AI coding agents and frontier AI labs are reshaping how companies think about databases, moving away from cloud-only strategies toward hybrid architectures that blend on-premises infrastructure with multi-cloud flexibility. According to MongoDB CEO CJ Desai, the shift is driven by two converging forces: hyperscaler capacity shortages and the unique demands of agentic artificial intelligence workloads that require real-time, operational data access rather than batch analytics.

What's Driving the Return to On-Premises Data Centers?

The conventional wisdom that all enterprise workloads would migrate to public cloud providers like Amazon Web Services, Google Cloud, or Microsoft Azure is colliding with physical reality. Desai shared concrete examples of major companies hitting capacity walls. A Fortune 100 company in Texas, ranked among a hyperscaler's top 50 customers, wanted to move additional artificial intelligence workloads to the public cloud but was told no capacity existed. The company had been decommissioning its own data center; it reversed course and now plans to run those workloads on-premises instead.

A large U.S. telecommunications company faced similar rejection for regional capacity, forcing it to sign with a second hyperscaler and manage a multi-cloud environment, adding complexity it had hoped to avoid. The drivers behind this shift include data sovereignty concerns, particularly in Europe where regulatory requirements like French data residency rules apply, the need to control proprietary data, and the acute shortage of graphics processing units (GPUs) dense compute infrastructure.

• Data Sovereignty: European companies face regulatory requirements that mandate data residency within specific countries, making on-premises or region-locked cloud deployments necessary.
• GPU Scarcity: The physical shortage of GPU-dense computing resources means hyperscalers cannot accommodate all customer requests, even for their largest accounts.
• Proprietary Data Control: Enterprises want to keep sensitive intellectual property and customer data within their own infrastructure rather than relying on shared cloud environments.

How Do AI Coding Agents Differ From Traditional Database Workloads?

The rise of AI coding agents like Devin, developed by Cognition, has exposed a fundamental mismatch between traditional database design and agentic artificial intelligence requirements. Desai draws a sharp distinction between operational databases like MongoDB, which handle real-time transactions, and analytical databases like Snowflake or Databricks, which process batch queries for business intelligence.

MongoDB serves three distinct customer tiers in the artificial intelligence space. Frontier labs, such as undisclosed companies working on inference and voice products, use MongoDB as a memory layer for agents handling unstructured data at extreme scale. AI-native companies including 11 Labs, which powers over 50 million agents, and Cognition, the creator of Devin, rely on MongoDB's document model flexibility for live coding platforms, voice agent orchestration, and robotics applications. Large enterprises like banks, insurance companies, and telecommunications firms are still mostly in the prototyping phase, using MongoDB for customer-facing agentic applications.

The document-oriented approach matters because artificial intelligence agents need to iterate rapidly without schema rewrites. When a coding agent like Devin makes a decision, that decision must be recorded, retrieved, and acted upon in milliseconds, not hours. Analytical databases excel at answering questions like "How many customers churned last quarter?" but struggle with "What did this agent decide three seconds ago?"

Why Enterprise AI Architectures Are Becoming Increasingly Complex

Desai recounted a revealing encounter with a large New York-based bank that illustrated how nascent enterprise agentic architectures remain. The bank showed him an architecture diagram for a single agentic application containing 55 separate components: multiple large language models (LLMs), both closed-source and open-source; agent runtimes; databases; vector databases for semantic search; embedding services; observability tools; security modules; evaluation frameworks; guardrails; and deployment orchestration systems.

"I said, 'How would this picture have looked a year ago?' They said, 'Half of these boxes would have not been there because we realized in trying to create agents, observability, security, how you do eval, how you move things in production, guardrails, whatever the cases are.'"

CJ Desai, CEO at MongoDB

The bank expects the diagram to continue expanding, not contracting, as new concerns emerge around cost management, multi-model routing, and compliance auditing. Desai's interpretation is sobering: this is the slowest the artificial intelligence adoption pace will ever be. The infrastructure layer is still being built; deterministic outcomes, human-in-the-loop requirements, and regulatory compliance force enterprises to add rather than subtract components.

What Happens When AI Startups Outgrow Their Infrastructure?

AI-native companies often outgrow their infrastructure faster than anticipated, creating urgent operational challenges. Desai recounted a vivid story of a prediction-market client running a World Cup soccer application that texted him in distress: the company was days away from exhausting MongoDB disk capacity, and the tournament still had weeks remaining. The CEO was stressed about disk space, providing specific terabyte counts via text message.

The episode revealed a critical gap: such customers do not have, and cannot afford to hire, traditional database administrators (DBAs). They need auto-scaling that is both predictive and proactive. Desai noted that historically, "the most expensive person you can get is called Oracle DBA," but now artificial intelligence companies are saying they do not have time to hire these specialists. If software vendors want to serve this market, their products must include machine-based database administration built in.

Desai

Another example illustrates the multi-cloud portability challenge: a frontier lab released a new model version and found its MongoDB capacity in Amazon Web Services insufficient. It wanted to seamlessly move some database workloads to Google Cloud Platform. Customers increasingly demand frictionless multi-cloud portability because hyperscaler capacity is no longer guaranteed.

How Should Organizations Prepare for Agentic AI Workloads?

• Evaluate Operational vs. Analytical Needs: Determine whether your artificial intelligence workloads require real-time transaction processing (operational) or batch analytics (analytical), then select database infrastructure accordingly. Coding agents and live decision-making systems need operational databases.
• Plan for Multi-Cloud and Hybrid Architectures: Do not assume all workloads will stay in a single cloud provider. Design systems that can migrate workloads between on-premises, AWS, Google Cloud, and Azure without major rewrites.
• Build Auto-Scaling Into Your Stack: Artificial intelligence workloads can spike vertically with new model releases or viral adoption. Invest in databases and infrastructure that scale automatically without requiring human intervention or database administrator expertise.
• Account for Observability and Compliance Complexity: Enterprise agentic applications require observability, security, evaluation frameworks, guardrails, and compliance auditing. Budget for these components from the start rather than bolting them on later.

The convergence of hyperscaler capacity constraints and the unique demands of artificial intelligence agents is reshaping enterprise infrastructure strategy. Companies that built their strategies around "all-in on public cloud" are discovering that approach no longer works. Instead, the future belongs to organizations that can seamlessly blend on-premises, multi-cloud, and edge computing resources while maintaining real-time operational databases that agents can query and update in milliseconds. For AI-native startups and enterprises alike, data is no longer a supporting player in the artificial intelligence era; it is the unsung hero.