Granite Raises €4M to Ship Owned Agentic Cloud

Granite raised €4 million on 28 September 2026 to sell a cloud built for AI agents that a buyer can own, not only rent. Bifrost Studios led the round, with unnamed private investors. The firm is a Copenhagen studio company, not IBM’s Granite language models.

The money is for bringing software and hardware together and putting both in front of customers. Granite says more than €30 million already went into the underlying technology over eight years, so this cheque is a go-to-market budget on a stack that, in the company’s telling, already exists.

A €4 Million Cheque on Eight Years of Work

Most cloud rounds of this size fund software. Granite is selling a physical unit as well: a modular, containerised data centre that holds compute, storage and networking, and that can be placed where power is available. Dedicated systems stay under the customer’s control. The same stack is also offered as a cloud service.

Bifrost Studios is not a classic venture firm writing a first cheque into a slide deck. Its public line is that it does not invest in companies, it builds them. The studio lists Granite as sovereign cloud infrastructure for Europe, in its AI and infrastructure line, and says it has launched seven studios and 33 companies.

THE ROUND IN FIGURES

• The cheque: €4 million, announced 28 September 2026, led by Bifrost Studios with unnamed private investors.
• Prior spend: More than €30 million over eight years, according to the company, on the technology now being sold.
• The studio: Bifrost lists 7 studios launched and 33 companies built, with Granite in the 2026 scale cohort.
• The cost claim: Granite puts “up to 70% lower cost” on the site, next to 100% data sovereignty and 99.99% availability, all as company figures.

Peter Damm, Granite’s co-founder and chief executive, is a four-time founder with one exit. Bifrost says he spent seven years as a COO and CFO in London, then co-founded a venture fund that still holds investments. His job now is to turn a long engineering cycle into orders.

Agents That Do Work Need a Place to Sit

Damm’s pitch starts with a workload change, not a funding gap. Chatbots answer questions in bursts. Agents carry out work: they call tools, keep memory, and touch internal systems for hours. That load is less like a search query and more like a plant that stays on.

In August 2025, Gartner predicted that 40% of enterprise applications would include task-specific AI agents by the end of 2026, up from less than 5% in 2025. If even part of that lands, the infrastructure question stops being “which model API” and becomes “where does the agent live, what may it touch, and who holds the keys.”

Most agent stacks still default to a hyperscale account, because that is where the GPUs and the identity tools already are. Granite’s wager is that memory, permissions and tools have to sit close to the data the agent acts on, on a fabric a European buyer can point at. The container is how that wager is delivered. Power availability, not a metro campus brochure, decides the site.

Grand View Research put the global sovereign cloud market at $96.8 billion in 2024 and projected $648.9 billion by 2033, a 23.8% compound annual growth rate from 2025 to 2033. Granite is a small vendor in that field. It is also one of the few at this cheque size that is trying to ship its own box rather than rent someone else’s hall.

What Sits Inside Granite’s Container

Granite’s site splits the offer in two. GraniteCloud is the unit a customer deploys and owns. GraniteNet is the same compute, storage and agent runtime sold as a network service. Both run on Mycelium OS, which the company describes as a stateless layer that makes every node disposable and rebuildable, with an encrypted mesh and a built-in VPN.

A GraniteCloud host needs a prepared pad, a grid connection and cooling. The unit is a self-contained, liquid-cooled container. The company says it deploys in-region in Frankfurt, Paris or Stockholm in months rather than years, with final power, GPU and footprint figures set against the site. That last clause matters: there is no public standard rack rating on the homepage.

GRANITECLOUD VERSUS GRANITENET

Storage can be mounted under a compute slice or exposed as WebDAV or S3. Inference is pitched as a single endpoint that routes a prompt to the best model at the best price. The runtime is model-agnostic: bring open-weight or licensed models and swap them. The line on the site is “Not your key, not your agent.”

The agent layer is the piece that makes this more than a pretty container. Granite says it gives agents memory, permissions and tools so a workload is not tied to a hyperscaler’s location or software. Open source, end to end, is the other claim. None of those figures, including the 70% cost cut, come with a published benchmark on the public site.

Own the Box or Use the Network

The commercial design is a ramp, not a binary. A buyer can consume GraniteNet now and drop a GraniteCloud unit later. Deployed units, the company says, interconnect and become part of the same network. That is how a seven-figure round can talk about hardware without pretending it will finance a national GPU fleet.

If the customer pays for the box and the pad, Granite’s near-term cost is integration, first units, and sales. That is why the €4 million can be enough as a commercialisation cheque and still look tiny next to hyperscale capex. The hardware still has to leave a factory. The site says the platform is live and in deployment across Europe, with the first sovereign container shipping in 2026.

THREE WAYS ONTO THE STACK

• Cities and operators: Keep sensitive workloads on infrastructure that is physically and legally local, on a pad the buyer controls.
• Regulated firms: Finance, defence, healthcare and other critical operators running agents under NIS2, DORA and the EU AI Act.
• AI builders: Teams that want sovereign compute and an agent runtime on demand, without buying or running any hardware.

The middle path is the one that will decide whether this is a product or a pamphlet. A scaleup renting a node is a familiar cloud motion. A city or a hospital taking delivery of a liquid-cooled container is a different sale, with site works, grid talks and a compliance file attached.

Europe’s Rules Are Now a Product Spec

Granite’s “why now” list is legal as much as technical. It names four pressures: compliance deadlines, the US CLOUD Act, pricing that was not built for always-on agents, and physical risk to hyper-centralised plants. The first two are the ones a procurement team can put in a tender.

The NIS2 cybersecurity obligations had to be written into national law by 17 October 2024 and have applied since 18 October 2024. The EU AI Act rules now in force became applicable on 2 August 2026, when the Commission’s AI Office and national authorities started enforcement. GraniteNet ships with a “Compliance Pack” the company says is built into the platform, and it is careful to add that the pack complements, rather than replaces, a buyer’s own legal review.

On the CLOUD Act, Granite’s own wording is blunt: any US-incorporated provider is, in its view, a legal backdoor, wherever the server sits. That is the company’s reading, not a court finding. It is also the argument European buyers already use when they ask for operational sovereignty rather than a residency clause in a US contract.

AI is moving from answering questions to carrying out work. That changes what businesses need from the infrastructure behind it. We are building Granite so organisations have a genuine choice about where their AI runs, who controls the underlying technology and how they develop their AI capabilities over time.

Peter Damm, CEO of Granite

Damm made a second point in the same burst of remarks: Europe does not lack AI ambition, but the pipes behind that ambition matter as much as the models. The work, he said, is turning years of technology into infrastructure companies can actually use, with sovereignty and robustness built in from the ground up.

A Co-Founder With Eight Cloud Exits

The person who makes the eight-year, more-than-€30-million claim easier to believe is not only Damm. Kristof De Spiegeleer is listed as Granite’s co-founder. Bifrost describes him as a serial cloud-infrastructure founder with eight exits across 25 years, including sales to Symantec, Sun Microsystems and Western Digital.

De Spiegeleer is also co-founder and chief executive of ThreeFold, a peer-to-peer cloud whose public materials describe Zero OS on bare metal, a Mycelium overlay network, and quantum-safe storage. Granite’s homepage uses two of those labels, Mycelium OS and quantum-safe storage, for the layer under both GraniteCloud and GraniteNet. Granite has not published a line-by-line map from ThreeFold’s grid to the new containers. The overlap in names, people and claims is sitting in plain sight.

That is the second-order fact hiding under a seed-sized headline. A small commercialisation round can talk about owned European boxes because the operating layer has been in the world for years, under another brand, with another go-to-market. The 2026 job is packaging: a pad, a grid feed, an agent control plane, and a sales motion aimed at buyers who now have NIS2 and the AI Act on the desk.

Bifrost’s own portfolio shows how tightly the studio is clustering around that job. Beside Granite it lists Vector of One, an applied-AI shop, and Permyt, described as a permissions layer for AI agents. Permissions are one of the three things Granite says its agent layer must provide. The studio is building the surrounding parts, not only a container company.

The First Container Is Due in 2026

The round carries no stage label. There are no named customers on the public site. GPU counts, kilowatts and prices are confirmed per site, which is honest and also convenient. The test that will settle the story is physical: a liquid-cooled unit on a European pad, running an agent with memory and keys the buyer holds.

FROM ENGINEERING TO A LOADING DOCK

1. Eight years of development: Granite says the infrastructure technology behind the products was built over this span, with more than €30 million invested.
2. 17 October 2024: EU member states’ deadline to transpose NIS2; the rules have applied since 18 October 2024.
3. 2 August 2026: The EU AI Act’s main duties become applicable and enforcement starts.
4. 28 September 2026: Bifrost Studios leads the €4 million commercialisation round with unnamed private investors.
5. 2026: Granite says the first sovereign container is shipping, while GraniteNet is already offered as reserved capacity.

A hyperscaler will not notice €4 million. A hospital, a city operator or a defence contractor that wants an agent on a floor it controls might. The company has given itself the rest of 2026 to put the first box on a truck. Until that unit is plugged in, the agentic cloud is still a site, a pad specification, and a well-funded operating system looking for a customer who will take the keys.