OpenAI

Full briefing

OpenAI was founded in 2015 as a non-profit and transitioned to a 'capped-profit' structure in 2019 to enable equity investment. Headquartered in San Francisco, the company had approximately $3.7 billion in annualized revenue in 2024 (primarily from ChatGPT Plus subscriptions and API revenue) and was pursuing a full conversion to a for-profit public benefit corporation. Its strategic partnership with Microsoft — which invested $13 billion across multiple tranches — makes Azure the exclusive cloud platform for all OpenAI workloads. OpenAI's product portfolio spans ChatGPT (the world's most widely used AI application, with over 200 million weekly active users), the GPT-4o and o3 families of language models, Sora (text-to-video), DALL-E (text-to-image), and the Whisper speech recognition system. The company also offers the OpenAI API, which allows developers to integrate these models into third-party applications via Microsoft Azure. The scale of compute required by OpenAI is extraordinary. Training a GPT-4-class model is estimated to require approximately 25,000 A100 GPUs running for 90–100 days, consuming in the range of $50–100 million in compute cost per training run. GPT-4o and o3 represent further scale-ups. ChatGPT inference, which serves hundreds of millions of queries daily, requires sustained GPU fleet deployment equivalent to a medium hyperscaler — all hosted exclusively on Microsoft Azure's AI HPC clusters. OpenAI's compute relationship with Microsoft extends beyond simple cloud rental. The companies co-developed Azure Maia 100 — Microsoft's custom AI ASIC (TSMC-fabbed) — with OpenAI providing workload requirements that shaped the chip architecture. This makes OpenAI an indirect customer of TSMC and the broader semiconductor supply chain even for custom silicon. OpenAI has been the most visible catalyst for the AI GPU supply crunch. The launch of ChatGPT in November 2022 triggered an immediate surge in hyperscaler GPU procurement that propagated through the entire NVIDIA supply chain — from TSMC CoWoS packaging to SK Hynix HBM allocation to OSAT capacity at ASE and Amkor. The company's continued frontier model development and the prospect of GPT-5 training runs continue to be a primary demand signal for the AI chip industry.