Software Is Disappearing. Your Data Centre’s Data Is What’s Left

Key takeaways

• For decades the software in a data centre set the workflow. DCIM, BMS and SCADA screens decided what operators saw, logged and could change.
• AI agents now write, test and rebuild code, so interfaces can be generated on top of a site’s data. Eurostat counts 19.95% of EU enterprises with 10 or more employees using AI in 2025, up from 13.48% in 2024.
• Agents are trained and tuned on data. Years of a site’s own power, cooling, alarm and maintenance records become the asset that lasts.
• Data sovereignty follows directly: where that data sits, which law can reach it and who may train on it decide who keeps its value.
• The EU Data Act has applied since 12 September 2025. It gives users of connected products a right to their data, and cloud switching charges end on 12 January 2027.

Software used to set the rules in a data centre, and the data followed. Now data sets the rules, and agents build the software on top of it.

The decision for operators: treat your facility’s operating history as a capital asset. Collect it, clean it, own it in your contracts and keep it under your jurisdiction.

Software used to set the rules

In most facilities the operator meets the building through screens. The DCIM tool shows racks, power chains and capacity. The building management system shows chillers, air handlers and alarms. SCADA watches the electrical plant. Each tool decided which points were logged, how often, which alarms existed and what a shift could change.

The operator learned the software’s workflow. The data it kept was mostly the data its own screens needed. That arrangement made sense while code was expensive. Writing, testing and integrating a control interface took years of engineering, so a vendor spread the cost over many sites, and every site fitted the product.

Coding agents now write code, run the tests, read the errors and try again. They port a script from one interface to another and write the glue between two systems. Those are the tasks that made integration projects long and licences sticky. When they cost less, a site can generate a dashboard, a report or an operator assistant on top of its own data. The fixed product becomes one option among several.

Agents learn from data, so the data is the asset

An agent that supports an operator is a general model plus what it has learned from one site. The general model can be replaced when a better one appears. The record of how one building behaves cannot be bought. It shows how the chilled water loop answered a heat wave, which UPS modules drifted before they failed, how load grew hall by hall, and what the site did during a grid disturbance.

For a data-centre operator, data to train models means three jobs in practice:

• Control tuning. Setpoints for supply air, chilled water and fan speeds, fitted to years of the site’s own weather, load and energy data.
• Failure prediction. Alarm histories and maintenance logs tied to the asset they describe, so a model can learn which patterns came before a fault.
• Capacity planning. Power and cooling headroom by hall and by phase, measured across seasons, to decide when the site needs more.

Each job needs long, clean and time-aligned history. A model fitted to another site’s data learns another building. This is why the history is scarce. It takes years to build and it cannot be recreated later: a sensor that was never logged in 2024 has no 2024 data.

Why data sovereignty follows

If operating data is the asset that lasts, control over it decides who captures its value. Four questions settle that control: where the data is stored, which jurisdiction can reach it, who can access it, and who may train on it.

The risks are practical:

• Jurisdiction. Data held by a cloud provider under a non-EU legal order can be the subject of an access order from that country’s authorities.
• Training rights. Service terms can allow a provider to use customer data to improve its own products, including models it then sells to other customers.
• Lock-in. History kept only in a proprietary format is hard to move, and an export that drops point names, units and asset links loses most of its training value.
• Switching losses. Changing provider without a full export can leave the training history behind with the old one.

Europe has started to answer each of these. Under Article 28 of the Data Act, cloud and other data processing providers must publish the jurisdiction their infrastructure is subject to, and the measures they take against foreign governmental access. Article 32 requires them to take technical, organisational and legal measures against third-country governmental access to non-personal data held in the EU where it would conflict with EU or national law. A foreign court or authority decision is recognised only through an international agreement, such as a mutual legal assistance treaty, or under strict conditions set out in the article.

Chapter VI of the same act deals with switching. A contract must allow a notice period of at most two months, a transition of at most 30 days and at least 30 days to retrieve the data (Article 25). Providers must point to a register of the data formats they export (Article 26) and offer open interfaces free of charge (Article 30). Switching charges may cover only the provider’s direct costs until 12 January 2027, and none are allowed after that (Article 29).

Where operating data includes personal data, such as access logs or staff on shift, the GDPR transfer rules apply as well: Article 44 and Article 46 allow transfers outside the EU only under the conditions of Chapter V, such as appropriate safeguards.

The Commission is also funding the alternative. Its common European data spaces programme lists projects for energy and for manufacturing. On 5 December 2023 it approved an Important Project of Common European Interest in cloud and edge computing, jointly notified by seven member states. Its Cloud and AI Development Act page sets out a plan to strengthen Europe’s sovereignty in cloud and AI.

The practical line for an operator is simple. Keep a full copy of the operating history in a format you control, under EU jurisdiction, with contract terms that forbid training on your data without your consent.

The EU legal frame

The Data Act, Regulation (EU) 2023/2854, has applied since 12 September 2025 (Article 50). It covers connected products: items that generate data about their use or environment and can communicate it (Article 2). Its recitals name commercial infrastructure and industrial machinery among them. Plant that reports its own condition, such as chillers, UPS systems and switchgear, can fit that definition. Whether a given item does depends on its design and its contract.

Where a user cannot reach the data directly, the data holder must make it available free of charge, in a structured, commonly used and machine-readable format, and continuously and in real time where feasible (Article 4). The user can also have the data sent to a third party of its choice (Article 5). The holder may use non-personal data only on the basis of a contract with the user, and may not pass non-personal product data to third parties beyond that contract (Article 4). Trade secrets are protected, and in exceptional cases a holder can refuse access to specific data. Products placed on the market after 12 September 2026 must be designed so that users can reach the data directly (Article 3 and Article 50).

Unfair data terms imposed by one company on another are not binding (Article 13). This applies to contracts concluded after 12 September 2025, and from 12 September 2027 to older long or open-ended contracts.

The AI Act, Regulation (EU) 2024/1689, has required providers of general-purpose AI models since 2 August 2025 to keep a copyright policy and to publish a sufficiently detailed summary of the content used for training (Article 53 and Article 113). The Commission published the template for that summary on 24 July 2025. Before buying a model, an operator can read what kind of content trained it.

The Copyright in the Digital Single Market Directive, Directive (EU) 2019/790, allows text and data mining of lawfully accessible works unless the rightholder has reserved that use in an appropriate manner, such as machine-readable means for content published online (Article 4). That matters for manuals, reports and studies you publish. Raw sensor readings are facts that copyright may not protect at all, so for operating data the contract carries the weight.

What to do

• Data inventory. List every system that logs data (DCIM, BMS, power monitoring, SCADA, maintenance management, metering), with its points, logging interval and retention.
• Retention. Keep raw time series at full resolution for years, with synchronised clocks and stable point names.
• Ownership clauses. Read the DCIM, BMS and cloud contracts for who owns the data, who may use it and whether the provider may train on it. Ask for a written access right and a clause that forbids training without your consent.
• Export formats. Run a full export now, in an open format that keeps point names, units and asset identifiers.
• Jurisdiction. Keep one full copy under EU law, and read each provider’s published statement on jurisdiction and governmental access.
• Model test. Before trusting a newer model, replay a past period of your own data and compare its advice with what actually happened. Adopt it only if it makes the operation better.

ScienceShot works the same way at a small scale. Our trackers of European markets, grid connections and rules are dated datasets on the Intelligence hub, and an AI assistant can read them through our AI connector. The data-centre design tool builds a site plan on top of the same data.

@online{scienceshot-software-is-disappearing-data-2026,
  author       = {Turner, Alex},
  title        = {{Software Is Disappearing. Your Data Centre's Data Is What's Left}},
  organization = {ScienceShot},
  date         = {2026-10-18},
  year         = {2026},
  url          = {https://scienceshot.com/post/software-is-disappearing-data-centre-data}
}