AI model compression startup Multiverse raises $570M at $1.7B valuation

Money is pouring into the idea that AI doesn't have to live exclusively in a hyperscaler's basement. Multiverse Computing SL announced Monday it closed a $570 million Series C, pushing its valuation to $1.7 billion pre-money. For context, this is the same company that raised $215 million at a much smaller number back in June. Five months, almost a fivefold step up. Forgepoint Capital International, BNPP SIVF, and Bullhound Capital co-led the round, with Santander, Tikehau, HP, Orange Ventures, Scania Invest, and NAventures among a dozen other backers writing checks.

What's actually being sold here is compression, not chatbots. Multiverse's product, CompactifAI, borrows tensor networks — a mathematical tool from quantum physics — to strip large language models down without gutting their accuracy. The company claims it can cut a model's hardware footprint by 80% to 95%. In one recent demo, it took Meta's Llama 3.3 70B and shrank its disk footprint from 130 gigabytes to 65, letting it run on an Intel Xeon 6 CPU instead of a fleet of GPUs. It still needs roughly a terabyte of RAM, so this isn't quite a smartphone trick yet, but it's a meaningful step away from the assumption that bigger models always need bigger iron.

CEO Enrique Lizaso frames the whole pitch as busting a myth the industry has quietly accepted for years — that raw power demands expensive infrastructure. Whether or not that's overstated marketing, the practical case is straightforward. Run inference locally and you cut latency, since the round-trip to a cloud server disappears. You also keep sensitive data off someone else's servers entirely, which matters a lot to banks, defense contractors, and hospitals. Multiverse's roster of customers — Allianz, Bank of Canada, Bosch, Iberdrola, Telefónica — reads like a list of industries that care deeply about both of those things.

The company isn't just selling a smaller model file, either. It's building a routing layer, CompactifAI Router, that decides on the fly whether a workload can stay local or needs the cloud, plus a broader software stack meant to let enterprises run their entire AI pipeline behind their own firewall, no hyperscaler required. That's an ambitious pitch — essentially offering companies an exit ramp from the Nvidia-and-cloud-provider toll road that dominates AI economics right now.

The fresh capital is earmarked for exactly that ambition: more R&D on compression algorithms, investment in what the company calls AI gigafactory infrastructure, and expansion into East Asia, Southeast Asia, the Middle East, Canada, and the U.S. If the technology holds up outside of curated demos, Multiverse is betting that the next phase of the AI buildout isn't about bigger data centers, but about needing far fewer of them.