Guaranteed Returns

OpenAI is offering private equity firms a guaranteed minimum return of 17.5% on their investment in a $10 billion joint venture designed to accelerate enterprise AI adoption. TPG is anchoring the deal, with Advent International, Bain Capital, and Brookfield Asset Management joining as co-founding investors. The PE firms would commit roughly $4 billion in preferred equity and receive board seats in exchange for deploying AI across the companies they control. (Source: Yahoo)

The guaranteed floor is significantly higher than typical preferred instruments and has drawn comparisons to the yield promises that sank Terra-LUNA in 2022. The deal structure reveals a fundamental problem OpenAI cannot solve with better models alone: MIT research found that only 5% of enterprise AI deployments achieve measurable business results, with an estimated $30 to $40 billion wasted on failed implementations. McKinsey's own data shows 88% of AI proofs-of-concept never reach production. (Source: BeInCrypto)

The strategic logic is blunt: private equity firms control portfolio companies top to bottom. They can force adoption that organic enterprise sales cannot. David Sacks, co-chair of the President's Council on Science and Technology, described the approach on the All-In Podcast this week: "They're betting they can own the change management around AI. Everyone assumes you throw AI over a wall and a business automatically knows how to use it. What we're seeing is it's pretty difficult." (Source: Sherwood)

Both OpenAI and Anthropic are aggressively courting PE firms because they control enterprise budgets and influence how companies adopt software. The race is growing more urgent as both companies prepare for potential IPOs this year. Goldman Sachs projected this month that by 2030, more than 60% of software industry operating profit could migrate to AI agent systems — a shift that would reshape every SaaS contract on a company's balance sheet. (Source: Bloomberg)

The Trump Administration released a National Policy Framework for Artificial Intelligence on March 20, urging Congress to adopt broad federal preemption of state AI laws and a "light-touch" regulatory approach. The framework explicitly calls for precluding states from regulating AI model development, while preserving states' traditional police powers only for general laws protecting children, preventing fraud, and safeguarding consumers. (Source: White House)

The framework arrives as a wave of state AI accountability laws are taking effect. Illinois's HB 3773, effective since January 1, prohibits the use of discriminatory AI in employment. Colorado's SB 24-205, the first U.S. state law specifically governing high-risk AI in financial services, takes full effect June 30, requiring impact assessments, bias audits, and consumer disclosures for AI-driven lending decisions. The White House framework would override these and similar efforts with industry-led standards and voluntary regulatory sandboxes. (Source: CNBC)

Seven priority areas anchor the framework: 1) child safety with age-assurance requirements, 2) economic growth through streamlined permitting, 3) intellectual property with deference to courts, 4) free speech protections against "government-driven censorship," 5) innovation through regulatory sandboxes, 6) workforce development in AI education, and 7) infrastructure safeguards against energy cost increases. The White House wants to codify the framework into law this year and believes it can generate bipartisan support. (Source: Sullivan & Cromwell)

Critics note the timing: the framework lands the same week as the IAPP Global Privacy Summit in Washington (March 30 - April 2), where FTC Commissioner Mark Meador is set to preview 2026 enforcement priorities across 60+ sessions on AI governance and the state privacy patchwork. The tension between federal preemption and state-level protections is expected to dominate the conference. (Source: Reed Smith)

Google expanded its Personal Intelligence feature to all free-tier U.S. Gemini users in March, less than two months after its paid-only launch in January. The feature connects Gemini to a user's Gmail, Google Docs, Photos, and Search history, allowing the AI to deliver personalized answers drawn from private data. It works across Search, Chrome, and the Gemini app. The feature is opt-in, requiring manual permission grants — but prompts sent to Google's servers may include details drawn from connected apps. (Source: WinBuzzer)

Google says it does not train Gemini on private emails or photos, though it does use prompts and responses to improve the model — meaning data from connected apps surfaces in prompts even if raw Gmail content is not used for training.

A Malwarebytes survey found that nine in 10 respondents expressed concern about AI using their data without consent. The Washington Post flagged the privacy risks when the feature launched for paid users in January, noting the scale of data exposure when an AI assistant has access to years of personal email, documents, and photos. (Source: Washington Post)

The expansion to free users dramatically increases the surface area. Google's free-tier user base dwarfs its paid subscribers.

The rollout is limited to U.S. users with personal Google accounts, with no international schedule confirmed. Regulatory scrutiny from the FTC and state privacy authorities is expected, arriving the same week as the Trump administration's push to preempt state AI privacy laws with a federal "light-touch" framework. (Source: HUMAI)

Meta's Fundamental AI Research team released TRIBE v2 on March 26, a foundation model that predicts human brain activity across vision, sound, and language. The model scales to approximately 70,000 brain voxels — a 70-fold increase over the original TRIBE — and can predict neural responses for new individuals and unseen languages without retraining. Meta open-sourced the model weights, codebase, and an interactive demo under a CC BY-NC license. (Source: MarkTechPost)

The training data came from over 700 volunteers who watched movies and listened to podcasts inside fMRI machines, producing more than 1,115 hours of brain activity recordings paired with the stimuli that produced them. TRIBE v2's predictions matched population-level brain activity better than most real scans, which are typically clouded by heartbeats, movement, and noise. The model effectively creates what researchers are calling "digital twins" of neural activity. (Source: Neuroscience News)

The research implications are significant for neuroscience, but the open-source release raises questions about downstream applications. A model that can predict how neurons fire in response to what people watch, hear, or read — and do so for individuals it has never scanned — has obvious applications in advertising, content optimization, and surveillance that extend well beyond the stated research goals.

The CC BY-NC license restricts commercial use but does not prevent modification or redistribution by non-commercial actors. (Source: Dig Watch)