A Spanish quantum-computing startup just raised $570 million at a $2.3 billion valuation on the back of 96x year-over-year sales growth — and it did it without building a quantum computer. Multiverse Computing’s product is quantum physics applied to a very unglamorous, very valuable problem: making today’s AI models drastically smaller and cheaper to run.

The raise

Multiverse, founded in San Sebastián in 2019, closed a $570 million Series C at a $2.3 billion post-money valuation — 5x its Series B — bringing total funding to $800 million. The round was co-led by Forgepoint Capital International, BNP Paribas’ Solar Impulse Venture Fund and Bullhound Capital, with Santander, HP, Orange Ventures and Qatar Development Bank among the participants.

The growth figures behind it are genuinely unusual: 96x sales growth in Q1 2026 year over year, and more than 10x annual revenue since the Series B.

What it actually does

The product, CompactifAI, is a compression engine that uses quantum-physics-based tensor networks to shrink large language models by 80–95% while, the company says, preserving accuracy. Compressed open-source models like Llama run 4–12x faster at 50–80% lower cost, and can operate on edge devices — phones, drones, factory equipment — with no cloud connection.

Note what’s clever here. This is not a quantum computer, and it doesn’t need one. It borrows the mathematics of quantum many-body physics — tensor networks, developed to represent quantum states efficiently — and applies it to the classical problem of representing a neural network with far fewer numbers. Quantum-inspired, running on ordinary hardware. That distinction is why it can ship to customers today rather than in a decade.

Why the timing is perfect

The reason a model-compression company is growing 96x is that the entire industry just discovered its AI bills are unsustainable. Running large models in the cloud costs per token, and those costs scale brutally with usage. There’s a live scramble for anything that makes AI cheaper to run — the same pressure driving vendors toward consumption pricing and enterprises toward local inference.

Compression sits exactly on that nerve. If you can run a model at 80% lower cost, or run it on a device you already own instead of renting cloud GPUs by the token, the economics of deploying AI change materially. And edge deployment unlocks cases the cloud can’t serve — a drone or a factory line that can’t depend on connectivity, or data that can’t legally leave the building. Customers already include Allianz, Bank of Canada, Bosch, Iberdrola, Telefónica and PwC.

The caution

Two things temper the enthusiasm. First, “80–95% smaller while maintaining accuracy” is a vendor claim, and compression always trades something — the interesting question is what degrades, on which tasks, and whether “maintaining accuracy” holds on the specific workloads a customer cares about rather than on benchmarks. Independent evaluation is what turns that claim from marketing into fact.

Second, model compression is an active research area with many approaches — quantisation, distillation, pruning — and the frontier labs building the models have every incentive to make their own models cheaper to run. A compression specialist’s moat depends on staying meaningfully ahead of techniques that are also improving fast and are partly commoditised in open-source tooling. 96x growth says the current lead is real and valuable; it doesn’t guarantee it’s durable.

The read

Multiverse is a bet on a genuinely important and unglamorous problem — AI is too expensive to run, and making it cheaper is worth a great deal — solved with a genuinely clever borrowing from quantum physics that, crucially, ships on today’s hardware. The growth is extraordinary and the customer list is real.

The open questions are the accuracy trade-off under real workloads and the durability of the lead against a fast-moving, partly-commoditised field. But the direction is dead-on: the AI story of 2026 is quietly shifting from “how capable” to “how affordable,” and the companies making models cheaper to run are riding the more durable half of that trade.

Reporting on a venture funding announcement as covered on 28 July 2026. Performance and compression figures are company-reported and not independently verified. Not investment advice.