As venture investors warn startups to lock in cloud-compute contracts amid rising default rates, one Spanish company is betting that the safest infrastructure is no infrastructure at all. Multiverse Computing is pushing a radically different vision of artificial intelligence: compressed models small enough to run directly on phones, laptops and edge devices, sidestepping hyperscalers and their financial and technical bottlenecks.

Multiverse built its reputation on quantum-inspired algorithms for optimization and simulation. Now it is applying that expertise to AI compression, shrinking models from OpenAI, Meta, DeepSeek and Mistral AI so they can run with far less memory and compute. The company argues that this is not just an efficiency play but a way to harden AI against supply-chain shocks and privacy risks.

Its consumer showcase is CompactifAI, an app that behaves like a standard AI chat assistant but hides an unusual architecture. At its core is Gilda, a compact model designed to run locally and offline on sufficiently powerful devices. When a user’s hardware cannot cope, an orchestration layer called Ash Nazg silently routes the request to cloud-based models instead. The result is a hybrid system that demonstrates what on-device AI can do today, while acknowledging that many existing phones are not yet up to the task.

CompactifAI’s modest download numbers suggest Multiverse is not chasing mass consumer adoption. The real bet is on enterprises. The company has opened a self-serve API portal that lets developers plug directly into its compressed models, monitor usage in real time and decide whether to deploy them on-premises, at the edge or in the cloud. For customers wrestling with soaring GPU bills, the promise is straightforward: smaller models, lower latency and lower cost.

Multiverse’s latest flagship, HyperNova 60B 2602, is derived from OpenAI’s gpt-oss-120b and is engineered to deliver faster, cheaper responses than the original, particularly for agentic coding workflows in which AI systems autonomously plan and execute multi-step programming tasks. That kind of workload magnifies even small efficiency gains.

The company already works with more than 100 customers, including financial institutions, industrial giants and energy utilities. Many of these clients care less about chatbots than about embedding AI into drones, satellites, factory equipment and other environments where connectivity is patchy and data sensitivity is high. For them, Multiverse’s compressed models are not a curiosity; they are a potential path to making advanced AI both deployable and dependable at the edge.