Capital

95 Percent Smaller Models Change Where AI Runs

Multiverse Computing is raising up to 500 million euros for software that shrinks AI models by up to 95 percent. The real consequence is where inference runs.

CapitalBy Servola Tech Desk2026-07-273 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

95 Percent Smaller Models Change Where AI Runs

Key takeaways

A Spanish company has announced a funding round of up to 500 million euros on the claim that large language models can be made 80 to 95 percent smaller without losing much accuracy, and if that claim holds on your workload, the cheapest inference is the kind you host yourself.
The round targets up to 570 million dollars, or 500 million euros, at a pre-money valuation of roughly 1.5 billion euros. That is a five-fold step up from the Series B of June 2025 and takes total funding raised to around 800 million dollars.
The customer list is European industry rather than Silicon Valley: Iberdrola, Bosch, Telefonica, Indra, Allianz and PwC, plus the Bank of Canada. Compression is being bought by operators who already own hardware, not by companies renting more of someone else's.
Before your next inference contract renews, take one real workload and measure whether a compressed model holds the accuracy you need. If it does, the hosting decision reopens, and the data residency question you closed when you signed reopens with it.

A funding round built on a claim about size

Enrique Lizaso runs an artificial intelligence company from Donostia, on the Basque coast, and on 27 July he put a number on a bet he has been making for years. Multiverse Computing announced a Series C targeting up to 570 million dollars, or 500 million euros, at a pre-money valuation of 1.7 billion dollars, roughly 1.5 billion euros. That is a five-fold step up from the Series B the company closed in June 2025, and it takes total funding raised to around 800 million dollars.

Why it matters: the money is not going into a bigger model. It is going into making models smaller. Lizaso stated the thesis plainly in the announcement, saying the industry has accepted a false constraint for years, that powerful models require expensive infrastructure. A round of this size is the first serious institutional test of whether that constraint is real.

The investor list reads like a map of European capital rather than a Silicon Valley syndicate. Forgepoint Capital International, the BNPP Solar Impulse Venture Fund and Bullhound Capital are leading, joined by Santander Alternative Investments, Tikehau Capital, HP Inc., Orange Ventures, Scania Invest, NAventures, Qatar Development Bank, Zouk Capital, SETT, the European Innovation Council Fund, the Basque government's Hazten Scale-Up Fund and Kutxa Fundazioa.

What tensor networks do to a model

CompactifAI, the company's product, borrows tensor networks from quantum physics and applies them to compression. Multiverse says the technique shrinks large language models by 80 to 95 percent with minimal loss of accuracy, and that the smaller models run faster and draw less power. In May the company said its CompactifAI interface was serving leading coding agents at up to 75 percent lower cost.

The deployment record is the part worth reading twice. Multiverse says its compressed models run on millions of devices, including drones, cameras, satellites, vehicles and telecommunications infrastructure. Those are not data centre workloads. They are places where a full-size model was never going to fit, and where sending the data somewhere else is either too slow or not permitted.

Yes, but: compression is lossy by definition, and minimal accuracy loss is the company's own characterisation measured on its own benchmarks. The claim that matters to a buyer is narrower than the headline. It is whether a compressed model holds accuracy on your task, not on a public leaderboard, and that is testable before you commit.

The part that lands in your budget

More than 100 organisations already buy this, and the names are industrial rather than experimental: Iberdrola, Bosch, Telefonica, Indra, Allianz, PwC and the Bank of Canada. Multiverse reports ten-fold annualised revenue growth since the Series B and first-quarter 2026 sales up 96 times year on year. Growth at that rate on that kind of customer base is a signal about procurement, not about research.

The consequence: for most European operators the binding constraint on deploying artificial intelligence has been access to compute they do not own. Compression attacks that constraint from the other side. A model small enough to run on your own hardware is a model whose data never leaves your network, which turns an architecture decision into a compliance decision you have already made.

The practical move is unglamorous. Before your next inference contract renews, take one real workload and measure whether a compressed model holds the accuracy you actually need. If it does, the hosting question reopens, and so does every question about jurisdiction that you settled the day you signed.

Frequently asked questions

Has Multiverse Computing actually raised the money?

The company announced a Series C targeting up to 570 million dollars, or 500 million euros. A target is not a close. Treat the figure as the size of the round being sought at a stated pre-money valuation of roughly 1.5 billion euros, and expect a final number when the round completes.

Is this a quantum computing investment?

No. The technique borrows tensor networks, a mathematical tool from quantum physics, but it runs on ordinary hardware. The product is model compression for classical computers, which is why the customer list is made up of industrial operators rather than research laboratories.

What should we test before buying compression?

Accuracy on your own task, measured on your own data, against the model you use today. Public benchmark results tell you the technique works in general. They do not tell you whether the specific capability your process depends on survived the compression, and that is the only result that governs the decision.

The interesting thing about this round is not the valuation. It is that half a billion euros of institutional money has been placed on the proposition that the expensive part of artificial intelligence was never necessary. Whether that proves right will be settled in procurement departments, one workload at a time, and the operators who run the test first will know the answer long before the market does.

Multiverse ComputingCompactifAImodel compressionSeries Cedge AIinference costs

A Leaked Call Paused a 480 Billion Yuan Round

DeepSeek suspended its second funding round on 25 July after its founder's remarks leaked. What he said matters more to a European buyer than the pause does.

3 min read

Capital

A Fourth DRAM Supplier Just Raised 8.6 Billion

CXMT listed in Shanghai on 27 July and rose 470 percent. The debut number is not the one that changes a European memory budget. The 57.92 billion yuan raise is.

5 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.