Meta's open-source Velox completed its repositioning from a query acceleration library to a “unified data execution layer” in 2026, aiming to address the bottlenecks of high latency and complex hardware coordination when Agents access data. This means competition in AI applications is moving down to the execution infrastructure, rather than remaining confined to model parameters and conversational performance.

What this is

Velox is an open-source C++ execution library initially built to accelerate Presto queries. As Agents—AI programs capable of autonomously invoking tools, reading data, and completing multistep tasks—became the primary consumers of enterprise data, traditional analytics engines had to handle high-concurrency queries, vector retrieval, and model inference while coordinating different hardware such as CPUs and GPUs.

Velox's approach is to provide a unified execution layer that lets different engines and hardware share optimization capabilities, reducing the cost of repeatedly moving data among Presto, Spark, and AI inference systems. VeloxCon China 2026, taking place in Shanghai on December 5, will focus on GPU acceleration, Agent infrastructure, and unified multiengine architectures.

Industry view

We believe this direction has practical value. When enterprises deploy Agents, the challenge is often not simply whether the model can produce an answer, but whether it can reliably retrieve and analyze internal data at low cost. A unified execution layer can improve latency, resource utilization, and operational complexity, making it particularly attractive to enterprises with large volumes of business data.

However, an execution layer does not automatically become infrastructure simply because the concept is advanced. A unified architecture must support Presto, Spark, GPU inference, and multiple data sources, and its engineering complexity may be substantially higher than that of optimizing a single query system. It is also difficult to fully reconcile differences in semantics, stability, and operational practices across engines. More importantly, Velox's effectiveness still needs to be validated in large-scale production environments; its status as a universal standard cannot be inferred from the technology narrative alone.

Impact on regular people

For enterprise IT: When purchasing Agent platforms in the future, evaluators may shift their focus from the results of model demonstrations to data connectivity, execution speed, access controls, and total cost of ownership. Data governance and system integration capabilities will become even more important.

For individual careers: Agents are expected to read enterprise reports, documents, and business systems faster, but whether companies can safely open these systems to Agents depends on permission isolation, result traceability, and clear accountability—not on model capability alone.

For the consumer market: As backend query and analysis speeds improve, applications that directly use data, such as customer service and shopping recommendations, may become more responsive. This is an infrastructure change, however, and the final user experience will still depend on whether product teams actually adopt Velox and whether they can control compute costs.