Reporting Year 2026 · The Company’s First Sustainability Report As of September 3, 2026

1. Strategy

basebox develops the legally compliant AI platform for healthcare and government. Its mission—to bring AI into regulated organizations in a way that is safe, manageable, and affordable—means making AI usable while maintaining full control. The responsible use of resources is part of this mission: for basebox, safety, control, and sustainability are not mutually exclusive but are interlinked.

This report is basebox’s first sustainability report. It lays the foundation for tracking progress in the coming years. basebox has deliberately avoided setting complex targets that cannot be reliably substantiated at this time and instead transparently outlines what has already been implemented and where there is room for improvement.

The report addresses two areas separately: basebox’s business operations (Section 3) and the operation of the AI platform, including the data centers and AI models used (Section 4). The second area is more relevant to customers because that is where the vast majority of energy consumption occurs.

A key aspect: basebox’s fundamental technological decisions—the exclusive use of Rust as the implementation language and intelligent model routing when deploying AI models—were made primarily for security and efficiency reasons. Both generate additional environmental benefits, which are described in Section 5.

2. Process Management

Responsibility for sustainability lies with René Herzer, CEO and founder of basebox.

The following guidelines have been in place since the company’s founding in 2022:

• Office electricity: 100% green electricity (EWS Schönau) since the start of business operations in 2022

• Travel: Business travel within the DACH region is generally by train; car-sharing (membership in Utting am Ammersee) only in exceptional cases

• Vehicle Fleet: No company-owned vehicle fleet, except for bicycles

• Working from Home: All employees can work from home

• Working Hours: Flexible scheduling based on individual performance; overtime is not recommended

• Data Center Selection: When selecting data centers and cloud service providers, the energy mix, PUE value, and applicable certifications are taken into account

3. Company Operations

3.1 Office Operations

The office in Utting am Ammersee (approx. 150 square meters) has been powered entirely by green electricity from the provider EWS Schönau since the company was founded in 2022.

3.2 Mobility

Business travel is primarily by train: approximately 2 trips per month, each with 2 people, across the DACH region. In addition, there are approximately 4 flights per year between Ireland and Germany, due to the Head of Engineering being based in Ireland. The company does not maintain its own vehicle fleet.

3.3 Resources

Paper consumption is minimal, as printing is rarely done in day-to-day operations. Waste is sorted according to German standards.

4. Operation of the AI Platform

This section concerns the service itself: the data centers in which the platform and AI models are operated, as well as the energy consumption and emissions generated during use.

4.1 Data Centers Used

The platform is operated exclusively in data centers within the European Economic Area. For each location, the PUE (Power Usage Effectiveness) value, the share of renewable energy, and the applicable certifications are reported.

Platform operation — Hetzner Online GmbH, Germany PUE 1.10 to 1.16 · 100% renewable energy · ISO/IEC 27001

In-house inference operations starting November 1, 2026 — noris network AG, Aschheim near Munich PUE below 1.2 · 100% renewable energy (CMS standard “Renewable Energy Generation”) · ISO/IEC 27001 based on IT-Grundschutz, ISO 14001, KyotoCooling

Model inference GPT-OSS-120B — STACKIT GmbH & Co. KG, Germany PUE below 1.3 (across the fleet) · 100% renewable energy · BSI C5 Type 2, ISO/IEC 27001, ISO 50001, EN 50600

Model Inference Claude Sonnet 5 — Google Cloud, Frankfurt (europe-west3) Fleet PUE 1.09 (TTM 2025) · 68% CO₂-free energy (CFE, 2024) · BSI C5:2020 certification, ISO/IEC 27001, ISO 50001

Model Inference Mistral Large 3 — Mistral AI SAS, France No published metrics from the provider

Image Generation FLUX.1 — IONOS SE, Germany PUE 1.23 · 100% renewable energy · ISO/IEC 27001, BSI C5

Web search, only when activated — European Search Perspective SAS, EU No published metrics from the provider, no certification

Change effective November 1, 2026: Starting on this date, basebox will operate its own inference infrastructure at the Munich East data center of noris network AG. This will shift part of the model processing from external providers to in-house operations. The location was selected based on factors including the energy mix, the PUE value, and KyotoCooling technology.

Transparency note regarding Mistral AI and web search: No published data center metrics are available for these two services. basebox discloses this fact rather than providing estimates. Both services are provided within the EU. Web search is only active when it is explicitly enabled.

4.2 Energy Consumption and Emissions per Query

The following figures show the estimated energy consumption and CO₂ emissions per query for the models used. All values refer to a response of approximately 400 tokens; for image models, they refer to a generated image.

Claude Sonnet 5 — Google Vertex AI, Frankfurt Per query: approx. 1.24 g CO₂eq Model training: not published by the manufacturer

Mistral Large 3 — Mistral AI, France Per query: approx. 1.14 g CO₂eq (manufacturer’s LCA) Model training: 20.4 kt CO₂eq over an 18-month life cycle. This value is taken from the manufacturer’s LCA (Mistral, 2025) for Mistral Large 2 and serves as a conservative reference value for the model family.

GPT-OSS-120B — STACKIT, Germany Per request: approx. 0.11 g CO₂eq Model training: open weights, training CO₂ not published

Kimi K2.6 — basebox in-house operation starting November 1, 2026 Per request: approx. 0.0042 kWh. This results in 0.15 g CO₂eq when powered by green electricity and 2.08 g CO₂eq when powered by standard German grid electricity. Model training: weights undisclosed, training CO₂ not published

FLUX.1 — IONOS AI Model Hub, Germany Per image: approx. 2.66 g CO₂eq Model training: not published by the manufacturer

How these values are calculated

The figures are estimates based on energy consumption, the data center’s PUE, and the emission factor of the electricity purchased. The calculation consistently follows this formula:

Energy consumption of the request × data center PUE × emission factor of the electricity mix

In detail:

• Claude Sonnet 5: approx. 0.003 kWh × PUE 1.09 × 380 g/kWh. The German grid mix is used for the emission factor rather than the provider’s offset, because grid electricity is physically available at the Frankfurt location. This is the more conservative estimate.

• GPT-OSS-120B: approx. 0.003 kWh × PUE 1.2 × 30 g/kWh. The low emission factor results from the actual use of green electricity at the location and explains the significant difference compared to the other text models.

• Kimi K2.6: approx. 0.0042 kWh, corresponding to about 15 seconds of computation time on two H200s and one RTX 6000. Two values are reported because the emission factor varies significantly depending on the electricity source.

• FLUX.1: approx. 0.007 kWh per image × 380 g/kWh. Image generation consumes significantly more energy per request than text generation.

Putting these numbers into context: what they do and don’t represent

These values are estimates, not measured consumption data. Reliable, independently verified figures on the energy consumption of LLM inference are limited across the industry; most model providers do not publish either training or inference data. basebox therefore discloses the basis for its calculations so that the figures are transparent and verifiable, rather than presenting them as measured values.

The actual consumption of a single query varies considerably—depending on the length of the input, the length of the response, the use of reasoning functions, and whether additional data sources are queried. The values above describe a typical case, not an upper limit.

basebox is developing its own energy and utilization monitoring system to gradually replace these estimates with measured values. Starting with in-house operations in November 2026, this will be technically feasible for self-operated inference.

5. Efficiency Through Product Decisions

Two technological decisions shape the platform’s energy requirements.

5.1 Rust as the Sole Implementation Language

The platform is developed 100% in Rust. This decision was made primarily for security and reliability reasons—key requirements when handling critical data from healthcare and public administration—but it also brings efficiency benefits:

• Native compilation to machine code without a runtime interpreter or virtual machine. Studies rank Rust on par with C/C++ in terms of energy efficiency, while offering additional memory safety.

• No garbage collector, eliminating recurring CPU and energy cycles during continuous operation.

• Lower resource requirements enable smaller hardware, reduced cooling needs, and a smaller hardware footprint.

Source: Studies on the energy efficiency of programming languages conducted by the Universities of Coimbra and Minho. The specific savings potential depends on implementation and workload.

5.2 Model Routing and Model Selection

Not every query requires the largest available model. The platform is designed to route queries to the smallest model capable of reliably solving the task:

• Right-sizing computing power: Simple tasks use smaller, more resource-efficient models; only complex queries require large models.

• Avoiding redundancy by caching previously answered or similar queries.

• Reduced need for energy-intensive accelerator hardware for the majority of data traffic.

The figures in Section 4.2 illustrate the magnitude of this effect: For a comparable query, there is roughly a tenfold difference between the most energy-efficient and the most energy-intensive text model. The choice of model is thus the most effective single lever for reducing energy consumption during operation.

Administrators can specify which models are available within their organization, taking energy efficiency into account as a criterion.

6. Reporting to Customers

basebox provides customers with annual sustainability reports on their use of the platform. These reports include:

• aggregated energy and CO₂ figures for platform usage during the reporting period

• a breakdown by models and services used

• the current data center metrics for the locations used: PUE value, energy mix, certifications

• year-over-year trends, once more than one reporting period is available

• measures implemented and planned during the reporting period to reduce energy consumption and emissions

The report is prepared for the first time for the first full contract year and is updated annually thereafter.

Measures for Reduction

Implemented:

• 100% green electricity for office operations since 2022

• Selection of data centers with a high share of renewable energy and a low PUE value

• Rust as the implementation language, resulting in lower computing requirements during continuous operation

• Caching to avoid redundant model calls

• Train-based business travel; no company-owned vehicle fleet

Planned:

• In-house energy and utilization monitoring to replace estimates with measured values (starting with in-house operations in November 2026)

• Expansion of model routing to increase the proportion of queries answered using small models

• Shifting additional inference to in-house operations with 100% green electricity

• Expansion of the data set to include providers for which no metrics are currently available

7. Company

7.1 Team Structure

basebox employs an international team, including members from Germany, Brazil, and India (based in Germany).

• Full-time: 4 women (3 of whom are in tech roles), 3 men (2 of whom are in tech roles)

• Part-time: 1 woman, 2 working students (male)

• Freelancers in tech roles: 2 men (through their own companies in Ireland and Estonia)

7.2 Working Conditions

• Very low sick leave rate; no significant instances of illness

• No workplace accidents during the reporting period

• 30 days of vacation per year

• Flexible work schedule; overtime is not encouraged

• Consistent option to work from home

7.3 Employee Turnover

During the reporting period, two employees left the company: a Senior Backend Developer (who became self-employed) and a Full Stack Engineer.

7.4 Professional Development

There is currently no separate budget for professional development. basebox is considering establishing one as the company grows.

7.5 Compensation

Compensation is based on transparent, role-specific criteria. All employees can organize their work hours flexibly.

8. Corporate Governance

8.1 Procurement

Work equipment is procured partly as refurbished devices and partly as new. New purchases are necessary in the context of product development, as up-to-date hardware is required for testing the latest devices.

8.2 Compliance

During the reporting period, there were no legal violations and no fines.

9. Summary

Strategy: First sustainability report; safety and efficiency as the unifying principle.

Processes: Responsibility lies with management; energy mix and PUE as selection criteria for data centers.

Platform Operations: All data centers are located within the EEA. For five of the seven sites in use, PUE values range from 1.09 to 1.23, and 100% of the electricity comes from renewable sources (with the exception of Google Cloud Frankfurt: 68% CO₂-free energy). Energy consumption and emissions are reported for each model with the calculation basis disclosed.

Efficiency: Rust and model routing as structural levers. There is approximately a tenfold difference between the most energy-efficient and the most energy-intensive text models.

Reporting: Annual sustainability reporting for customers, broken down by models, data center metrics, and trends over time.

Company: 100% green electricity in the office since 2022; rail-based mobility; diverse team; no incidents during the reporting period.

Responsible for the content: René Herzer, Managing Director, basebox GmbH. Inquiries regarding this report: support@basebox.ai