Quantum-Inspired AI and Tensor Network Compression with Román Orús

Román Orús is one of the rare physicists who built a foundational mathematical tool — tensor networks — and then watched it become the engine of a unicorn. His 2013 introduction to tensor networks has been cited over 2,000 times; his company, Multiverse Computing, just announced a $570 million Series C at a $1.7 billion pre-money valuation. That arc — from condensed matter theory to Europe's largest quantum software company — is worth understanding on its own terms. But what makes this conversation particularly timely is a May 2026 paper Orús co-authored demonstrating that individual layers of Meta's Llama 3.1 8B language model can be encoded as quantum circuits and executed on IBM's 156-qubit Quantum System Two while the model generates text. It's a proof of concept, not a product — but it's a real result, and Orús is honest about what it does and doesn't prove.

This episode is for listeners who want a technically grounded, hype-free account of the quantum-AI intersection: what tensor networks actually are, why they keep getting rediscovered across different fields, where classical simulation of quantum systems genuinely competes with quantum hardware, and what it looks like to build a company at the boundary between those two worlds.

Sponsor Message

The Capital of Quantum is a people story. Built on a top-five quantum PhD program  and 35-plus years of quantum research leadership. It's the billion-dollar initiative behind Discovery Center, launching this month with Microsoft, IQM, and Quantum Motion inside. That's why IonQ was born and is headquartered here, and why global companies keep choosing a spot minutes from Washington, D.C. This is where quantum is transforming the world. Come see it at the Quantum World Congress, September 23rd through 25th, College Park, Maryland. CapitalOfQuantum.com.

What We Get Into

• What tensor networks actually are — Orús explains the core idea without equations: tensors as the "DNA" of a quantum state, and how a network of them lets you see and quantify the internal correlations (entanglement) that matter versus the ones you can safely ignore.
• Why the same math keeps appearing in different fields — from condensed matter simulation to quantum computing simulation to machine learning, and why Orús sees that recurrence as a sign of something deep rather than a coincidence.
• How ChatGPT changed Multiverse's trajectory — the company was already applying tensor networks to machine learning before 2022; the emergence of large language models gave them a problem where the fit was obvious and the market was enormous.
• What "90–95% compression with minimal accuracy loss" actually means — Orús explains the overparameterization problem in current AI models and why he believes tensor networks address a genuine structural inefficiency, not just a tuning opportunity.
• The IBM kicked Ising model episode — Orús describes how his team rapidly produced a classical tensor network simulation of an experiment IBM had presented as evidence of quantum utility, and what that kind of competition between classical and quantum methods actually does for the field.
• The Cayley Unitary Adapter experiment — how Multiverse sliced individual layers out of Llama 3.1 8B, encoded them as quantum circuits, ran them on a 156-qubit IBM processor, and achieved a 1.4% perplexity improvement — and why Orús argues the improvement-per-parameter ratio is the number that matters, not the headline percentage.
• Why edge deployment is the real commercial driver — drones, satellites, vehicles, and industrial devices that cannot rely on cloud connectivity are the market pulling Multiverse toward smaller, more efficient models, not just benchmark competition with frontier labs.
• How Orús thinks about Multiverse's identity — he calls it a "quantum AI company," not a quantum company or an AI company, and explains what that distinction means for how they allocate research effort and where they expect to be when fault-tolerant quantum hardware matures.
• What he'd tell a PhD student today — a genuinely honest answer about the trade-offs between academic research and deep-tech industry, from someone who has lived both simultaneously.

Resources & Links

Guest & Company

• Román Orús — Personal Site — Lists talks, reviews, affiliations, and awards including the 2024 Physics, Innovation and Technology Prize from the Royal Spanish Society of Physics.
• Multiverse Computing — Official Website — Home page for CompactifAI, Singularity, and Multiverse's full product portfolio.

Papers & Articles Discussed in This Episode

• arXiv 2605.05914 — "Quantum-enhanced Large Language Models on Quantum Hardware via Cayley Unitary Adapters" (May 2026) — The paper at the center of the episode: Llama 3.1 8B layers running on IBM's 156-qubit Quantum System Two. Start here if you want the technical details behind the quantum-in-an-LLM result.
• Multiverse Computing — "Talking to a Quantum Computer" (May 2026) — The accessible blog-post version of the Cayley Unitary Adapter experiment; a good entry point before tackling the arXiv paper.
• arXiv 2401.14109 — CompactifAI: Extreme Compression of Large Language Models Using Quantum-Inspired Tensor Networks — The foundational peer-reviewed paper introducing Multiverse's tensor network compression method.
• arXiv 2509.06653 — "Classical Neural Networks on Quantum Devices via Tensor Network Disentanglers" (Sep 2025, rev. Apr 2026) — Orús, Singh, and Aizpurua on hybrid classical-quantum execution for bottleneck neural network layers; the research underpinning the longer-term hybrid architecture vision.

Models & Products

• HyperNova 60B on Hugging Face — Open-source 50%-compressed model derived from GPT-OSS-120B; context for Multiverse's open model strategy.
• Pulsar 16B Launch with NVIDIA (June 2026) — Announcement of Multiverse's open reasoning model, scoring 87.22 on AIME 2025 at 16B parameters.

Funding & Company Context

• Series C Announcement — GlobeNewswire (July 2026) — $570M / €500M round at $1.7B pre-money valuation; covers investor lineup and deployment verticals.

Key Quotes & Insights

> "We are using atomic bombs to kill a mosquito." Orús on the overparameterization of current large language models — and why he believes the transformer-attention paradigm, however successful, ...