Ben Horowitz 长文解析 Databricks 与智能体时代的公司形态
Key Highlights
a16z's Ben Horowitz wrote a long piece on why Databricks is becoming the infrastructure for the agent era. The numbers are striking: a $7B revenue run-rate, 80% year-over-year growth, and a $190B valuation, as the company evolves from a data platform into an agent substrate. The argument reframes Databricks not as a warehouse vendor but as the bedrock on which autonomous software will be built — a far larger addressable market than data storage alone.
The reframing is the interesting part. If agents are the next compute wave, then the platform that owns enterprise data and governance owns the on-ramp to that wave.
What the Argument Is
Horowitz's core thesis is that as enterprises hand core business processes to agents, they need a unified data, governance, and compute substrate — and Databricks' lakehouse precisely plays that role. Agents that call private enterprise data while staying compliant and auditable inherently depend on such a platform. The bet is that value in the agent economy accrues not to the model with the best demo but to the layer that safely connects models to a company's actual data.
That is a defensible claim. Models are increasingly commoditized and interchangeable; the durable moat is the governed data they run on, which is expensive and risky to move.
Technical Details
Databricks converges data engineering, ML training, model inference, and governance into one platform, so the path from data to agent is no longer fragmented across a half-dozen tools. Products like Mosaic AI push inference and agent orchestration, letting enterprises deploy models securely on their own data instead of moving it to external APIs. That "keep the data where it is" posture is increasingly important as regulators and boards scrutinize how sensitive information flows into third-party model endpoints.
Consolidation reduces both cost and risk. Every tool boundary is a place where governance can break and latency can creep in, so collapsing the stack into one platform is as much a compliance strategy as a technical one.
Versus the Competition
Against Snowflake's data-warehouse tilt and Azure or AWS's cloud-resource tilt, Databricks' strategy is "data plus AI as one." In the agent wave, whoever seamlessly feeds enterprise data to agents with governance controls the entry point. That makes its valuation logic closer to an "AI infrastructure monopolist" than pure SaaS, because the switching costs of moving both data and agent logic are enormous once a company is fully onboarded.
The comparison is instructive: Snowflake owns the warehouse, clouds own the compute, but Databricks is arguing it owns the operating layer where data becomes action — and action is where the value is.
Industry Impact and Use Cases
The piece signals a trend: the agent race is not about a single model but about who provides a trustworthy, compliant, governable data and execution substrate. For CIOs, choosing a platform is choosing the future agent "operating system." For startups, it means big opportunity in building agent middle layers around the lakehouse ecosystem — the connectors, evaluators, and guardrails that sit between raw data and shipped autonomous workflows.
The startup opportunity is real but parasitic: it thrives only where the platform thrives, which is the risk of building on someone else's substrate.
Broader Context
Horowitz's essay captures a broader truth: in the agent economy, infrastructure wins. The models may grab headlines, but the durable value sits in the governed data layer that makes agents safe to deploy in enterprises. Investors and builders should watch who controls that layer, because it is likely to capture a disproportionate share of the value the agent wave creates.
What to Watch
Horowitz's essay captures a broader truth: in the agent economy, infrastructure wins. The models may grab headlines, but the durable value sits in the governed data layer that makes agents safe to deploy in enterprises. Investors and builders should watch who controls that layer, because it is likely to capture a disproportionate share of the value the agent wave creates, while the models themselves become a commodity priced toward marginal cost.
The lesson for builders is simple: own the governed data layer, and the models on top become interchangeable commodities you can swap without losing your moat, because the durable value lives in the data and the rules around it rather than in any single model.