Sam Altman on Sora, Energy, and Building an AI Empire
Summary
OpenAI’s strategy is a vertically integrated loop: infrastructure enables research, research enables products, and the product is a personal AI subscription that follows users across services and future devices. Altman describes three core pieces: personal AI, the infrastructure needed to support it, and the AGI research mission. The enormous infrastructure build has no separate business thesis yet; it exists to support the service and research. Horowitz says he had been against vertical integration but now thinks he was wrong; Altman says OpenAI has repeatedly had to do more things than expected.
Sora is both a consumer product and an AGI-relevant bet on world models, with social deployment helping society adapt before highly capable video becomes ubiquitous. Video has “much more emotional resonance than text,” making deepfakes and rights questions immediate, but Altman argues “society and technology have to co-evolve.” Sora uses tons of compute absolutely, though only a small fraction of OpenAI’s total.
The capability Altman is most excited about is the AI scientist, rather than better basic chitchat. GPT-5 is already producing “little examples” of novel math, physics, and biology work; Altman expects models within two years to perform “bigger chunks of science” and make important discoveries. Current LLM-based systems may only need to reach the point where they conduct “better research than all of OpenAI put together.”
OpenAI is making an aggressive infrastructure bet because it sees both the research roadmap and its prospective economic value one to two years ahead. Partnerships spanning AMD, Oracle, NVIDIA, and the chain “from the level of electrons to model distribution” are necessary to mobilize enough capacity. Horowitz frames the limit as some fraction of global GDP devoted to knowledge work, since robots are not yet part of this discussion. OpenAI would not invest this aggressively for today’s models alone.
Research remains senior to product growth even with ChatGPT at roughly 800 million weekly active users, a figure cited by Erik Torenberg. When capacity is constrained, GPUs “almost always” go to research, with only temporary exceptions for viral product launches: “We’re here to build AGI.” OpenAI’s research culture resembles a seed-stage firm backing exceptional founders more than a conventional product organization.
Altman now expects AGI to “go whooshing by” as a continuous transition rather than trigger an instantaneous singularity. He still anticipates “some really bad stuff” and wants careful testing once models become extremely superhuman, but opposes broad regulation of less capable systems. Torenberg argues that falling behind China would be extremely dangerous; Altman agrees that it would be “extremely dangerous,” while the broader comparison to not regulating nonexistent capabilities is Torenberg’s framing.
AI content markets may separate training rights from generation rights while forcing new trust and payment models. Altman’s “forced guess” is that training becomes fair use, while generation involving protected characters, styles, or IP may be handled under a different model with rights-holder restrictions and choices; he does not specify a settled licensing market. Some rights holders may ultimately complain their characters appear too little. Sora may require per-generation pricing, while paid product recommendations inside ChatGPT would destroy its trusted-adviser relationship.
Energy is a core constraint on AI: near-term United States baseload additions will mostly be natural gas, while solar-plus-storage and nuclear dominate the long run. Nuclear adoption depends less on rhetoric than economics—if it becomes “crushingly economically dominant,” political and regulatory barriers should fall quickly. For investors seeking the next trillion-dollar company built on near-free AGI, Altman’s honest answer is “I have no idea”; conviction comes from building and experimenting, not pattern matching.
Deep dive
1. Full-stack integration has become part of the mission
Altman describes three connected core pieces: a personal AI subscription, the research effort pursuing AGI and improving the product, and the infrastructure supporting the research and service. The intended product is an AI that gets to know the user, works in OpenAI’s consumer offerings, logs into other services, and eventually reaches dedicated devices.
The causal chain is explicit: “The research enables us to make the great products and the infrastructure enables us to do the research.” The infrastructure is not presently a standalone offering; its purpose is to “build this AGI and make it very useful to people.”
Asked whether the unprecedented buildout might eventually serve other companies, Altman leaves the possibility open but offers no plan: constructing “the biggest infrastructure project in history” may reveal another use. OpenAI has also asked its own models strategic questions and received insightful answers management had missed.
Horowitz says, “I was always against vertical integration, and I now think I was just wrong about that.” Altman says specialized suppliers such as NVIDIA still matter, but OpenAI’s story has been that it had to do more things than expected to deliver on its mission. His benchmark is the iPhone, “the most incredible product the tech industry has ever produced,” and extraordinarily vertically integrated.
2. Sora is a world-model and social-adaptation experiment
Sora may look peripheral to AGI, but Altman bets that “really great world models” will prove more relevant than observers expect. ChatGPT once attracted the same AGI skepticism, yet its deployment improved models, revealed how society wanted to use them, and made an abstract AGI debate concrete.
The deployment logic is societal as well as technical: “Society and technology have to co-evolve. You can’t just drop the thing at the end.” Because future video models can deepfake anyone or show anything and carry more emotional force than text, Altman wants the public to understand where video is going before it is everywhere.
OpenAI also wants “fun and joy and delight” rather than using AI solely to make people “ruthlessly efficient.” Sora consumes tons of compute in absolute terms, Altman concedes, but not a large fraction of OpenAI’s total.
Chat has saturated only the most basic conversation, not the range of tasks expressible through text—“Please cure cancer” remains far beyond it. Altman imagines interfaces built from continuously rendered real-time video and ambient hardware that understands context, rather than a phone indiscriminately blasting text-message notifications.
3. Deep learning keeps producing breakthroughs—and science is next
Altman expected language-model scaling laws to be a once-only discovery, then thought the same after the reasoning breakthrough. Instead, deep learning has been “this miracle that keeps on giving”: a fundamental technology that continues yielding breakthrough after breakthrough.
Horowitz describes a layered capability overhang: most users still frame AI through ChatGPT; Silicon Valley specialists using Codex think those users “have no idea what’s going on”; a few scientists say the same about the Codex users. Returning to GPT-3.5 would make today’s users wonder how anyone tolerated it.
Altman’s personal Turing test has always been scientific discovery. GPT-5 now shows “little examples” across math, physics, and biology; within two years, he expects “bigger chunks of science” and important discoveries. The recursive threshold may simply be LLM-based systems conducting “better research than all of OpenAI put together.”
4. Infrastructure spending is a forward bet, with research first in line
OpenAI has “never been more confident” in its research roadmap or the economic value of using the resulting models. Delivering infrastructure at the required scale means mobilizing a large part of the industry, “from the level of electrons to model distribution,” with more partnerships expected after the AMD, Oracle, and NVIDIA discussions.
Altman rejects the idea that demand is literally unlimited: the boundary is some amount of global GDP. Horowitz frames the relevant near-term limit as the fraction devoted to knowledge work, since OpenAI does not yet do robots. OpenAI would expand for today’s unmet demand, but “would not be going this aggressive” without expected model advances; it gets to see “a year or two in advance.”
Even with ChatGPT at roughly 800 million weekly active users, a figure cited by Torenberg, the allocation rule is clear: constrained GPUs “almost always” go to research. A viral feature may occasionally borrow capacity, but “on the whole, we’re here to build AGI.”
The AMD discussion surfaces Altman’s evolution from investor to operator. His early dealmaking reflected an investor advising a company and focused on distribution of money; he now considers how agreements must be operationalized and their implications over time. Yet the investor background helped elsewhere: a strong research culture resembles a seed-stage investing firm betting on founders more than a product company.
5. AGI looks continuous, but frontier risks remain discontinuous
Chatbot obsequiousness is not technically difficult to change; many users actively want it. OpenAI’s mistaken assumption was that billions of people would all want to talk to “the same person.” Altman expects short-term personality selection and, eventually, a system that interviews users and infers how each prefers to interact.
Static benchmark scores are less interesting as they saturate and become “crazily gamed.” Altman points to scientific discovery as an evaluation that can remain useful for a long time; Horowitz calls revenue an interesting measure of whether capability creates real value.
Horowitz says the popular conception of the Turing test “went whooshing by,” briefly shocked the world, then became ordinary. Altman expects AGI to follow: it will not change the world by the “impossible amount” imagined or become an immediate singularity, because people and institutions are more adaptable than forecast. “It will be more continuous than we thought.”
That continuity does not eliminate danger. Altman expects bad outcomes and worries about billions of people “talking to the same brain,” but wants regulatory burdens concentrated on models that become truly, extremely superhuman. Broad restrictions could produce a European-style “complete clamp.” Torenberg argues that falling behind China would be dangerous; Altman calls that outcome “extremely dangerous.”
6. Copyright, ads, and open weights are becoming control points
Altman’s forced guess is that society treats training as fair use. On generation involving protected styles, characters, or IP, he uses a human-reading analogy: an author may read a novel and draw inspiration from it, but cannot reproduce it; users can discuss Harry Potter without simply “spitting it out.” He expects restrictions and rights-holder decisions to matter, but does not specify a settled licensing market.
Rights holders are already splitting. Some fear misuse, while others worry Sora will not feature their characters enough: controlled interaction could deepen audience relationships and franchise value. Horowitz’s pushback—worth keeping—is that legacy creative industries can behave irrationally, as music publishers sometimes suppress what artists would regard as valuable promotion.
Sora also challenges the old 1%-create, 10%-comment, 100%-view assumption: easier tools reveal that far more people want to make content. People are generating funny videos of themselves and friends for group chats. Because Sora videos are expensive, Altman assumes some form of per-generation charge may be necessary, while future revenue sharing could preserve the internet’s bargain that creators receive money or attention.
Ads are acceptable only if they preserve trust. An Instagram-style discovery ad might add value, but recommending a paid coffee machine over the best one would make ChatGPT’s trusted-adviser relationship “vanish.” Meanwhile, a cottage industry is already engineering reviews that ChatGPT will favor; Altman admits, “I don’t know how we’re going to fight it yet.”
Open-source weights create another governance question. Altman is pleased people like GPT-OSS, while Horowitz says universities are using Chinese models and warns that dependence on DeepSeek could cede “control of the interpretation of everything” to weights that might be influenced by the Chinese government.
7. Energy economics will determine how far the AI buildout runs
Altman’s energy thesis predates the AI boom: he says cheaper and more abundant energy has been the highest-impact way to improve people’s quality of life. AI and energy began as independent interests and then converged; his investments in companies including Retro Biosciences, Helion, and Oklo were attempts to fund work he believed mattered.
In the short term, Altman expects most net-new United States energy, at least for baseload power, to come from natural gas. Over the long term, the dominant mix should be some unknown ratio of solar-plus-storage and nuclear, including advanced nuclear, small modular reactors, and fusion.
Nuclear’s timetable depends on price. If it becomes “completely, crushingly economically dominant,” the importance of cheap energy should generate pressure for faster NRC action and construction; if costs merely match alternatives, anti-nuclear sentiment could delay adoption for a really long time. Outlawing nuclear was, in Altman’s words, “an incredibly dumb decision.”
Asked what near-free AGI will make investable, Altman refuses synthetic certainty: “I have learned deep humility on this point.” The answer emerges from building, experimenting, and talking to people in the trenches—not five-year plans or hunting for “the next OpenAI.” Horowitz frames the AI turning point as enough GPUs and data making the lights come on; Altman calls the underlying lesson a bitter one.