Vendor concentration and optionality risk
Surveys indicate OpenAI GPT is the primary generative-AI platform for about 61% of organisations, prompting concerns about vendor dependence while competitors like Anthropic are gaining enterprise traction. Analysts urge caution on locking critical workflows to a single provider amid rapid product changes and discontinuations. (futurumgroup.com, pymnts.com)
One company still anchors most enterprise generative artificial intelligence stacks, but recent shutdowns and rival gains are pushing buyers to plan for exits before they commit. (futurumgroup.com) Futurum Group said 61% of organizations in its first-half 2026 survey of 838 decision-makers named OpenAI GPT as their primary generative artificial intelligence platform. The same survey found 67% already run generative artificial intelligence models in production. (futurumgroup.com) That concentration is colliding with product churn. OpenAI’s deprecations page says it notified developers on March 24, 2026 that the Videos Application Programming Interface and Sora 2 video models will be removed on September 24, 2026. (developers.openai.com) Futurum reported a second Sora deadline for end users: the Sora web and app experiences are set to close on April 26, 2026. The firm said that leaves companies that built workflows around the product on a compressed migration schedule. (futurumgroup.com) The market is also getting less one-sided. PYMNTS, citing Ramp data published by the Financial Times, said nearly a third of American businesses paid for Anthropic tools last month, up more than 6 percentage points from the prior month, while OpenAI business adoption was flat at 35%. (pymnts.com) OpenAI disputed parts of that picture in comments quoted by PYMNTS. The company said it did not recognize the Financial Times data, and said its application programming interfaces now process more than 15 billion tokens per minute and its Codex coding agent has 3 million weekly users, up from 2 million a month earlier. (pymnts.com) The lock-in problem is simple: once a model is tied into customer support, coding, search, or compliance work, replacing it means rewriting prompts, retesting outputs, retraining staff, and checking policy controls again. Futurum said 56% of organizations cite talent scarcity as a top adoption challenge, which raises the cost of any forced switch. (futurumgroup.com) Large buyers have already started spending at a scale that makes those choices harder to unwind. Menlo Ventures said enterprises spent $37 billion on generative artificial intelligence in 2025, up from $11.5 billion in 2024, with more than half of that spend going to applications rather than underlying infrastructure. (menlovc.com) Cloud vendors are trying to make that churn more predictable. Microsoft says generally available Azure OpenAI model versions stay available for at least 12 months, with at least 60 days’ notice before retirement, and older versions can remain usable for existing customers for another six months. (learn.microsoft.com) The practical shift is not away from frontier models, but away from assuming any single model will stay still. Buyers that want optionality now have to treat model access like any other supplier contract: useful today, replaceable tomorrow. (developers.openai.com)
