OpenAI Leads U.K. AI Provider Mentions, GreenPT Study Finds

By Amara Osei

OpenAI accounted for 670 of the 1,377 documented links between U.K. companies and AI providers in GreenPT's November study, equal to 48.7% and about twice Anthropic's total. It was the most frequently named provider in GreenPT's review of public documents linking British companies to AI technology.

In GreenPT's ranking, Anthropic ranked second with 322 links, or 23.4%. Open-source models followed with 235, or 17.1%, while Google had 150, or 10.9%. GreenPT's main ranking put OpenAI and Anthropic together at 72%.

What did GreenPT actually count?

GreenPT, a European AI platform that markets privacy and sustainability services, collected the links from public material. Its sources included company statements, supplier case studies and privacy notices.

GreenPT also reviewed supplier and customer disclosures, privacy notices and lists identifying companies that process data for a business. Each counted entry, which the report called a reference, linked one U.K. company with one AI model provider.

GreenPT counted each company once for every provider connected to it. A company naming both OpenAI and Anthropic appeared under both, so the 1,377 total does not represent 1,377 separate businesses.

Because GreenPT reviewed documents rather than survey a representative group of British businesses, companies that had not discussed AI suppliers publicly were absent. GreenPT said the findings cannot establish market share, customer spending, national adoption rates or actual AI workloads.

How strong is the evidence behind the rankings?

GreenPT split its research into a main ranking and a second dataset assessing 734 company-provider links across 15 providers. It classified an entry as supported when the documents disclosed use or explicitly identified an AI supplier and corroborated a connection to a U.K. company.

A statement from the company or its supplier was sufficient. GreenPT did not independently check whether the technology was running inside the business.

Other links involved proposed projects, offered services, outdated statements or conditional language. The weaker set also covered pilots, plans, research, historical use, unclear company identities, offered integrations and attributions GreenPT rejected.

GreenPT estimated that about 31% of links involving OpenAI, Anthropic and Google lacked evidence meeting its standard. The supported rates were 69.9% for OpenAI, 68.5% for Anthropic and 65.7% for Google.

GreenPT said those differences fell within the samples' statistical uncertainty. It concluded that OpenAI's larger total separated it more clearly than the strength of evidence behind each connection.

Researchers then applied those rates to the larger totals and combined them with the open-source result. GreenPT estimated that 1,023 of 1,377 links, or 74.3%, met its evidence standard.

That figure was a projection because researchers did not classify every main-ranking entry. GreenPT said extending the smaller dataset's results to the full ranking required an assumption.

All 235 links in the open-source category met GreenPT's standard. The result lifted the estimated evidence-backed share above the roughly 69% combined rate for OpenAI, Anthropic and Google.

GreenPT's open-source category covers open-weight models used directly or through services that route requests to different models. It defines weights as the numerical settings a model learns during training; releasing them lets others run and adapt the model.

The category groups several technologies, unlike the main ranking's individual suppliers. Its third-place finish therefore reflects the standing of a technology group rather than one company.

Which industries show the clearest support?

GreenPT also assembled 969 industry entries by combining the 734 evidence-assessed links with the 235 open-source links. Its breakdown reported these totals and supported counts:

• GreenPT counted 392 software, IT and AI services entries, about 40% of the dataset, including 278 that met its standard.
• GreenPT recorded 138 financial-services entries, with 109 supported, and 95 business and professional-services entries, with 81 supported.
• GreenPT found education had the highest supported proportion at 93.7%, covering 74 of 79 entries.
• GreenPT found marketing, media and creative businesses had the lowest supported proportion at 63.5%, covering 47 of 74 entries.

GreenPT found that open models made up 28% to 33% of evidence-backed entries in each of the nine industry groups. OpenAI led the named providers in most groups.

Anthropic led in retail, food and hospitality, with 16 supported entries against OpenAI's 10. Anthropic tied OpenAI in marketing, media and creative businesses and in healthcare and life sciences.

GreenPT recorded Google's strongest industry showing in financial services, where it supplied about 16% of evidence-backed entries. GreenPT cautioned that healthcare, with 31 total entries, and legal services, with 47, were too small for firm provider comparisons.

What changes when smaller providers are included?

GreenPT's main ranking omitted some suppliers examined elsewhere in its research. Twelve additional providers appeared in the evidence-assessment dataset: Mistral, Amazon, IBM, xAI, Meta, Alibaba's Qwen, Cohere, Stability AI, DeepSeek, Midjourney, Stable Diffusion and Voyage AI.

GreenPT counted 81 entries for those suppliers, of which 60 met its evidence standard. Adding them would lift the main total to 1,458 and reduce OpenAI's share to 46%, Anthropic's to 22.1%, open source's to 16.1% and Google's to 10.3%.

All 25 Amazon entries met the standard. GreenPT identified Amazon as the only smaller provider whose statistical range clearly exceeded the combined evidence-support rate for OpenAI, Anthropic and Google, while most smaller suppliers had too few entries for reliable comparisons.

GreenPT's methodology makes corroborated deployment the practical measure for future disclosures, rather than pilots, plans or offered integrations. The distinction matters because public mentions did not establish live AI workloads.