Spanish startup Multiverse Computing has secured a landmark €189 million (about $215 million) Series B funding round to scale a technology that could fundamentally alter the economics of artificial intelligence. The company’s quantum-inspired software, CompactifAI, is designed to compress the size of powerful Large Language Models (LLMs) by up to 95%, a move aimed at slashing the immense operational costs of AI and enabling advanced models to run on everyday devices far from the cloud.
The investment, detailed in an announcement from Multiverse Computing, was led by Bullhound Capital with significant participation from strategic investors including HP Tech Ventures and Santander Climate VC. With a portfolio of 160 patents and 100 customers globally, including Bosch and the Bank of Canada, Multiverse is tackling one of the biggest hurdles in the AI industry: the staggering cost of running models, a process known as inference. The company’s technology promises to make these models 4x to 12x faster, which could translate to an inference cost reduction of 50% to 80%.
This infusion of capital is aimed squarely at what Polaris Market Research estimates is a $106 billion AI inference market. By dramatically shrinking models like Llama and Mistral, Multiverse intends to make it feasible to run sophisticated AI not just in data centers, but on PCs, smartphones, cars, and even tiny computers like a Raspberry Pi. “The prevailing wisdom is that shrinking LLMs comes at a cost. Multiverse is changing that,” said Enrique Lizaso Olmos, the company’s CEO.
The Technology: Quantum Physics Meets Neural Networks
At the heart of Multiverse’s technology is a radical departure from conventional compression techniques. Instead of using standard methods like quantization or pruning, CompactifAI employs tensor networks, a mathematical framework inspired by quantum physics, to remap a neural network’s complex architecture. This approach, pioneered by Multiverse co-founder and Chief Scientific Officer Román Orús in his earlier research, allows the company to identify and eliminate billions of redundant parameters within a model.
“For the first time in history, we are able to profile the inner workings of a neural network to eliminate billions of spurious correlations to truly optimize all sorts of AI models,” said Orús. The result, as detailed in a company paper published on arXiv, is a compression ratio that far exceeds traditional methods.
Multiverse says it achieves up to 95% compression while retaining 97-98% of the original model’s accuracy—a significant improvement over industry-standard techniques that can cause a 20-30% accuracy loss for a much smaller size reduction.
This “lossy” but highly efficient approach stands in contrast to “lossless” methods like the DFloat11 technique. As outlined in its research paper, the creators of DFloat11 argue that for sensitive applications, guaranteeing bit-for-bit accuracy is critical, as lossy methods can introduce unpredictable variables that require extensive testing. Multiverse is betting that for most applications, its near-perfect accuracy at a fraction of the size is a winning formula.
A $215 Million Bet on AI Efficiency
The massive funding round is not just a vote of confidence in Multiverse’s technology but also a significant event for the European tech scene. According to the recent ‘State of Quantum 2025’ report, European quantum startups historically receive a disproportionately small share of global funding, making this €189 million round a major outlier.
Per Roman, Managing Partner at lead investor Bullhound Capital, noted that the company’s ingenuity is “accelerating European sovereignty” by introducing “material changes to AI processing.”
The business strategy focuses on the most expensive part of the AI lifecycle: inference. A recent market report from SNS Insider on Edge AI chips found that the inference segment accounted for roughly 75% of the market’s total revenue in 2024. By making this process cheaper and more efficient, Multiverse positions itself as a foundational layer in the AI infrastructure stack, offering its compressed models on its CompactifAI product page.
The Crowded Race to Shrink AI
Multiverse is entering a competitive landscape where various companies are tackling the AI efficiency problem from different angles. While Multiverse focuses on compressing the static model for inference, other innovations target different parts of the AI lifecycle.
For instance, Alibaba’s ZeroSearch framework was developed to slash the costs of training an AI by simulating search engine interactions, as detailed in a scientific paper on arXiv. Meanwhile, Sakana AI’s memory optimization system improves efficiency on long-context tasks by dynamically managing the model’s active memory, or KV cache.
The push toward the edge is also seeing new collaborations. In a new collaboration, Nota AI and Wind River are working to deploy generative AI models directly on edge devices in the automotive and IoT sectors. Avijit Sinha of Wind River affirmed the importance of this trend, noting that “AI model optimization and software-defined automation will be key” to unlocking new applications at the edge.
Unleashing AI from the Data Center
The ultimate goal of these efficiency gains is to power an “AI on the Edge” revolution, moving intelligence from centralized cloud servers onto local devices. This shift is driven by several key benefits, including the need for low-latency responses, enhanced user privacy by keeping data on-device, and reduced operational costs.
This is the future that investors like HP Tech Ventures are banking on. Tuan Tran, President of Technology and Innovation at HP, explained that by making AI more accessible at the edge, Multiverse’s approach has the potential to bring benefits like enhanced performance and privacy to companies of any size.
The commercial opportunity is enormous, with the Edge AI market projected to soar from $21.4 billion in 2024 to over $221 billion by 2032. By enabling complex AI to run on a simple device, Multiverse isn’t just cutting costs—it’s aiming to redefine where and how artificial intelligence can be used.
