Multiverse Computing Is Raising $570 Million to Make AI Smaller. The Data Center Still Wants a Flag.

There is a particular kind of startup sentence that begins with “AI on a smartphone” and ends with “sovereign AI gigafactory infrastructure.” It is the linguistic equivalent of a bicycle attached to a battleship. Multiverse Computing has now raised $570 million to make both halves of that sentence real.

The San Sebastián, Spain-based company just announced the Series C at a $1.7 billion pre-money valuation, a fivefold step up from its Series B. The round is co-led by Forgepoint Capital International, BNPP Solar Impulse Venture Fund, and Bullhound Capital, with participation from Santander Alternative Investments, Tikehau Capital, HP, Orange Ventures, Scania Invest, Qatar Development Bank, the European Innovation Council Fund, Spain’s SETT, and other strategic and public investors. The company’s announcement says the financing is expected to bring total funding to $800 million and may remain open to selected strategic investors, because apparently $570 million is now the opening bid.

Multiverse builds software that compresses AI models so they require less compute, memory, and energy. Its flagship technology, CompactifAI, uses tensor networks—a mathematical framework with quantum-physics ancestry—to shrink large language models by as much as 80 to 95 percent with, the company says, immaterial loss of accuracy. The pitch is straightforward: powerful AI should not require every camera, factory, vehicle, and national data center to phone home to a hyperscaler.

This is not merely a funding story about “efficient AI,” a phrase currently doing the work of a whole business plan. Multiverse says annualized revenue has grown more than tenfold since its Series B in June 2025, sales were up 96 times year over year in the first quarter of 2026, and it has more than 100 customers. Reported customer examples include the Bank of Canada, Bosch, Telefónica, and Allianz, while the company says its models are being deployed across devices and systems including drones, cameras, satellites, vehicles, and telecom infrastructure.

The Model Has Been Put on a Diet

Multiverse is betting that the next advantage comes from subtraction. Compress the model, and it can run on hardware that is closer to the user: a smartphone, an AI PC, a factory floor, an autonomous vehicle, or a satellite that cannot casually send a terabyte of sensor data to California every time it sees a suspicious cloud.

The company’s CompactifAI Router is meant to decide whether a workload runs locally or goes to the cloud. That hybrid approach is strategically sensible. Not every task needs a data center, and not every device should be asked to perform an entire frontier model while also taking a phone call. The plumbing is the point: model compression, deployment, orchestration, and governance have to line up before “AI everywhere” becomes more than a keynote prop.

Sovereignty Is the New Cloud Preference

The round’s investor list is a small map of the current AI anxiety. There are venture investors, industrial companies, telecom players, sovereign and public funds, and a national development ecosystem that would very much like Europe to have more than excellent research papers and a polite relationship with American cloud bills.

“Sovereign AI” can mean a government keeping sensitive workloads inside its borders, a company controlling its data and deployment stack, or a procurement committee discovering that dependence on three hyperscalers is a strategic risk.

Multiverse’s product sits neatly inside that concern. Compressing models reduces the hardware required, while its planned software layer combines model compression, GPU orchestration, AI services, and sovereign-grade controls. This is where the pitch expands from “we make models smaller” into “we are the operating system for a national AI metabolism.”

That ambition is not random feature confetti. Enterprises really do need a way to decide which workloads belong on a device, in a private cluster, or in the public cloud. Governments really do care about power, supply chains, and control. And manufacturers would love to use AI without sending every operational secret through an infrastructure stack owned by somebody else.

It is also a crowded conversation. Etched’s inference machine is attacking the cost and speed of serving models with specialized hardware. Multiverse is attacking the same bill from the software side, by making the workload smaller before it reaches the hardware. The winner may not be the company with the best slogan about efficiency. It may be the one that can make a customer’s existing infrastructure less embarrassing.

The Round Is Big Enough to Have a Foreign Policy

A $570 million Series C is not a normal “hire a few engineers and improve onboarding” financing. Multiverse plans to expand its efficient-model library, continue algorithm research, invest in sovereign AI gigafactory infrastructure, and establish stronger commercial operations in East Asia, Southeast Asia, the Middle East, Canada, and the United States.

That is a capital-intensive plan hiding inside a software-company wrapper. Even if the core product is code, the market around it includes AI factories, hardware partners, certifications, integrations, and long enterprise procurement cycles. The customer may want a compressed model. Its legal department wants a deployment architecture, a security review, an indemnity clause, and three workshops about data residency.

There is another risk: compression is valuable only if it preserves what the customer cares about. “Up to 95 percent smaller” is a beautiful number, but production workloads contain awkward exceptions, changing hardware, and accuracy thresholds that live in contracts rather than demos. The demo is never the hard part. The hard part is proving the model survives a Tuesday.

Europe Gets to Build the Smaller Robot

Europe has spent years worrying that it is behind in frontier AI because it lacks the hyperscaler budgets, chip supply, and giant model labs of the United States and China. Multiverse is making a different argument: perhaps Europe can win by making powerful models practical where cost, energy, privacy, and sovereignty matter more than leaderboard theater.

That is close to the logic behind CuspAI’s materials-discovery foundry, where the startup is also trying to turn scientific infrastructure into a platform rather than a single clever demo. It is the same late-stage instinct: build the connective tissue around a technology, then raise enough money to make the connective tissue impossible to ignore.

The danger is that “infrastructure” can become a flattering word for “we intend to participate in every budget.” Multiverse wants to sell to device makers, data centers, governments, and enterprises across multiple regions while funding deep research and its own infrastructure layer. That is a lot of nouns for one company to carry. Sila’s battery-scale-up story is a useful reminder that strategic importance does not make manufacturing, deployment, or timelines behave politely.

Verdict: Serious Breakout, With a Gigafactory Habit

Multiverse Computing looks more like a serious breakout than a capital furnace with good branding. The technical problem is real, the customer categories are concrete, the reported growth is substantial, and the product aligns with several forces that are not going away: AI costs, energy limits, privacy requirements, device-side intelligence, and governments discovering that “just use the cloud” is not a sovereignty strategy.

But this is also a beautiful overreach in the most late-stage way. The company has raised enough money to turn model compression into a global operating doctrine, complete with gigafactories, sovereign controls, and a map of strategic regions. I mean that as both a joke and a compliment. If Multiverse can make AI smaller without making it less useful, it may become a quiet winner of the infrastructure cycle.

For now, the verdict is cautiously enthusiastic: a real technical wedge, a timely market, serious customers, and a round large enough to test whether efficiency can scale without becoming another form of excess. The models may fit on a phone. The business plan still needs its own continent.