Multiverse Computing Lines Up a $570M Series C at $1.7B to Shrink LLMs With Quantum Math

TL;DR

Multiverse Computing, the San Sebastián startup that compresses large language models with quantum-inspired tensor networks, announced a Series C with commitments targeting up to $570 million (€500 million) at a $1.7 billion (€1.5 billion) pre-money valuation, per its July 27 press release. Forgepoint Capital International, BNPP Solar Impulse Venture Fund, and Bullhound Capital co-led. The product is CompactifAI, which the company says cuts model size by 80-95% with "immaterial accuracy loss" so inference can move out of hyperscaler data centers and onto edge devices and your own hardware. If the round closes at target, total funding reaches roughly $800 million.

The round

The valuation is a 5x step-up from the June 2025 Series B, a $215 million (€189 million) round led by Bullhound, per The Next Web. Beyond the three co-leads, the roster runs deep: Santander Alternative Investments, Tikehau Capital, HP Inc., Orange Ventures, Scania Invest, NAventures, Qatar Development Bank, Zouk Capital, SETT, the EIC Fund, the Basque Government's Hazten Scale-Up Fund, and Kutxa Fundazioa.

Read that list again: a PC maker, a telecom, a truck maker, and two banks. Those are strategics buying a supplier, not tourists chasing a logo. HP putting money into a company whose whole pitch is "run capable models on modest hardware" tells you where it thinks the edge-AI market is going.

One wording detail matters. The company says the round is "targeting up to" $570 million in commitments and "may remain open to select additional strategic investors," a nuance Tech.eu flagged while half the internet's headlines went with "raises." Commitments are not a wire transfer. The number is real enough to report; it is not yet money in the bank.

Quantum math, no quantum computer required

Multiverse was founded in 2019 and its chief scientific officer, Román Orús, is a physicist who built his career on tensor networks: the toolkit many-body quantum physicists use to represent absurdly large systems compactly. CompactifAI applies that math to LLM weight matrices, factoring giant matrices into networks of small interconnected tensors that keep the strongly correlated structure and throw away the redundancy.

If that sounds abstract, picture a 10,000-row spreadsheet where most rows are really combinations of the same 40 underlying patterns. You do not store the 10,000 rows; you store the 40 patterns plus the recipe for recombining them. Tensor networks do that to a model's weights, layer by layer, which is a different axis of attack than quantization (fewer bits per weight) or pruning (deleting weights outright), and it can stack on top of both.

The company's own product page claims the compressed models deliver 50-80% lower inference costs and up to 2x faster inference, and it sells them two ways: an inference API on AWS, or private deployment in your own cloud, on-prem, or on edge devices. It packages compressed versions of open models such as Meta's Llama, plus its own headline model, HyperNova 60B, which it calls "the world's most efficient model in its category."

Why this lands on a blog about running your own AI

Because the entire thesis is that inference belongs where homelabbers already put it: on hardware you control. The press release says CompactifAI models are deployed across millions of devices, including drones, cameras, satellites, vehicles, and telecom infrastructure, for customers like Allianz, Bank of Canada, Bosch, Iberdrola, Indra, PwC, and Telefónica. Forgepoint's managing director describes the company as having evolved "from being the leading downstream LLM compression technology to becoming a complete AI foundry and Operating System."

Europe has struggled to mint AI companies at frontier-lab valuations, and its newest unicorn is, in very European fashion, a company devoted to using less. But the strategy is coherent: skip the capex war entirely, let American and Chinese labs spend billions training open-weight models, then make money making those models cheap to run. In a month when the industry is debating $250 billion data center backstops, a $1.7 billion bet on needing fewer GPUs is a genuinely contrarian position.

The self-reported growth numbers back the demand story, with an asterisk. The company claims 10x annualized revenue growth since the Series B and Q1 2026 sales up 96x year-over-year. A 96x multiple mostly tells you how small the starting quarter was; the 10x figure is the one doing real work.

The caveats

Four things to keep straight. First, this is a round announcement, not a closing: commitments "targeting up to" $570 million, with the round possibly staying open. Second, the revenue multiples are self-reported and unaudited. Third, every compression and accuracy number here is the company's own; "immaterial accuracy loss" is a marketing phrase, not a benchmark table, and no independent evaluation is cited in the release. Fourth, the competition is partly free: the open-source ecosystem already quantizes models at zero cost, and compression research (low-bit, ternary, distillation) is one of the most crowded corners of the field. What enterprises are paying Multiverse for is the packaging, support, and compliance wrapper, which is a real business, but a different one from owning irreplaceable math.

Key Takeaways

• Multiverse Computing announced Series C commitments targeting up to $570 million (€500 million) at a $1.7 billion pre-money valuation, co-led by Forgepoint Capital International, BNPP Solar Impulse Venture Fund, and Bullhound Capital.
• That is a 5x valuation step-up from its $215 million Series B in June 2025, and takes lifetime funding to roughly $800 million if the round closes at target.
• CompactifAI uses quantum-inspired tensor networks to shrink LLMs by a claimed 80-95%, with 50-80% lower inference costs and up to 2x faster inference, per company figures.
• Strategic backers like HP, Orange, Scania, and Santander signal demand for inference on edge devices and owned hardware rather than hyperscaler data centers.
• The round is "targeting up to" $570 million and may remain open; growth figures are self-reported and compression claims lack cited independent benchmarks.
• The bet is contrarian in the best way: while rivals finance gigawatt data centers, Multiverse is priced at $1.7 billion for making models need less.

Sources: GlobeNewswire press release, Tech.eu, The Next Web, Multiverse Computing: CompactifAI