Table of Contents
There is no single price. An industrial AI project's cost is mostly people's time: scoping, preparing data, integrating with plant systems, and keeping the model accurate after go-live. Hardware, cloud fees and licences are often the smaller part. Budget in phases, each ending in a go/no-go decision, and price three years of running costs before you start.
Key takeaways
- The model code is a small part of a production machine learning system; data collection, data verification, serving infrastructure and monitoring make up most of it (Google Cloud).
- Data scientists in Anaconda's 2020 survey reported spending about 45% of their time loading and cleaning data (InfoWorld). Budget data preparation as its own line.
- Compute for a pilot is often cheap. Owning an on-premise LLM server costs about $48,000 in hardware before customs, shipping and VAT (Thunder Compute, Exxact).
- Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, partly because of escalating costs (Intelligent CIO).
- In Algeria, Law 23-12 requires the total need to be set before procurement and forbids splitting it, but allows firm and conditional tranches.
What are you actually paying for?
A vendor quote usually shows licences and hardware. The full cost has more lines. Google Cloud's MLOps guide lists configuration, automation, data collection, data verification, testing and debugging, resource management, model analysis, metadata management, serving infrastructure and monitoring around the ML code, and states that "the real challenge isn't building an ML model, the challenge is building an integrated ML system" and operating it continuously in production (Google Cloud).
| Cost line | What it covers | What drives it | Public reference price (November 2026) |
|---|---|---|---|
| Scoping and readiness | Use case choice, data audit, baseline, business case | Number of sites and systems involved | Priced in person-days |
| Data preparation | Extraction, cleaning, labelling, time alignment | Data quality, gaps in maintenance or lab records, scanned documents | Priced in person-days |
| Sensors and data acquisition | Instruments, gateways, network, where data does not exist yet | Number of assets | Wireless triaxial vibration sensor listed at $399.95 (Erbessd) |
| Model development | Features, training, testing against a baseline | How new the problem is | Rented GPU at a median $3.03 per hour (Thunder Compute) |
| LLM compute | API fees or on-premise GPUs | Users, tokens per question, data rules | $2 input / $10 output per million tokens for gpt-6.1-sol (OpenAI) |
| Integration | Historian, SCADA, ERP, document systems, single sign-on | Number of systems, OT security rules | Priced in person-days |
| Adoption | Training operators and engineers, procedures | Number of users and shifts | Priced in person-days |
| Operation | Monitoring, retraining, support, hardware warranty | Process changes, model updates | Annual |
Half of these lines have no list price. They are your staff's time and the vendor's person-days, and they are where estimates go wrong. In Anaconda's 2020 survey, data scientists reported spending 19% of their time loading data and 26% cleaning it, against about 11% each for model selection, training and deployment (InfoWorld). These are self-reported figures from data scientists in general, not industrial projects, but industrial data rarely makes this work shorter.
Which phases should the budget follow?
Split the budget so that spending grows only with evidence:
- Scoping and readiness. Choose one use case, measure the current baseline, and check that the data exists. Our AI readiness assessment guide lists what to check.
- Proof of concept on historical data. Work offline to answer one question: does the data contain the signal? The output is a metric compared with the baseline.
- Pilot in real conditions. Connect to live data and real users, without letting the model act alone. Agree the pass mark in writing before it starts.
- Industrialisation and rollout. Integration, security review, production hardware, training and documentation. Most of the hardware and integration lines fall here.
- Operation. Monitoring, retraining and support, budgeted per year for at least three years.
Each phase ends with a go/no-go decision. This matters because many projects stop after the proof of concept. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing "poor data quality, inadequate risk controls, escalating costs or unclear business value". Its analyst Rita Sallam added that GenAI "costs aren't as predictable as other technologies" (Intelligent CIO). We look at the other causes of failure in why industrial AI projects fail.
How do costs differ by type of project?
| Project type | Main cost driver | Hardware | Common surprise |
|---|---|---|---|
| Document assistant (LLM with retrieval) | Cleaning documents, access rights, testing answers | GPU server on site, or API fees | Scanned and outdated documents, not compute |
| Predictive maintenance | Sensors where none exist; failure history | Sensors, gateways, standard servers | Too few recorded failures to learn from |
| Soft sensor or process model | Aligning process and lab data | Existing historian and a server | Lab timestamps that do not match sampling time |
| Visual quality inspection | Image capture and labelling of defects | Cameras, lighting, edge computer | Rare defects mean slow labelling |
| Formulation optimisation | Structuring past formulations and test results | Server | Knowledge spread across files and notebooks |
For a document assistant, see our guide to RAG on maintenance manuals. For process models, soft sensors explains why lab data alignment takes so much effort.
What do real prices look like? Three worked examples
The examples below use published prices checked in November 2026 and stated assumptions. They show orders of magnitude, not a quote.
1. API fees for a document assistant pilot. Assume 50 users asking 20 questions per working day over 22 days: 22,000 questions per month. With 6,000 input and 500 output tokens per question, the same assumptions as our on-premise vs cloud cost comparison, gpt-6.1-sol at $2 / $10 per million tokens (OpenAI) costs $0.017 per question, or about $374 per month. Indexing 10,000 pages at an assumed 500 tokens per page with text-embedding-3-small, at $0.02 per million tokens, costs about $0.10. Compute is not what makes this pilot expensive. If your documents cannot leave the site, the API is not an option anyway.
2. Hardware for an on-premise LLM. OpenAI's open-weight gpt-oss-120b is released under Apache 2.0 and is documented to run on a single 80 GB GPU (Hugging Face). Two 96 GB RTX PRO 6000 Blackwell cards at NVIDIA's $16,000 marketplace price (Thunder Compute) plus a reference 4U host server listed at $16,005 without GPUs, with a three-year parts and labour warranty (Exxact), come to about $48,000. Add 19% VAT if you cannot recover it, plus customs duties and shipping. Our GPU sizing guide explains how to size the server for your users.
3. Sensors for predictive maintenance on 12 machines. One manufacturer lists a 12-unit kit of wireless triaxial vibration sensors at $5,199.95, about $433 per sensor, and a higher-end triaxial wireless unit at $3,699.95 (Erbessd). Both were shown as sold out when we checked. Hardware is the easy line to price. Installation, network work and the months of data collection before a model has enough to learn from are not in the list price.
Why do running costs matter more than the build?
A model that worked at handover degrades when feedstock, equipment or operating modes change. Google researchers wrote that "it is common to incur massive ongoing maintenance costs in real-world ML systems" (Sculley et al., 2015). Each year, budget for:
- monitoring of model accuracy against a reference, such as lab results or confirmed failures;
- retraining after process changes, new product grades or new equipment;
- software and model updates, and the testing they require;
- hardware support after the warranty ends;
- a named owner on your side. Our cost comparison assumed a quarter of an engineer's time for one LLM server.
Ask every vendor for a three-to-five-year price schedule, not only a first-year price. Our vendor selection checklist includes this question.
Which costs are specific to Algeria?
Imported hardware. The standard VAT rate is 19%, and VAT is collected on imports as well as on business transactions (Grant Thornton). Customs duties, shipping, delivery lead times and foreign-currency payment come on top. Ask your customs broker and finance department before fixing the budget.
Personal data. Law 18-07, amended by Law 25-11 of 24 July 2025, requires controllers to appoint a data protection officer and subjects transfers of personal data abroad to prior authorisation from the national authority, the ANPDP, save for exceptions in the law (Gide). Maintenance reports and incident files often name employees, so a foreign API can add compliance work. Confirm your case with counsel.
Public procurement. Law 23-12 of 5 August 2023 shapes how public buyers can phase a project:
- Article 16: needs are determined before any procedure starts, the total amount sets which procurement committee is competent, and splitting needs to escape tender procedures or controls is forbidden.
- Article 30: a contract may have a firm tranche and one or more conditional tranches when economic or financial conditions justify it. A pilot followed by a conditional rollout fits this structure.
- Article 13: public economic enterprises are not bound by the law's award procedures, but must adopt their own, based on its principles.
This is a summary, not legal advice. Check the current texts with your procurement department and counsel.
How do you keep the budget under control?
- Start with one use case and one baseline metric.
- Price each phase separately, with person-days for your staff as well as the vendor's.
- Make data preparation and integration explicit lines, not part of "implementation".
- Decide early whether data may leave the site; it changes the hardware line completely.
- Link each phase's budget release to a go/no-go result.
- Turn the results into a business case with the method in how to build an AI business case.
What we have seen at KABAS AI
Our consulting service includes readiness assessment and use case identification. We recommend starting there, because the cheapest phase is the one that stops a weak project early.
The published ENAD Shymeca case study shows where effort goes in practice. The Virtual Lab implementation followed five steps: discovery, platform configuration, data integration with migration of historical formulation data, training and adoption, and continuous optimisation. Only one of them is configuring software. At ADFC, decades of formulation experience existed but were spread across formats and locations, and structuring that knowledge was part of the work. When you budget, give data integration and adoption their own lines.
For document assistants, DALILE runs on your own infrastructure, which moves the cost from token fees to hardware and operation. If you want a phased estimate for your own use case, contact us.
Sources
Prices checked in November 2026. They change often; recheck them before you decide.
- Google Cloud, MLOps: Continuous delivery and automation pipelines in machine learning
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015
- InfoWorld, For data scientists, drudgery is still job one (Anaconda 2020 State of Data Science survey)
- Intelligent CIO, Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025
- OpenAI API pricing
- Hugging Face, openai/gpt-oss-120b model card
- Thunder Compute, NVIDIA RTX PRO 6000 Blackwell pricing
- Exxact, TensorEX 4U configurator, TWS-173594376
- Erbessd Instruments store, wireless vibration sensors
- Grant Thornton, Indirect tax guide: Algeria
- Gide, Protection of personal data: focus on Algerian regulations
- Law 23-12 of 5 August 2023 on public procurement, Journal officiel no. 51 (FAOLEX copy)
Frequently asked questions
How much does an industrial AI proof of concept cost?
Mostly the time of the people involved: your engineers, your IT and OT staff, and the vendor's team. Compute is often small at this stage; in our example, 50 pilot users of a document assistant cost about $374 per month in API fees. Ask vendors to price each phase in person-days and agree the pass mark before the work starts.
Do we need to buy GPUs for an industrial AI project?
Not always. Many models on sensor or process data run on ordinary servers. A large language model running on your own site does need GPUs: OpenAI's open-weight gpt-oss-120b is documented to fit on a single 80 GB GPU, and two 96 GB cards cost about $32,000 at NVIDIA's November 2026 marketplace price.
What are the running costs of an AI system after go-live?
Budget every year for monitoring, retraining when the process or data changes, software and model updates, hardware support, and the staff who own the system. Google researchers described the maintenance costs of real-world machine learning systems as massive and ongoing, so price at least three years of operation before you approve the build.
Is it cheaper to buy an AI product or build a custom model?
A product spreads its development cost across customers, so you mainly pay for licences, integration, data preparation and training. A custom model fits unusual problems but you pay for all the development and maintenance yourself. Compare both on a three-year total cost, including your own staff time.
How should an Algerian public company budget a phased AI project?
Law 23-12 on public procurement requires needs to be set before the procedure starts and forbids splitting them to avoid tender procedures. It allows contracts with a firm tranche and one or more conditional tranches, which suits a pilot followed by a rollout. Public economic enterprises apply their own procedures based on the law's principles. Confirm your case with your procurement department and counsel.



