The Superintelligence roadmap

"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."

From a $500 billion restructuring and a 27% Microsoft stake to a five-layer safety architecture and a $1.4 trillion compute build-out, OpenAI has laid out its most detailed roadmap yet for moving beyond AGI to superintelligence—a path measured not in philosophical debates but in concrete milestones: an intern-level AI researcher by September 2026, a fully autonomous researcher by March 2028, and superintelligence within the decade.

The Route to Superintelligence

On October 28, 2025, Sam Altman and OpenAI Chief Scientist Jakub Pachocki delivered a rare roadmap livestream that revealed the company's internal timeline for the path to superintelligence. The milestones are unusually concrete: an AI research intern capable of end-to-end research (designing experiments, executing analyses, interpreting results, and proposing follow-up work) is expected by September 2026, and a fully autonomous AI researcher by March 2028.

Pachocki's rationale is grounded in a simple but powerful metric: "A good way to judge this is to look at how complex a task the model can solve, and how long it can work continuously". GPT-3 handled tasks lasting tens of seconds; GPT-4 can manage minutes to five hours. The next step is systems that can think for days, mobilising entire data centres to solve a single scientific problem.

The company's timeline for superintelligence itself is even more ambitious. Jakub Pachocki has indicated that deep learning systems could reach superintelligence within a decade, while SoftBank CEO Masayoshi Son has suggested a far more aggressive timeline of two years. Son, whose firm is a major OpenAI investor, has stated that "AI is already designing OpenAI's next model" and that "engineers will eventually no longer be capable of designing the next generation of models".

The Structure: A Nonprofit Foundation, A Public Benefit Corporation

The restructuring completed in October 2025 created a dual-layer structure that Altman described as "unprecedented in its scale, and not just in one dimension". At the top sits the OpenAI Foundation, a nonprofit with an equity stake valued at approximately $130 billion, making it one of the wealthiest nonprofits in the world. The Foundation "remains in control of the for-profit", with control over the board and the power to ensure the company's mission "benefits all of humanity".

Below it sits the OpenAI Group PBC, a public benefit corporation valued at $500 billion, in which Microsoft holds a 27% stake worth roughly $135 billion. The deal keeps the two companies intertwined until at least 2032, with OpenAI committing to purchase $250 billion in Azure cloud services. As Bret Taylor, chair of the OpenAI Foundation's board, put it: "The nonprofit remains in control of the for-profit, and now has a direct path to major resources before AGI arrives".

The Safety Architecture: Five Layers of Alignment

OpenAI's approach to superintelligence safety is not a vague commitment but a five-layer framework designed to make AI thinking traceable and controllable. The layers are:

  1. Value Alignment addresses the foundational question: where do the AI's values come from, and can it understand high-level human goals?

  2. Goal Alignment tests whether the AI accurately interprets and executes specific tasks.

  3. Reliability ensures stable output on simple tasks and honest acknowledgement of uncertainty on complex ones.

  4. Adversarial Robustness maintains stability in the face of malicious attacks.

  5. System Safety defines clear boundaries for data access and device control.

The most distinctive feature is the emphasis on Chain of Thought Faithfulness—making the model's internal reasoning process traceable without distorting it. As Jakub explained: "We believe that preserving the privacy of the model's chain of thought is one of the best ways to understand it". The design is deliberately restrained: rather than forcing models to produce "perfect" reasoning, they are allowed to think privately before being reviewed, with summary mechanisms providing visibility without interference.

The Compute Commitment: $1.4 Trillion

OpenAI has outlined plans for a staggering $1.4 trillion investment in computational infrastructure. This includes the "Stargate" data centre project in Abilene, Texas, with a target of reducing per-gigawatt compute costs to $20 billion within five years. As Altman put it: "We are building a factory that can produce one gigawatt of compute per week. Not a virtual platform—real infrastructure. Land. Workers. Energy. Cooling systems".

The company has already committed to more than 30 gigawatts of compute capacity, with financial obligations approaching $1.4 trillion . Altman believes that if they can provide this at sufficiently low cost, "we can support the entire society's demand for AI".

The GFN Context

For Global Future Nexus, OpenAI's superintelligence roadmap underscores both the promise and the governance challenge of the AGI era. The shift from nonprofit to public benefit corporation, the five-layer safety architecture, and the unprecedented compute build-out all point to a future where superintelligence is not a distant horizon but a planned engineering milestone. Yet the concentration of capability in a single entity, the $1.4 trillion infrastructure bet, and the timeline measured in months rather than years all raise questions that GFN's mission is designed to address: who controls the infrastructure of intelligence, and who benefits from its outputs? The roadmap is not a prediction—it is a plan. The question is whether the governance frameworks can keep pace.

Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)

Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)

Nicolas de Loisy

Advisory specialized in logistics, transportation, and supply chain management.

http://www.scmo.net
Previous
Previous

The AGI definition consensus

Next
Next

AGI's 10 cognitive domains