OpenAI said to post higher margins on business sales

OpenAI has reportedly improved its compute margin for paid users to 70% as of October, up from 52% at the end of 2024.

Compute margin measures the share of revenue left after the cost of running AI models.

Despite the improvement, OpenAI has not yet posted a profit and continues to face high computing and infrastructure costs.

The company is said to be in early talks with Amazon to raise at least US$10 billion and use Amazon’s chips, in a deal that could value OpenAI at over US$500 billion.

OpenAI faces rising competition from Google and Anthropic, while most ChatGPT users remain unpaid.

Its compute margins for paid accounts are higher than Anthropic’s, although Anthropic is reportedly more efficient in server spending.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Margin gains need context on scale and profitability

OpenAI moved compute margin from 52% to 70%. Without paid subscriber counts, enterprise Annual Recurring Revenue (ARR) or losses the 18‑point gain may not shift profitability.
In July 2025, ChatGPT had 35 million paid users 1 and an estimated $12 billion ARR 2. OpenAI lost $11.5 billion over the year 1, so margin gains may not cover compute and data center costs.
Annualized revenue could reach $20 billion by late 2025 1. Monthly user growth slowed from 42% early 2025 to 13% by September 1, which makes margin gains central to profit as growth cools.

Migration to Amazon chips meets cost pressure

OpenAI may shift workloads to Amazon chips under a reported $10 billion deal, and enterprises could follow to cut per‑token inference costs. Amazon Web Services (AWS) Trainium2, second-generation AI training chip, delivers 30–40% better price‑performance than GPU-based Amazon Elastic Compute Cloud (EC2) P5e and P5en instances 3.
Cloud consultancies plus systems integrators offer Large Language Model (LLM) benchmarks on AWS Inferentia (AWS’s inference chip) and Trainium 4 5. vLLM is an open-source high-throughput inference engine, while lm-evaluation-harness is an open-source benchmarking suite.
Trainium3, AWS’s third-generation training chip, delivers 5× more output tokens per megawatt than Trainium2 5. It supports frontier models, which are the largest and most advanced AI systems 3. This enables Return on Investment (ROI) case studies from moving inference workloads, especially for enterprises spending heavily on Nvidia GPUs.