OpenAI's growth is straining its model

OpenAI is big enough to add a million ChatGPT users in an hour, but it is still small enough to run out of the chips that make its system work. In February 2025, Sam Altman said OpenAI was “out of GPUs,” meaning the company had to stagger the release of GPT‑4.5 because it did not have enough graphics processing units, the specialized chips used to train and run artificial intelligence models. (techcrunch.com) A month later, the pressure got worse after OpenAI put image generation directly inside GPT‑4o, the model that powers ChatGPT. OpenAI said on March 25, 2025 that GPT‑4o could generate photorealistic images, render text inside pictures, and edit uploaded images in the same chat window. (openai.com) That feature spread faster than OpenAI’s hardware could keep up. On April 1, 2025, Altman said the image tool’s popularity would delay new releases, break some features, and slow the service, and he said ChatGPT had added one million users in a single hour while the company worked through “capacity challenges.” (techcrunch.com) This is the awkward part of the artificial intelligence business in 2026: the product looks like software, but the bottleneck looks like heavy industry. Every new chatbot answer and every new generated image depends on rooms full of graphics processing units, power equipment, cooling systems, and fiber links, so a viral feature can hit the company like a factory suddenly taking twice as many orders. (techcrunch.com) (openai.com) OpenAI has tried to solve that by getting much bigger, much faster. The company announced a $40 billion financing round on March 31, 2025, and SoftBank separately said on March 27, 2026 that it had secured a $40 billion bridge loan to support further investment in OpenAI. (cnbc.com) (msn.com) OpenAI also tied itself to Stargate, the giant data center buildout with SoftBank and Oracle. In Port Washington, Wisconsin, that translated into a proposed $15 billion campus linked to a 1.3 gigawatt artificial intelligence data center project, which is roughly the scale of utility infrastructure, not a normal office park. (broadbandbreakfast.com) Then the local politics arrived. On April 8, 2026, Port Washington voters approved a referendum requiring voter approval before city leaders can offer tax incentives for projects worth more than $10 million, after a fight over the OpenAI and Oracle project. (broadbandbreakfast.com) The margin was not close. Broadband Breakfast reported that 2,710 voters backed the measure and 1,371 opposed it, with more than half of the city’s 8,257 registered voters participating, while Oracle had pitched the project as bringing $175 million in infrastructure upgrades, more than 4,000 construction jobs, and 1,000 permanent jobs. (broadbandbreakfast.com) OpenAI is also fighting on another front at the same time. Reuters reported on April 6, 2026 that OpenAI asked the California and Delaware attorneys general to investigate Elon Musk’s alleged anti-competitive behavior ahead of an April trial tied to Musk’s 2024 lawsuit over OpenAI’s shift toward a for-profit structure. (reuters.com) So the company now has three problems that feed each other: a product that can grow faster than its chip supply, an infrastructure plan that depends on winning over towns that may not want giant data centers, and legal fights that consume management attention while the system is already under strain. The strange part is that all three problems come from success: more users, bigger projects, and higher stakes. (techcrunch.com) (broadbandbreakfast.com) (reuters.com)