Who Can Stop the Job? Northern Data Software GmbH and the Control Layer Beneath 22,000 GPUs

Summary

• Northern Data Software GmbH is a real German legal entity with a Norwegian foreign enterprise registration, but current public evidence supports treating it as a group services company—not as the owner of every GPU, data centre, cloud contract or software component marketed by Northern Data.
• The usable product is assembled across multiple control surfaces: Taiga Cloud contracts and provisions capacity, national subsidiaries operate the hardware, Ardent develops physical sites, Gcore provides significant orchestration software, and Quake AI is the new post-acquisition commercial banner.
• Northern Data has converted expensive GPUs into revenue unevenly: cloud utilisation and revenue from its dominant customer deteriorated in 2025 while customer concentration remained high and debt, electricity commitments and rapid hardware obsolescence continued to exert pressure; stronger utilisation reported in early 2026 remains a company metric, not an independent capacity audit.
• A buyer must procure the service by control boundary, requiring named counterparties, a service-specific architecture and responsibility matrix, measured availability, software partner continuity, data jurisdiction analysis and a tested exit plan rather than relying on group-level claims.

The Ticket That Crosses Five Boundaries

Imagine an AI training job stopped at 03:10. Eight H100s are allocated, the application can still reach object storage, but a GPU has disappeared from the node and the distributed job will not recover. The customer's urgent question is not 'Who owns the accelerator?' It is 'Who can restore my workload?'

The answer crosses several organisations. Taiga Cloud's public documentation tells the user to restart the affected instance and, if the problem persists, to submit a support request. A cloud operations team may need to inspect the virtual or bare-metal instance. The platform software may need to reschedule it. Someone with facility access may need to power-cycle or replace a physical component. A network or storage provider may own the actual failure domain. If the incident becomes a service credit dispute, the contracting company and precise service-level terms become decisive.

This sequence is the best way to understand Northern Data Software GmbH. It is tempting to collapse the name into a simple proposition: it must be the company that wrote the software controlling Northern Data's GPUs. The public record does not support that shortcut. The GmbH exists, and it has had a real operational role. Yet the current cloud, hardware and facilities stack is spread across several group companies and a significant external technology partner. Since Northern Data's acquisition by RUM Group, the customer-facing future is also narrated through Quake AI. None of these facts make the GmbH fictional or unimportant.

They make it a boundary company—an entity whose significance depends precisely on the decisions, people, contracts and systems that sit inside it at any given time.

This distinction is commercially material. A logo may suggest a vertically integrated provider while responsibility is split among a cloud vendor, an equipment-owning subsidiary, a data centre operator, a software supplier and a financed holding structure. The provider may well coordinate these pieces. But a buyer cannot infer coordination from a common brand. It must test it in contracts and operations.

The central thesis, therefore, is not that software is more important than hardware in the abstract. It is narrower and more practical: the party that controls admission, scheduling, identity, networking, storage, telemetry, incident response and workload recovery controls the customer's usable capacity. A GPU owner controls an asset. A data centre operator controls a physical environment. A lender can constrain asset sales and new debt. The customer experiences all of them through a control plane that may not belong to any single one.

One GmbH, Two Norwegian Registrations, No Dual Domicile

The identity bridge starts with unusually helpful public registry evidence. The Brønnøysund Register Centre in Norway listsNORTHERN DATA SOFTWARE GMBH, organisation number 928 496 384, as a foreign entity. The record gives its address as An der Welle 3, 60322 Frankfurt am Main; its original register as HRB 119359 at the Frankfurt local court; its legal form as a German GmbH; and Germany as the jurisdiction governing the enterprise. It also references northerndata.de. This is not a Norwegian company sharing a name.

A second official entry isNORTHERN DATA SOFTWARE GMBH NUF, organisation number 928 338 010. A NUF is a foreign enterprise registered in Norway, not a change of nationality of the foreign enterprise. Its registered activity is computer programming, with a description covering development, production, licensing and distribution of software products and IT services. It is registered for Norwegian value-added tax. Both entries point back to Frankfurt.

These records prove a bridge between the exact German company and a Norwegian registration. They do not prove that the GmbH owns Northern Data's Norwegian data centres, employs the engineers who operate them, signs every Norwegian electricity contract, or sells all Taiga services. Nor do they turn "Germany / Norway" into a claim of two headquarters. Germany is the legal home; Norway is a foreign registration and operating jurisdiction connection.

The distinction becomes sharper in Northern Data's own reports. The group's2021 annual reportdescribed Northern Data Software GmbH as broadly responsible for operational tasks, including software development and the sale of computing power. The same report separately described Northern Data NOR AS, based in Notodden, as a company providing hosting services to affiliates. At that time, a broad description of "software and computing power operations" for the GmbH had first-party support, while Norway's hardware hosting role was already allocated elsewhere.

The language has changed. The group's2025 annual reportsays Northern Data Software GmbH provides "a range of central and corporate services" to affiliates. It identifies Taiga Cloud Limited in Ireland as the company selling cloud services to external customers and executing the group's cloud strategy. It places Norwegian site operations and management under Northern Data NOR AS. It attributes technology and infrastructure support for Taiga entities to another German subsidiary, ND CS (Services) GmbH. Ardent companies own or operate facilities in multiple jurisdictions.

This is not simply a cosmetic reorganisation on this analysis. It changes what can safely be attributed to the directory entity. The strongest current public statement is that Northern Data Software GmbH is a wholly owned and consolidated group services company. The evidence does not disclose a standalone product catalogue, a direct customer list, headcount, code repository, cloud contract, intellectual property schedule or operational service-level agreement for the GmbH. It may employ people who affect the platform or hold rights that are not publicly listed. The absence of disclosure is not proof it does not.

It is a limit on what an outsider can assert.

The same report lists the GmbH with negative equity of approximately €203.8 million and a 2025 loss of approximately €2.9 million, while noting the use of a German law exemption to publish separate annual financial statements. These figures are group disclosure about a subsidiary, not a standalone economic account. They nevertheless refute the idea that the directory line can be treated as a clean, standalone software vendor with publicly observable unit economics.

An honest boundary therefore has three parts. The legal subject is Northern Data Software GmbH. Its verified contemporary role is central and corporate services within the group, with an earlier disclosed role in software development and computing sales. The operating system around it belongs to Northern Data Group and now RUM Group, with specific functions assigned to Taiga, national hardware operators, Ardent, Gcore and Quake. The analysis can follow that infrastructure because it explains the GmbH's context; it must not silently rename the entire system as the GmbH.

What "Vertical Integration" Omits

Northern Data presents itself as covering the chain from data centres to a software layer. Itscompany descriptionnames hardware and infrastructure relationships with NVIDIA, AMD, HPE, Gigabyte and VAST Data and describes a vertically integrated model. The strategic logic is credible: securing land, power, cooling, accelerators, networking and cloud software can eliminate margins and reduce dependence on a single external host. But vertical integration is a spectrum, not a binary condition.

At the physical level, control means title or lease, power rights, grid connection, transformers, backup generation, cooling, fire suppression, physical security, operator access and technicians. At the equipment level, it means ownership or financing of GPUs, servers, switches and storage. At the infrastructure software level, it means firmware, drivers, hypervisors or container runtimes, provisioning, schedulers, tenancy, identity, metering, billing and observability. At the customer level, it means portal, API, support, contract and the ability to migrate or delete data.

Northern Data does not publicly claim that a single legal entity holds all these rights, and its documents show it does not perform every function without third parties.

This matters because integration can fail at the joints. A facility may be technically ready while GPUs are delayed. GPUs may be installed while the cloud stack is being migrated. A cluster may be schedulable but not allocated to a paying customer. A contract may reserve capacity while a customer's deployment ramps slowly. A cloud interface may show a resource even if a physical fault requires local intervention. "Capacity" can mean authorised megawatts, powered megawatts, installed GPUs, available GPUs, allocated GPUs, or GPUs actually producing recognised revenue. They are not interchangeable.

The most important joint in 2025 was the software-defined operation. Northern Data reported that it upgraded its Taiga fleet during the second and early third quarters to support on-demand utilisation via technology provided by Gcore. The migration temporarily reduced availability and utilisation. In October, the company said more than 15,000 GPUs were allocated and fleet utilisation exceeded 60 per cent. The annual report says this helped the company achieve a monthly cloud revenue run rate of approximately €10 million in the fourth quarter.

These are management metrics, not an independent audit of scheduler telemetry, and the reported transition itself demonstrates the point: installing accelerators did not make them immediately saleable under the desired business model.

"Software" in this context is not a decorative layer on top of infrastructure. It determines whether the same H100 can be sold for a long reservation, an on-demand VM, a bare-metal node, a managed Kubernetes worker, or a scheduled HPC job. It decides which tenant can see which network, which image boots, which metering event becomes an invoice, how faults are isolated, and whether a customer's checkpoint can be restored elsewhere. A control plane change can improve addressable market and utilisation. It can also strand capacity during migration, add vendor dependency, or complicate an exit.

How a Customer Reaches a GPU

Taiga's documentation provides a more concrete architecture than the group's investor language. A user starts with an account and tenant, authenticates via Okta, uploads an SSH key, selects a cluster, creates a network, and deploys a machine via a portal or API. Thefeatures guidedescribes concepts including tenants, clusters, flavours, images, network credentials and machine UUIDs. It says a typical H100 configuration exposes eight GPUs and uses NVLink within a node, InfiniBand between nodes, and separate north-south connectivity. Thegetting started guideoffers virtual machines and bare-metal resources and directs customers to a status page and support.

This reveals several control layers.

First, identity is centralised. Taiga'suser management documentationsays new users and deletions go through support rather than a self-service administrator. It also warns that deleting a user does not stop running resources in the tenant. This design can protect workloads from accidental deletion, but it means identity termination and workload termination are separate operational acts. A buyer must ask who can suspend a compromised identity, who can stop its instances, how fast each action propagates, and what logs prove it.

Second, provisioning is cluster-specific. Images, machine types, networks and available capacity can differ by cluster. A nominally portable VM may still depend on a particular GPU model, driver, CUDA version, network fabric, storage attachment, and orchestration configuration. The API reduces manual work; it does not by itself make a workload portable across providers or even Taiga locations.

Third, support is part of the control plane. Taiga's publichelp policystates standard support hours of 09:00 to 18:00 Central European Time on weekdays, excluding public holidays, and provides an escalation path if a case remains unresolved for two weeks. This may be adequate for development or non-critical offline workloads. It is not automatically equivalent to 24/7 incident response for a production inference service. Paid support or negotiated enterprise terms may offer more, but their content and pricing are not publicly specified. Buyers must obtain them rather than assume them.

Fourth, physical intervention remains unavoidable. Taiga's documentation acknowledges that a GPU can "fall off the bus", describes the error as potentially random, and advises a restart, a ticket, and if necessary, on-site investigation. Its portal guide says status incidents or planned maintenance can make services unavailable. This candour is useful: no orchestration layer abolishes hardware failure. The operational question is whether monitoring detects the fault, a scheduler drains the node, capacity exists elsewhere, state is recoverable, and the support chain can act within the customer's recovery objective.

Finally, geography changes the chain. Taiga'slocations guidelists Boden in Sweden, London in the UK, and Norwegian capacity at Lefdal and Kristiansand. A Norway selection can therefore affect latency, energy mix, applicable contracts, physical operator, and data transfer design. This does not mean the customer has contracted with Northern Data Software GmbH NUF or that the GmbH owns the Norwegian hardware. Location is an attribute of the service; legal responsibility must be traced independently.

Gcore Is Not a Footnote

Northern Data's transition from predominantly reserved bare-metal capacity to more flexible cloud consumption depends materially on Gcore. In itsQ3 2025 operational update, Northern Data described a software stack managed by Gcore and managed Kubernetes services and presented the migration as the path to on-demand commercialisation. AGcore product announcementdescribes its own stack as providing GPU virtual machines, bare metal, spot instances and managed Kubernetes. These overlap with the features that turn Northern Data's hardware into cloud products.

The relationship is deeper than a conventional software subscription. Northern Data's 2025 annual report says it held a 2.13 per cent stake in Gcore and had provided a €10 million interest-free loan, with governance arrangements supporting the accounting conclusion that Northern Data had significant influence. This makes the dependency partly strategic and financial. It may align incentives, but it also complicates a simple "internal platform" story.

The important unknowns are contractual. Public documents do not disclose which party owns customer tenancy data, metering logic, custom integrations or operational telemetry; the duration and termination rights of the Gcore arrangement; source or configuration escrow; change control authority; disaster recovery obligations; or the time and cost to replace the software stack. They also do not specify which employees of Northern Data Software GmbH, ND CS, Taiga or Gcore have privileged access to each layer.

A procurement team should therefore map at least four failure modes. Gcore itself could suffer an outage or security incident. The business relationship could end. A control plane upgrade could temporarily remove capacity, as the 2025 migration did. Or post-acquisition integration could move workloads to Quake while existing Taiga services remain in place. For each case, the buyer needs a sustained operational state, not a strategic promise.

This is where the precise entity boundary proves its value. Calling the entire platform "Northern Data Software" hides the external dependency. Calling Gcore the platform owner goes too far in the other direction. The evidence supports a composite stack: physical infrastructure owned and financed by the group, Taiga's commercial and service layer, Gcore technology, local operating subsidiaries, and group central services. Control is shared and must be assigned function by function.

A Price Is a Bet on Utilisation

Taiga's public pricing makes the economic proposition readable while leaving significant costs open. On 17 July 2026, thepricing pagedisplayed eight-GPU H100 bare-metal nodes from $2.50 per GPU-hour and H200 nodes from $2.95, with on-demand H100 and H200 offerings from $2.70 and $3.10 respectively. The page also listed CPU instances and named payable or configurable components including storage, public IP addresses, managed Kubernetes, container registry, platform subscriptions, Direct Connect, and several support tiers. Many of these items required a quote. The page is dynamic; it is evidence of an observed offer, not a permanent rate or total cost guarantee.

For a customer, the headline GPU-hour is only the numerator in a more complicated calculation. The useful cost includes reserved but idle hours, failed jobs, checkpoint storage, data transfer, premium support, networking, orchestration, engineering time, and the performance difference between configurations. Eight poorly connected H100s may be worth less than fewer accelerators with proper networking. A cheaper spot instance may be unsuitable if preemption destroys an expensive run. A long reservation may make sense for stable training demand and become costly when model architecture or chip generation changes.

For the provider, the same price is a bet that utilisation will cover power, facility expense, support, software, financing and obsolescence before the hardware loses competitiveness. Northern Data's 2025 results show how unforgiving this equation can be. Revenue from continuing operations fell to €80.0 million from €121.1 million in 2024, while adjusted EBITDA went from a profit of €32.5 million to a loss of €37.9 million. The net loss was €482.7 million, though that number includes major impairments and other items and should not be confused with cash burn alone. Year-end cash was €57.6 million.

The segment picture is more revealing. Taiga generated approximately €79.6 million in external revenue in 2025 and an EBITDA loss of approximately €70.8 million. Ardent reported only approximately €0.4 million in external revenue and a small EBITDA loss, as its business still largely supported Taiga rather than selling mature colocation capacity to external customers. Northern Data attributed the cloud shortfall in part to the software migration, delayed commercial ramp, and evolving market pricing. It withdrew its 2025 guidance in October after initially anticipating revenue of €240–320 million and adjusted EBITDA of €80–130 million.

These are group and segment numbers. They are not Northern Data Software GmbH's revenue, and they do not reveal the GmbH's transfer pricing arrangement. They nevertheless explain the environment in which a central services subsidiary operates. When GPU utilisation disappoints, pressure runs through the entire chain: pricing becomes more aggressive, support and software investment compete with liquidity, expansion plans may be phased, and the assumed useful life for hardware becomes harder to defend.

Northern Data reported a much stronger position in early 2026. ItsQ1 updatesaid fleet utilisation had reached approximately 85 per cent in March and Q1 revenue was €43 million, of which Taiga contributed €42.4 million. The company defined utilisation as revenue-generating GPUs divided by total GPUs. This definition is better than an undefined marketing percentage, but it leaves questions: whether every revenue-generating GPU was busy for the entire period; whether utilisation was committed, spot, or intercompany; what net price applied; and how the metric treats maintenance or unavailable capacity. A buyer or investor should consider the figure as a management operational indicator, not an independent workload trace.

The Customer That Moves the Entire Fleet

Utilisation becomes more valuable when diversified. Northern Data was not diversified in 2024. Its 2025 annual report says one Taiga customer accounted for €108.8 million in revenue the prior year, approximately nine-tenths of the group's continuing revenue. In 2025, revenue attributed to the same unnamed customer fell to €38.2 million—still almost half of the group's €80.0 million total. The decline was large enough to shift group economics even as Northern Data tried to broaden the customer base.

The customer's identity should not be guessed. Northern Data and counterparties have announced large relationships, including aCore42 framework for up to 10,000 GPUs, but a public framework announcement is not proof that Core42 is the customer in the concentration note or that the full framework became recognised revenue. The transaction documents for Rumble also describe a separate Tether agreement under which Tether may buy up to $75 million of GPU services in each of two years. This does not identify the historical customer either.

Concentration is a control risk because a large buyer can influence the operating model without owning the infrastructure. It can reserve entire clusters, shape location and security requirements, negotiate tailored pricing, alter deployment timelines, and leave a substantial hole when consumption drops. Conversely, moving from a single high-reservation customer to multiple on-demand tenants increases sales reach but also increases demands on identity, billing, isolation, support and scheduler quality. The software transition was therefore not only a product improvement; it was an attempt to change who could buy the fleet.

Allocated and reported utilisation must also be separated. An allocated GPU may be contractually assigned but idle. A GPU generating any revenue qualifies under Northern Data's disclosed utilisation definition even if its duty cycle is low. A fully occupied GPU may still generate thin returns if its price is heavily discounted or its power and financing costs are high. The most useful dashboard for decision-making would reconcile installed, available, allocated, billed and physically used GPU-hours by model and location, then link them to net revenue and service credits. Northern Data does not publish this reconciliation.

Capital Has Its Own Control Plane

The other party able to influence the fleet without scheduling a single job is the capital provider. Northern Data built and bought infrastructure faster than its 2025 revenue could comfortably finance. At the end of that year, the group reported €718.9 million in non-current liabilities, including a €614.9 million shareholder loan balance, against €57.6 million in cash. The original Tether facility bore interest at EURIBOR plus three percentage points and had been fully drawn.

Northern Data said it had breached financial covenants during 2025, obtained waivers, and relied on financial and operational measures to conclude that no material going concern uncertainty existed. It also disclosed that a demand for immediate repayment could not have been satisfied without new financing or an asset sale. The auditor issued an unqualified opinion; this is not a guarantee that future cash flow forecasts will be met.

The cost of financing is tangible. Northern Data reported €24.7 million in interest paid in 2025, while other interest was capitalised or accrued. Interest consumes cash whether GPUs are busy or not. A fleet that is economically attractive at high utilisation can become fragile when deployment is delayed, prices drop, or a major customer reduces usage. Hardware also ages while debt remains: a new accelerator generation can compress rental prices for the previous one before original financing is amortised.

The post-acquisition structure makes capital control more explicit. Rumble'sSEC transaction prospectusin April 2026 described the €575 million principal facility as fully drawn, interest capitalised, covenant history, and planned treatment of the Tether receivable. After completion, RUM Group's18 June current reportsaid half of the relevant receivable had been transferred in exchange for a pre-funded warrant and the other half became a five-year secured facility of €317.5 million due from a new Irish holding company. This facility retains the EURIBOR plus three percentage point rate. Subject to legal and agreed security principles, Northern Data shares, holding company interests and borrower group assets are to support the financing; covenants restrict matters including additional subsidiary debt, liens and disposal of substantially all assets.

These provisions do not put Tether in the cloud console. They can nevertheless affect strategic freedom. A highly financed group may need lender consent or covenant headroom for a disposal, refinancing or reorganisation that a customer assumes is operationally straightforward. The same counterparty is also a significant shareholder and potential compute customer. Interests may align—funding expansion and providing demand—but concentration between capital, ownership and sales increases the importance of conflict governance and transparent related-party terms.

The exact effect on Northern Data Software GmbH cannot be inferred from group documents alone. The public transaction documents do not provide a simple schedule indicating which specific assets or rights of the GmbH were pledged, what warranties are in force, or what internal service contracts might move. A customer does not need privileged financing documents to ask a reasonable question: if a covenant event, asset sale or restructuring occurs, which legal company is supposed to keep operating the portal, licences, support and workload data, and what contractual rights preserve continuity?

A Ranked Machine, An Unfinished Domain

Northern Data's installed compute capacity is not purely promotional. The independentTOP500 entry for Njoerdidentifies an HPE Cray XD670 system using NVIDIA H100 SXM5 accelerators, NDR400 networking and Ubuntu, with 273,280 cores and a measured Rmax performance of 78.2 petaflops. It ranked 26th in June 2025 and 31st in November. The result confirms the group could assemble and operate a technically substantial high-performance system. It does not measure cloud availability, customer job throughput, security, support or profitability.

A benchmark also does not convert a development pipeline into ready capacity. RUM Group'sclosing announcementsaid the acquired company had approximately 22,000 H100 and H200 GPUs and over 200 megawatts of "unmonetised" data centre capacity, with approximately 250 megawatts current and planned across ten sites by 2027. It said four sites were owned. These are post-transaction company assertions combining present assets and plans. The word unmonetised is significant: electrical potential is not the same as powered halls, installed servers or contracted cloud revenue.

The Ardent Pittsburgh site illustrates the steps between property and product. Itsfacility pagecurrently advertises 2.4 megawatts available, a path to 12 megawatts by 2027, rack densities above 135 kilowatts, liquid cooling capacity, redundant power and cooling, carrier-neutral connectivity, and a power usage effectiveness (PUE) target of 1.15. Each item is relevant to AI hosting. Several are plans or design claims, not independently verified operating statistics. Earlier ambition can also change as engineering, capital and demand mature; procurement must use current energisation timeline, utility evidence and acceptance tests rather than an older headline.

Power quality matters as much as quantity. Northern Data's 2025 sustainability report says the group consumed approximately 1,036 gigawatt-hours of energy. On a location-based measure, 48 per cent was renewable, 41 per cent fossil, 8 per cent nuclear and 3 per cent unknown. The same report cites a PUE of 1.15 at Boden. The group-level mix complicates any overall claim that the fleet is run on clean energy, while a single facility PUE cannot be applied to every site. The environmental section was not in the financial auditor's assurance scope.

For customers, the practical distinction is between capacity attributes and workload guarantees. A site may offer favourable climate, low-carbon electricity or high-density cooling and still lack a second network path, a particular storage class, on-call technicians, or spare nodes for failover. Conversely, a third-party facility may offer mature resilience while reducing Northern Data's control over maintenance and recourse. The SEC prospectus says the company uses owned and third-party facilities and may have limited recourse under third-party colocation agreements.

This is why facility owner, hardware operator, cloud vendor and software controller must be named for each selected cluster.

Ninety-Five Is Not 99.99

Reliability claims become dangerous when numbers from different services are placed side by side without their denominators. The SEC prospectus says nearly all of Northern Data's customer agreements require at least 95 per cent availability. Five per cent downtime in a 30-day month is approximately 36 hours. The document does not say customers regularly experience this amount, and an individual agreement may be stronger. It shows that the contractual floor described to investors may be well below the four-nines language now displayed on theQuake AI site.

Quake's 99.99 per cent statement is a current marketing claim for its platform; it is not, on available evidence, a historical availability metric for the entire Taiga GPU fleet or a term automatically inserted into every existing contract. Even a four-nines commitment needs a precise service definition. Does availability measure the portal, API, a VM, an entire cluster, storage, or network reachability? Does it exclude planned maintenance, customer configuration, upstream carrier incidents, capacity exhaustion, or early feature rollouts? Is the remedy a credit, a termination right, or an actual recovery obligation?

An availability percentage without these clauses cannot assess the risk of a failed training run.

Northern Data itself paints a more grounded risk picture in the prospectus. It reports that services have experienced interruptions and lists possible causes, including hardware and software defects, power or mechanical system failures, limited public evidence redundancy, and third-party site failures. It also warns of cyberattacks, supply chain constraints and dependence on a limited number of equipment suppliers. The disclosure does not provide a public incident timeline, affected customers, duration or service credits. This gap prevents a reader from independently calculating fleet availability.

Cybersecurity evidence is also limited. The 2025 annual report says the group operates an information security management programme aligned or certified with ISO 27001, conducts internal and external audits, and was not aware of any data breaches in 2024 or 2025. These are company statements. The public documents reviewed here do not expose a certification scope that maps every Taiga region, Ardent facility, Gcore component and Quake service to the same audited boundary. "No known breach" is also different from "no failure", "no vulnerability" or "no customer-impacting incident".

A buyer should request the current certificate and statement of applicability; penetration test and remediation summaries; subcontractor and privileged access lists; vulnerability and patch turnaround times; tenant isolation design; key management options; log retention; incident notification timeline; and the trailing twelve months of service-specific availability. For hardware faults, it should test detection, node quarantining, replacement stock and job recovery.

For control plane faults, it should ask whether existing instances continue running if the portal or scheduler fails, and how administrators regain access without weakening tenant separation.

The distinction between prevention and recoverability is particularly important for AI/HPC. A 48-hour training run can fail near the end even if monthly availability appears high. Checkpointing can limit loss, but only if the application writes it frequently, storage remains consistent, and recovery has been tested on another node. The provider may keep infrastructure available while the customer's distributed framework crashes. Contract language must therefore separate infrastructure availability, job-level responsibility and optional managed services rather than claiming one percentage covers all three.

European Locations Under a US Parent

Data locality is one of Taiga's clearest selling points. Its locations guide says data stays within the selected European region and characterises the service as outside the reach of the US CLOUD Act. The first claim can be tested via architecture, contracts and logs. The second is a legal conclusion whose assumptions may have changed.

On 17 June 2026, Rumble completed its acquisition of Northern Data, and the parent company was renamed RUM Group. The current report establishes that a US-listed group now controls Northern Data via an Irish holding structure. The acquisition does not automatically prove that US authorities can obtain all European customer data, just as storing data in Europe does not automatically eliminate all cross-border legal exposure. Applicability may depend on possession, control or custody, corporate relationships, service design, encryption key custody and conflicts with European law.

The appropriate response is neither to repeat the old sovereignty slogan nor to declare it false without legal analysis. A regulated buyer must obtain a jurisdiction memorandum that is dated and service-specific after the acquisition. It must identify the contracting entity, data and metadata locations, administrators capable of remote access, parent company rights, subcontractors, government request process, encryption ownership and procedure for contesting conflicting demands. It must also cover telemetry, support attachments, identity data and billing records, not just model weights in object storage.

Norway is relevant but not magical. Lefdal and Kristiansand may provide physical locality and, depending on the actual electricity contract, energy advantages. The NUF registration creates a real Norwegian compliance footprint for the GmbH. Yet the hardware operator disclosed by Northern Data is Northern Data NOR AS, and some capacity is hosted with other infrastructure providers. A "Norway region" must therefore be converted into a named responsibility map: who owns the servers, who controls the cages, who provides power and networking, who administers the platform, and which company owes the customer a remedy.

Quake Is a Destination, Not Yet a Map

RUM Group says Quake AI combines Northern Data's acquired infrastructure with Rumble's cloud business. Quake'shomepagealso says Rumble Cloud has become Quake AI, while existing Rumble Cloud accounts, logins and services continue unchanged. It advertises single H100 or H200 GPUs as well as clusters. This is evidence of a new umbrella and continuity for Rumble Cloud users. It is not evidence that every Taiga customer, contract, tenant, region or support process has already migrated.

Product surfaces remain visibly different. Taiga publishes GPU node and on-demand pricing along with infrastructure add-ons. Quake'spricing pageprominently lists fixed monthly CPU plans and shared cloud plans and does not expose directly comparable public GPU pricing for every Taiga configuration. Different catalogues may coexist during integration. They may also reflect different backends, customer segments or billing systems. Public pages do not yet explain a single entitlement model between Taiga and Quake.

Integration could create real benefits. Rumble brings an existing cloud customer interface, payments and adjacent storage or delivery capabilities. Northern Data brings GPU capacity, facilities and operational experience. Gcore brings a mature orchestration layer. If RUM Group unifies identity, inventory, billing and support without destabilising workloads, it can market capacity more effectively on a broader base. If it simply places a new brand on separate systems, customers will still encounter fragmented consoles and responsibility.

The post-acquisition question is who owns the target control plane. Will Taiga's Okta tenants migrate to Quake identity? Will Gcore remain the scheduler and Kubernetes layer? Will customers receive new contracts from a RUM Group company? Who will maintain Taiga's APIs, images and network definitions? Will model and data transfers be optional, and can a customer remain in a particular European region? What happens to support histories and security evidence? Public documents do not yet fully answer these questions.

Northern Data Software GmbH could be important in this transition if its central services mandate includes integration, group systems, licensing or operational governance. It could also be peripheral to the customer-facing migration. Current evidence does not permit a stronger allocation. Anyone contracting on the assumption that the GmbH is "Quake's software company" should ask for the document that says so.

The Substitute May Be a Different Operating Model

The competitive set is broader than specialised GPU clouds. Rumble's SEC prospectus names CoreWeave, Applied Digital, Nebius, NScale and Lambda and points to hyperscale clouds. It also recognises the customer option to build its own infrastructure. These substitutes sell different allocations of control.

A hyperscaler offers deep managed services, mature identity and global compliance programmes, but may charge more, restrict specialised configurations, or expose the buyer to a large proprietary ecosystem. A GPU specialist may offer attractive accelerator economics and faster access but have a shorter operational track record or thinner support catalogue. A colocation or bare-metal arrangement gives the customer more control over the stack and more responsibility for it. An owned cluster maximises direct operational authority but requires scarce engineering, power procurement, spares, security and capital.

A broker may aggregate capacity while adding another contractual layer.

Taiga's strongest proposition is not simply "European GPUs". It is the potential combination of European locations, dense H100/H200 infrastructure, bare-metal and on-demand choice, high-performance networking, and a specialised operations team. Its weaknesses are the inverse of that specificity: concentrated demand, financial pressure, a significant software partner, a changing parent company, and less public operational evidence than larger clouds.

A useful competitive test uses the customer's actual workload. It fixes the model, dataset, precision, framework, node count, checkpoint interval and target region; runs an agreed benchmark; measures training time or inference throughput; records failed job frequency and operator intervention; then totals the full bill. The same test should be repeated after a maintenance event and with a node removed. A provider with the lowest headline GPU-hour may lose once idle time, recovery and engineering effort are included.

The test should also compare control, not just speed. Can the customer bring its own image and encryption keys? Can it inspect network and access logs? Can it force identity lifecycle without a support ticket? Can it reserve a failure-isolated cluster? Can it export snapshots in a documented format? Can it reproduce the environment elsewhere? These questions reveal whether the provider supplies a convenience accelerator or an environment the customer cannot easily leave.

Leaving Starts Before Signing

AI infrastructure creates switching costs through ordinary engineering decisions. A workload accumulates machine images, drivers, container registries, Kubernetes manifests, Slurm scripts, network policies, object storage paths, access roles, monitoring rules, checkpoints and support knowledge. Nothing needs to be deliberately proprietary to become sticky. Large datasets make transfer slow and potentially expensive. Reserved capacity commitments may overlap ramp at a new provider. A regulated user may need to requalify the new environment.

Taiga's API and standard technologies may help, but standards do not erase implementation details. CUDA code may run on another NVIDIA environment while performance changes with interconnect. An S3-compatible interface may ease application migration while authentication, lifecycle policy or edge behaviour differ. Kubernetes entities may be portable while storage classes, GPU operators, load balancers and identity integrations are not. A bare-metal image may not boot on different firmware and network design.

There is also enterprise transition risk. If Taiga services are folded into Quake, a customer could face new terms, credentials, endpoints or support processes even without choosing to change provider. A provider-controlled migration must have the same disciplines as a customer-led exit: inventory, compatibility testing, rollback, data validation, security review and agreed downtime.

The contract should establish exit assistance before dependency grows. Minimum terms include data and image export formats; network and identity configuration export; exit pricing schedule; continued access during transition; backup deletion and retention rules; deletion certificate; post-termination support; survival of audit and incident obligations; treatment of prepaid reservations; and rights if a significant subcontractor or control plane provider changes. Source code escrow may be irrelevant for a standard cloud service, but configuration continuity and a documented fallback operating mode are not.

The exit test should be empirical. A customer can provision a small equivalent environment elsewhere, restore a checkpoint, redirect a non-critical job and compare results. It can repeat user revocation, log export and tenant deletion. The test will discover hidden state well before a covenant event, outage or commercial dispute forces the issue.

The Procurement Test for This Particular Chain

Northern Data's platform should not be rejected because its legal and technical structure is complex. Most serious clouds rely on subsidiaries, equipment vendors, data centre partners and software components. It should be purchased with a diligence set designed for its actual chain.

Identity and authority.The proposal must name the contracting vendor, invoice issuer, data subcontractor, hardware owner, facility operator, first-line support provider and software operator for each selected region. It must state the exact duties of Northern Data Software GmbH, if any, rather than rely on its name. Enterprise warranties and assignment provisions should cover the entities whose performance matters. A current group chart should distinguish Taiga Cloud Limited, Northern Data NOR AS, the relevant Ardent company, ND CS, Gcore and RUM Group.

Service architecture.The buyer needs a diagram showing identity, portal and API, orchestration, hypervisor or bare-metal provisioning, Kubernetes or Slurm, network fabrics, storage, metering, logging and support. Each component should have an owner, hosting location, privileged access group, recovery design and planned post-Quake state. The diagram should identify what continues working when the Gcore-managed control plane, customer portal or external networking fails.

Capacity proof.The provider should separate authorised, contracted, powered and occupied megawatts; installed, available, allocated and billed GPUs; owned and third-party facilities; and current capacity from 2027 plans. For a reservation, it should identify serialised or otherwise reserved equipment, acceptance criteria, maintenance spares and replacement rules. A TOP500 result is useful engineering evidence, but the buyer still needs its own workload benchmark.

Commercial reconciliation.A quote should include GPU-hours, reservation minimums, storage, snapshots, data transfer, public IPs, Direct Connect, orchestration, platform subscriptions, premium support and professional services. It should specify whether a stopped instance continues to bill, how failed hardware is treated, what happens during planned maintenance, and how spot interruption works. Renewal pricing, hardware generation substitution and early termination merit explicit formulas.

Reliability and support.The applicable service level must identify the measured component, calculation period, exclusions, maintenance notice, credit and termination threshold. Support hours and severity response times should match the workload. The buyer should see recent incident and availability evidence for the selected region, plus escalation through Taiga, Gcore, hardware operator and facility. A public two-week escalation path is not an incident response target for a critical training or inference service.

Security and jurisdiction.The diligence package should map the ISO 27001 evidence scope, tenant isolation, administrator access, encryption keys, vulnerability management, logging, incident notice and subcontractors. The jurisdiction opinion should be updated after the RUM Group acquisition and cover control as well as storage location. Norway or another European region should be verified through the actual service chain, not inferred from the GmbH's NUF registration.

Continuity and change.The contract should address loss or replacement of Gcore technology, changes in contracting entity, migration into Quake, sale of a data centre asset, insolvency or enforcement, and migration between locations. Material changes in identity, orchestration, data access or service levels should require notice and, when risk changes substantially, an exit right. The buyer should receive a current roadmap but rely on binding continuity terms.

Exit.Before production, the customer should perform a small restore elsewhere and document every dependency. The agreement should preserve sufficient access, support and export capability to repeat it at termination. The goal is not zero switching cost; it is a known, funded and timebound exit.

This test also protects Northern Data. Clear boundaries reduce disputes caused by customers assuming a group brand guarantees a functionality or remedy that no contract contains. Workload benchmarks distinguish real performance from headline GPU counts. A documented operational chain makes the value of central services visible even when the legal entity performing that work is not the cloud vendor.

The Control Signals to Watch

The next phase should be judged by a small number of observable changes rather than another large capacity number.

First, watch whether RUM Group publishes a consistent product and legal migration from Taiga and Rumble Cloud to Quake. A unified brand is the starting signal. Common identity, APIs, regional terms, service levels, support and billing would be stronger evidence that the control plane has effectively converged. Until then, existing service documentation must be presumed to describe the environment it names, not the entire combined group.

Second, watch utilisation with price and cash. The reported 85 per cent in March 2026 is encouraging, but sustainable economics require net revenue per available GPU-hour, controlled fault and support cost, and cash conversion after power and interest. RUM Group raised 2026 revenue guidance for the acquired operation to €170–190 million in its closing announcement. Delivery against this range, without another major working capital or debt intervention, will reveal more than allocation alone.

Third, watch the customer mix. Replacing a single dominant reservation customer with multiple sustainable customers would reduce negotiation and ramp risk. New frameworks should be judged by recognised revenue and cash collection, not by maximum GPU language. Related-party demand must be understood separately because Tether is simultaneously a financier, a shareholder and a potential customer.

Fourth, watch the software dependency. Gcore's continued involvement may make sense; stability matters more than ideological ownership. What matters is whether Northern Data or RUM Group can demonstrate change control, incident accountability, continuity rights and a credible migration path. If Quake introduces a new layer, buyers should ask whether it replaces, wraps or simply renames the existing control plane.

Fifth, watch the physical conversion funnel. Ardent's announced megawatts must progress through permitting, grid connection, construction, energisation, equipment installation, customer acceptance and external revenue. Reporting each step would make the group's "unmonetised" capacity understandable. Claims about power source, PUE and water should remain site-specific and, where possible, independently assured.

Finally, watch Northern Data Software GmbH itself. Current public reports have narrowed its description from software development and computing sales to central and corporate services. A new statutory filing, group reorganisation, employee disclosure, licence agreement or integration role could refine the boundary. Until such evidence appears, the company should be neither inflated into the sole author of Northern Data's cloud platform nor dismissed as an empty registration.

The decisive question of the failed job at 03:10 remains the right conclusion. The GPU may sit in Norway, the contract may come from Ireland, the facility may be operated by another group company, orchestration may depend on Gcore, central services may pass through Frankfurt, and the ultimate parent may be in the US. The customer does not need all these functions under one legal roof. It needs the party at each delivery point to be named, capable and bound to act. The Northern Data group controls an impressive compute domain.

Whether Quake can turn that domain into a reliable cloud will be decided by who can stop, move and recover the job—and whether the customer can prove it before the job matters.