Perplexity CEO Aravind Srinivas: Neocloud Must Own a Complete Agent Inference Platform, Not Just Rent GPUs
Perplexity CEO Aravind Srinivas stated that Neocloud cannot just rent GPUs; it needs to have a complete Agent inference platform, including open-source weight models (SGLang), fine-tuned models (Miles, River), and a secure sandbox (SPACE).
He mentioned looking forward to collaborating with partners such as Coreweave, Akamai, and Crusoe; this statement referenced reports about Perplexity's open-source SPACE sandbox.
SPACE is a sandbox platform based on the Firecracker micro-VM, which has been used in production for Perplexity Computer, with a median startup latency of about 60 milliseconds, and is now open-sourced under Apache 2.0.
Miles is a high-performance reinforcement learning framework used in conjunction with SGLang for large-scale model post-training; Srinivas also previously mentioned vertically integrating the agent infrastructure and optimizing the sandbox for the best silicon.
Neocloud, like Coreweave and Crusoe, primarily provides GPU rental and clusters, and has been named a potential full-stack partner.
In market mechanisms, the buyer is the AI application party that requires secure Agent execution, while the seller is Neocloud's GPU capacity and Perplexity's software layer; the event is driven by the open-source sandbox, with resources flowing from pure hardware rental to software + hardware full-stack, benefiting partners that provide a complete inference platform, while putting pressure on providers that only rent GPUs.
This viewpoint expands Agent inference from GPU rental to the full-stack demand for model services and secure execution.
Source: Public Information
ABAB AI Insight
Aravind Srinivas has long promoted Perplexity's expansion from a search engine to an Agent platform, having open-sourced the SPACE sandbox and used it in production; he has now clearly stated that Neocloud needs to have a complete inference layer consisting of SGLang, Miles/River, and SPACE.
The capital path reflects the combination of secure sandbox and model service layer with GPU rental, collaborating with Coreweave, Akamai, Crusoe, etc., allowing pure hardware providers to extend into a full-stack inference platform, thereby capturing the software value of Agent execution.
This is similar to the historical evolution of cloud computing from hardware rental to platform as a service, and also resembles the trend of AI infrastructure expanding from training clusters to inference and Agent execution layers; currently, it is transitioning from GPU capacity competition to secure inference full-stack.
Essentially, this represents a restructuring of the industry chain: Agent inference is being restructured from simple GPU rental to a complete platform that includes open-source model services, fine-tuning frameworks, and isolated sandboxes, shifting value from hardware utilization to secure execution and model optimization.
ABAB News · Cognitive Laws
Just renting GPUs is not enough; full-stack is needed to retain Agents
The open-source sandbox turns the security layer into a collaboration entry point
The inference platform can determine customer stickiness more than computing power itself