Multiverse Computing, a Spanish scaleup operating at the intersection of AI and quantum physics, has set its sights on up to $570 million in a funding round that would value it at $1.7 billion. A significant leap for a company focused on compressing Large Language Models by 80-95% without meaningful accuracy loss, just as AI inference moves decisively toward edge devices.

The core technology, CompactifAI, takes a quantum-physics-inspired approach to reshaping model architecture, drastically reducing footprint and enabling direct execution on drones, cameras, satellites, vehicles, and telecom gear. Unlike conventional quantization, which typically trades precision for lighter compute, CompactifAI promises a nearly painless compromise between performance and size. Early adopters — including Allianz, Bosch, Bank of Canada, Indra, and Telefónica — suggest the bet is more than theoretical.

The round, still open and co-led by Forgepoint Capital International, BNPP SIVF, and Bullhound Capital, with commitments from Santander Alternative Investments, Tikehau Capital, Orange Ventures, and Scania Invest, could bring the company’s total funding to around $800 million. A fivefold increase in valuation from its previous round signals how strongly investors are backing a paradigm shift: AI no longer confined to data centers but distributed across an ecosystem of local, often disconnected devices.

The analysis must go deeper. If CompactifAI delivers, the implications cut to the heart of AI infrastructure debates. First, squeezing models down to 5-20% of original size makes them viable on hardware with minimal VRAM, reducing dependence on high-end GPUs and cloud services. Enterprises with data sovereignty requirements — defense, finance, healthcare, telecom — can run inference locally without sacrificing quality, lowering TCO and exposure risk. Second, the balance of power shifts from hyperscalers toward system integrators and embedded device makers, who can now differentiate with on-device AI without prohibitive costs. Third, it puts competitive pressure on GPU vendors for inference: as models demand fewer resources, the appetite for costly accelerators may shrink.

A caveat, however. The technology invokes quantum physics, but it’s unclear how much is genuinely quantum and how much is advanced tensor algebra. The accuracy drop is called “immaterial,” but independent testing across diverse benchmarks remains limited. Early adoption in critical sectors could surface hidden pitfalls in specific domains.

For those evaluating on-premise or edge deployment today, the Multiverse Computing story is a bellwether. Aggressive compression solutions could tip the calculus between cloud and local, making dedicated GPU clusters unnecessary. At AI-RADAR, we regularly unpack these trade-offs: the /llm-onpremise framework helps weigh latency, privacy, cost-per-token without chasing fads.

The round isn’t closed yet, but a $1.7 billion valuation suggests a strong bet on a future where AI doesn’t need a permanent connection to a data center. If that future arrives, the industry’s deck will have been well and truly shuffled.