Work AGI is the Only AGI that Matters

March 25, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
March 25, 2026

HEADLINES

OpenAI's Risk Disclosure Doc Leaked — Here's What's Inside

CNBC obtained documents that resemble an OpenAI IPO prospectus, though OpenAI says it's a standard private fundraising disclosure, not IPO prep. The document flags heavy reliance on Microsoft as a key risk: "Microsoft is responsible for a substantial portion of our financing and compute," and warns that a termination or modification of that relationship could be devastating — particularly pointed given reports that Microsoft is weighing a lawsuit over OpenAI's AWS partnership. Additional disclosures cover capital expenditure, compute reliance, ongoing Elon Musk litigation, the unusual public benefit corporation structure, and geopolitical risk related to Taiwan. Sources said OpenAI is also seeking a further $10B from investors on top of the $110B already raised, with the round expected to close by month end.

SpaceX Could File IPO Paperwork This Week

SpaceX is finalizing its prospectus and could file with the SEC as soon as this week, with the stock expected to begin trading in June — making xAI the first of the three major AI startups to go public. SpaceX is targeting $75B, up from a planned $50B, which would be the largest IPO in history by a wide margin and exceed all IPO capital raised last year combined. At the last private valuation of $1.25T, the debut would be as roughly the 12th largest company in the world. Unconventional features include 20% of shares going to retail (double the usual 10%), no standard six-month insider lockup, and all five major investment banks preparing pitches with none officially hired. The skeptical read: xAI is a fourth-rate lab and SpaceX is raising primarily to buy GPUs.

Pre-IPO AI ETF Trading at 1,200% Premium

Fundrise's innovation fund — holding shares in SpaceX, Anthropic, and OpenAI — is up 1,500% since listing, saw a 64% single-day jump, and was halted twice for volatility. By end of day it was valued at more than 16x the actual value of its holdings, implying an Anthropic valuation near $5 trillion against a February raise at $380B. This is purely a market structure problem: no new shares can be created, so the fund completely detaches from underlying value. Jack Shannon of Morningstar: "With the implied valuations when you have this premium, your upside is gone." Matt Malone of Opto Investments noted it shows exactly why staying private so long is rough on retail investors — by the time the public gets access, the price "doesn't really make sense."

SoftBank Is Betting the Company on OpenAI — Literally

The FT reports SoftBank is testing its own self-imposed 25% loan-to-value borrowing limit after committing another $30B to OpenAI. Last year's $22.5B investment already required selling all NVIDIA holdings and taking out margin loans against ARM stock. CFO Yoshimitsu Goto: "I don't deny the possibility in the future that we may temporarily go beyond 25%." With $50B in total commitments this year, SoftBank stock has fallen 45% and credit default swaps are at their highest in almost a year. More than ever, Masayoshi Son is betting the company on OpenAI.

Sam Altman Steps Down from Helion Board as OpenAI Energy Deal Takes Shape

Altman is stepping down as chairman and board member of Helion Energy as the fusion startup works on a supply deal with OpenAI. Axios reports the deal would guarantee OpenAI 12.5% of energy initially produced, scaling to 5 gigawatts by 2030 and 50 gigawatts by 2035. Altman personally led Helion's $500M Series E in 2021 — the largest ever venture investment in a nuclear fusion startup at the time — and says he's recusing himself from negotiations while retaining a financial interest.

Judge Calls Pentagon's Treatment of Anthropic "Troubling"

Federal Judge Rita Lin heard Anthropic's injunction application in Northern California and was unimpressed. She called the Pentagon's actions "troubling," saying it appeared the government was punishing Anthropic for bringing public scrutiny to the dispute — a First Amendment violation. The Pentagon's lawyer tried to walk back Hegseth's tweet (which said no contractor "may conduct any commercial activity with Anthropic"), claiming the designation only applied within Pentagon systems. The judge pushed back sharply on the "kill switch" argument too: if being a stubborn IT vendor who "insists on certain terms and asks annoying questions" is enough to be designated a supply chain risk, "that seems a pretty low bar." An order is expected in the coming days.

MAIN STORY

Work AGI Is the Only AGI That Matters

The hallmark of 2026 has been inflection-point-scale change, and nowhere has the AI race sharpened more acutely than at OpenAI, which is watching an insurgent Anthropic dominate the enterprise and coding conversation. For most of 2025, OpenAI let a thousand flowers bloom — competing more like Google than a focused lab. That changed with the code red in December, and this week brought the clearest signal yet of how serious the refocus is: a major reorg, a new model in the pipeline, Sora shut down, ads overhauled, and the product division renamed "AGI Deployment." Taken together, it's a declaration that for the frontier labs right now, the only type of AGI that matters is Work AGI.

SIDE QUESTS GIVE WAY TO AGI DEPLOYMENT

The Reorg
The product division is now called "AGI Deployment."
Sam Altman told staff he's narrowing his own role to raising capital, supply chains, and the data center buildout. Safety moves into the research organization under Chief Research Officer Mark Chen; security moves to the scaling organization under President Greg Brockman. With Altman stepping back from day-to-day commercial strategy, Fidji Simo is effectively in the driver's seat — and her team has been renamed AGI Deployment. Last week she confirmed the next project: combining ChatGPT, Codex, and the Atlas browser into a desktop superapp. Her framing: "When new bets start to work, like we're seeing now with Codex, it's very important to double down on them and avoid distractions."

Spud
"Things are moving faster than many of us expected."
Altman told staff pretraining is complete on a new model codenamed Spud, and that they expect "a very strong model in a few weeks" that "can really accelerate the economy." The phrase drew immediate attention — one observer called it "either AGI or a really confident marketing team." Given how carefully OpenAI has avoided bombastic over-promising since GPT-5's misaligned launch expectations, it reads as much as a rallying call to internal staff during a moment of intense transition as it does an external signal.

Sora Is Dead
"The first time OpenAI publicly chose not to do something it clearly wanted to do."
The mandate to end side quests has its first and most visible victim: the Sora app and all OpenAI video model products. The WSJ reported staff were surprised by how compute-hungry Sora was relative to demand — the freed compute goes to Spud. The Sora research team isn't being cut: they're pivoting to "world simulation research, especially as it pertains to robotics," with the prize described as "automating the physical economy." One casualty: the billion-dollar Disney partnership. Disney had chosen to partner rather than sue after Sora's October launch; their statement was gracious. Worth noting: the death of Sora is not an indictment of AI video. When Sora launched it was just Midjourney and Runway in the space — now there are over a hundred companies with marketing departments, agencies, and studios fully committed.

Ads and Instant Checkout
"$680 billion market vs. $40 trillion of automatable knowledge work."
Sora is gone but ads are still happening. OpenAI has hired former Meta executive Dave Dugan as VP of Global Ad Solutions; the pilot is over and ads roll out to all free and Go subscribers soon. The problem: no modern ad sales platform, sales run via phone calls and spreadsheets, and multiple agency executives unable to prove ChatGPT ads delivered measurable results for clients. On shopping, Instant Checkout has been shelved. OpenAI's statement: "We've found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide." The implicit question from observers: when does OpenAI kill the ad side quest too? The $680B ads market is dominated by incumbents; the roughly $40 trillion market of automatable knowledge work is largely untapped.

What AGI Actually Means Right Now
"The odds of 100,000 agents building Nvidia is zero percent."
The renaming of the product division to "AGI Deployment" triggered a fresh round of debate about what AGI actually means. On the Lex Fridman podcast, Jensen Huang was asked when AI could start, grow, and run a billion-dollar technology company. His answer: "I think it's now. I think we've achieved AGI" — meaning OpenClaw could plausibly generate a dotcom-era novelty app that briefly gets a few billion users. But when Fridman pressed him, Jensen acknowledged the limits: "The odds of 100,000 of those agents building Nvidia is zero percent." Benjamin Todd at 80,000 Hours offered a more careful answer: current AI is superhuman at some cognitive tasks but still worse than most humans at others — "impressive, general, but not yet AGI." The more useful frame: we effectively have task-level AGI. Almost anything specific and discrete, AI does very well. The breakdown comes with long chains of tasks requiring minimal human oversight — exactly what Work AGI is trying to solve. As Ethan Mollick suggested, maybe we should just retroactively agree with Tyler Cowen that o3 was AGI and stop arguing — because AGI alone is clearly not enough for transformation. The stories of OpenAI and Anthropic partnering with consulting and private equity firms to help enterprises actually deploy AI make that obvious.