As OpenAI restructures, Microsoft locks in long-term gains

Hello, Beatrice Nolan here, filling in for Jeremy Kahn, who is traveling back from the Coins2Day Global Forum in Riyadh today. In big AI news, OpenAI and Microsoft announced that they had reached an agreement on the future of their partnership that allows OpenAI to complete a long-awaited corporate restructuring.

The arrangement converts OpenAI’s previous for-profit arm into a public benefit corporation that can issue traditional equity and will give shareholders a potentially more prominent voice in OpenAI’s governance—two changes that were seen as critical for OpenAI to continue to raise the billions of dollars of capital it will need to build more advanced AI models, construct massive datacenters, and continue its push to become a key technology platform for consumers and enterprises.

Under the deal, the new OpenAI Group PBC will remain controlled by the nonprofit OpenAI Foundation. The nearly year-long negotiations with Microsoft, which reportedly caused significant tension between the two companies, had been OpenAI’s main obstacle to completing the restructuring. And at first glance, Microsoft appears to have extracted significant concessions from the AI lab.

The tech giant—which has poured more than $13 billion   into OpenAI since 2019—will take a 27% stake in OpenAI. Given OpenAI’s current valuation, Microsoft’s position will be worth about $135 billion. It will also retain access to OpenAI’s technology through 2032, including any models that reach the milestone of artificial general intelligence (AGI). Whether models have passed this threshold—which OpenAI had previously defined publicly as an AI system capable of performing most economically-valuable cognitive tasks as well or better than a human—will also now be verified by an independent expert panel.

Previously, OpenAI alone could decide when AGI had been reached, which was seen as a possible leverage point to end or change Microsoft’s rights under their partnership. This arrangement reportedly raised tensions, with Microsoft reportedly worried that OpenAI could prematurely declare AGI, using a high-performing AI model as the milestone, which would have major financial and IP implications for both companies. So this decision independent panel is a win for Microsoft.

The deal also lets Microsoft pursue AGI independently, or with third parties, while still requiring OpenAI to share many of its research techniques and breakthroughs. Under the new deal, Microsoft retains access to much of OpenAI’s underlying research methods and systems, although the company will not have access to OpenAI’s consumer hardware, or the model weights and core architectural details of any models considered “research.” (It will retain rights to these key technical details for OpenAI’s production models until 2032.)

Still, this gives Microsoft visibility into things like OpenAI’s model training infrastructure and optimization methods, along with the opportunity to take what it’s learned from OpenAI’s research methods and apply that knowledge to develop its own AGI models. Potentially complicating this further is Microsoft’s expanding partnership with Anthropic, with Claude now available in Microsoft 365 and Excel.

Microsoft did give up its cloud exclusivity with OpenAI, which technology analyst Zeus Kerravala called a “major concession on its part.” However, he also noted that the company had secured several critical, structural concessions from OpenAI in return that outweighed this.

“These concessions ensure the longevity and value of Microsoft’s investment,” he told Coins2Day. “Essentially, Microsoft traded cloud compute exclusivity, something it was struggling to meet anyway, for technological certainty and long-term IP access.”

Investors seemed to agree as the company’s stock rose 2%, pushing its market valuation past the $4 trillion mark again the day before the company reports its Q3 earnings.

A regulatory win for OpenAI

For its part, OpenAI will now have access to the full funding promised by investors, including SoftBank, Thrive, and other venture capital firms that had been contingent on the restructure. The move also positions OpenAI to raise additional capital more easily in the future.

The company also appears to have cleared a critical regulatory hurdle. Following the news of the deal, the Attorney General of Delaware, Kathy Jennings, announced that her office has issued a “Statement of No Objection” to the proposed corporate recapitalization.

The Attorney General of California, Rob Bonta, told Coins2Day in a statement that it had “secured concessions that ensure charitable assets are used for their intended purpose, safety will be prioritized, as well as a commitment that OpenAI will remain right here in California.” As a result, Bonta said his office would “not be in court opposing OpenAI’s recapitalization plan.”

This is a blow for several nonprofits that have been campaigning against the restructuring, arguing that OpenAI had drifted from its core mission of developing AGI in a way that “benefitted all humanity” and that it had prioritized shipping products over AI safety.

These non-profit groups had been lobbying the Attorney Generals’ offices to block the deal. Advocacy groups, nonprofits, and some former OpenAI employees, as well as OpenAI co-founder-turned-bitter-commercial-rival Elon Musk, have openly opposed the restructuring on various ground. Musk has argued that the move is evidence that OpenAI CEO Sam Altman and cofounder Greg Brockman deceived him when he made the initial multi-million dollar donations that established the lab. Others have argued that the restructuring risks channeling profits that should have had a charitable public purpose into the pockets of OpenAI’s venture investors, cofounders, and employees.

OpenAI has tried to dampen some of these objections by pledging that its non-profit OpenAI Foundation will make large grants for charitable purposes, including a just-announced $25 billion commitment to projects that aim to improve health and cure diseases or that aim to increase societal resilience to some of the disruptions AI is likely to cause, including potential widespread job loss.

Nonetheless, with regulatory approval secured and Microsoft’s concerns addressed, OpenAI finally has its for-profit structure—and billions of dollars in capital. And Microsoft keeps most of what it found valuable in the OpenAI partnership, while securing a big equity stake that could prove a windfall for its own shareholders.

With that, here’s more AI news.

Beatrice Nolan

[email protected]

FORTUNE ON AI

Open-source AI is ‘China’s game right now’—and that’s a problem for the U.S. And its allies, Andreessen Horowitz partner says — Beatrice Nolan

Now we know that AI won’t take all of our jobs, Silicon Valley has to fix its fundamental mistake: Automation theater has to end — Joel Hron

Qualcomm CEO warns that ‘everybody’s playing to win’ when it comes to an AI bubble—but it’s still too early to tell who will succeed — Beatrice Nolan

After Microsoft invested $13 billion into OpenAI, its AI chief is slamming erotica features like ChatGPT’s: ‘This is very dangerous’ — Sasha Rogelberg

EYE ON AI NEWS

Elon Musk launches Grokipedia as a new rival to Wikipedia. Elon Musk has touted his new venture, Grokipedia, an AI-driven encyclopedia, as an unbiased alternative to Wikipedia. Early pages resemble Wikipedia’s format but, according to the Washington Post, are presented with a more right-leaning tone. The outlet also identified multiple factual errors. Musk has raised issues with Wikipedia before, publicly calling out what he considers the site's leftward shift. The site, which launched this week, currently hosts around 885,000 articles, far less than Wikipedia’s more than 8 million. Read more in the Washington Post.

Claude adds Excel integration and real-time market data tools. Anthropic has rolled out a finance-focused upgrade for its Claude chatbot, adding direct integration with Microsoft Excel and seven new real-time data connectors. Anthropic's Claude can now analyze, edit, and generate spreadsheets directly within Excel, offering financial professionals a more interactive way to work with data. The new connectors link Claude to key financial platforms—including Moody’s, the London Stock Exchange Group, and MT Newswires—allowing it to access live market updates, earnings call transcripts, and investment research. Anthropic also introduced six new “Agent Skills” tailored to finance, enabling Claude to produce reports, model cash flows, and generate company profiles automatically. Read more here.

Qualcomm enters AI chip race with data center processors. Qualcomm has announced a major move into the AI data center market with two new accelerator chips designed to challenge Nvidia and AMD’s dominance. The chips, designed to power AI inference rather than training, can fill a full, liquid-cooled server rack and are built on Qualcomm’s Hexagon neural processing units, which are used in its smartphone chips. With demand for AI computing expected to drive up to $6.7 trillion in data center spending through 2030, Qualcomm’s entry marks a significant expansion beyond its traditional mobile chip business. Shares of Qualcomm surged 11% following the announcement. Read more in CNBC.

OpenAI urges U.S. To boost energy output to stay ahead in AI race. OpenAI is calling on the U.S. Government to dramatically expand national energy production, warning that the country risks falling behind China in the global AI race without massive new power investments. In an 11-page submission to the White House Office of Science and Technology Policy, the company urged the U.S. To commit to building 100 gigawatts of new energy capacity each year — nearly double what the nation added in 2024. OpenAI noted that China added 429 gigawatts last year compared with the U.S.’s 51 gigawatts, creating what it called an emerging “electron gap.” Read more in CNBC.

EYE ON AI RESEARCH

Another MIT report says the benefits of AI may not be as clear cut as forecast. This time in coding. A new MIT Technology Review Insights report, conducted in collaboration with Snowflake, has found that 77% of data engineers are facing heavier workloads despite the widespread adoption of AI tools meant to boost productivity. The survey of 400 senior technology executives found that 83% of organizations have already deployed AI-based data engineering tools, but 45% cite integration complexity and 38% report tool sprawl and fragmentation as major adoption challenges. While AI is automating many data tasks, the proliferation of disconnected systems has created a productivity paradox where individual tasks are faster, but overall workflows are slower, according to the report. Data engineers now spend 37% of their time on AI-related projects, up from 19% two years ago, and expect that to reach 61% within two years. 

AI CALENDAR

Nov. 10-13:  Web Summit, Lisbon. 

Nov. 26-27:  World AI Congress, London.

Dec. 2-7: NeurIPS, San Diego.

Dec. 8-9:  Coins2Day Brainstorm AI San Francisco. Apply to attend here.

BRAIN FOOD

Are AI web browsers a security risk? After OpenAI launched its much-anticipated web browser last week, I wrote about some of the risks around prompt injections. Since then, perhaps unsurprisingly, more security risks have emerged. Security firm LayerX discovered a potentially major vulnerability in the ChatGPT Atlas browser that allows attackers to inject malicious instructions into ChatGPT’s memory. Using a Cross-Site Request Forgery (CSRF) attack, hackers can exploit a logged-in user’s session to implant hidden instructions that persist across devices and browsers. Once infected, ChatGPT may unknowingly execute these instructions, potentially deploying malware. LayerX’s tests showed that the ChatGPT Atlas browser is especially vulnerable, blocking only about 5.8% of phishing attacks they tested, making users up to 90% more exposed compared to Chrome or Edge, which blocked roughly half of such attacks. Cybersecurity is always somewhat of a cat and mouse game, with companies identifying and then patching security flaws, but the scale of the security risks begs the question if AI browsers are just too risky to trust with the kind of deep system access they require to be useful.