reach.

That gap is the real bottleneck in enterprise AI right now, and it is why the conversation needs to move past the “AI data platform” as a static repository and toward something more demanding: a semantic, self-operating data plane built specifically for agents. Call it the agent-ready data platform.

Agents Don’t Just Query Data. They Depend On It, Continuously.

A chatbot that answers a question once is forgiving of a messy data estate. An agent that plans a task, retrieves context, calls three tools, checks its own work, and repeats that loop dozens of times before finishing is not. Every one of those steps re-reads accumulated context, which is why, as Clarke’s team has noted, most of the token load in agentic workloads lands on the input side, not the output. If the retrieval underneath that loop is slow, ungoverned, or blind to half the enterprise’s unstructured content, the agent either fails, hallucinates, or burns tokens compensating for a weak foundation. Clarke’s own framing at Dell Technologies World was direct: the architectural principle has to be “move AI to the data, not data to the AI”.

This is also, increasingly, a cost problem. As agentic systems take on more cognitive work, spend migrates from headcount to tokens. A model that has to re-retrieve, re-rank, and re-reason because the underlying data plane is slow or fragmented produces worse answers and more expensive ones, at a moment when global token consumption is projected to grow roughly 3,400% by 2030. Enterprises that solved this with brute-force cloud compute are already seeing the bill: Dell has shown that a single developer running ten agents can generate a $3,400 cloud bill in a single day, workloads that cost effectively nothing when the data and compute sit closer together.

Two Planes, One Discipline

Inside Dell’s own architecture for enterprise AI, one framework captures where this problem actually lives. Enterprise AI runs on two distinct planes: a data plane, where enterprise data is discovered, prepared, governed, and retrieved to create trusted context, and an execution plane, where agents plan, act, and orchestrate work across the business. The two are tightly connected, but they have to be built and optimized separately, because AI only succeeds when the data underneath it is trustworthy and the execution on top of it can scale. Optimize the data path and you optimize the agentic outcome sitting on top of it. The reverse doesn’t hold.

That reframing matters because most enterprise AI investment to date has gone into the execution plane: models, agent frameworks, orchestration logic. The data plane has been treated as plumbing. An agent-ready data platform inverts that assumption. It treats the data plane as the product.

What “Agent-Ready” Actually Requires

Three characteristics separate a data plane that merely stores enterprise data from one that is ready to sit underneath autonomous agents.

It has to be semantic, not just searchable. Keyword search finds strings; agents need meaning. A data plane built for agentic AI has to unify structured and unstructured retrieval through hybrid search that blends lexical matching with vector similarity and learned sparse retrieval, so an agent can pull the right contract clause, the right sensor log, and the right customer record in a single pass, regardless of format. Ninety percent of enterprise data is unstructured, and that’s not a footnote; it’s the majority of the value an agent has to reason over.

It has to be self-operating. Enterprises cannot hand-wire a new pipeline every time a team wants to stand up an agent. Orchestration needs to automate the ingest-to-enrich-to-retrieve-to-infer flow end to end, managing lifecycle, scaling, and data movement, so data engineers stop rebuilding plumbing for every use case.

It has to be secured and accountable by default. Once agents start calling the CRM, the ERP, and financial systems on their own, teams need to log every touchpoint and be able to reverse it. As Clarke put it, “when an agent takes an action on your behalf, you need to know what it did, why it did it, and how to undo it if it got it wrong”. In an AI workforce, every action needs a receipt, and that discipline has to be built into the data layer, not bolted onto the agent.

A Concrete Example: Dell AI Data Platform With NVIDIA

This isn’t theoretical. It’s the design principle behind the Dell AI Data Platform with NVIDIA, and it’s a useful proof point for where the industry needs to go.

AIDP separates GPU-accelerated data engines from storage so each can scale independently, then unifies them under end-to-end orchestration. Its Data Search Engine, the semantic retrieval layer, combines BM25 keyword search, HNSW vector search, and ELSER learned sparse retrieval into one hybrid engine, and it runs GPU-accelerated vector indexing powered by NVIDIA cuVS. Its Data Processing Engine runs Apache Spark accelerated by NVIDIA cuDF for batch and real-time analytics and ETL, and its Federated SQL Engine, powered by Starburst, lets teams query across data sources without moving the data first. On top of that, NVIDIA NIM and NeMo Retriever bring the model-serving and retrieval microservices that let agents actually act on what the data plane surfaces.

The performance case for that architecture is measurable, not aspirational. In Dell’s own testing, AIDP delivers up to 12x faster vector indexing with NVIDIA cuVS, up to 6x faster SQL queries, and up to 19x faster time to first token for inference workloads using the vLLM engine, LMCache, and the NVIDIA NIXL library on NVIDIA H100 GPUs, plus up to 150 GB/s of throughput per rack unit, and NVIDIA GPU Direct support, which matters because it means the co-engineering between Dell and NVIDIA happened at the architecture level, not as an afterthought integration. That combination, semantic retrieval fast enough for agent-speed reasoning, orchestration that doesn’t require a new pipeline for every use case, and governance embedded in the platform itself, is what “agent-ready” looks like in practice, and it’s the reason Dell and NVIDIA jointly engineer this as one stack rather than a set of adjacent products.

Every enterprise racing to deploy agents will eventually run into the same wall: agents are only as capable, as fast, and as trustworthy as the data plane feeding them. Organizations that treat that data plane as core infrastructure, building it semantic, self-operating, and secured by design, will be the ones whose agents actually scale in production instead of stalling in pilot. The ones that keep bolting agents onto legacy storage and hoping retrieval keeps up will find out the hard way that tokenomics punishes weak foundations twice: once in latency, and once on the bill.

The AI-native enterprise isn’t defined by which model it uses. It’s defined by whether its data plane was built to operate itself, understand what it holds, and keep every agent action accountable before the agents showed up expecting it to.