AI Implementation Cost in Iran (2026): Toman Estimates and a Budget Guide

There is no single figure for the cost of implementing AI in an Iranian company, but realistic ranges can be built in toman from prices that are public in the Iranian market. A supplier who names one firm number before seeing your workflow, your data, and your systems is guessing; engineers' salaries, server rents, and token prices in Iran, on the other hand, are published and can be cited. This guide is written for the managing director, CFO, or CTO who has to put an AI line into the 1405 budget and defend it later. It gives you a toman estimate table for five common options, with every figure derived from dated, sourced inputs; it explains the eight cost lines; it provides a worksheet to fill with your own figures; and it shows how to judge whether a pilot has earned the next tranche of spending.

The central recommendation is simple. Do not budget a programme from a vendor headline. Fund a small, measured pilot, and let every later stage be justified by what that pilot measured. In the Iranian market the largest line in that pilot is usually people's time, not the model bill.

Commercial-interest disclosure: ZharfAI, the publisher of this guide, sells AI consulting and implementation. We therefore benefit when companies decide to spend on AI. Apply every test in this article to our proposals as strictly as to anyone else's. We quote no ZharfAI prices here, and we make no claim about the return any specific company will achieve.

Why the honest answer is a range, not a price

AI spending is harder to forecast than a conventional software licence because the expensive parts depend on things nobody has measured yet: how clean your data is, how many systems the output must reach, how often a person has to correct the result, and how much usage grows once people trust it. Gartner made the point bluntly in July 2024, when it 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. The same release quotes its analyst saying that costs "aren't as predictable as other technologies."

That release also published typical cost ranges for five ways of deploying generative AI, for large organisations outside Iran. Its dollar figures do not transfer to an Iranian company's budget, so we do not reproduce them here, but the ratios between the rows do transfer. Moving from a ready tool to retrieval over your own documents (RAG) multiplies the upfront cost by roughly four to ten, and moving from retrieval to fine-tuning multiplies it again by roughly five to nine. Recurring cost per user in the fine-tuning row is roughly fifteen to forty times the ready-tool row, so a design that looks affordable with ten pilot users can become expensive with a thousand. Most first projects we see in Iranian organisations, such as document extraction, internal knowledge search, or a Persian assistant for staff, belong in rows 2 and 3 of the toman table below. Fine-tuning or training a model is rarely the right first step.

AI implementation cost in toman: a Mehr 1405 estimate

The table below is a dated estimate for a small or mid-sized Iranian company, built from public Iranian market prices on 3 Mehr 1405 (2026-09-25). All figures are in million toman; they move with inflation and the exchange rate, and they do not replace a real quote. Each range comes from the inputs listed under the table, so you can repeat the arithmetic with your own figures. USD equivalents in parentheses use the free-market rate of 234,700 toman per dollar on the same date.

Dated estimate, Mehr 1405; figures in million toman

The monthly column excludes the time your staff spend reviewing outputs, value-added tax, and contingency; the worksheet further down this page counts those separately. If you hand the build to an outside partner, expect their price to sit above the person-month cost in this table, because it also covers their overhead, margin, and risk.

Inputs to the estimate

1. People. The Jobvision 1405 salary report, published in Ordibehesht 1405 (April 2026) from the responses of more than 100,000 people, gives the middle half of requested monthly net salaries for 1405 as follows: a data science and AI specialist in Tehran, 37 to 71 million toman (median 50; about USD 160 to 300); a senior specialist in the same group nationwide, 62 to 94 million toman (median 73; about USD 260 to 400); and a mid-level Python developer in Tehran, 45 to 78 million toman (median 58; about USD 190 to 330). These are net salaries. To reach the employer's full cost (employer insurance share, tax, year-end bonus, severance accrual, and equipment) we multiplied them by 1.4. That multiplier is our assumption, not a sourced figure, and you should replace it with your own payroll number. The result: a mid-level engineer-month costs about 50 to 100 million toman (USD 210 to 430) and a senior specialist-month about 90 to 130 million toman (USD 380 to 550).
2. Ready-made Persian chatbots. IlaChat's pricing page showed monthly plans from 1,100,000 toman (USD 4.7) for 1,000 AI messages to 7,700,000 toman (USD 33) for 10,000 messages, and MuChat's pricing page showed plans of 2 to 10 million toman a month (USD 8.5 to 43) for 1,000 to 10,000 messages. MuChat states that its messages are answered by OpenAI's 4o-mini model, so at least in this example the "buy" option also depends on a foreign model.
3. Tokens on a domestic API. ArvanCloud's AI service price table listed these prices in toman per million tokens: GPT-5 mini 55,200 input and 441,600 output (USD 0.24 and 1.88); GPT-5 276,000 input and 2,208,000 output (USD 1.18 and 9.41); the open-weight GPT-OSS-120b 22,080 input and 110,400 output; and the BGE-M3 embedding model just 2,208 toman. ArvanCloud charges pay-as-you-go with no fixed fee; if you choose a gateway with a monthly subscription, add that too.
4. An ordinary server for the application and database. An ArvanCloud cloud server with 2 cores and 4 GB of memory costs 1,972,800 toman a month (USD 8.4), and one with 4 cores and 8 GB costs 4,161,600 toman (USD 17.7).
5. GPU servers inside Iran. IranServer's GPU server page, with servers located in Shiraz or Tabriz, showed these monthly rents: an RTX 3090 with 24 GB of GPU memory at 40.5 million toman (USD 173), an RTX 4090 with 24 GB at 86 million (USD 366), an RTX A6000 with 48 GB at 117 million (USD 499), and an L40S with 48 GB at 182 million toman (USD 775). ArvanCloud's AI server offers one A30 with 24 GB for 273.89 million toman a month (USD 1,167) and says hosting can be in a domestic or a foreign data centre.
6. Exchange rate. Alanchand showed a free-market selling rate of 234,700 toman per US dollar on the morning of Friday 3 Mehr 1405. We use it only to translate dollar-priced items into toman and to give the USD equivalents in this article.

All prices in inputs 2 to 6 were read from each provider's public page on 2026-09-25. Naming these providers is neither a recommendation nor a quality comparison; each one is simply a dated public price.

How each row is calculated

Each line below is the multiplication and addition behind the table, and the table figures are these results rounded. All figures are in million toman.

• Row 1: one-time equals 0.2 to 0.5 person-months to prepare questions and documents and test the answers, times 50 to 100, which gives 10 to 50. Monthly equals the subscription of 1.1 to 10 (input 2) plus 0.05 to 0.1 person-months to maintain the knowledge base (2.5 to 10), which gives about 4 to 20.
• Row 2: one-time equals 2 to 3 engineer-months for the build, a light integration, and a Persian evaluation set, plus 0.5 to 1 person-month of a domain expert writing reference answers; 2.5 to 4 person-months in total, times 50 to 100, which gives 125 to 400. Monthly equals tokens of 3.9 to 19.3 (worked example below) plus one server at 2 to 4.2 (input 4) and 0.1 to 0.25 person-months of support (5 to 25), which gives about 11 to 48, rounded to 10 to 50.
• Row 3: one-time equals two engineers for 3 to 4 months (6 to 8 person-months times 50 to 100) plus 1 to 2 person-months of a senior specialist for security, access control, and evaluation (times 90 to 130), which gives 390 to 1,060. Monthly equals tokens of 12 to 60 plus one or two 4-core servers for the application and vector database (4.2 to 8.3) and 0.25 to 0.5 person-months of operations (12.5 to 50), which gives about 30 to 120.
• Row 4: one-time equals the row 3 build plus 1 to 2 senior person-months to deploy and tune the model server (90 to 260), which gives 480 to 1,320. Monthly equals one GPU at IranServer, from an RTX 4090 to an L40S (86 to 182), plus the application server (4.2 to 8.3) and 0.5 to 1 person-month of operations (25 to 100), which gives about 120 to 290. The RTX 3090 at 40.5 has the same 24 GB of memory and is enough for this volume, but it is an older, slower generation; we used the RTX 4090 for speed headroom, and with an RTX 3090 the monthly floor of this row drops to about 70 (USD 300). High availability means paying for two GPUs.
• Row 5: we found no public Iranian price for fine-tuning, so this row is built only from the inputs and is a realistic floor, not a ceiling. One-time equals the row 4 build (480 to 1,320) plus 2 to 4 senior person-months (180 to 520), 2 to 4 person-months of domain experts labelling data (100 to 400), and 1 to 2 months of a 48 GB GPU for training (117 to 364), which gives about 900 to 2,600. Monthly equals row 4 plus 0.25 to 0.5 senior person-months for evaluation and retraining (22 to 65), which gives about 140 to 360.

Worked example: the token bill

Suppose the pilot processes 5,000 invoices a month and each task uses about 6,000 input tokens (the document text and instructions) and 1,000 output tokens: 30 million input tokens and 5 million output tokens a month. With GPT-5 mini on ArvanCloud the cost is 30 × 55,200 plus 5 × 441,600, about 3.9 million toman (USD 16). With GPT-5 the same work costs 30 × 276,000 plus 5 × 2,208,000, about 19.3 million toman (USD 82). For RAG with 20,000 questions a month, each using 6,000 input and 600 output tokens, one question costs about 600 toman with GPT-5 mini and about 3,000 toman with GPT-5, or 12 to 60 million toman a month (USD 51 to 256). Building a vector index for ten thousand documents of two thousand tokens each with BGE-M3, 20 million tokens in all, costs about 44 thousand toman. The lesson is plain: at these volumes the model bill is the smallest line, and the largest line is people's time. Put your own volumes in place of these assumptions to rescale the figures.

Why the cost structure in Iran is different

In Iran, people are relatively cheap in toman and anything tied to the dollar is expensive. The median requested net salary of an AI specialist in Tehran, 50 million toman, is under USD 220 a month at about 235,000 toman per dollar, while renting one RTX 4090 in Shiraz at 86 million toman comes to about USD 370 a month. A rented GPU costs more than the engineer who works with it. Token prices are dollar prices even when they are quoted in toman: ArvanCloud's toman price for GPT-5 mini is exactly 220,800 times the dollar price Liara lists for the same model, a rate below today's market that will rise if the list is reset at the day's rate. Buying hardware is priced in dollars plus the cost of importing under sanctions: in a Torob search on the same date, RTX 4090 listings ranged from 749 million to 1.3 billion toman (USD 3,200 to 5,500). Meanwhile the dollar rose about 47% in six months according to the same Alanchand page, and in the Jobvision report the median requested salary increase for 1405 in the job groups used here is 42% to 56%. Sanctions also create risk costs: the largest foreign providers do not support Iran, and every foreign dependency needs a costed fallback path.

The eight cost lines of an AI implementation

Every proposal, internal or external, should be broken into the same lines so that you can compare them. We use eight, plus contingency. In the toman table above, most of the one-time column is people's time spent on these lines.

The line most often underfunded is the first one. Gartner's February 2025 research on AI-ready data found that 63% of organisations either do not have, or are unsure whether they have, the right data management practices for AI, in a survey of 1,203 data management leaders conducted in July 2024. Gartner predicts that through 2026 organisations will abandon 60% of AI projects that are not supported by AI-ready data. In practice this means the data line is not a one-off tidy-up before the "real" work. It is often the real work.

Integration deserves the same honesty. Connecting to an Iranian accounting package or ERP can mean a documented interface, direct database access, a scheduled file export, or, in the worst case, nothing but the user screen. Reading data is far cheaper than writing it back, because every write into a financial or operational record needs validation, an audit trail, and a tested way to reverse it. Our guide to AI in accounting and financial document automation walks through that boundary for finance teams.

Model access in Iran: domestic API, self-hosted GPU, or foreign API

For an Iranian company, model access is not only a price question. It is a continuity question, and the cheapest option on paper can be the most expensive one to lose.

Foreign APIs. The largest foreign providers do not list Iran as a supported country. OpenAI's supported countries and territories page states that accessing or offering access to its services outside the listed countries may result in an account being blocked or suspended, and Iran is not on the list. Anthropic's supported regions list does not include Iran either. We do not advise on ways around these terms. For budgeting, the consequence is concrete: any design that depends on a foreign model needs a costed fallback path, and payment continuity has to be treated as a risk, not an assumption.

Domestic gateways. Several Iranian platforms sell model access in toman, but in every example we checked the underlying price is in dollars. As dated examples, not recommendations: Liara's AI service page displayed two plans on 2026-09-25, 825,000 toman per month (USD 3.5) or 1,145 toman per hour with access to 66 language models, and 1,475,000 toman per month (USD 6.3) or 2,048 toman per hour with 113 models. On both plans, token costs are charged separately "according to the model provider's tariff," and the model price list on the same page is in US dollars per million tokens, for example USD 0.25 input and USD 2 output for GPT-5 Mini. ArvanCloud quotes the same kind of price directly in toman (input 3), and AvalAI's pricing documentation states that per-million-token prices are in dollars and that the toman charge is the dollar cost multiplied by the Tether-to-toman rate. The budgeting lesson is general: the platform fee may be small and local, but the usage line still moves with the exchange rate. The model lists on these gateways also include providers that do not list Iran as supported, so ask any gateway how it secures that upstream access and what happens to your service if it stops.

Self-hosted open models. Running an open-weight model on GPUs in an Iranian data centre, or in your own server room, removes per-token billing and keeps data inside your perimeter. It replaces a variable bill with a fixed one. A transparent way to estimate it is:

• monthly fixed cost = GPU lease (or purchase price divided by useful months) + hosting and power + storage and backup + operations hours × loaded hourly cost;
• cost per completed task = monthly fixed cost ÷ tasks completed in the month;
• break-even against an API = the monthly volume at which tasks × average tokens per task × price per token equals the monthly fixed cost.

The inputs above put numbers into this formula. Renting an RTX 4090 from IranServer costs 86 million toman a month. If one RAG question costs about 600 toman in tokens with GPT-5 mini, a dedicated server only breaks even on token cost at about 144 thousand questions a month; against GPT-5 at about 3,000 toman per question, the figure is about 29 thousand questions. Even that comparison ignores the server's operating time and the quality gap between a small open model and a large commercial one. The sound reason for a domestic GPU is usually keeping data inside the perimeter, or very high volume, not low cost. Buying a card instead of renting is possible too: with Torob listings for the RTX 4090 at 749 million to 1.3 billion toman, the purchase price equals about 9 to 15 months of rent, before the host server, power, data-centre space, and maintenance.

The inputs to that formula change quickly. The Stanford AI Index 2025 reports that the inference cost of a system performing at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024, while hardware costs declined about 30% a year and energy efficiency improved about 40% a year. Those declines are measured in dollars, and in Iran a rising exchange rate can cancel them out. Treat any multi-year hardware purchase or fixed per-token commitment with care, and write a review clause into it. The comparison between local and foreign tools is covered in depth in our guide to choosing between Iranian and foreign AI tools, and the technique of sending easy tasks to cheaper models is explained in AI model routing for cost and quality.

The stages of implementation and what each one costs

The phrase "implementation stages" hides the most important budgeting fact: the cost lines do not arrive at once. Each stage has a different dominant expense and ends with a decision.

A pilot should use the cheapest integration that still produces honest evidence, often a controlled file export rather than a live connector, and it should run in assisted mode, where a person approves each output. The rollout budget is where the organisation changes, not just the software. McKinsey's state of AI survey, published in August 2026, found that nearly three quarters of the respondents it classes as AI high performers report fundamentally redesigning workflows because of AI, compared with one quarter of other respondents. If the rollout budget contains no time for redesigning the process, it is budgeting for the tool and not for the result.

What pushes the cost up, and what brings it down

The same workflow can cost several times more in one company than in another. These are the drivers we see most often.

Cost rises when:

• the output writes into a financial, legal, or customer record instead of producing a draft;
• a wrong answer has high consequences, so every output needs human review;
• inputs are scans, handwriting, stamps, or dozens of different supplier formats;
• the data must stay on premises and the organisation has no suitable hardware;
• the service must answer in real time, at high volume, across several channels;
• there is no single owner who can approve acceptance criteria and data access;
• the plan includes fine-tuning or training a model before a simpler design has been tried.

Cost falls when:

• the first scope is one workflow, one user group, and one document type;
• the system reads from existing systems and hands its result to a person;
• clean exports or a documented interface already exist;
• a ready model is combined with retrieval over approved documents;
• simple, high-volume tasks go to a smaller model and only hard cases reach a large one;
• batch processing overnight replaces real-time calls where users can wait;
• the Persian evaluation set built for the pilot is reused for every later change.

Hidden costs and common failure modes

The gap between the quote and the final bill usually comes from costs that nobody wrote down. The failure patterns behind them are well documented. RAND's 2024 study, The Root Causes of Failure for Artificial Intelligence Projects, interviewed 65 experienced data scientists and engineers and identified five leading causes: stakeholders misunderstanding or miscommunicating the problem, lacking the data needed to train an effective model, focusing on the latest technology instead of the users' real problem, lacking the infrastructure to manage data and deploy models, and applying AI to problems that are too difficult for it. RAND also notes that by some estimates more than 80% of AI projects fail, about twice the rate of IT projects without AI. That figure is an estimate RAND cites from other work, not its own measurement, but its five causes map directly onto budget lines.

Hidden costs to put on the worksheet explicitly:

• Human review that never ends. If a person checks 20% of outputs for the life of the service, that time is a permanent operating cost.
• Exceptions. The cases the system cannot handle still need a person, a queue, and a response time.
• Re-evaluation. Every upstream model update, prompt change, or new document type needs the Persian test set run again.
• Exchange-rate movement. Any usage priced in dollars moves with the rate. According to the Alanchand page, the free-market dollar moved between about 106 thousand and 238 thousand toman in the year to 3 Mehr 1405; budget for that range, not a point.
• Wage inflation. In the Jobvision report, the median requested salary increase for 1405 in the job groups used in this estimate is 42% to 56%; a project that crosses a fiscal year needs the pay rise in its estimate.
• Security testing. Language-model applications need testing for prompt injection, sensitive information disclosure, and excessive tool authority, as listed in the OWASP GenAI LLM Top 10 2026.
• Expert time. Accountants, lawyers, or engineers who label data and judge outputs are not free because they are internal.
• Parallel running. The old process usually runs alongside the new one for a period, which doubles part of the workload.
• Exit. Moving data, configuration, and evaluation sets to another supplier or model has a cost; price it before you sign.

Operating cost is not a theoretical worry. In McKinsey's 2026 survey, about one in five respondents said their organisation is limiting AI use because of operating costs, including token costs.

Build, buy, or partner

The build, buy, or partner decision changes which lines you pay and who carries the risk.

Buy a ready product or platform when the workflow is common and the product already fits your data rules. Upfront cost is lowest and time to value is shortest; in the Iranian market a ready-made Persian chatbot subscription costs about 1 to 10 million toman a month (row 1 of the table). But customisation, data location, and integration are limited to what the product offers, and per-user or per-message fees grow with adoption.

Build in-house when the workflow is a source of advantage, the data is highly sensitive, and you can staff data, application, and security roles for the long term. Building software has become cheaper: McKinsey's 2026 survey found that 32% of respondents say their organisation decided against buying at least one software product or feature because it could be built internally with agentic coding tools. Evaluation, security, and operations, however, do not get cheaper at the same pace.

Partner with an external team when you need speed and experience but want to keep ownership. The cost is the partner's fee plus your own owner's time. Ask the partner to state person-months and the rate for each role separately, so that you can compare them with the people input in the table above. The risk is dependence, which you manage in the contract: ownership of code, configuration, prompts, and evaluation sets, documentation, knowledge transfer, and a tested exit. Our guide on how to choose an AI company in Iran sets out how to compare partners on evidence rather than presentations.

Many organisations end up with a hybrid: a bought platform for general tasks, a partner-built service for one high-value workflow, and an internal team that owns evaluation and operations.

A budgeting worksheet you can fill in

Copy this table into a spreadsheet. Every line is in million toman, and wherever a figure appears in parentheses it is a default taken from the sourced Mehr 1405 inputs, which you should replace with your own number. The hourly cost is the person-month cost divided by about 176 working hours a month, and the day rate is the person-month cost divided by 22 working days. Where a line is not needed at a stage, write zero and say why.

Fill it three times: a low, an expected, and a high scenario. Vary volume, share of outputs needing review, exchange rate, and number of integration surprises. Given the dollar's 47% rise over the past six months, run the high scenario at an exchange rate at least 50% above today's. The spread between the low and high totals is itself useful information: a wide spread means the pilot has not yet answered enough questions to justify a large commitment.

How to read a vendor quote

A quote is comparable only when it is broken into the same lines as your worksheet. Ask every supplier to map their price onto those lines and to mark each one as included, excluded, or your responsibility.

Warning signs in a quote:

• one lump sum with no breakdown;
• a return on investment promised before anyone has seen your data;
• "unlimited" usage without a fair-use limit or an explanation of who pays for model calls;
• no evaluation plan and no acceptance criteria linked to payment milestones;
• no clause for exchange-rate movement on dollar-linked components;
• silence on ownership of code, configuration, and evaluation data, or on exit.

Questions to ask before you sign:

1. Which of the eight cost lines does this price cover, and which remain with us?
2. Is the price fixed, time and materials, per user, or per usage, and which parts are linked to foreign currency?
3. Who pays for model usage, and what happens to the price if the upstream model or API changes or becomes unavailable?
4. What Persian test set will you use, who builds it, and do we own it afterwards?
5. What is the integration method for each system, and who fixes it when it breaks?
6. How many hours of human review per month should we expect in the first quarter of operation?
7. What does support cost after go-live, and what response times does it include?
8. What will it cost us to leave, and what will we receive when we do?

How to measure the return on a pilot

A pilot earns the next budget only if it is measured against a baseline recorded before it started. Gartner's 2024 release notes that productivity gains are often difficult to translate directly into financial benefit, so write the arithmetic down in advance rather than reconstructing it afterwards. Three formulas are enough for most pilots:

• Net monthly value = hours removed × loaded hourly cost + errors avoided × average cost of an error − review hours × loaded hourly cost − monthly run cost.
• Payback period in months = one-off cost of the next stage ÷ net monthly value.
• Cost per accepted task = total monthly run cost ÷ tasks accepted without major rework.

Hours saved are not cash unless they are redeployed or avoid a hire, so record what happened to the time. Measure the review and correction effort honestly; our guide to measuring AI productivity after review and rework shows how easily that effort disappears from a report. Keep expectations grounded: in McKinsey's 2026 survey, 37% of respondents attributed at least some EBIT impact to AI, and only about 6% qualified as high performers attributing 5% or more of EBIT to it.

Two tools on this site can help you prepare. The AI pilot readiness check asks five questions about problem clarity, data, integration, access rules, and measurement, and suggests whether you are at discovery, pilot design, or controlled validation. For finance teams, the financial document automation capacity estimator estimates the monthly hours released after subtracting human review time. It reports hours, not money: multiply by your own loaded cost, and treat the result as an assumption to test, not a saving.

Decide in advance what result means go, what means revise, and what means stop. Stopping a pilot that missed its criteria is a successful use of a small budget.

A sensible next step

If you are preparing an AI line for this year's budget, start with discovery rather than a purchase: one workflow, its owner, a data sample, the systems involved, and the constraints you already know. Then fill in the toman table and the worksheet on this page with your own figures. Our AI consulting and implementation service begins at exactly that point and can end with an honest recommendation not to proceed. If you would like a scoped discussion, use the contact page, and hold our answer to the same worksheet and questions you would use for any other supplier.

Source Notes (reviewed 2026-09-25)

• People costs: Jobvision, 1405 salary report, file published in Ordibehesht 1405 (April 2026) from more than 100,000 responses; downloaded and read on 2026-09-25. Page 165 (Python, mid-level, Tehran), 166 (Python, senior, Tehran), 396 (data science and AI, specialist, Tehran), and 397 (data science and AI, senior specialist, nationwide). Figures are requested monthly net salaries for 1405 and the middle half of the distribution; the 1.4 multiplier for the employer's full cost is our assumption.
• Token prices in toman: ArvanCloud AI service price table, read on 2026-09-25. The 220,800 ratio is the table's toman price for GPT-5 mini divided by the dollar price of the same model on Liara's page.
• Cloud server: ArvanCloud cloud server, monthly prices of the basic packages on 2026-09-25.
• GPU servers: IranServer GPU servers, monthly rents and Shiraz and Tabriz locations on 2026-09-25; ArvanCloud AI server, monthly rent on the same date, with a choice of domestic or foreign data centre.
• Ready-made Persian chatbots: IlaChat pricing and MuChat pricing, monthly plans on 2026-09-25.
• Gateway billing methods: Liara AI service page, plans and the dollar model price list on 2026-09-25; AvalAI pricing documentation, the dollar-to-toman conversion at the Tether rate on the same date. These are dated examples of billing structure, not endorsements.
• GPU purchase price: Torob search for the RTX 4090, shop listing prices on 2026-09-25; these are asking prices, not transaction prices.
• Exchange rate: Alanchand, US dollar price, selling rate of 234,700 toman at 09:05 on Friday 3 Mehr 1405; the one-year low and high of 106,100 and 238,100 toman and the six-month change of 47.10% are quoted from the same page.
• Cost ratios by deployment approach and abandonment risk: Gartner, "Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025," July 29, 2024. The ratios are calculated from Figure 1 of the release; its dollar figures describe large organisations outside Iran and are deliberately not quoted. The 30% figure is a prediction.
• Data readiness: Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk," February 26, 2025. Survey of 1,203 data management leaders, July 2024; the 60% abandonment figure is a prediction through 2026.
• Enterprise returns and operating-cost constraints: McKinsey, "The state of AI in 2026: On the road to ROI," August 25, 2026. Online survey of 1,719 participants in 97 countries, May 4 to June 8, 2026, weighted by national GDP; self-reported.
• Root causes of failure: RAND, "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed," August 13, 2024. Based on interviews with 65 practitioners; the more-than-80% failure figure is an estimate RAND cites, not its own measurement.
• Falling inference and hardware costs: Stanford HAI, AI Index Report 2025. Figures cover November 2022 to October 2024.
• Foreign API availability: OpenAI supported countries and territories and Anthropic supported countries and regions, both checked 2026-09-24; Iran was not listed on either page.
• Security testing scope: OWASP GenAI LLM Top 10 2026. A threat-discovery resource, not a certification.