Microsoft Told Enterprise Buyers Its New AI Was Trained on ‘Clean Data.’ Its Own Paper Reveals 24.2 Billion Pages From Common Crawl.

Twenty-four point two billion. That is the number of deduplicated web pages from Common Crawl that Microsoft fed into the training pipeline for MAI-Thinking-1, according to the company’s own technical preprint published alongside the model’s debut at Build 2026. Hours earlier, from a stage in San Francisco, Microsoft AI CEO Mustafa Suleyman had introduced the same model to enterprise buyers as built on “clean, commercially licensed data” with “zero distillation.”

Common Crawl provides no licensing guarantees. It provides no author consent mechanisms. It is the exact repository named as Exhibit A in New York Times v. OpenAI, Authors Guild v. OpenAI, and Getty Images v. Stability AI. What Microsoft sold as a compliance breakthrough on Tuesday morning was, by the technical paper published that same afternoon, indistinguishable from the data practices the company was implicitly criticizing.

Seven Models, One Breakup

Build 2026 was supposed to be a declaration of independence. After years as OpenAI’s distribution layer and compute bank, Microsoft unveiled seven in-house AI models built from scratch by its MAI Superintelligence Team, a unit formed in November 2025 under Suleyman’s leadership:

Model	Parameters	Key Claim	Status
MAI-Thinking-1	35B active / ~1T MoE	97% AIME 2025, matches Claude Opus 4.6 on SWE-Bench Pro	Private preview
MAI-Code-1-Flash	5B	51% SWE-Bench Pro at fraction of cost	Rolling out in VS Code
MAI-Image-2.5	Undisclosed	#2 on Arena image editing leaderboard	Live in PowerPoint
MAI-Transcribe-1.5	Undisclosed	SOTA across 43 languages, 5x faster than rivals	GA in Foundry
MAI-Voice-2	Undisclosed	Natural prosody, 15 languages	GA in Foundry

The strategic logic is clear enough. Microsoft owns roughly 27% of OpenAI and holds an exclusive IP license through 2032, but OpenAI has been quietly distributing through AWS and Google Cloud since its contract renegotiation in April. Suleyman told The Verge that the renegotiation was “the pivotal moment,” enabling Microsoft to pursue frontier AI “with our own IP, with our own data, no distillation, training from scratch.”

The numbers back the ambition. MAI-Thinking-1 is a 35-billion-active-parameter Mixture of Experts architecture, approximately one trillion parameters total, with a 256,000-token context window capable of ingesting roughly 600 pages in a single pass. Microsoft claims it scores 97% on AIME 2025, 94.5% on AIME 2026, and 53% on SWE-Bench Pro. In blind evaluations conducted by Surge, an independent human rating partner, evaluators preferred it over Claude Sonnet 4.6 for overall quality.

None of these benchmarks have been independently reproduced.

What the Preprint Actually Says

Within hours of the keynote, researchers reviewing the accompanying technical paper found what the marketing had omitted. MAI-Thinking-1’s data pipeline began with a proprietary web crawler that ingested approximately 1.2 trillion open-web pages. After filtering out piracy domains, adult content, and exact duplicates, the corpus shrank to 794 billion pages. On top of that filtered web crawl, Microsoft layered 24.2 billion deduplicated pages from Common Crawl.

That is the contradiction distilled to its essentials: a keynote promising “clean, commercially licensed data” and a paper describing a pipeline that started from 1.2 trillion scraped pages plus 24.2 billion from a repository offering zero licensing of any kind.

Microsoft’s defense, buried in the technical brief, is that its proprietary crawler respects the Robots Exclusion Protocol, an opt-out mechanism where publishers who do not specifically configure their servers to block Microsoft’s crawler are treated as having consented by default, effectively shifting the burden of action onto millions of individual website owners rather than placing it on a company worth three trillion dollars.

This is not a novel data strategy. It is the industry standard, the same approach deployed by every major lab from OpenAI to Anthropic to Google, and what made Microsoft’s version newsworthy was not the practice itself but the chasm between what the sales team said and what the research team published on the same afternoon.

Why Enterprise Buyers Care

Microsoft’s “clean data” pitch was not accidental positioning. It was the core selling proposition for compliance-sensitive industries, a deliberate play for finance, healthcare, and government procurement teams who were told, explicitly, that unlike the competition, Microsoft carried no copyright exposure. Publication of the preprint demolished that claim within hours.

Enterprise risk calculus works in binaries. Either a model’s training data has been licensed, or it hasn’t. Common Crawl is unambiguously unlicensed, and the “fair use” doctrine that the entire AI industry relies on remains unsettled in federal court, which means a procurement officer at JPMorgan or the Department of Veterans Affairs cannot sign off on “we believe fair use will eventually be upheld” any more than they could sign off on a vendor whose financial audit was pending.

Several prominent technology commentators who had praised the “clean data” claim as a competitive breakthrough retracted their initial assessments within days of the preprint’s publication. Procurement teams at multiple Fortune 500 companies have begun re-evaluating their MAI deployment roadmaps, according to industry analysts.

Independence Has a Price Tag

Strip away the data controversy, and the financial incentive for in-house models is enormous. Microsoft pays OpenAI for both compute access and model licensing, and in Q3 FY2026, while Azure revenue grew 18.3% year-over-year to $82.89 billion in total company revenue, margin pressure from third-party model costs remained a persistent drag that in-house models could eliminate. Analysts at TD Cowen, which maintained a Buy rating with a $540 price target after Build, estimate that self-built models could save Microsoft $2-4 billion annually in licensing fees alone once they reach production parity.

Efficiency numbers are striking if they survive external scrutiny. Microsoft claims MAI-Thinking-1, tuned via its new “Frontier Tuning” capability with enterprise-specific data, outperforms GPT-5.5 on McKinsey’s internal benchmarks at one-tenth the token cost. MAI-Code-1-Flash, at just 5 billion parameters, achieves 51% on SWE-Bench Pro while consuming a fraction of the compute of larger coding models, and both run on Microsoft’s custom Maia 200 silicon, which Suleyman’s team claims delivers 1.4 times the performance per watt of Nvidia’s GB-200.

Teams of fewer than ten engineers built each model, using roughly 50% fewer GPU hours than competing labs. If those numbers hold under external scrutiny, Microsoft has done something genuinely impressive: built a competitive model family at a fraction of the resource cost, a real engineering achievement that was then wrapped in a marketing claim its own researchers contradicted before the keynote audience had left the building.

Strongest Counterargument

Every frontier AI model trains on web-scraped data. OpenAI’s GPT series used Common Crawl, Google’s Gemini used web crawls, Anthropic’s Claude used web crawls, and if fair use is eventually upheld in court, as the legal trend lines in Thomson Reuters v. Ross Intelligence and Andy Warhol Foundation v. Goldsmith suggest it might be, then Microsoft’s data pipeline carries no additional legal risk beyond what every competitor already bears. Marketing was premature, not fraudulent. And Microsoft’s filtering pipeline, which reduced 1.2 trillion pages to 794 billion after removing piracy and adult content, arguably represents more curation than most competitors have disclosed to date.

Fair point. But the controversy is not whether Microsoft’s data practices are worse than the industry’s; it is that the company marketed them as categorically better. “Commercially licensed” and “Common Crawl” cannot coexist in the same data lineage without an explanation that Microsoft has not yet provided.

Limitations

All figures for the training data, including the 1.2 trillion page count and the 24.2 billion Common Crawl component, are self-reported by Microsoft in its technical preprint with no independent audit conducted to date. Benchmark claims, including AIME and SWE-Bench Pro scores, have not been independently reproduced, and Surge evaluations were commissioned and paid for by Microsoft. No formal response from Microsoft has addressed the discrepancy between its marketing and its preprint, and it is possible that “commercially licensed” was intended to describe a subset of the training data rather than its totality, a definitional distinction that would narrow the contradiction without eliminating it.

What It Means

Microsoft built something real at Build 2026: seven in-house models, trained from scratch, competitive on the benchmarks that matter, running on custom silicon at a fraction of the cost, representing a genuine strategic pivot away from OpenAI dependency. But then the company chose to sell this achievement with a claim its own researchers knew was false, or at minimum misleading, and published the proof on the same afternoon.

What you can do: If you are evaluating AI model vendors for regulated industries, request the full training data lineage documentation before procurement. Ask specifically whether Common Crawl or equivalent web-scraped repositories were used. Do not accept “commercially licensed” without a precise definition of what that term covers and what it excludes. If you are a publisher or content creator, check whether your robots.txt blocks Microsoft’s AI training crawlers; if it does not, your content may already be in the training set. And if you are an investor: the MAI models represent a genuine margin opportunity for Microsoft, but the data provenance controversy introduces regulatory risk that has not been priced in.