OpenAI

Theory of Change

OpenAI's stated theory of change has shifted repeatedly. The current mission (from the 2024 990 filing) is: "ensure that artificial general intelligence benefits all of humanity" -- notably removing the word "safely" that appeared in every previous filing.

The original 2015-2017 mission was far more specific: "advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return" with commitments to "openly share our plans and capabilities" and "build safe AI technology." The 2018 Charter added the "assist clause": if a safety-conscious project gets close to AGI first, OpenAI would "stop competing with and start assisting this project."

In practice, Sam Altman describes OpenAI as a consumer technology company. Asked whether a 1-billion-user destination site or the state-of-the-art model is more valuable in five years, he answered "the 1-billion user site." His vision: bundled subscriptions, sign-in identity, hardware devices, an expanding consumer bundle. He compares the AI model to the transistor: "it on its own will not be a differentiator."

OpenAI's own explanation for the evolution: "The words and approach changed to serve the same goal -- benefiting humanity." Altman on the nonprofit structure: "If I knew everything I knew now, of course we would have set it up differently... We literally had no idea we were ever going to become a company."

What They Do

Products: ChatGPT (500M+ users, 20M+ paid subscribers), GPT model family (API), DALL-E, Sora (video generation), Operator (AI agent), Prism (scientific writing), deep research agent. Revenue approximately $12.7B annualized (2025).

Safety research: Collaborated with Apollo Research on anti-scheming training (reduced covert actions from 13% to 0.4% in o3, though rare serious failures persist and researchers conclude the method "is not sufficient for future models"). Published deliberative alignment work. Maintains a Preparedness Framework (updated April 2025).

Safety framework changes: The Preparedness Framework v2 (April 2025) dropped persuasion and manipulation as tracked risks. Lowered the deployment threshold: now considers releasing "high risk" models and even "critical risk" models if a competitor has already released a similar model. GPT-4.1 shipped without any safety report. GPT-4o safety testing was squeezed into one week.

Military: $200M one-year DoD contract (July 2025). Classified-network deployment (February 2026), announced the same day Anthropic was labeled a "supply-chain risk" for refusing to permit mass surveillance and autonomous weapons use. OpenAI's contract includes anti-surveillance language, but the contract text has not been released. Former Pentagon officials and national security lawyers describe the published excerpts as having "many potential loopholes" -- the words "intentionally" and "deliberately" track intelligence community legal definitions that permit the surveillance they appear to prohibit. OpenAI originally banned all military use in its terms of service; this prohibition was silently removed in January 2024.

Infrastructure: Stargate Project ($500B joint venture with SoftBank/Oracle/MGX), $300B Oracle cloud commitment, and approximately $1.4T total infrastructure commitments over 8 years.

Acquisitions: io (Jony Ive's AI hardware, $6.5B), Statsig ($1.1B), Torch (healthcare, $60M), and others. Building products to compete with GitHub, LinkedIn, and other established platforms.

Key People

Sam Altman (CEO, co-founder): Former Y Combinator president. Fired by the board in November 2023 for not being "consistently candid," reinstated 5 days later after 738 of 770 employees threatened to resign. Multiple former colleagues -- including CTO Mira Murati, Chief Scientist Ilya Sutskever, and board members -- described patterns of dishonesty and manipulation. Signed equity clawback provisions threatening departing employees, then publicly denied knowledge of them. The Economist: "a visionary with a trustworthiness problem."

Safety team exodus (2024-2025): The single most revealing finding. Within 18 months, nearly every senior safety-focused leader departed:

Ilya Sutskever (co-founder, Chief Scientist) -- founded Safe Superintelligence Inc.
Jan Leike (superalignment co-lead) -- joined Anthropic. Public statement: "safety culture and processes have taken a backseat to shiny products"
John Schulman (co-founder, alignment lead) -- joined Anthropic
Mira Murati (CTO), Bob McGrew (CRO), Barret Zoph (VP Research) -- all departed the same day
Miles Brundage (Head of AGI Readiness): "Neither OpenAI nor any other frontier lab is ready"
Daniel Kokotajlo (governance): forfeited ~$2M equity rather than sign NDA. Lost confidence OpenAI would "behave responsibly around the time of AGI"

Both the Superalignment team (May 2024) and AGI Readiness team (October 2024) were dissolved.

Current leadership: Jakub Pachocki (Chief Scientist, replacing Sutskever) and Mark Chen (CRO) are described as capabilities-focused. Board chair Bret Taylor (former Salesforce co-CEO). Board includes former NSA Director Paul Nakasone. CPO Kevin Weil holds a military commission as lieutenant colonel in the U.S. Army.

Money and Incentives

Revenue: ~$3.7B (2024), ~$12.7B annualized (2025). ChatGPT subscriptions dominate revenue. Loss of ~$5B (2024), projected $8B operating loss (2025). Cash-flow-positive targeted for 2029. Revenue target: $200B by 2030.

Valuation and funding:

Oct 2024: $6.6B raise at $157B valuation
Mar 2025: $40B at $300B (SoftBank $30B lead)
Oct 2025: $10B employee share sale at $500B
Feb 2026: $110B at $840B (Amazon $50B, SoftBank $30B, Nvidia $30B)

The October 2024 round came with a condition: the money converts to debt unless OpenAI restructures into a for-profit entity. This is the mechanism that forced the governance change.

Infrastructure commitments: ~$1.4T over 8 years. Carnegie analysis: to justify these, OpenAI needs revenue growth comparable to early Google (190% annualized), reaching ~$2T by 2030 -- nearly 2% of global GDP. This is far from guaranteed given competition from Google, Anthropic, Meta, and others.

Microsoft relationship: $13.75B+ invested, 27% equity stake. Azure is OpenAI's primary compute provider and largest cost. OpenAI committed to $250B of Azure purchases. Revenue sharing: 20% to Microsoft until AGI achieved. The relationship creates deep mutual dependence.

Corporate structure: Nonprofit (2015) -> Capped-profit subsidiary with 100x cap (2019) -> Cap quietly increased 20%/year (revealed by press in 2023, not by OpenAI) -> Cap proposed for elimination (2024) -> PBC with nonprofit retaining 26% (October 2025). Each transition reduced governance constraints. Profit caps have been eliminated. The PBC structure has no legal precedent for enforcing mission obligations over profit.

Foundation: Holds 26% stake (~$130B+). Pledged $1B in grants for 2026 -- less than 1% of assets, well below the standard 5% endowment payout. From 2019-2024, the nonprofit was essentially dormant ($44K revenue in 2022). Appoints all PBC board members but actual exercise of governance authority is untested.

Incentive analysis: Every major financial relationship -- Microsoft ($250B Azure contract), SoftBank ($30B+ invested), Amazon ($50B invested), Oracle ($300B cloud contract), Disney ($1B), military ($200M+), and 20M paying subscribers -- creates pressure toward faster deployment and against safety-motivated slowdowns. The $840B valuation implies expectations of massive revenue growth that cannot be achieved while exercising significant caution. Altman's personal equity position remains undisclosed.

What Others Say

Jan Leike (former superalignment co-lead, now at Anthropic): "Building smarter-than-human machines is an inherently dangerous endeavor. OpenAI is shouldering an enormous responsibility on behalf of all of humanity. But over the past years, safety culture and processes have taken a backseat to shiny products... Over the past few months my team has been sailing against the wind. Sometimes we were struggling for computing resources and it was getting harder and harder to get this crucial research done."

Helen Toner (former board member): Altman "withholding information, misrepresenting things, and in some cases outright lying to the board" for years. Two executives described "psychological abuse." "How scared people are to go against Sam" -- employees feared retaliation.

Former Pentagon officials (The Intercept, March 2026): Brad Carson (former Under Secretary of the Army): "I'm not confident in the language at all." A former Pentagon AI official argued the safeguards amount to trusting OpenAI's and the Pentagon's leadership. Alan Rozenshtein (former DOJ): "It's quite possible that OpenAI understands that these red lines are fake, but has written a contract to give them some PR coverage."

Stuart Russell (UC Berkeley, Nobel-adjacent AI researcher): "AI CEOs claim they know how to build superhuman AI, yet none can show how they'll prevent us from losing control... they admit the risk could be one in ten, one in five, even one in three."

Ed Zitron (critic): Makes the financial argument that OpenAI's business model is unsustainable, that Altman cannot articulate how it works, and that safety rhetoric masks purely commercial motivation.

In defense: Adam Holter (drawing on YouTuber Theo's analysis) argues most OpenAI Files accusations are overblown: Y Combinator title was a clerical error, indirect equity stakes are meaningless, Anthropic criticism is conflicted. Acknowledges equity clawback was genuinely bad but notes it was fixed. The Bret Taylor board response did not dispute Toner's specific claims but stated the independent review concluded the firing "was not based on concerns regarding product safety."

What's Absent

No published safety budget or safety-to-capabilities spending ratio
No safety headcount data (absolute or proportional)
No published text of the DoD contract
No accounting of what happened to the superalignment team's work or the promised 20% compute allocation
No independent audit of safety claims
No candid explanation from current leadership of why nearly every senior safety leader departed
No public information about Altman's current equity position
OpenAI.com entirely inaccessible (403 errors on all pages) -- all characterizations come from third-party descriptions
FLI Safety Index Existential Safety grade: D (all labs fail this domain)