Summary

• Multiverse Computing is targeting up to €500 million for technology designed to reduce model size and computing demand.
• The financing may remain open to additional investors, while performance figures remain company claims.
• Compression reinforces an established move towards efficient, local, and edge inference rather than creating a new trend.

A Spanish AI company has secured commitments for a funding round of up to €500 million, giving model compression a level of financial backing more commonly associated with foundation models and hyperscale infrastructure.

Multiverse Computing announced a Series C round of up to $570 million, equivalent to around €500 million, at a pre-money valuation of $1.7 billion. The company says the financing may remain open to selected strategic investors, so the headline ceiling should not be treated as a final closed amount.

The capital will support CompactifAI, a model compression technology that applies tensor network techniques to reduce the memory and computing resources required by large models. Multiverse claims that it can cut model size by between 80% and 95% with limited loss of accuracy, although results will vary according to the model, task, hardware, and acceptable performance threshold.

Forgepoint Capital International, the BNP Paribas Solar Impulse Venture Fund, and Bullhound Capital are co-leading the round. Other commitments named by the company include Santander Alternative Investments, Tikehau Capital, HP, 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.

Efficiency has become a commercial layer

Reducing the cost of AI inference is not a newly emerging field because quantisation, pruning, distillation, sparse architectures, and smaller task-specific models have been part of machine learning engineering for years. Generative AI has broadened the commercial demand for those techniques as companies encounter the cost, latency, energy use, and governance burden attached to larger systems.

Using the biggest available model for every task can produce weak economics when a smaller system performs a defined function adequately. Local inference may also reduce network latency, continue operating without a cloud connection, and keep some sensitive information within a device or site.

CompactifAI is intended for cloud, on-premises, and device deployment. Multiverse says its software can determine whether a task should be handled locally or sent to a remote system, creating a route towards hybrid installations across vehicles, industrial equipment, telecommunications networks, cameras, satellites, and other edge environments.

Such an architecture carries a particular European appeal because it can reduce dependence on constant hyperscale cloud access and give customers more control over data location. Local execution does not automatically provide sovereignty, however, when the original model, chips, operating software, or management layer may still come from outside Europe.

Headline reductions need workload evidence

Compression percentages cannot describe the trade-offs involved in a production deployment. A technique that works well for text classification may behave differently in medical imaging, industrial controls, code generation, multilingual support, or fraud detection.

General benchmark scores can also hide failures in rare or difficult cases that carry disproportionate commercial or safety consequences. Enterprise evaluation therefore needs to examine latency, energy use, memory footprint, throughput, error rates, performance drift, and the cost of recompressing a model after an upstream update.

Governance becomes more complicated when a compressed derivative sits downstream from a foundation model. Customers need to understand how the smaller version was produced, whether licensing permits the modification, which safety controls remain intact, and how vulnerabilities or material model changes will be communicated.

Multiverse names Allianz, Bosch, Iberdrola, Indra, PwC, Telefónica, and the Bank of Canada among customers or partners. It also claims rapid revenue and sales growth, although its financing announcement does not provide detailed audited accounts against which those figures can be assessed.

Large capital brings a wider execution test

A round approaching €500 million would give Multiverse resources beyond those of a conventional optimisation software vendor. The company plans to expand its model library, increase research spending, enter additional markets, and participate in the software layer supporting European sovereign AI and gigafactory programmes.

That breadth increases the operational challenge because compression tools must work across model families, chips, deployment environments, and regulated sectors. Strategic investors may provide distribution and customer access, but the company will still need repeatable evidence that savings achieved in trials survive production workloads.

Competition is also developing from several directions. Foundation model providers are releasing smaller versions of their own systems, chipmakers optimise software around their hardware, and open-source communities continue to improve quantisation and inference frameworks.

A specialist supplier must therefore provide enough additional efficiency, automation, assurance, or portability to justify another layer in the enterprise technology stack. Customers will be reluctant to add dependency unless the tool can remain useful when hardware, models, and operating requirements change.

The scale of Multiverse’s intended financing indicates that investors expect efficiency technology to capture a meaningful share of AI spending. Commercial proof will arrive when customers can show that compressed systems reduce cost and energy use while retaining the reliability, security, and governance required in production.