Table of Contents
Before a public body or state-owned company tenders for on-premise AI, it should confirm which procurement text applies, choose lots and tranches that match the project's phases, specify hardware by performance rather than brand, require offline operation with evidence, define provisional and final acceptance tests, and price at least three years of support and exit.
Key takeaways
- Law 23-12 governs procurement by the State, local authorities and public-law establishments. Public economic enterprises adopt their own procedures based on its principles (Article 13). Confirm which applies to you with counsel.
- An on-premise AI purchase mixes supplies and services. Under Article 26, when the supplies are worth more than the services, the contract counts as a supply contract.
- Specify GPU memory, users and response time, not a brand: Article 16 forbids specifications oriented towards a product or supplier.
- Ask for evidence that the system runs without internet, plus a software bill of materials and the origin of every model, as recommended by joint guidance led by CISA.
- Use provisional and final reception (Article 86) as two separate tests, and give support years their own price lines.
Which procurement rules apply to your company?
This decides how much of the rest is a legal obligation. The table summarises Law 23-12 of 5 August 2023 (FAOLEX copy). It is not legal advice; confirm with your procurement department and counsel.
| Buyer | Rule under Law 23-12 |
|---|---|
| State institutions and administrations, local authorities, public establishments under public law | The law applies in full (Article 9) |
| Public establishments under commercial rules | The law applies to operations financed directly, wholly or partly, from the State or local budget (Article 9); otherwise they adopt internal procedures derived from it (Article 12) |
| Public economic enterprises, defined as commercial companies majority-owned, directly or indirectly, by the State or another public body (Article 4) | Not bound by its award procedures, but must adopt their own, based on free access, equal treatment and transparency (Articles 5 and 13) |
| Any other body using public funds | Must adopt procedures based on the same principles (Article 14) |
The rest of this checklist cites Law 23-12, since enterprise procedures are based on its principles; yours may differ in detail. Our RFP guide covers the problem statement, data annex and scoring grid, and the sovereign AI on-premise guide the technical choices behind the specifications.
How should the contract be structured?
On-premise AI combines GPU servers, licences, installation, integration, training and years of support.
| Decision | What Law 23-12 says | What it means for on-premise AI |
|---|---|---|
| Type of contract | Supplies and services together count as a supply contract when the supplies are worth more (Article 26) | A hardware-heavy purchase may be a supply contract even with substantial services, and general administrative clauses differ by type of contract (Article 17) |
| Lots | Separate lots, evaluated lot by lot, or a single lot when services cannot be separated; the choice is the buyer's (Article 29) | Separate hardware and software lots widen competition but split responsibility for performance |
| Phasing | A firm tranche and one or more conditional tranches (Article 30); splitting needs to avoid tender procedures is forbidden (Article 16) | Pilot server and pilot in the firm tranche, rollout as a conditional tranche |
| Price | Lump sum, unit price, mixed or controlled expenses; firm or revisable (Articles 73 and 74) | Hardware is easy to price as a lump sum; support and extra users suit unit prices |
| Subcontracting | At most 40% of the contract; foreign companies bidding alone must, unless shown to be impossible, subcontract at least 30% to companies under Algerian law (Article 82) | Ask who installs, who supports and who supplies the hardware, by name |
| National preference | National call when national production can meet the need (Article 59); preference margin for products of Algerian origin and companies under Algerian law majority-owned by resident nationals (Article 62) | Check with procurement how these apply to imported GPUs and local integration |
| Disputes | Settled under Algerian law, with an amicable settlement clause in the specifications (Article 87) | Matters most for foreign licensors and hardware makers |
Also note Article 71: a contract holder with information that could advantage it in another tender may take part only if it proves this does not distort competition. Our comparison of local and international AI vendors covers the preference rules in more detail.
What should the hardware clauses cover?
- Performance, not brand. State the model size to run, GPU memory, concurrent users and target response time. For scale, NVIDIA lists 80 GB of memory and up to 700 W for the H100 SXM, and 94 GB at 350 to 400 W for the H100 NVL (NVIDIA). Our GPU sizing guide shows how to turn users into memory.
- Site readiness. One reference 4U GPU server is listed with two redundant 3,000 W power supplies at 200 to 240 V (Exxact). Ask bidders for power, cooling and rack requirements, and check them against the room before award.
- Warranty and repair. The same server is listed with a three-year parts and labour warranty. Ask who repairs it on site in Algeria, and how long a failed GPU takes to replace.
- Import costs. Algeria's standard VAT rate is 19% and VAT is collected on imports (Grant Thornton). Ask who imports, who clears customs and who bears delivery delays.
- A test environment. Budget a smaller second environment for testing updates.
What should the software and model licences say?
- Name every model and its licence. Licences differ. gpt-oss-120b is published under Apache 2.0 and fits on a single 80 GB GPU. The Llama 3.1 licence requires you to display "Built with Llama" and to start the name of any distributed model built from it with "Llama", and requires a separate licence above 700 million monthly active users.
- Keep a licence inventory. OWASP warns that AI work combines many software and dataset licences and recommends an inventory of all licence types, with regular audits (OWASP LLM03:2025).
- Licence term. Perpetual or annual? If an annual subscription is not renewed in time, does the system stop, and does licence activation need an internet connection?
How do you make sure it really runs offline?
The CISA-led guidance tells operators to ask whether vendors "can operate on premises or without continuous access to the internet/vendor cloud" (CISA et al.). Put the answer to the test.
- List every outbound connection. Open-source components can call out by default. vLLM "collects anonymous usage data by default", and setting
VLLM_NO_USAGE_STATS=1turns this off (vLLM). Hugging Face Transformers needsHF_HUB_OFFLINE=1to stop HTTP calls to the Hub when loading a model (Hugging Face). - Updates. How do new versions arrive: removable media, a one-way transfer, a controlled connection? Who checks them? See air-gapped AI deployment.
- Remote support. Decide whether it is allowed at all. If it is, the same guidance says new connection paths such as remote access should be approved and pass through your existing DMZ or jump host.
What security evidence should bidders provide?
- SBOM and model origin. CISA's guidance asks vendors for a software bill of materials and to "clarify the supply chain of AI models used in their products (e.g., where models are hosted)".
- Model files. OWASP recommends using models only from verifiable sources, with integrity checks using signing and file hashes. Prefer safetensors: Hugging Face warns that loading pickle files can lead to arbitrary code execution (Hugging Face).
- Operator control. The CISA-led guidance recommends that the operator, not the vendor, controls when AI features are enabled or disabled.
- Industrial security practice. If the system touches plant networks, ask which parts of ISA/IEC 62443 the vendor follows; the series includes requirements for IACS service providers (62443-2-4) and for secure product development (62443-4-1).
- Personal data. Law 18-07, amended by Law 25-11, requires the national authority's authorisation for transfers of personal data abroad and adds impact assessments for high-risk processing (CMS). Ask counsel how logs and support access fit these rules. Our questions on whether a vendor trains on your data apply here too.
How should acceptance and reception work?
Article 86 of Law 23-12 provides for provisional and/or final reception. Use the two steps for different questions.
- Write the tests into the specifications. The metric, the pass mark and test data the vendor has never seen.
- Test first outside production. The CISA-led guidance recommends testing AI systems on infrastructure designed for testing before moving them to production.
- Provisional reception. Check the delivered items against the contract, run the test, and run it again with outbound internet blocked.
- Operating period. Real users, real documents, with a log of incidents and response times.
- Final reception. Confirm stable performance, completed training and handed-over documentation.
Define performance in measurable terms, so that late or non-conforming delivery can trigger the financial penalties provided by Article 84 without debate.
What about training, support and exit?
- Know-how transfer. Article 60 requires the specifications to provide for training and the transfer of technology and know-how related to the contract. Name who is trained, on what, and how you check it.
- Support years. Price each year of support separately for at least three years. Our vendor selection checklist lists the support questions to ask.
- Exit. Require export of your document indexes, configurations and any fine-tuned models in open formats, and full documentation, so another contractor can take over.
What does the checklist look like on one page?
| Area | Ask in the specifications | Check at reception |
|---|---|---|
| Rules | Which procurement text applies | Not applicable |
| Structure | Lots, tranches, price form, subcontractors by name | Subcontractors match the offer |
| Hardware | Performance, power, warranty, repair in Algeria | Delivered configuration and power draw |
| Licences | Every model and component, with licence and term | Licence inventory complete |
| Offline | List of outbound connections, update method | Test run with internet blocked |
| Security | SBOM, model origin, file format, operator control | Hashes, SBOM and settings match |
| Acceptance | Metric, pass mark, unseen test data | Provisional and final test results |
| Exit | Export formats, documentation, training plan | Export tested once |
What we have seen at KABAS AI
We are a local vendor that sells on-premise AI, so we answer these clauses as a bidder; put us through the same checklist.
DALILE is deployed fully on premises and is designed to work without an internet connection; run the offline test on it all the same. The ENAD Shymeca case study describes the first Open Innovation contract between an AI startup and a national Algerian industrial company, a contract form worth raising with your procurement team. The ADFC case study describes a deployment that keeps data control within Sonatrach's IT environment, the kind of requirement that belongs in the specifications from the start.
One practical point from Article 71: if a vendor, including us, helps you scope the project under a contract, decide early whether it may bid on the follow-on tender, and how you will show that the specifications treat all bidders equally.
Our consulting team can help define the use case and technical requirements. To discuss a tender you are preparing, contact us.
Sources
- Law 23-12 of 5 August 2023 on public procurement, Journal officiel no. 51 (FAOLEX copy)
- CISA, ASD's ACSC et al., Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- OWASP, LLM03:2025 Supply Chain
- Hugging Face, Pickle scanning
- Hugging Face Transformers, Installation: offline mode
- vLLM documentation, Usage stats collection
- Hugging Face, openai/gpt-oss-120b model card
- Meta, Llama 3.1 Community License
- NVIDIA H100 Tensor Core GPU
- Exxact, TensorEX 4U configurator, TWS-173594376
- Grant Thornton, Indirect tax guide: Algeria
- ISA, ISA/IEC 62443 series of standards
- CMS Expert Guide to data protection and cyber security laws: Algeria
Frequently asked questions
Does Law 23-12 apply to public economic enterprises?
Not its award procedures. Article 13 says public economic enterprises, the commercial companies majority-owned by the State or another public body, must adopt their own procurement procedures based on the law's principles of free access, equal treatment and transparency. The exception is when they act as delegated project owner for the State or a local authority. Confirm your case with your procurement department and counsel.
Should we buy the GPU servers and the AI software in one lot or separate lots?
Article 29 of Law 23-12 allows either, and the choice belongs to the buyer. Separate lots open the hardware to more bidders but split responsibility when the system is too slow or fails its acceptance test. A single lot makes one contractor answer for the whole result. If you separate them, write the hardware lot from the software bidder's sizing and test both together at reception.
Can we name a specific GPU or server brand in the tender?
For buyers under Law 23-12, Article 16 requires specifications based on standards, performance or functional requirements that are not oriented towards a given product or supplier. Describe the need instead: GPU memory, the model size to run, concurrent users, response time and power limits. If you need compatibility with existing equipment, ask counsel how to word that requirement.
How can we buy hardware now and support over several years?
Law 23-12 offers several structures: a contract with a firm tranche and conditional tranches (Article 30), a contrat-programme spanning two or more budget years through application contracts (Article 32), and a marché à commandes for routine, repetitive services, lasting one renewable year (Article 33). Which one fits depends on your budget rules, so decide with your procurement and finance departments.
What should provisional acceptance of an on-premise AI system include?
At least a check that the delivered hardware and software match the contract, a test on questions or data the vendor has never seen against a pass mark written in the specifications, and proof that the system runs with its outbound internet access blocked. Leave real-use criteria, such as stability and performance with actual users, for final acceptance.
Does running AI on our own servers remove data protection obligations?
No. Law 18-07 on personal data, amended by Law 25-11, applies to processing on your own servers too, including logs of questions that name employees. On-premise hosting mainly avoids transfers abroad, which need authorisation from the national authority. Ask counsel whether remote support access by a foreign vendor raises the same question.



