October 10th 2025
OpenAI’s evolution has entered a new phase. In a recent a16z discussion with Ben Horowitz and Erik Torenberg, Sam Altman outlined how the company now functions as a vertically integrated AI ecosystem — research, infrastructure, and consumer products all under one roof.
What began as a non-profit lab in 2015 has become the most valuable startup in the world, unified by a single mission: to deliver a “personal AI” that learns users’ habits, lives across devices, and becomes an indispensable part of everyday life.
Altman admits he was once “wrong” about avoiding vertical integration. OpenAI’s trajectory now mirrors Apple’s — a full-stack approach where hardware, software, and infrastructure evolve together. That coherence, he argues, is essential to scaling both the technology and its societal impact.
“The story of OpenAI has been realizing that to deliver on our mission, we have to do more things than we thought,” Altman said.
OpenAI’s new structure ties its three pillars — research, infrastructure, and product — into one reinforcing system. The company is now building out what Altman calls “the largest infrastructure project in human history,” partnering with AMD, Nvidia, and Oracle to secure compute and power for the coming decade.
Much of the conversation focused on Sora, OpenAI’s text-to-video model. Altman sees it as both a scientific enabler and a social rehearsal for what’s next.
World-model understanding — the ability for AI to simulate physics, causality, and intent — is, he believes, essential to AGI. But Sora also gives society a chance to adapt to a world where synthetic video becomes indistinguishable from reality.
“Technology and society have to co-evolve,” Altman explained. “You can’t just drop the thing at the end — it doesn’t work that way.”
The unexpected early use cases — personal memes, inside jokes, and short clips between friends — have already forced a rethinking of monetization. Unlike ChatGPT’s subscription model, Sora may move toward usage-based pricing or new creative economies altogether.
Altman suggests that the next wave of AI interaction will move far beyond text-based chat. Future interfaces could take the form of context-aware devices that understand user intent, ambiently respond, and even render real-time video scenes as the interface itself.
He envisions an AI that naturally adjusts tone, depth, and complexity without explicit prompts — a system that understands how a user wants to be spoken to, not just what they’re asking.
OpenAI’s forthcoming GPT-5, Altman revealed, is beginning to display early signs of independent scientific reasoning. Researchers have observed models proposing mathematical proofs, biological hypotheses, and small-scale discoveries.
This, Altman says, marks the beginning of the “AI scientist” era — a world in which large models accelerate the pace of discovery across every discipline. “Scientific progress is what makes the world better over time,” he said. “If we get a lot more of it, that’s a big deal.”
Altman expects that training AI models on public data will ultimately be defined as fair use, while new frameworks will govern output tied to specific IP. Interestingly, he says, many rights holders now want more representation in generative systems, not less — seeing AI as a new channel for discovery and fandom.
OpenAI’s decision to open-weight models like GPOSS, meanwhile, reflects a pragmatic stance: encouraging transparency and academic access, but wary of ceding influence to opaque open-source systems outside the U.S.
Perhaps the most consequential section of the conversation centered on energy. Altman describes AI and energy as now inseparable: “They were two independent interests that really converged.”
He expects short-term growth to rely heavily on natural gas while long-term supply shifts toward solar + storage and advanced nuclear. If nuclear power becomes “crushingly cheaper,” he predicts, political resistance will fade and adoption will accelerate rapidly.
In Altman’s view, energy availability is becoming the real bottleneck in AI progress — a constraint that will define the next decade of global industrial policy.
Altman’s regulatory stance is measured: focus oversight on superhuman-capable systems, not everyday AI tools. He cautions that sweeping regulation could suppress innovation while authoritarian states surge ahead.
He draws a parallel to the early internet — transformative, chaotic, and too dynamic to be governed by static frameworks. The goal, he says, should be safety testing at the edge, freedom at the core.
Altman’s discussion revealed not just OpenAI’s strategy, but its expanding worldview. The company is building toward a future where compute, energy, and content are interdependent — where infrastructure, creativity, and cognition converge.
