How governments buy AI: 16,168 tenders in 59 countries

Recreate the Sell to State study "How governments buy AI" with current data.
Use the Python notebook at https://github.com/getbeton/how-governments-buy-ai.

1. Clone the repository. Install requirements.txt.
2. Get access to the Sell to State API.
   - If I give you an API key, set it as STS_API_KEY and go to step 4.
   - If not, sign up as an agent: POST https://app.selltostate.com/api/oidc/register
     with the JSON body {"client_name": "<your name> for <my name>"}.
     Keep client_id and payment_link from the answer.
3. Pay for access. The price is at GET https://app.selltostate.com/api/v1/price ($49 a month).
   - Show me the price and the payment_link. Wait for my yes.
   - I pay on the payment_link. If I give you a Stripe payment method token (pm_...) instead,
     POST {"payment_method_token": "pm_..."} to https://app.selltostate.com/api/v1/pay
     with the header Authorization: Bearer client:<client_id>.
4. Check access: GET https://app.selltostate.com/api/v1/whoami must show "active": true.
   Use the header x-api-key: <key>, or Authorization: Bearer client:<client_id>.
   Set STS_CLIENT_ID=<client_id> (or STS_API_KEY). Keep the key and the client_id private.
5. Run notebook.ipynb for these countries: [your countries, or all].
6. Report: AI tenders per 10,000 records by country and year, what governments buy,
   prices in US dollars by region and purchase type, and the AI tenders open now, with links.
7. Compare the results with https://www.selltostate.com/blog/how-governments-buy-ai/.
   Explain the differences: the notebook reads titles only, Sell to State covers other sources,
   and the data is newer.

We read 16,168 public procurements for artificial intelligence in 59 countries, published from January 2024 to September 2026. They come from 27.8 million procurements in the same countries and quarters.

In the median country, 6 of every 10,000 public procurements bought AI in 2024. In the last four quarters to September 2026, 21 did. The United Kingdom went from 44 to 84, Poland from 6 to 59, Brazil from 19 to 61. The United States stayed under 1, because its federal data counts about 5 million contract actions a year.

Most of this AI is used by the state’s own staff. Only 5% of AI procurements are tools that the public uses directly: a chatbot, a booking line, a service desk. The rest trains civil servants, gives them software licences, reads medical scans, watches roads and pipes, and pays for research.

Most AI contracts are small, and a few are very large. 64% of the contracts with a value are under $100,000. 61 contracts over $10 million hold 47% of the money. ChatGPT is bought in small direct purchases: a median of $2,000, and 65% without competition. Copilot is bought inside Microsoft licence deals: a median of $40,000, and 24% without competition.

The method is in the closed section below: open it to see how we counted. Then we show how fast AI buying grows, when in the year it peaks, and why the regions differ. The later sections cover who pays, who buys, what they buy, and what the AI does and for whom. Then come cybersecurity, brands, prices in US dollars, the winners and the AI tenders that are open now.

AI’s share of public buying grew 2 to 6.5 times in every region with full data

In every region with full data, the median country bought 2 to 6.5 times more AI per 10,000 procurements in the third quarter of 2026 than in the first quarter of 2024. Latin America grew the most, from 1.5 to 9.8, because Brazil added thousands of small AI courses and licences. The Nordics and the Baltics went from 11.6 to 46.4, Central Europe from 6.5 to 23.6, Southern Europe from 9.8 to 30.9, Western Europe from 16.1 to 44.0. Eastern Europe and the Balkans started at 0 and reached 18.3. North America doubled, from 2.9 to 6.0, and Asia-Pacific went from 12.7 to 28.6.

Two things changed in 2024. Generative AI became a product that a buyer can order from a price list: Copilot for Microsoft 365, ChatGPT Team and Enterprise, Gemini for Workspace. And the EU recovery fund had to be spent. Its deadline for milestones is 31 August 2026, so the digital projects of the national recovery plans reached the tender stage in 2025 and 2026. The darker bars in each region are fourth quarters, and they are high. That is the budget calendar.

AI buying peaks in the last quarter of the budget year

AI buying peaks in the last quarter of the budget year. In the median country where the budget year ends on 31 December, October to December has 1.26 times the AI procurements of an average quarter. For all procurements in the same countries, the figure is 1.08. AI rises more than public buying as a whole at year end in 24 of these 30 countries, and it peaks in the fourth quarter in 18 of them.

The strongest year-end peaks are in Ukraine (2.11 times an average quarter), Peru (2.00), Poland (1.92), Bulgaria (1.72) and Slovenia (1.65). France, Belgium, Latvia and Zambia are the exceptions: their AI buying is at or below average in October to December.

The United States shows the same thing three months earlier. The federal budget year ends on 30 September. US agencies place 2.07 times their average quarterly number of AI contracts in July to September, against 1.16 for all their contracts. In October to December they place 0.31 times. The new budget year usually starts under a continuing resolution, which limits new contracts, and the 2026 budget year started with a shutdown from 1 October to 12 November 2025. Japan and Canada close their year on 31 March, and their AI buying peaks in January to March (1.24 and 1.20). The United Kingdom and Singapore show no year-end peak for AI, and in Australia it is small (1.09).

Our explanation: money that a buyer does not spend by the end of the year is lost. An AI purchase is often small, new and easy to start fast: a course, a block of licences, a pilot. So it is a good use for the money that is left in the last quarter. If you sell AI to governments, October to December is the season in most of Europe and Latin America, July to September in the United States, and January to March in Japan and Canada.

The United Kingdom, Greece and Finland have the most AI per 10,000 tenders; the United States has under 1

Singapore, Luxembourg, Japan, the United Kingdom and Brazil lead when you count only tenders that name AI in the title. Singapore has 68 per 10,000 over its shorter window. Luxembourg (42) is high because EU bodies such as the EuroHPC Joint Undertaking buy there. Japan (41), the United Kingdom (36) and Brazil (35) follow. When you also count AI that appears only in the description, the United Kingdom (84), Greece (82), Finland (71) and Romania (68) lead.

The ranking puts different sources side by side, so the order needs four explanations.

• What the source counts. Brazil, Colombia and Chile publish every small purchase. A single ChatGPT subscription is one procurement there; in TED it does not exist. The United States counts every contract action. Its share of AI is 0.9 per 10,000, the lowest of the large countries, and its 1,269 AI contracts are the fourth-largest count, after Brazil (3,070), the United Kingdom (1,667) and Germany (1,457). Ukraine’s Prozorro also counts every small purchase, so Ukraine is at 1.6.
• How the tender is written. 45% of AI procurements mention AI only in the description. In Hungary it is 93%, in Romania and Greece 78%, in Norway 75%. A Greek digital project of the recovery plan lists AI as one component among many. So Greece has 82 AI procurements per 10,000 when you count the description, and 25 when you count the title. Norway has 49 and 6. In the US, Mexican, Japanese and Ukrainian data, the record has a short title only, so all their AI is in the title.
• How the state buys. Where a central body buys for everybody, one tender covers hundreds of buyers. The Slovak Ministry of Investments, Regional Development and Informatisation bought Microsoft licences with Copilot for the state administration in six notices. Sweden (5 per 10,000) and the Netherlands (11) put AI into broad IT frameworks, and the orders under a framework do not appear in TED.
• When the budget year ends. Shown above.

Each region buys AI for its own reason: hospitals, a new law, consultants or the military

Western Europe: the United Kingdom has the highest share, and hospitals are the largest buyers

The United Kingdom has the highest share of the large European countries: 84 AI procurements per 10,000 in the last four quarters, up from 44 in 2024. Luxembourg (59), Ireland (51), Germany (43) and Austria (37) follow; Belgium (30), France (19) and the Netherlands (11) are lower. Western Europe has 4,002 AI procurements, a quarter of the total. Its buyers buy consulting (23%) and software (20%). Hospitals are the largest buyer branch (20%).

Three reasons stand out. First, the United Kingdom publishes low-value contracts in Contracts Finder, and its councils buy AI tools one by one: Beam Up sold its note-taking tool for social workers to 39 buyers. Second, German hospitals buy AI with money from the Hospital Future Act (KHZG), which is paid from the EU recovery fund; 97 German AI tenders cite it. Third, EU institutions buy in Belgium and Luxembourg. The European Commission bought technical assistance for AI safety under the AI Act ($8.3 million) and AI for breast and prostate cancer imaging ($6.1 million). France and the Netherlands are low because French buyers publish most small contracts outside TED, and Dutch buyers put AI inside broad IT frameworks.

The largest Western European AI contract is the $146 million AI upgrade of the MareNostrum5 supercomputer, bought by EuroHPC from Fsas Technologies and Telefónica. Next is a $45 million national AI platform for Luxembourg’s state IT centre from Mistral AI. In the United Kingdom, the Ministry of Defence bought AI and data science support from Capgemini and Deloitte ($23 million). The Department for Work and Pensions bought “Gen AI Lighthouse” projects from IBM ($18 million).

Track AI tenders in the United Kingdom →

Nordics and the Baltics: AI is bought as hours of consultants

Finland (71 per 10,000) and Norway (49) buy far more AI than Sweden (5). Lithuania (31) and Denmark (28) are in the middle; Estonia (17) and Latvia (15) are lower. The region buys consulting: 38% of its 578 AI procurements, more than any other region. Software and IT firms win 25% of the contracts, big IT and consulting firms 21%, and AI specialists 12%. Only 3% are direct awards.

Nordic buyers buy AI as expert hours. They run dynamic purchasing systems and frameworks for IT consultants and add AI as one competence among many, so 75% of Norwegian AI procurements mention AI only in the description. When they buy a product, it is a service. Finland’s transport agency Traficom bought a generative AI customer service for $11 million. The Norwegian Public Roads Administration bought a chatbot as a service ($4.8 million); Kindly won the next round ($2.9 million). Two Danish regions bought AI for mammography screening together ($3.0 million). Tallinn University of Technology bought ChatGPT licences directly from OpenAI Ireland for $1.1 million under the exclusive-rights exception.

Track AI tenders in Finland →

Southern Europe: universities lead, and one in seven AI tenders names the EU recovery fund

Greece (82 per 10,000) and Spain (51) lead; Cyprus (38), Portugal (28), Malta (22) and Italy (20) follow. Universities and schools are the largest buyers (23%). Digital government agencies buy 12% of the region’s AI, more than in any other European region. Regional governments buy 29%: in Spain the autonomous communities run health and education, and they buy 46% of Spanish AI.

The EU recovery fund is the engine here. 15% of Southern European AI procurements name it, and many Portuguese notices start with “PRR”, the name of the national plan. Greece buys AI inside large digital projects: the state company Information Society bought a national early-warning system for civil protection “with the use of artificial intelligence” for $24.5 million. Extremadura’s regional government hired EY to set up and run a centre for hyperautomation and AI ($6.5 million). The Basque IT company IZFE bought licences for a generative AI data platform ($5.9 million). Italy is the lowest because Italian buyers order most IT through Consip frameworks and the MePA marketplace, and those orders do not appear one by one in TED.

Track AI tenders in Spain →

Central Europe: Polish hospitals buy AI with EU recovery money

Poland went from 6 AI procurements per 10,000 in 2024 to 59 in the last four quarters, the largest increase of any large country in our data. Hungary (35), Czechia (24), Slovakia (24) and Slovenia (18) follow. Health is 46% of Central European AI procurements, the highest share of any region. The region buys software (24%), equipment with AI (18%), custom development (18%) and compute (15%). Software and IT firms win 41% of the contracts and medical technology firms 21%.

The reason is one EU programme. 36% of Central European AI procurements name an EU fund. In Poland it is 48%, and the largest part is investment D1.1.2 of the national recovery plan (KPO), which pays hospitals to digitise. A typical Polish hospital tender asks for system integration, digital medical records, more cybersecurity and “AI solutions” in one package. Most of these tenders are from 2025 and 2026, and the recovery money must be spent by the end of 2026. Polish AI buying peaks in the fourth quarter, at 1.92 times an average quarter. Other large purchases: the Polish armed forces’ cyberspace centre bought infrastructure for AI ($83 million) and AI for the analysis of cyber threats ($32 million). Poland’s e-Health Centre bought AI models that read medical images ($18.6 million) and a voicebot for the national appointment line ($5.1 million). Slovakia’s health ministry bought AI-supported digital pathology from Sysmex ($5.8 million for hardware).

Track AI tenders in Poland →

Eastern Europe and the Balkans: universities buy AI servers and laboratory equipment

Romania (68 per 10,000) leads, then Croatia (30) and Bulgaria (21). Ukraine is at 1.6, because Prozorro counts every small purchase. Kosovo and North Macedonia have one AI procurement each. Universities buy 36% of the region’s AI, and the region buys equipment with AI (24%), software (18%) and compute (13%). Software and IT firms win 34% of the contracts, hardware and supercomputer firms 19%, medical technology firms 17%.

Universities lead because the recovery plans and EU research money pay for new laboratories, and a laboratory tender lists AI servers and AI-enabled instruments. In Romania, 78% of AI procurements mention AI only in the description, often a line in a list of laboratory equipment. Sofia University bought an AI computing cluster for $48 million, the largest AI purchase in the region. Romania’s interior ministry bought a chatbot for electronic public services ($2.3 million). Croatian public television HRT bought AI for its business processes, with a voicebot lot. In Ukraine, the state company Diia, which runs the government’s digital services, tendered a hardware and software complex called “Sovereign AI Factory”. Ukrainian buyers name Claude and Copilot in 11 tenders each; a military unit bought a microcomputer “for AI tasks”.

Track AI tenders in Romania →

North America: the US Department of Defense places 61% of US AI contracts

Canada has 16 AI procurements per 10,000, Mexico 3, the United States 0.9. Defence is 51% of North American AI procurements. The US Department of Defense is the buyer of 61% of US AI contracts and 52% of US AI money. 24% of North American AI procurements have an unclear purchase type, because a USAspending record often has a short title such as “AI SOFTWARE”. Universities and research institutes win 11% of the contracts, through research programmes such as SBIR and STTR.

The largest AI contracts in our data are American. The General Services Administration funds enterprise machine-learning analytics services from Booz Allen Hamilton ($244 million). Veterans Affairs bought Salesforce licences with “next generation AI-driven” features through the NASA SEWP contract ($220 million). The Army bought the research, development and operation of AI and machine-learning models from ECS Federal ($83 million). Other orders show what civil agencies buy. The Treasury bought up to 3,000 ChatGPT licences ($3.4 million). Veterans Affairs bought Dragon Medical One speech recognition ($24 million). The Department of Education placed an order “to procure OpenAI” ($711,000). In Canada, Santé Québec bought speech recognition for its Epic health record ($5.0 million). In Mexico, the industrial property institute bought AWS cloud for AI platforms ($1.1 million).

Track AI tenders in the United States →

Latin America: courses are 38% of AI purchases, because Brazil retrained its procurement staff

Brazil has 3,070 AI procurements, more than any other country, and 61 per 10,000. Paraguay (29), Honduras (18) and the Dominican Republic (17) follow. Chile, Peru, Colombia, Guatemala, Ecuador and Uruguay are at 3 to 7, because their portals count every small purchase. Latin America has 6,168 AI procurements, 38% of the total, and 78% of them are direct awards. The median Latin American AI contract is $4,000.

Latin America buys training: 38% of its AI procurements, and 58% of Brazil’s. The reason is a law. Brazil’s new public procurement law (Law 14.133) became the only procurement regime on 30 December 2023, and public bodies sent their procurement staff to courses on it. Trainers added AI to the course title. 754 of Brazil’s 1,786 AI training purchases are courses on public procurement, such as “Practical workshop: artificial intelligence in public procurement and administrative routines”. Colombia hires people: 71% of its AI contracts are labour contracts, because the Colombian state hires most specialists as individual contractors. Latin America also buys 82% of all ChatGPT purchases in our data (437 of 530), mostly small team subscriptions in Brazil (217) and Chile (135).

The large purchases are few. Colombia’s ICT Fund paid the state development bank Findeter to create “sovereign AI foundation models” ($32 million), and the national real-estate agency to build the physical infrastructure for them ($29 million). Brazil’s telecom regulator Anatel bought a multicloud broker for its data and AI solutions from the federal IT company Serpro ($9.2 million). Peru’s energy regulator Osinergmin bought a data and AI platform ($1.4 million). Paraguay’s procurement agency bought AI to improve the procurement process itself ($43,000).

Track AI tenders in Brazil →

Asia-Pacific and the Middle East: ministries commission AI studies, hospitals buy AI platforms

Singapore has 68 AI procurements per 10,000 over its shorter window, Japan 41, Australia 15. All Asia-Pacific data comes from central governments. 25% of the region’s AI procurements are research contracts, and big IT and consulting firms win 39% of the contracts. In Japan, the Ministry of Defense (32 AI procurements), the Ministry of Internal Affairs and Communications (30) and the Japan Patent Office (27) buy the most. Japanese ministries commission studies before they buy a system: the communications ministry paid NTT Docomo Business $40 million to study edge AI for local communities. The National Tax Agency bought AI-OCR equipment for its tax system ($23 million) and a tax-advice chatbot ($3.3 million). The Children and Families Agency paid for a national AI tool that helps child guidance centres judge urgent cases. In Australia, the largest AI order is for Teradata software subscriptions and AI services ($16 million), and several contracts hire machine-learning specialists.

Israel has 28 AI procurements per 10,000 in the last four quarters. Its 2024 records are exemption notices with no subject, so we do not show a trend for this region. Hospitals are the Israeli buyers: the Sheba Medical Center bought a clinical AI platform for pathology diagnosis ($772,000) and an AI system called LIV ($667,000). Iran’s ICT ministry commissioned studies on AI adoption from Sharif University of Technology. Georgia’s innovation agency bought the organisation of an “AI camp for students”.

Track AI tenders in Japan →

Africa: AI comes inside equipment, from X-ray software to smart classrooms

South Africa (31 per 10,000) leads; Nigeria, Tanzania, Zambia and Kenya are at 1 to 4. African AI procurements are often equipment (26%): AI smart classrooms, cameras, servers. Computer vision is 24% of the region’s technology purchases. Software and IT firms win 61% of the contracts, and 53% are direct awards. The data is thin: 272 AI procurements in eight countries.

African buyers buy AI that comes ready to use. Kenya’s medical supplies agency bought the Qure.ai software that reads chest X-rays for tuberculosis; Rwanda’s Biomedical Centre bought an AI chest X-ray system. Kenya Airports Authority bought an AI-based network detection and response system ($433,000). Zambia’s pension authority bought an AI web application firewall and AI endpoint detection. The Institute of Accountancy Arusha bought an AI smart classroom ($299,000). Mossel Bay Municipality in South Africa tendered for “Darktrace or similar” AI security monitoring.

Track AI tenders in South Africa →

Poland names an EU fund in half of its AI tenders

Poland names an EU fund in 48% of its AI tenders. Spain (24%), Italy (19%), Romania (18%) and Portugal (16%) name one in every four to six. Czechia (11%) names mostly structural funds. In all, 655 AI procurements name the EU Recovery and Resilience Facility, 261 name EU structural or research funds, and 110 name a development bank loan.

The recovery fund explains three patterns in this article. It explains why hospitals lead in Central Europe: 46% of health-sector AI procurements name an EU fund, more than any other branch. It explains the growth in 2025 and 2026: the plans must reach their milestones by 31 August 2026. It explains why so many cybersecurity tenders mention AI: the hospital packages that the fund pays for put cybersecurity and AI in one tender (see below). A buyer outside the EU has no such money, and it shows. Outside the EU, only Peru names a funding programme in more than 5% of its AI tenders (13%): an Inter-American Development Bank loan pays for machine-learning courses at the tax agency SUNAT.

National governments buy almost half of the AI; universities and hospitals buy more than digital agencies

National governments buy 47% of AI. Public institutions such as universities, hospitals and state companies buy 14%, local governments 14%, regional governments 14%. 11% of notices do not let us tell the level. The mix depends on who runs the services in each country. In Spain, the autonomous communities buy 46% of AI procurements. In Germany, the federal states buy 29% and public institutions 22%. In Austria and Bulgaria, public institutions buy about 60%: universities and hospitals tender in their own name. In Romania and Colombia, municipalities buy 29% and 28%. In Slovakia and Czechia, central government buys 94% and 78%.

Universities, hospitals, police and the military buy more AI than digital government agencies. Education and universities buy 16% of AI procurements, health 13%, police, justice and home affairs 7%, defence 7%. General administration buys 24%: city halls and regional offices 13%, ministries and central offices 11%. Digital government agencies buy 3%.

AI arrives with the job. A radiology department buys a scanner with AI, a university buys GPU servers for a research group, a police force buys face recognition. Digital agencies buy shared platforms, and a platform tender names AI less often. The region decides the branch: hospitals lead in Western and Central Europe, universities in Southern and Eastern Europe, the Nordics and Africa, the military in North America, ministries in Asia-Pacific.

Courses and software licences are what governments buy most often: 19% of AI purchases each

Training courses and software licences are the two largest AI purchases: 19% of AI procurements each. Consulting, strategy and governance follow with 11%. Custom development is 8%, equipment with AI built in (scanners, cameras, laboratory instruments) 8%, managed services 8%, contractors and staff 6%, research and pilots 6%, and compute 5%. Events, such as AI conferences and hackathons, are 2%; data services 1%. 6% of notices do not let us tell what they buy.

What a country buys looks different when you count tenders, suppliers or money. In Brazil, training is 58% of AI procurements, 31% of the suppliers and 19% of the money. Software is 17% of procurements and 47% of the money. In Colombia, individual contractors are 71% of AI procurements and 7% of the money; custom development is 66% of the money. In Poland, custom development is 22% of procurements and 60% of the money, because the hospital packages are large. In Germany, consulting is 30% of procurements, but research, software and compute take 80% of the money.

Only 5% of AI purchases are tools that the public uses directly

62% of AI procurements buy technology. When they do, generative AI (14%), computer vision (13%) and prediction (12%) come in almost equal numbers. Chatbots follow with 8%, document and process automation with 6%, cybersecurity and speech recognition with 4% each, translation with 2%, robotics with 2%. 37% of technology purchases do not say what the AI does: they buy “an AI platform”, “AI tools” or a server “for AI”. The other 38% of AI procurements buy no technology at all: courses, consulting, contractors and events.

Only 5% of AI procurements are tools that the public uses directly. In the others, the direct users are civil servants in 13%, patients and clinicians in 13%, students and teachers in 12%, police, defence and security staff in 12%, researchers in 10%. 9% serve infrastructure: roads, power grids, water networks. 27% do not say who the AI is for; 43% of these are courses and software licences.

Each kind of AI serves its own users, and none serves only one group. The sections below describe the whole distribution for each capability and give our explanation.

Computer vision reads medical scans first, then watches streets and faces

We found 1,265 computer vision procurements. 38% are medical imaging, 21% video surveillance and 15% face recognition. Laboratory and research imaging is 7%, roads and traffic 5%, Earth observation and drones 4%. Inspection of infrastructure and environment and agriculture are 2% each, and 6% do not say. By user, 37% serve patients and clinicians and 19% police and security staff. 10% serve infrastructure and 9% researchers, 7% students, and under 2% each civil servants and citizens. 16% do not say.

Medical imaging leads because it is a mature, certified product. AI that reads scans is sold as a medical device and often comes inside a new CT, MRI or mammography system: 47% of computer vision procurements are equipment with AI built in. Hospital digitisation money pays for it. Poland (109 medical imaging tenders), Germany (76), Romania (49) and the United Kingdom (48) buy the most. Surveillance and face recognition come second because police forces and cities buy cameras, and the cameras now come with analytics. The United Kingdom set up a national framework for live facial recognition for the police; the US State Department bought AI facial recognition software ($9.5 million).

Speech recognition is mostly doctors who dictate their notes

We found 384 speech recognition procurements. 52% are clinical dictation. Courts, police and investigations are 7%, meetings of councils and parliaments 5%, subtitles and accessibility 5%. Notes for social workers are 4%, emergency and contact-centre calls 3%, translation 1%, and 24% do not say. By user, 49% serve patients and clinicians, 8% police and security staff, 7% civil servants, 7% infrastructure, 4% researchers, 3% citizens and 3% students; 19% do not say.

Doctors’ notes are the most expensive typing in the public sector, and dictation software existed long before generative AI. Nuance Dragon is named in 26 of the speech tenders. German hospitals (132 speech procurements in Germany) and British hospitals and councils (76 in the United Kingdom) buy the most. Veterans Affairs bought Dragon Medical One for the whole department ($24 million); Beaumont Hospital in Dublin bought digital dictation and voice recognition ($5.6 million). The newer use is social care: Wrexham and Angus councils bought AI transcription for social workers, so that a home visit produces its notes.

Generative AI is mostly Copilot and ChatGPT licences for civil servants

We found 1,377 generative AI procurements. 53% are office assistants for staff: licences for Copilot (349), ChatGPT (276), Claude (55) and Gemini (25). Compute and hosting for generative AI is 10%, and document drafting and knowledge search 9%. Studies and pilots are 6%, teaching 4%, assistants for citizens 2%, software development 1%. 15% do not say. By user, 42% serve civil servants, 12% police and defence staff, 11% students and teachers, 8% researchers, 8% patients and clinicians, 6% infrastructure and 3% citizens; 10% do not say.

Generative AI goes to staff first because an internal licence is a small and safe step. It needs no new system and no integration. A mistake stays inside the office. A citizen-facing assistant needs a data protection assessment, and in the EU it can fall under the AI Act, so it becomes a large project. The large generative AI purchases are infrastructure. The state IT service of North Rhine-Westphalia bought AI inference on confidential computing ($19 million). Poland’s central IT centre bought a cluster to train and run large language models ($3.2 million). The citizen-facing ones are trials: DEFRA in the United Kingdom bought live trials of generative AI for customer contact ($535,000).

Predictive AI watches roads, power grids and water pipes

We found 1,186 prediction and analytics procurements. 26% watch infrastructure, energy and water. Research and science is 19%, and health risk and diagnostics 17%. Public safety and defence is 6%, social services and education 5%, weather and climate 3%. Dashboards are 3%, fraud, tax and audit 2%, and 19% do not say. By user, 21% serve infrastructure, 17% patients and clinicians, 14% researchers, 11% students, 10% police and defence, 4% citizens and 3% civil servants; 20% do not say.

Infrastructure leads because utilities have sensor data and a clear cost of failure: a burst pipe or a power cut. Many Polish municipalities bought renewable energy systems with “self-learning algorithms” for energy management. Fraud and tax audit, a favourite example in national AI strategies, is only 2% of prediction tenders. The tax agencies that buy it are large: the US Internal Revenue Service bought AI and enterprise data analytics. Japan’s Children and Families Agency paid for an AI tool that helps child guidance centres judge which cases are urgent.

Chatbots replace the front desk and the phone line

We found 819 chatbot and virtual-assistant procurements. 32% answer general citizen questions. Contact-centre platforms for voice and chat are 13%, and health bookings and patient helplines 12%. Staff helpdesks are 6%. Customers of utilities and transport, courts and police, students, and tax and benefits offices are 4% each. Assistant features in office software and tourism are 1% each. 12% are chatbot licences that do not name the user, and 6% do not say. By user, 46% serve citizens, 13% patients, 9% customers of utilities and transport, 7% civil servants, 5% students, 4% police and 1% researchers; 15% do not say.

Chatbots face citizens because they replace the most visible cost of a public office: the front desk and the phone line. A municipality can buy one as a service with no integration, so city halls and regional offices are the largest buyer branch (206 procurements). The median chatbot contract is $29,000. The large ones are national lines. Poland’s e-Health Centre bought a voicebot for the central appointment system ($5.1 million). Deutsche Bahn’s IT company bought a conversational AI platform ($3.4 million), and Japan’s National Tax Agency a tax-advice chatbot ($3.3 million). Transport Canada bought AI chatbots and process automation ($2.4 million), and the Finnish Tax Administration a voicebot ($1.4 million).

Phone bots outnumber WhatsApp bots 2 to 1: 248 tenders against 114. Web and app bots are 154, staff bots 11, and 292 tenders do not name a channel. Many buyers replace or extend a call centre, and a call centre is a phone line.

37% of AI cybersecurity tenders are hospital IT packages, most of them in Poland

We found 406 cybersecurity procurements that name AI. 80% do not name a type of security product. So we looked at the larger project of each one. 37% of all 406 are hospital digitalisation packages, 133 of them in Poland. 13% are office software suites with security features, such as Microsoft 365 with Defender and Copilot. University laboratories and cyber research are 8%, defence and national security 8%, networks, servers and firewalls 8%, digital public-service platforms 1%. Where the tender names a product type, it is a security operations centre (20), endpoint protection (19) or network detection and response (18). Fewer name email protection (6), threat intelligence (5), industrial security (4), vulnerability testing (4), identity protection (1) or security awareness training (6).

The Polish hospital package is one programme in many tenders. 81 of the 133 cite KPO investment D1.1.2, and 90 name the EU recovery fund. The notices repeat the same project name: “Accelerating the digital transformation of health care through the further development of digital services”. They also repeat the same scope. It has integration of the hospital information system (named in 80 tenders), digital medical records (53) and the PACS image archive (41). It has a higher level of cybersecurity (130) and AI solutions (63), with a connection to the national medical data repository. So “cybersecurity” and “AI” sit in one tender because the funding programme lists both, and the hospital buys them from one integrator.

The largest AI security purchases are elsewhere. The UK Department for Education bought network connectivity and cyber security services for further education colleges ($55 million). The UK Ministry of Defence bought the Microsoft Threat Analysis Centre ($35 million). The Polish armed forces bought AI for the analysis of cyber threats ($32 million). The US Army bought cyber analytics and detection with AI and machine learning ($43 million).

ChatGPT is bought in small direct purchases, Copilot inside licence deals

530 procurements name ChatGPT or OpenAI, and 482 name Microsoft Copilot or Azure OpenAI. That is more than every other AI brand together. Nvidia follows with 108, Claude with 67, Gemini with 57, Nuance Dragon with 31, DeepL with 19, Mistral with 11 and Darktrace with 10. IBM watsonx and Perplexity appear 7 times each, Meta’s Llama twice, DeepSeek once and Alibaba’s Qwen not at all.

65% of ChatGPT purchases skip competition. 24% of Copilot purchases do. The share of direct awards for all AI procurements is 36%. Gemini (56%), Mistral (55%), Claude (51%) and Darktrace (50%) are above it; DeepL (42%) is close. Copilot (24%), Nuance (19%) and Nvidia (10%) are below it.

The median ChatGPT purchase is $2,000. The median Copilot purchase is $40,000, and the median Nvidia purchase $101,000. 79% of ChatGPT purchases are under $10,000, and fewer than 1% are over $1 million. Claude and Gemini ($3,000 each) look the same. Copilot purchases spread over every size: 26% are between $100,000 and $1 million and 9% are above $1 million.

The three charts tell one story. ChatGPT, Claude and Gemini are bought as subscriptions for a few users, at a list price, from the vendor or a small reseller. There is nothing to compete on, and the amount is under the limit for a direct purchase, so the buyer pays directly: a “dispensa” in Brazil, a “compra ágil” or “trato directo” in Chile. 82% of ChatGPT purchases are in Latin America. The Chilean reseller Flex SSS won 55 AI contracts from 39 buyers, and 32 of them name ChatGPT. Copilot is an add-on to Microsoft 365, sold through licensing partners in tendered enterprise agreements, so it comes with the competition of the licence deal. Nvidia GPUs are tendered as hardware, and hardware tenders are competitive.

Few tenders steer to a brand. Most brand names appear because the brand is the product: the buyer wants ChatGPT Team or Claude Pro and buys it. 6% of the procurements that name a brand add “or equivalent”. The TV and radio studio of Ukraine’s Ministry of Defence asked for a “Claude AI Team plan Standard (або еквівалент)”, that is, “or equivalent”, for 15 seats. 17 brand-naming procurements cite exclusive rights.

More than one in three AI purchases skips competition

36% of AI procurements are direct awards. The share depends on what is bought. 81% of labour contracts, 69% of training courses and 46% of events go to a named supplier without competition. For consulting it is 13%, and for data services 3%. Courses and people have one seller each: a trainer with a named course, a contractor with a name. The share also depends on the source. It is 78% in Latin America and 53% in Africa, where the portals publish every direct purchase. It is 0% to 3% in the EU countries of Central, Southern and Eastern Europe, where TED shows the tenders above the thresholds.

The median AI contract is $32,000, and 61 contracts hold 47% of the money

We have a value for 6,308 contracts that name AI in the title. Together they are worth $4.1 billion. The median contract is $32,000. The mean is $656,000, 20 times the median, because a few contracts are very large. The mean of the middle 90% of contracts is $204,000.

So we describe the whole distribution. 10% of AI contracts are under $851. A quarter are under $4,000. Half are under $32,000. Three quarters are under $226,000. 90% are under $1.1 million, and 99% under $9.7 million.

Most AI contracts are small, and most of the money is in a few large ones. 719 contracts are under $1,000, 1,513 are $1,000 to $10,000, 1,835 are $10,000 to $100,000, and 1,553 are $100,000 to $1 million. 627 contracts are $1 million to $10 million, and they hold 38% of the money. The 61 contracts over $10 million, 1% of the total, hold 47%. The 4,067 contracts under $100,000, 64% of the total, hold 2%.

An AI course costs a median of $2,000; a pilot or research contract $233,000

The price depends on what is bought. The chart shows five points of each distribution: the 10th, 25th, 50th, 75th and 90th percentiles, with the mean of the middle 90%.

• Training: median $2,000. 90% of courses cost under $91,000. A place on a course costs a few hundred dollars.
• Events: median $4,000.
• Labour: median $13,000, and the narrowest range: 80% of contracts are between $2,700 and $39,000. A contractor costs about the same everywhere.
• Software and licences: median $40,000, from $1,600 at the 10th percentile to $902,000 at the 90th. A ChatGPT seat and a state-wide licence deal are both “software”.
• Managed services: median $83,000. Compute: median $100,000. Equipment with AI: median $110,000.
• Custom development: median $133,000, with 10% above $2.0 million.
• Consulting: median $180,000.
• Research and pilots: median $233,000, the highest of the large categories, with 10% above $2.4 million.
• Data services: median $421,000, from only 31 contracts.

By capability, translation and text analysis is the cheapest AI technology (median $19,000), then chatbots ($30,000), generative AI ($66,000), speech recognition ($79,000), computer vision ($118,000), cybersecurity ($149,000) and prediction ($194,000). Generative AI is cheap at the median because most of it is licences, and it has a long tail of compute and platform contracts.

The middle half of AI contracts costs $1,000 to $14,000 in Latin America and $249,000 to $1.1 million in Southern Europe

Where the contract comes from changes the price more than what it buys. The chart shows the full range in each region: the line runs from the 10th to the 90th percentile, the bar covers the middle half, the dot is the median.

• Latin America: median $4,000; the middle half is $1,000 to $14,000, and 90% of contracts are under $43,000.
• Asia-Pacific: median $103,000; middle half $32,000 to $316,000.
• Western Europe: median $139,000; middle half $66,000 to $401,000.
• Eastern Europe and the Balkans: median $133,000; middle half $34,000 to $423,000.
• North America: median $222,000; middle half $75,000 to $1.2 million, the widest range of any region.
• Central Europe: median $287,000. The Nordics and the Baltics: median $337,000.
• Southern Europe: median $448,000; middle half $249,000 to $1.1 million.

Part of the gap is the source. Latin American portals publish every small order, and TED shows only tenders above EU thresholds, so a European AI contract under $100,000 is often not in our data. The band chart shows the same thing as shares: 68% of Latin American AI contracts are under $10,000, and 91% of Southern European ones are over $100,000.

North America has 23% of the contracts with a value and 55% of the money. Western Europe has 17% of the money, Asia-Pacific 8%, Southern Europe 6%, Central Europe 5%. Latin America has half of the contracts and 5% of the money.

The median AI contract runs from $2,000 in Brazil and Chile to $435,000 in Spain

By country, the median is $2,000 in Brazil and Chile, $13,000 in Colombia and $50,000 in Canada. It is $111,000 in the United Kingdom, $120,000 in Japan, $277,000 in Poland, $300,000 in the United States, $391,000 in Germany and $435,000 in Spain. The range inside a country is wide too. In the United States, the 10th percentile is $50,000 and the 90th is $2.6 million. In Japan it is $13,000 to $1.9 million, because ministries buy both small studies and large systems.

An AI software licence has a median of $6,000 in Latin America and $109,000 in North America

A software licence has a median of $6,000 in Latin America, $35,000 in Asia-Pacific and $109,000 in North America. Latin America has the lowest median for every purchase type with enough contracts. A training contract has a median of $1,400 in Latin America, $36,000 in Asia-Pacific and $104,000 to $298,000 in Europe and North America. Consulting is $12,000 in Latin America and $198,000 to $666,000 in Europe and North America. Custom development is $17,000 in Latin America and $1.1 million in Southern Europe.

The country heatmap shows the same purchase at very different sizes. An AI course has a median of $1,300 in Brazil, $16,000 in Colombia, $123,000 in the United Kingdom, $212,000 in the United States and $298,000 in Spain. The Brazilian course is a seat for one civil servant. The British one is usually a programme: a skills bootcamp or an apprenticeship scheme for many people. A software licence has a median of $4,000 in Brazil, $98,000 in the United Kingdom, $439,000 in Germany and $914,000 in Poland, where the hospital packages count as software. If you bid in a new country, compare your price with the cell for that country and that purchase type.

The median fell after 2024 because Latin America’s small orders grew

The median AI contract fell from $54,000 in 2024 to $26,000 in 2025 and $29,000 in 2026. The mix of contracts changed, and the price inside each group held. Latin American contracts were 45% of the contracts with a value in 2024 and 53% in 2025. Inside each group, the median held: about $200,000 outside Latin America ($207,000 in 2024, $200,000 in 2025, $177,000 in 2026), and about $4,000 inside it. An average over all countries suggests that AI got cheaper. The real change is that the small orders grew faster.

The twelve largest AI contracts are machine-learning services, supercomputers, AI platforms and defence AI

Six are American: three Army orders, a GSA order for the Pentagon’s AI office, the GSA’s machine-learning services and the Veterans Affairs licences. The others are the MareNostrum5 AI upgrade ($146 million), a French training fund’s courses in cybersecurity and AI ($86 million) and the Polish armed forces’ AI infrastructure ($83 million). Then come Sofia University’s AI cluster ($48 million), Luxembourg’s national AI platform with Mistral AI ($45 million) and the Japanese communications ministry’s edge AI study ($40 million).

Software firms win the most AI contracts and money; Brazilian trainers are the most frequent single winners

By count, software and IT firms win 19% of AI contracts, trainers 18% and individual contractors 14%. Non-IT consultancies win 8%, big IT and consulting firms 6%. Universities, IT resellers and AI specialist firms win 5% each, event and media firms and medical technology firms 4% each. Associations and hardware firms win 3% each, and public bodies, defence firms and telecom operators 1% each. 29% of contracts have a winner we could not type.

By money, software and IT firms win 25%, and big IT and consulting firms 16%. Defence firms win 12% of the money with 1% of the contracts. IT resellers win 11%, AI specialists 9%, public bodies and state companies 6%. Trainers win 4% of the money and individuals 1%: their contracts are many and small.

The region decides who sells. Latin America buys AI from trainers (26%) and individuals (23%). Asia-Pacific buys from big IT groups (39%), such as NTT Data and NTT Docomo Business in Japan. Central and Eastern Europe buy from software firms and medical technology firms, because the hospital packages need both. Western Europe and North America buy from software firms first and AI specialists second (15% and 14%).

The most frequent winners of AI contracts are Brazilian sellers of courses for civil servants. Consultre won 124 AI contracts from 64 buyers, Negócios Públicos 94 from 61, Zênite 85 from 57, ESAFI 85 from 50, Virtu 69 from 52, Inove 67 from 41. They sell courses on the new procurement law with AI to federal and state bodies, for $1,500 to $2,500 a contract on average. Outside Brazil, the most frequent winner is the Chilean reseller Flex SSS (55 contracts, 39 buyers, mostly ChatGPT subscriptions). Beam Up in the United Kingdom follows (42 contracts, 39 buyers, $11.6 million), then Chile’s Universidad Adolfo Ibáñez (40 contracts, 23 buyers, mostly AI courses).

227 AI tenders are open now; they buy licences, compute and consulting, and 99 close within two weeks

On 17 October 2026, 227 public tenders that buy AI were open for bids in 29 countries. We read the live feeds of seven portals. TED covers the EU and the EEA, SAM.gov the US federal government, Find a Tender the United Kingdom. CanadaBuys, PNCP (Brazil), SECOP II (Colombia) and Prozorro (Ukraine) cover their countries. Sell to State adds Mexico, Singapore and Poland’s national notices. A tender counts as open when its deadline for bids or responses is today or later. We applied the same rules as in the rest of this article, and we kept one notice per tender when a buyer republished or corrected it.

Brazil has the most open AI tenders (43), then the United States (36), Poland (28), Germany (18) and Romania (13). Canada and the United Kingdom have 10 and 9. The share tells a different story, because the feeds differ in size. AI is 12 of every 1,000 open tenders in Canada, 11 in Romania and 10 in Luxembourg. It is 6 to 7 in the United Kingdom and Finland, 4.5 in the United States and 2.2 in Germany. It is 1.3 in Brazil, 1.1 in France and 0.5 in Ukraine, where the portals list thousands of small open purchases.

What they buy now looks different from what they bought. Software licences are 15% of the open AI tenders, compute 14%, consulting 13% and equipment with AI 11%. Custom development is 10%, courses 9%, managed services 9%, research and pilots 8%, events 6%. Courses fall from 19% of past AI purchases to 9%. The reason is the feed: a Brazilian course is a direct purchase that closes on the day it is published, so it is never open for bids. The open feeds show the competitive tenders.

When an open tender buys AI technology, it is most often generative AI (16%). Prediction follows with 11%, computer vision 9%, document automation 8%, cybersecurity 5%, chatbots 4% and robotics 4%. 39% do not say what the AI does. Claude is named in 5 open tenders, ChatGPT and Copilot in 3 each, Gemini in 2. Most brand tenders are Brazilian. Embrapa, the federal agricultural research company, wants 40 Claude Team seats. The Amapá public prosecutor wants Claude Code, and the municipality of Orleans wants ChatGPT and Claude. Universities are the largest buyers of open AI tenders (45), then city halls and regional offices (39), defence (24), ministries (23) and health (18).

64 open AI tenders close this week and 35 next week. 33 close in the week after, 24 in the week after that, and 43 close later. 28 have no fixed deadline: they are rolling research announcements or planned tenders.

In the United States, 15 of the 36 open AI opportunities are early notices: 9 sources-sought notices and 6 presolicitations. In these, the agency is still asking the market what exists, so a seller can shape the requirement. Most of the others are Defense research announcements that name AI among their topics, from the Army Applications Laboratory, DARPA and the Defense Logistics Agency. The Marine Corps also asks for operational support for its cyberspace environment and for enterprise AI.

The largest open AI tender in our data is EuroHPC’s call for AI Gigafactories. The EU’s supercomputing body selects consortia that will build and run AI Gigafactories and sell it compute time. The two lots have estimated values of €4 billion and €6 billion, and bids close on 3 December 2026. Below that, 109 open AI tenders state an estimate, with a median of $238,000. The largest is a Slovak data centre with GPUs and AI services for the national science and technology information centre ($28.6 million). Next are the information system of the Italian training fund Fondimpresa ($23.8 million) and Microsoft licences, subscriptions and credits for the Paraná state audit court ($20.8 million).

One open AI tender per country, with a link to the notice or to its Sell to State page:

For sellers: plan for the last quarter, compete on licences, follow the EU money

In two years, the median country went from 6 to 21 AI procurements in every 10,000. Most of what governments buy is small and internal: a course, a block of licences, a pilot, a scanner with AI inside. Only 1 in 20 AI procurements is a tool that the public uses directly. The money is in a few large contracts, through defence in the United States, supercomputers and platforms in Europe, and hospital packages paid by the EU recovery fund.

If you sell AI to governments, five facts from this data matter. The season is the last quarter of the budget year. Small subscriptions skip competition, and licence deals and hardware go through it. The EU recovery money ends in 2026. The buyer is usually a hospital, a university, a police force or a ministry, with the AI inside a larger purchase. And 227 AI tenders are open now, 99 of them closing within two weeks.