Microsoft says its new model beats OpenAI's on a quarter of the compute, bad for AI's $625 billion debt

Microsoft's AI chief says MAI-Base-2 beats the base model behind OpenAI's GPT 5.6 on a quarter of the training compute. If the claim holds, it breaks the premise most of the AI buildout's debt is priced on.

Richard TangRichard Tang · 3 min read
Share
Mustafa Suleyman, CEO of Microsoft AI, wearing a cream shirt and round glasses, photographed in April 2024
1 / 6Slide 1 of 6
Microsoft AI chief Mustafa Suleyman made the efficiency claim in remarks to employees on Thursday.

Microsoft+0.55% — Microsoft, up 0.55 percent today just put a number on the question the AI financing trade has been avoiding: does every frontier model really need more compute than the last? Its AI chief Mustafa Suleyman told employees on Thursday that MAI-Base-2, announced the same day, is a stronger pre-trained model than the base behind OpenAI's$1.18T — OpenAI, private, latest valuation $1.18T GPT 5.6, and that it took one-quarter of the compute to train 1.

The comparison is Microsoft's own and private 1. Even as a claim, it landed on a market priced for the opposite: lenders this week demanded a 14.8% yield to hold the 2032 notes of CoreWeave+3.04% — CoreWeave, up 3.04 percent today, a "neocloud", the market's term for a firm that rents out AI computing power 23.

One benchmark is published; the other is Microsoft's word

Microsoft's published post gives the claim concrete backing. MAI-Base-2 uses a mixture-of-experts design, which activates only a portion of the model's parameters for each task, and Microsoft trained it on 16,000 Nvidia GB200 chips over 38 days 4. It says the run was 4 to 12 times more compute-efficient than DeepSeek-V4-Pro-Base, from the Chinese lab DeepSeek, across four evaluations 4.

The OpenAI comparison is different in kind: Suleyman said Microsoft matched its model against OpenAI's privately, using technical information it holds about them, and no third party has checked the result 1. Microsoft also sits on both sides of the claim. It hosts and resells OpenAI's models on its Azure cloud, puts them in Copilot, and holds an OpenAI stake likely worth more than $100 billion, so a better Microsoft model dents an asset Microsoft itself carries 1.

The debt was priced on compute compounding

Neocloud loans and hyperscaler leases assume demand for raw compute compounds, generation after generation. This week lenders repriced that premise. CoreWeave's shares fell 15% on Wednesday and its unsecured 2032 notes dropped below 80 cents on the dollar 5. Nebius−0.15% — Nebius, down 0.15 percent today, another neocloud, fell 10.1% to $186.01, breaking the $200 level a chart analyst had called the line between a rebound and a faster slide 3.

CoreWeave's debt has grown nearly fivefold in six quarters

$0B$20B$40B$60BQ4 '24Q1 '25Q2 '25Q3 '25Q4 '25Q1 '26Q2 '26$51.61B
Data
Total debt
Q4 '24$10.62B
Q1 '25$11.9B
Q2 '25$14.56B
Q3 '25$18.81B
Q4 '25$29.82B
Q1 '26$35.15B
Q2 '26$51.61B
Total debt at each quarter end, US$ billions, from CoreWeave's SEC filings. This week lenders demanded a 14.8% yield on its 2032 notes.10

Nebius's share price is also a funding gauge: the August series of its roughly $10 billion in convertible bonds converts near $313 a share, and the further the stock sits below that price, the more those notes behave like plain debt 3. Microsoft has spent against the same premise: Azure's leadership has discussed cutting its software group's costs by at least a fifth to help carry the data-center bill, according to a report Microsoft disputes 6.

Microsoft's quarterly capital spending has more than tripled since 2023

$0B$10B$20B$30B$40BQ4 '23Q2 '24Q4 '24Q2 '25Q4 '25Q2 '26$35.8BThe AI data-center ramp
Data
Capital expenditure
Q3 '23$9.92B
Q4 '23$9.74B
Q1 '24$10.95B
Q2 '24$13.87B
Q3 '24$14.92B
Q4 '24$15.8B
Q1 '25$16.75B
Q2 '25$17.08B
Q3 '25$19.39B
Q4 '25$29.88B
Q1 '26$30.88B
Q2 '26$35.8B
Capital expenditure per calendar quarter, US$ billions, from Microsoft's SEC filings.9

Cheaper intelligence is the point; the risk sits in the contracts

Efficiency is exactly what AI adoption needs: every training dollar buys more capability, and capability gets cheaper to build into products people use. The exposure sits with whoever lent against compute staying scarce, from neocloud bondholders to anyone counting Microsoft's OpenAI stake as an asset 1. Meta−0.72% — Meta, down 0.72 percent today pushed the same direction this week, unveiling Muse Spark Open, an open model under a new partnership with Apple whose published setup runs on two paired Mac desktops, pressure on rented compute from the other end 78.

The MAI-Base-2 claim is one company's private benchmark today. The line that settles it, and prices Microsoft's stake and its leases at once, is what Azure customers spend on OpenAI models at the next report.

Deepdive

AI-generated from this story and its cited sources. Not investment advice.

Reader comments

0 comments

    Sign up

    Get your curated digest

    After email confirmation, you will receive a daily digest of the most relevant news that matter to your portfolio