Microsoft AI chief says the company beat OpenAI at its own game, using a quarter of the computing power

Microsoft AI CEO Mustafa Suleyman said the company’s newly announced MAI-Base-2 outperformed OpenAI models on tests of pattern understanding and prediction, while using one-quarter of the compute used to train the OpenAI model he referenced. He also claimed MAI-Base-2 used 12 times less computation than an unspecified DeepSeek model; Microsoft’s comparisons relied on technical information it has about OpenAI models. The efficiency claims support Microsoft’s effort to develop its own AI models, but the article notes the strategic tension with OpenAI, a major partner and an investment Microsoft’s stake is likely worth more than $100 billion.

Analysis

The strategic value is reduced dependence on OpenAI, not a demonstrated near-term earnings uplift. A credible in-house base model could improve Microsoft’s negotiating leverage, protect Azure/Copilot continuity if the partnership frays, and retain more model economics internally. But pre-training benchmark claims do not establish comparable performance in deployed products, inference cost, reliability, or customer willingness to switch. The comparisons are company-reported and use confidential information, so treat them as a signal of capability—not independent validation.

There is also a balance-sheet and ecosystem trade-off: more substitution of OpenAI models may strengthen Microsoft’s product control while putting pressure on the value of its OpenAI stake and the partnership’s commercial value. Lower training compute per model is potentially margin-positive, but not automatically negative for accelerator demand: cheaper training can enable more experiments and larger workloads, while inference requirements remain unaddressed.

Near term (days), the announcement is a modest strategic positive for MSFT, but unlikely to warrant a large repricing on its own. Over 1–3 months, the key catalyst is evidence of deployment, external evaluations, and cost/performance in Copilot or Azure. Over 6–18 months, successful substitution could improve bargaining power and resilience; failure to translate pre-training results into products would leave the core economics largely unchanged. The contrarian risk is over-reading efficiency claims as proof of lower total AI capex or a near-term OpenAI displacement.

Key Decisions for Investors

• No standalone trade on the announcement: keep MSFT exposure unchanged absent independent evidence of product performance or a guidance-level economic impact.
• Set an MSFT-positive watch trigger for verified deployment in Copilot/Azure with measurable adoption, serving-cost, or retention benefits; reassess relative exposure against other cloud platforms only if those metrics emerge.
• Track MSFT’s disclosures on the OpenAI relationship and investment. A deterioration in partnership economics could offset the strategic benefit of in-house models; reassess if MSFT signals reduced OpenAI usage or a material change in the investment’s value.
• Falsify the efficiency thesis if independent evaluations show no product-level advantage, or if subsequent disclosures indicate no meaningful serving-cost or customer-adoption benefit. Do not infer reduced accelerator demand from pre-training compute claims alone.