The recent restructuring of the OpenAI–Microsoft relationship is, in one sense, just a business story: a major AI company loosening an exclusive cloud dependence, capping revenue-share payments, and diversifying toward other hyperscalers. Viewed from the standpoint of the path to beneficial superintelligence, however, it is much more than that. It is a live demonstration of various deeper governance issues. An organization attempting to create AGI cannot safely be structured like a clever variation on a normal venture-backed tech startup, because the gravitational pull of capital, compute, cloud distribution, and enterprise revenue will gradually force it to become exactly that.
According to what we see in the news, the revised agreement allows OpenAI to serve customers across any cloud provider, while Microsoft remains the primary one. Microsoft’s license to OpenAI model IP continues through 2032 but is no longer exclusive. OpenAI will continue paying Microsoft a revenue share through 2030, reportedly at the same 20% rate but now subject to a total cap, while Microsoft will no longer pay OpenAI a revenue share when customers access OpenAI models via Azure. OpenAI had already been diversifying through major Amazon arrangements, including an expanded AWS agreement and Amazon’s broader strategic investment.
This all totally makes sense from a standard tech industry perspective… HOWEVER, this shouldn’t blind us to the fact that it’s not merely a tactical adjustment. It is the next step in the unwinding of a structure whose central weakness was already visible: dependence on one dominant strategic partner for compute, distribution, capital, and market access as regards what aims to be the first human-level AGI and then superintelligence.
In the short run, the Microsoft relationship gave OpenAI the resources to scale faster than nearly anyone else. In the long run, it created precisely the dependency problem any serious AGI governance structure must avoid. If the goal is to build a giant AI product company, such dependencies are normal. If the goal is to create the technological basis for the Singularity in a way that benefits humanity broadly, they are obviously extremely risky.
One striking thing is how predictable all of this was. (And not just theoretically – a lot of us actually DID predict this sort of unwinding would happen, in one form or another, several years ago….)
OpenAI began with a nonprofit mission to develop AGI for the benefit of humanity. It then created a capped-profit structure intended to reconcile mission protection with capital formation. But … well…
The cap was so high that, in practical terms, it behaved much like ordinary startup upside.
The nonprofit lacked independent operational capability.
The core technology ceased to be meaningfully open. (There was never any clear or practical dedication to open source in the first place.)
The organization became tightly intertwined with a single corporate partner.
The governance documents may have described a special structure, but the economic reality increasingly resembled a standard proprietary AI company with a very large strategic cloud investor.
The new arrangement looks like a move toward a more conventional mature-startup posture: more cloud flexibility, more partner optionality, capped legacy obligations, less dependence on any one platform, more room to optimize enterprise distribution.
From a business point of view this is utterly sensible. OpenAI needs to serve customers where they are, avoid being boxed in by a single cloud, and preserve optionality against Amazon, Google, Oracle, and others.
Well, yes, yes, yes. Once the mission has been subordinated to conventional business pressures, the organization naturally converges toward conventional business logic. It stops being a constitutional experiment for beneficial AGI and becomes a high-growth AI platform company optimizing capital, distribution, revenue share, and strategic leverage.
There is nothing wrong with that if you are building enterprise software, or retail web services, or whatever. But AGI is not just another kind of software.
AGI, and especially the path from AGI to superintelligence, does not fit comfortably inside a normal business-shaped box.
A superintelligence-shaping organization is not just deciding pricing, channel strategy, and cloud margins. It is shaping who has access to transformative cognitive capability, who controls the infrastructure of cognition, how safety decisions get made under pressure, and whether the benefits of the technology diffuse broadly or concentrate in a handful of corporate and geopolitical power centers.
How Could One Do Things Differently?
Take the lessons of the OpenAI trajectory seriously and a different kind of AGI organization starts to come into focus. It’s not thaaaaaat hard to think of a concrete institutional architecture that would let you have a for-profit AGI company along with a nonprofit foundation, and yet would also treat AGI as the civilizational technology it actually is.
As it happens this is something I’ve been thinking through quite a lot in the last few months, as I work with a few colleagues on a new project called OpenBGI … which (yeah, what a coincidence!) is indeed architected to put together for-profit and non-profit structures, with a goal of working toward beneficial AGI and superintelligence leveraging the Hyperon architecture and associated methods like predictive-coding neural networks.
We are still working through the formal details of setting up OpenBGI, but there are a few core principles we’ve had in mind since we started the process.
First commonsensical principle:
structural separation.
Mission control, the open scientific core, commercial execution, and compute partnerships should each occupy their own clearly bounded layer rather than being fused into a single legal-financial entity whose internal contradictions resolve themselves, over time, in favor of whatever is most commercially convenient.
A nonprofit foundation can hold the mission, the open core R&D, the safety and alignment work, and the governance machinery. A for-profit company can hold the deployable systems, products, commercial IP, and investor returns.
A broader consortium of academic, governmental, corporate, and nonprofit collaborators can contribute to a shared open-source AGI stack. Each layer has its own legitimate logic, and none is allowed to swallow the others.
(In an OpenBGI context, we are thinking of these as OpenBGI Labs (the for-profit), OpenBGI Foundation (the non-profit) and a broader OpenBGI Consortium and network.)
Second commensensical principle:
irrevocable openness for the cognitive core.
The foundation must own the core AGI technology and releases it under an irrevocable open-source license. The for-profit company builds proprietary production systems, enterprise tooling, managed services, integrations, and applications on top. This creates a clean division of questions.
The foundation asks whether the AGI approach works, whether it is safe, whether it is scientifically sound, and whether it remains broadly beneficial. The for-profit asks whether it can be deployed reliably, securely, scalably, and profitably. Bug fixes and safety-critical discoveries in the core flow back to the foundation; production engineering and commercial features stay with the company. Both sides win, and neither has to pretend its incentives are something they are not.
Third commonsensical principle:
the nonprofit foundation must be operationally and economically real, not just ceremonial.
This is one of the central lessons from OpenAI. Nominal nonprofit control means little without independent technical staff, independent funding, independent governance, and enough operational capability to keep going if the commercial company or a major partner drifts away from the mission.
A foundation that exists only as a board with a vague oversight role and a financial dependency on the commercial entity will, in any sufficiently high-pressure moment, be governed by the entity it nominally governs.
Fourth commonsensical principle:
negotiate compute partnerships without selling your soul.
The bargain on offer from a hyperscaler is always going to be tempting in the same direction: maximum compute and capital in exchange for deep cloud dependence and strategic entanglement.
A better approach prefers bounded, plural, nonexclusive, and mission-preserving partnerships.
A large cloud or enterprise partner can have a defined joint venture. It can take a standard minority stake in the commercial company. It can hold one board seat with ordinary rights. It can have a time-limited channel exclusivity in a narrow market segment. What it should not have is veto power over technology direction, open-source commitments, future partnerships, financing, or M&A. The foundation must remain untouched. The open core must remain outside the reach of any single partner. A partnership may accelerate the mission; it must not become the mission.
Fifth commonsensical principle:
The path to beneficial superintelligence cannot rely principally on the good intentions of executives, investors, or partners.
Good intentions are not enough once billions of dollars, national strategies, corporate rivalries, and existential technological leverage are in play. The structure adopted has to make capture hard, make mission drift visible, make some commitments irreversible, and allow commercial success without allowing commercial logic to become absolute.
Summing up How One Could Do It Right:
Concretely, an organization following these commonsensical principles could commit to something like the following:
The core AGI substrate is a public-benefit scientific and engineering project, not a proprietary asset to be captured.
The commercial layer is legitimate and important, but it exists to deploy, support, scale, and fund the mission, not to own the mission.
Compute partnerships are necessary, but none may become sovereign.
Investors may receive upside, but not control over the civilizational direction of the technology.
The nonprofit foundation must be operationally real, not ceremonial.
The open-source core, safety discoveries, core bug fixes, and mission-critical alignment mechanisms must survive leadership changes, investor pressure, partner pressure, and market temptation.
This sort of approach, which we are keeping in mind as we finalize the details of OpenBGI, is not anti-business; it is just a more honest and direct form of business. In this sort of arrangement, the for-profit can generate substantial value through production engineering, enterprise deployment, managed services, vertical products, integrations, security, compliance, orchestration, training, support, and domain-specific applications. That is a much cleaner story than “we are a nonprofit-controlled capped-profit entity that is nevertheless mostly closed, dependent on a single giant partner, and forced by events to behave like a normal startup.”
A Note on the Crypto-Tokenomics Path
Everything above is about how a nonprofit/for-profit AGI organization with traditional business structure and traditional corporate partnerships could be structured to remain mission-aligned. It is worth being explicit that this is a different exercise from the one I have been undertaking for the last years with SingularityNET and the ASI Alliance. The ASI:Chain token economy we are gearing up to launch is asking a complementary question: how does one foster beneficial AGI within a crypto tokenomics ecosystem, where coordination, ownership, and incentives are mediated by a decentralized protocol rather than by a constellation of foundations, companies, and investors?
These are not competing approaches. They can and should exist in the same ecosystem, and arguably this is a high-value strategy, because no single institutional form is going to be sufficient to channel something as multidimensional as the emergence of AGI toward broadly beneficial outcomes. A well-architected nonprofit/for-profit hybrid with disciplined corporate partnerships addresses one set of risks. A well-architected token economy addresses a different set: it disperses ownership across a global community, ties incentives to verifiable contributions to the network, and creates economic primitives that do not require a single legal entity to act as steward. The two paths face different failure modes, attract different kinds of participants, and produce different kinds of resilience. They are best understood as complementary experiments inside a larger pluralistic landscape, each illuminating problems and possibilities the other cannot see clearly from the inside.
And of course, if one has a crypto ecosystem and a more traditional for/non-profit business operating co-developing the same open source AGI software, there are tremendous potential for mutual benefit – potentially, far more energetic and productive flywheels than exist even in the Big Tech universe.
Learning from OpenAI’s Mistakes
Technology-wise, there is a lot we can learn from OpenAI. They didn’t invent transformer neural nets, Google did (standing on the shoulders of so many others) – but they had the courage to try to scale them up before anyone understood much about how they worked, and they had the balls to roll them out even when they hallucinated like mad and had all other sort of obvious shortcomings. They did “move fast and break things” way better than any of the recent versions of Mark Zuckerberg, and they changed the world as a result.
I don’t see OpenAI as being on a path to making a genuine human-level AGI or ASI breakthrough, but they have accomplished amazing things. And I am a grateful user of their tools every day – while Claude is better at many things, GPT5.4-Pro is the smartest LLM at math and science I’ve ever used, and I have a couple hard-core science queries spinning on it waiting to converge right now while I write this essay.
On the organization and governance side, on the other hand, OpenAI’s path teaches by negative example. Its original mission was too dependent on trust, too weakly embodied in durable institutional machinery, too vulnerable to compute dependency, and too easily converted into a conventional commercial growth machine. The current Microsoft restructuring is a rational correction within that commercial trajectory; OpenAI is trying to claw back optionality after having given too much of it away. But the fact that the optionality needs clawing back is one among numerous pieces of evidence that the original structure was poorly suited to AGI-scale governance.
The OpenAI–Microsoft shakeup is therefore a useful public teaching moment. AGI governance cannot be retrofitted after the economic dependencies are already in place. It has to be designed before the money, compute, and distribution deals create their own institutional gravity. One example of this broader point is: Once a single partner has become too central, an organization may still be able to renegotiate, but only from inside the logic of the dependency it already accepted.
In figuring out the details of the OpenBGI structure, we are therefore being very careful to think through all the organizational and governance issues very carefully right from the start. It is totally possible to combine technical ambition, real commercial execution, and real compute partnerships, while refusing the premise that the Singularity should be governed by the default dynamics of a Silicon Valley platform company. If one’s goal is not merely to build a better AI product, but to shape the emergence of beneficial superintelligence – this requires not just smarter algorithms but smarter institutional architecture … and probably more than one kind of architecture, deployed in parallel, by groups willing to take the structural questions as seriously as the technical ones.
Building an appropriate organizational and business architecture for AGI work is not quite as difficult as actually building the AGI … but it’s still far from trivial. Fortunately at this stage the world provides us plentiful examples of things that do work and things that don’t work, which we can leverage as we collectively craft the best path forward in what may well be the last few years before the Singularity.
These are such amazingly exciting times to be in the AI field, but they are also critical times to get things right organizationally as well technically. But I’m sure we are up to this ;-)
What the Microsoft/OpenAI dynamic keeps revealing is that "aligned AI" tends to mean aligned with whoever controls the infrastructure. The values getting embedded aren't abstract human values, they're the values of specific institutions with specific interests at specific moments in history. That's not a conspiracy. It's just how power works. Which is exactly why the decentralization argument matters: not because distributed systems are more efficient, but because no single point of failure should get to define what human flourishing looks like for everyone else.
This piece hits the structural problem exactly: AGI governance cannot depend on good intentions sitting on top of a normal corporate incentive engine. Once mission, capital, compute access, commercial revenue, and strategic partnerships are fused together, the mission eventually gets interpreted through whoever controls the infrastructure.
The lesson I take from OpenAI/Microsoft is that the split has to be designed before the pressure arrives, not repaired afterward.
The structure I keep coming back to is not simply “nonprofit versus for-profit.” It feels more like a layered stewardship architecture.
One layer protects the mission: open research commitments, safety direction, public-interest governance, continuity, and the rules that prevent capture.
One layer handles execution: products, managed services, enterprise deployments, support, hardware, integrations, and the practical work needed to make the system sustainable.
But I think there needs to be a third layer underneath both: a protocol layer that records contribution, compute, provenance, and data rights separately.
That distinction matters because value in an AI ecosystem does not come only from code or capital. It also comes from the people who contribute ideas, build tools, provide compute, validate outputs, maintain infrastructure, and generate the data the system depends on.
Those should not all collapse into one ownership bucket.
Contribution should be recorded as contribution.
Investment should be recorded as investment.
Compute should be recorded as compute.
Data should remain under the custody of the person or community it comes from.
That last part may be one of the most important missing pieces. The current internet mostly treats user data as something platforms quietly collect, centralize, and monetize. A healthier AI economy would invert that. Individuals, communities, institutions, and tribes should be able to keep their data in their own custody and license controlled access under explicit terms.
A company might pay to query, analyze, or train on approved data, but the owner should decide what is available, what is restricted, what expires, what can be copied, what can only be used locally, and what is never for sale.
In plain terms: commercial AI should be able to make money by helping people control value, not by quietly taking the value.
So I think the governance problem is bigger than nonprofit versus for-profit. The deeper issue is whether the whole ecosystem can separate mission, execution, compute, contribution, and data custody before any one layer becomes powerful enough to dominate the rest.
A foundation without operational power becomes ceremonial.
A company without mission constraints becomes extractive.
A compute partner without boundaries becomes sovereign.
A token system without contribution validation becomes speculation.
A data economy without custody becomes surveillance capitalism with better branding.
The future structure probably needs all of these pieces separated clearly enough that no single actor can rewrite the whole system from one control point.
That, to me, is the real warning in the OpenAI/Microsoft story: not that commercial execution is evil, but that mission and infrastructure cannot be allowed to merge into the same command structure.
Once they do, the mission no longer governs the system.
The system governs the mission.
This was a great read. Thank you for everything you do.