Through software optimization alone, Microsoft delivers 90% more GPT-4o-family tokens per GPU than a year ago.
Resolved true The record shows it happened or was so.
Factual: Statements about the present or the past, checked against the record: did they hold up?
The evidence is the person's or their organisation's own records.
Said for an organisation they lead or work for; it counts on both records.
What would count as true or false
Judged as of July 2025. True: Microsoft's official records, such as the FY25 Q4 earnings call transcript or investor materials, or other Microsoft technical disclosures, state that software optimization alone produced about 90% more tokens per GPU (for GPT-4o-family models, or as the claim describes) than a year earlier. A figure within a few points (e.g., 85–95%) counts as true. Mostly true: the improvement is confirmed at about 90% but the model-family scope differs slightly, or hardware played some part. Mixed: a substantial gain is confirmed but it is materially different from 90% (e.g., 50–75%), or the gain is not attributed to software alone. Mostly false: the figure is drawn from a narrow benchmark and presented as general. False: the evidence contradicts it, for example a much smaller gain. The claim needs at least a matching Microsoft statement; independent support strengthens it.
Set before the research began.
Evidence (3)
-
E1 Microsoft (MSFT) Q4 2025 Earnings Call Transcript & Audio StockAnalysis.com · 30 July 2025 Own record
“Take, for example, the GPT-4o family of models, which have the highest volume of inference tokens. Through software optimizations alone, we are delivering 90% more tokens for the same GPU compared to a year ago.”
-
E2 Microsoft (MSFT) Q4 2025 Earnings Call Transcript & Audio StockAnalysis.com · 30 July 2025 Own record
“That GPT-4o example I gave is all software, right?”
-
E3 Microsoft conference call: Order surge, Azure in short supply, data center tightness expected to last until 2026 Longbridge (Wallstreetcn article) · 30 October 2025
“For example, this quarter we increased the token throughput of GPT-4o and GPT-5 (the two most widely used models) by over 30% per GPU.”
Reasoning
The transcript of Microsoft's FY25 Q4 earnings call (E1) records the statement that, for the GPT-4o family of models, software optimizations alone delivered 90% more tokens for the same GPU than a year earlier. In the same call (E2), the gain is described as "all software." That is the matching official Microsoft record the resolution criteria require. E3 reports a separate later quarter's gain of more than 30%, which covers a different period and does not contradict the figure. No independent source verifies the 90% figure itself.
Cite this claim
Satya Nadella, “Through software optimization alone, Microsoft delivers 90% more GPT-4o-family tokens per GPU than a year ago.” (said July 2025). Verdict: resolved true. Reliability API, claim c_de3a64440738, method 0.1.0+28da144b, dataset release v1.1.2. https://epistacklabs.com/claims/c_de3a64440738
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Satya Nadella
-
Mixed
“same thing in inferencing as well.”
-
Mostly true
“we’ve been at this in a variety of different forms for a long, long time in terms of building our own silicon.”
Microsoft has been building its own silicon in various forms for a long time.
-
Mostly true
“right as we speak, we have some of the Maia stuff powering GitHub copilot for example.”
As of November 2023, some Maia chips are powering GitHub Copilot.
-
Resolved true
“we built an entire rack which is liquid cooled for Maia.”
Microsoft built an entire liquid-cooled rack for its Maia AI chip.
-
Mixed
“I was able to use a model that is using 500B, I was able to use a 5B, and have it really deliver the same outcome, why would I not use that?”
Verdicts are made automatically; how verdicts are made.