Appendix 4.1 · Data Centre Specification

General provisions

1 Purpose and scope

2 Initial conditions and assumptions

3 Adopted architectural decisions

Section I Hardware

4 PROTO-01 main compute workstation

5 Workstation and client devices

6 Data storage

7 Network equipment and spare parts

8 Camera system prototype

9 GPS trackers for freight vehicles

10 Variable frequency drive small test rig

11 Power supply and infrastructure

12 Rack cabling and installation

13 Other spare parts tools and scanner

14 Hardware summary for stages 0 and 1

15 Cloud site equipment

Section II Software

16 Operating systems and infrastructure

17 AI data and agents

18 Data layer

19 Business systems

20 Engineering software

21 Monitoring

22 DevOps security and backup

23 Software summary

24 Cloud site software

Section III Allocation of functions between the local site and remote cloud

25 Allocation principle

26 Links and throughput

27 Overall allocation of functions

28 Two options for local AI and engineering software

29 Model allocation

30 Hybrid operating patterns

31 Backup and failover

32 Allocation of the site link

33 Data and security

Section IV Summary budget

34 CAPEX for stages 0 and 1

35 OPEX for year 1

36 Deferred costs for stage 2

37 Items excluded from the budget

Section V Decisions and risks

38 Decisions on open items

39 Risks and pre-purchase checks

40 Outstanding clarifications and next steps

Annex A Price sources

Annex B Transfer time calculation method

General provisions

1 Purpose and scope

This specification combines and replaces Hardware and Software v0.8 and Transfer of Functions to the Company Cloud v1.0. It covers the complete data centre structure: the local site, the remote cloud site, hardware, software and the division of functions between them. Section III compares two ways to use local AI: Option A runs Micromine and Metashape at the local site; Option B places those applications in the cloud, subject to licence and network checks.

Changes from versions 0.8 and 1.0:

Two documents are combined into one, the sections are renumbered.

The design now includes a cloud internet connection of at least 100 Mbit/s. Site-to-cloud transfer remains limited by the site's 10 Mbit/s link.

Cloud-site hardware and software are included as a separate design area. The working assumption is a configuration at least as capable as the local site, including a 96 GB GPU; it requires confirmation.

The earlier decision to omit cloud backup is reversed. A separate VPS relay for trackers is unnecessary.

Section III compares two ways to use local AI and states the conditions for selecting one.

The budget is recalculated for both options.

Commercial purpose. This specification estimates the technical cost of establishing one additional Oristella project centre for a contracting partner. Oristella already holds the core software licences for its central system; their acquisition is not charged again in the centre estimate. Additional site-specific licences, seats, subscriptions, transfers and renewals are treated as budget items or exclusions only where expressly identified below. The price payable under a project-centre agreement will reflect the selected configuration, the separately quoted site-specific costs and Oristella's contracted services.

2 Initial conditions and assumptions

A published price comes from a public source listed in Annex A. An estimate is an engineering assumption without a supplier quotation. The totals combine ranges and are not supplier offers.

3 Adopted architectural decisions

Section I Hardware

4 PROTO-01 main compute workstation

Role: AI/LLM, CV/OCR, embedding and reranker, Docker, PostgreSQL/TimescaleDB/Qdrant, Redpanda, Odoo, Traccar, Forgejo, Frigate, agent orchestrator, KVM/QEMU, Windows Server Evaluation VM (AD/DNS), Windows 11 VM (engineering software, GPU on demand; option A only).

Price risk. The source records an RTX PRO 6000 Blackwell price of USD 13,250, 55% above its USD 8,565 launch price. Secondary-market listings are stated at USD 17,860–21,800. Purchase only through authorised channels; availability and the actual GPU quotation strongly affect the entire project budget.

5 Workstation and client devices

This line remains the v0.6 engineering estimate.

6 Data storage

The NAS stores SeaweedFS data, a seven-day Frigate archive of approximately 3.6 TB, database dumps, pgBackRest and Acronis copies. The range is an estimate. Before purchase, confirm Synology's policy on third-party drives and the compatibility list for this model.

7 Network equipment and spare parts

Prepare a VLAN and port diagram, including Industrial VLAN 50 for PROFINET, and calculate the PoE budget. Six cameras and two access points fit within 140 W only if each camera draws no more than 12–15 W.

8 Camera system prototype

Frigate records and archives video; go2rtc relays streams to DeepStream. Keep a seven-day archive on the NAS. Frigate recording continues without GPU access while a Windows VM occupies the GPU, but GPU detection stops. Check applicable video-surveillance rules.

9 GPS trackers for freight vehicles

Trackers send data to a publicly reachable Traccar instance in the cloud, which replicates to the local site over WireGuard. No separate VPS relay is required. Tracker flash memory buffers data when the cloud is unavailable. SIM tariffs are included in OPEX.

10 Variable frequency drive small test rig

Only a qualified electrician may connect the variable frequency drive to mains power. Route all VFD access through the Tool Gateway. Before the first motor run, verify that no command can bypass the gateway. Provide an emergency stop in the power circuit.

11 Power supply and infrastructure

Provide automatic graceful shutdown on a UPS signal, using NUT or an equivalent, and calculate battery runtime.

12 Rack cabling and installation

Rack installation is deferred to stage 2. The v0.6 estimate was USD 1,120–2,240. In stages 0 and 1, network equipment can be placed on a table or wall and PROTO-01 remains in a tower case. Stage 2 adds a 4U rack case for PROTO-01, a redundant PSU, an 18–24U rack, patch panel, KVM console, cable management, labelling and server-room cooling. Confirm prices before procurement.

13 Other spare parts tools and scanner

Spare RAM may cost considerably more than budgeted; see the price risk in section 4. Do not purchase the drone or GNSS equipment for Metashape and TBC in stage 1. Decide after the use cases are defined.

14 Hardware summary for stages 0 and 1

15 Cloud site equipment

The remote cloud is treated as company-owned, so its hardware cost is excluded from the stages 0 and 1 budget. Its assumed configuration is at least as capable as the local site; verify actual specifications.

Under Option B, the cloud GPU is scheduled between the Windows VM during the day and batch level 2 LLM jobs at night (section 28.5). Under Option A, the cloud GPU can serve level 2 throughout the day.

Section II Software

16 Operating systems and infrastructure

The Windows Server license counts for all physical host cores: 24 cores is a 16-core license plus 8 cores (the published prices of the 16-core license are about 1,000–1,330 USD), plus about 265 USD for 5 custom CALs. RDS CAL is not needed as long as the engineering software runs on a single Windows 11 VM. Accessing Windows 11 VM from a MacBook usually requires a VDA license: check Microsoft's terms.

17 AI data and agents

18 Data layer

19 Business systems

Published Odoo prices vary by region and promotion. The Custom plan is approximately USD 49 per user per month with annual payment and approximately USD 61 on renewal. The Standard plan is hosted by Odoo; self-hosting requires Custom.

20 Engineering software

Under Option A, all engineering applications run in one Windows 11 VM with local GPU access, one session at a time. Under Option B, Micromine 2016 and Metashape Professional must run in a cloud Windows VM; Civil 3D, QGIS and TBC are also assigned there by default. Engineers connect by RDP over WireGuard, subject to the checks in section 28.

21 Monitoring

All Open Source, 0 USD.

22 DevOps security and backup

Acronis licensing assumptions: the Linux host and Windows Server VM are each treated as servers at approximately USD 620–780 per year; a Windows 11 VM is treated as a workstation at approximately USD 85–109. Licensing the KVM host as a virtualisation host would require the Advanced edition. Confirm the terms with the supplier.

23 Software summary

24 Cloud site software

Budget cloud software licences separately, including Windows Server for the AD replica, Windows for the engineering VM and Acronis for cloud workloads. Company ownership of the cloud does not remove software licence costs.

Section III Allocation of functions between the local site and remote cloud

25 Allocation principle

Keep functions at the local site when they must work offline, control equipment or generate more data than the link can carry. Assign latency-tolerant batch work and tasks that benefit from higher compute capacity or off-site placement to the cloud.

Assess five characteristics for each function:

26 Links and throughput

The cloud has an Internet channel of at least 100 Mbit/s, and the site-cloud communication is limited to the site channel, 10 Mbit/s. About 75% of the speed is actually available: ~0.94MB/s (~81GB/day) on-site and ~9.4MB/s (~810GB/day) in the cloud. This means that heavy downloads from the Internet (weight models, images, updates) are performed in the cloud, and copied to the site at night or by physical media.

Site to cloud link at 10 Mbit/s

Cloud internet link at a minimum of 100 Mbit/s

For interactive work, latency can matter more than transferred volume. Twenty sequential tool calls with 0.5 seconds of cloud latency add 10 seconds. Keep those chains local and invoke the cloud LLM in one larger step.

27 Overall allocation of functions

The following tables show functions common to both options. The only difference is the location of the engineering Windows VM and how the GPU is used (section 28).

Functions retained at the local site

Functions assigned to the cloud

Functions replicated between sites

28 Two options for local AI and engineering software

The options differ in where Micromine and Metashape run and how the site's GPU is used. The other functions in section 27 remain allocated in the same way.

28.1 Option A engineering software at the local site

Micromine 2016, Metashape after purchase, Civil 3D and QGIS run in a Windows 11 VM on PROTO-01. GPU passthrough to the VM is enabled when required.

While engineers use the GPU, vLLM, DeepStream and other GPU containers stop. Frigate continues recording video without the GPU, but GPU-based detection is unavailable.

Level 2 AI runs in the cloud around the clock; a cloud GPU is not occupied by Windows VM.

The engineering work does not depend on the channel; the geology and survey data remain on the site, a copy of the critical set goes to the cloud.

Risks: GPU reset during VFIO (reset) on Blackwell, AI level 1 interruption when engineers work, Micromine 2016 compatibility with Windows 11.

28.2 Option B Micromine and Metashape in the cloud

Under Option B, Micromine 2016 and Metashape Professional licences are placed at the remote site if vendor terms allow it. The Windows VM with those applications, and by default Civil 3D and QGIS, runs in the cloud.

Engineers connect via RDP via WireGuard; one session takes 2-5 Mbps, that is, up to half the site channel.

The local GPU remains assigned to level 1 AI without switching. The available capacity and uninterrupted operation support longer context, more agents and continuous computer vision.

The cloud GPU follows a schedule: Windows VM by day; batch level 2 LLM jobs by night (section 28.5).

The primary copy of geological and survey data resides in the cloud. Local exports remain at the site for work during a link outage.

Prerequisites: vendor terms permit cloud placement of the licences and keys; Micromine 2016 works on cloud-hosted Windows; and measured link availability and latency support RDP.

Risk: When the channel breaks, work in Micromine and Metashape stops.

28.3 Comparison of options

28.4 Recommendation and selection conditions

Option A is the baseline recommendation for stages 0 and 1 because it does not depend on the link or licence transfer. Select Option B only after all three checks below have been completed successfully:

Micromine and Agisoft have provided written confirmation of cloud licence terms, including key type and any USB-dongle requirements.

Two weeks of link measurements demonstrate acceptable working-hour availability and RDP latency.

Micromine 2016 has passed a cloud Windows test, and an RDP session has passed a test over the 10 Mbit/s site link.

If any condition is not met, retain Option A. Option B removes local VFIO reset risk and frees the site's GPU for continuous AI work.

28.5 Cloud GPU schedule under Option B

29 Model allocation

The models are divided into two levels. Level 1 is on-site and should always be available. Level 2 works in the cloud for heavy and non-urgent tasks. The cloud configuration is no worse than the site, that is, one card accounts for 96 GB of VRAM.

GPU use under each option

Level 1 VRAM estimate: approximately 55 GB for the LLM, including FP8 weights and KV cache; 16 GB for embeddings and reranking; and 25 GB for OCR, DeepStream and computer vision. Together they use nearly all of a 96 GB card. A smaller local GPU would require a reduced level 1 workload.

A single 96 GB card can host compressed models up to approximately 120B parameters. A 235B-class model needs a second cloud GPU. If the actual cloud configuration is smaller than assumed, reduce the level 2 workload rather than changing the overall allocation of functions.

30 Hybrid operating patterns

Route simple or urgent LLM requests to the local model and complex, non-urgent requests to the cloud. On a link outage, fall back to the local model and label the result as a simplified response.

Queue long-running jobs and return their results later; do not keep a connection open throughout processing.

Send derived information, such as descriptions, embeddings and metadata, instead of raw video and large files when sufficient for the task.

Use Redpanda and PostgreSQL logical replication with store-and-forward delivery. Assign one authoritative writer for each dataset to avoid divergence.

The local site is authoritative for operational data; the cloud is authoritative for analytics and backup copies.

Download model weights and images to the cloud, then copy them to the local NAS overnight. Keep local mirrors so a new model test does not consume the site link for half a day.

31 Backup and failover

Back up only the critical dataset to the company cloud. Restoring an entire 96 TB site over this link would take more than three years. Emergency operation therefore starts critical services in the cloud from replicas.

Transfer an initial dataset of 500 GB or more on a physical drive, followed by rate-limited overnight increments. Confirm whether Acronis can send copies to company storage by network share or SFTP. Keep an offline external-SSD copy as the third copy in the 3-2-1 scheme.

32 Allocation of the site link

Reserve a guaranteed, highest-priority share of the link for control and telemetry. Schedule bulk transfer overnight with a rate limit, using QoS queues on the MikroTik RB5009.

Parsec with a standard bitrate on such a channel is not suitable, RDP remains for remote work.

33 Data and security

Encrypt sensitive geological data, diagrams and database dumps over WireGuard when sending them to the company cloud. Before transfer, confirm that the cloud's jurisdiction permits storage of those data.

The restriction on external transfer applies to third-party services: do not send identifiable data to third-party APIs without anonymisation and permission.

Equipment control stays at the local site. Cloud agents may prepare recommendations and requests, but the local Tool Gateway issues VFD commands only after human confirmation.

The firewall denies cloud access to Industrial VLAN 50 and the Tool Gateway. Use separate WireGuard keys for the cloud and local site.

A compromised cloud account must not provide a route to equipment control. This is a network-design requirement.

Option B: The main copy of the geologic data is in the cloud, so the requirements for cloud protection (access via MFA, logging, separate Windows VM accounts) are higher than in option A.

Section IV Summary budget

34 CAPEX for stages 0 and 1

Local-site CAPEX is the same under both options because the 96 GB GPU remains in Option B. PROTO-01, particularly GPU and RAM pricing, drives most of the estimate range. Cloud hardware is excluded because the design assumes an existing company cloud.

Cost basis for one project centre. The stages 0 and 1 site hardware is estimated at USD 40,243–62,734, or USD 46,279–72,144 including the 15% reserve, under either option. The one-time software licence purchase in these stages is zero because the existing core licences are reused. Year-one operating costs, stage-2 additions and the exclusions in section 37 are shown separately. This range supports the centre quotation; the final purchase price depends on the contracted scope and site-specific costs.

35 OPEX for year 1

Cloud electricity, administration and workload licences are excluded from this OPEX estimate.

36 Deferred costs for stage 2

37 Items excluded from the budget

Not included: Geobank, drone and GNSS equipment (according to scenarios); VAT, delivery and duties; Startdrive / TIA Portal software for VFD; Internet channel and its redundancy; work of the administrator and integrator, implementation of Odoo; Micromine 2016 (available) and extension of its support; the cost of transferring Micromine and Metashape licenses to the cloud (option B); equipment, electricity, administration and cloud licenses.

Section V Decisions and risks

38 Decisions on open items

39 Risks and pre-purchase checks

40 Outstanding clarifications and next steps

1. Measure site-link latency, packet loss and availability during working hours over two weeks.

2. Confirm the actual cloud configuration: GPU, RAM, disk, network, administration, SLA.

3. Ask Micromine and Agisoft for license terms in the cloud.

4. Obtain supplier quotations for the GPU, RAM, CPU, motherboard, NAS, HDDs and switch.

5. Pilot the software stack on Blackwell and test GPU switching between host and Windows VM, including the scheduled cloud arrangement.

6. Create a diagram of VLAN, ports and calculation of the PoE budget.

7. Identify drone, GNSS, Geobank, Metashape and TBC use cases.

8. Then select Option A or B, prepare procurement version 2.0 and finalise the hardware bill of materials.

Document status: planning estimate until supplier quotations are received.

Annex A Price sources

Public sources used in price reconciliation (October 2026) Prices change and a quote is required before purchase.

Annex B Transfer time calculation method

Site: 10 Mbit/s = 1.25 MB/s theoretically; 75% of the usable speed is accepted, that is ~0.94 MB/s (~81 GB / day).

Cloud: 100 Mbit/s = 12.5 MB/s theoretically; accepted 75%, that is ~9.4 MB/s (~810 GB / day).

Transmission time = volume/speed. On the site: 20 GB ≈ 5.9 hours; 30 GB ≈ 8.9 hours; 60 GB ≈ 17.7 hours; 500 GB ≈ 6.2 days; 2 TB ≈ 24.6 days; 96 TB ≈ 3.2 years. In the cloud: 30 GB ≈ 53 minutes; 60 GB ≈ 1.8 hours; 500 GB ≈ 14.8 hours; 2 TB ≈ 2.5 days.

Video: 6 cameras of 8 Mbps = 48 Mbps, which is 4.8 times higher than the site channel.

Events: 100 clips of 10 s at 8 Mbps (~10 MB each) ≈ 1 GB per day.

LLM context: 100 thousand tokens at ~4 bytes per token ≈ 0.4 MB.

RDP: 2–5 Mbps per session; with a 10 Mbps channel, one session takes 20% to 50% of the channel.

All calculations are idealized and do not take into account latency, packet loss, and encryption overhead; with real measurements, they are refined.

← Return to main appendix