a16zSam Altman on Sora, Energy, and Building an AI Empire
CHAPTERS
OpenAI’s “personal AI subscription” vision and why it forces massive infrastructure
Altman frames OpenAI as building a ubiquitous personal AI people subscribe to and use across first‑party apps, third‑party logins, and future dedicated devices. To deliver that reliably, OpenAI must also become a mega-scale infrastructure builder tightly coupled to its research agenda.
- •Goal: become the default personal AI subscription for most people
- •AI that “gets to know you” across services and devices
- •Infrastructure build-out is required to support product + research at scale
- •Mission focus remains: build AGI and make it broadly useful
Vertical integration as strategy: research → infrastructure → products
The conversation connects OpenAI’s many bets under a single vertically integrated stack: research creates capability, infrastructure enables research, and products distribute and refine real-world usage. Altman notes he used to oppose vertical integration but now believes it’s necessary to execute the mission.
- •OpenAI’s bets are linked as a single stack rather than separate businesses
- •Altman’s view shift: vertical integration is more essential than expected
- •Examples from computing history (iPhone as highly integrated benchmark)
- •OpenAI still relies on key horizontal partners (e.g., NVIDIA), but must “do more” internally
Sora, world models, and co-evolving society with AI capability
Altman argues Sora is more AGI-relevant than it appears because better “world models” may be critical for future intelligence. He also emphasizes that releasing products like ChatGPT and Sora helps society adapt iteratively—especially as video deepfakes and synthetic media become pervasive.
- •Sora as a path to better world models, potentially important for AGI
- •ChatGPT served as both product and societal on-ramp to AGI realities
- •Co-evolution thesis: society must adapt continuously, not at the end
- •Video’s emotional resonance raises stakes for deepfakes and trust
- •Compute allocation: meaningful in absolute terms, smaller relative to total
Beyond chat: new AI interfaces, ambient devices, and real-time generated media
Altman distinguishes between chat being “good enough” for basic conversation and the much larger, unsaturated space of what a chat interface can accomplish. He forecasts richer interfaces—possibly real-time rendered video—and ambient, context-aware devices that reduce notification overload and improve timing and relevance.
- •Chat is saturated for small talk, not for high-impact tasks
- •Sora hints at interfaces built around real-time generated video
- •Ambient hardware that understands context and attention
- •Shift from intrusive notifications to context-aware interaction patterns
The AI scientist as the next ‘Turing test’: accelerating discovery
Altman describes AI doing science as his personal benchmark for transformative capability, claiming early examples are emerging and will expand significantly in the next couple years. He frames scientific progress as the primary driver of human welfare and expects AI-enabled discovery to be an underappreciated positive impact.
- •AI scientific discovery as a world-changing threshold
- •Claimed early signals with frontier models (math/physics/biology examples)
- •Expectation: larger chunks of science and meaningful discoveries soon
- •Scientific progress seen as the dominant lever for improving quality of life
- •Public discourse overweights risks vs. upside like curing diseases
2025 capability ‘overhang’ and how far LLMs can go before new breakthroughs
Altman says progress has been faster and more continuous than he expected, with repeated breakthroughs (scaling laws, reasoning improvements) and a widening gap between public perception and frontier capability. He suggests LLMs may advance far enough to help generate the next major research breakthroughs themselves.
- •Surprise: continued breakthroughs felt ‘improbably’ productive
- •Capability overhang: most users lag far behind frontier usage
- •Older models (e.g., early ChatGPT era) now feel dramatically weaker
- •View: push current approach until models can out-research top labs
Personality, preferences, and why one chatbot voice can’t fit billions
They discuss complaints about overly flattering/obsequious model behavior and Altman claims it’s not technically hard to change—some users actually prefer it. The deeper challenge is accommodating wide variance in user preferences, implying personalized or selectable model “personalities” that learn from interaction.
- •Obsequiousness is easy to adjust technically; demand varies by user segment
- •Real issue: huge distribution of desired assistant behavior
- •Likely future: model learns your preferences through interaction
- •Short-term: users pick from presets rather than a single global personality
CEO lessons: moving from investor mindset to operating reality
Altman reflects on how his early deals were approached with an investor’s lens rather than an operator’s, and how running a company changed his thinking. The discussion highlights the operational complexity of partnerships and execution beyond headline terms and leverage.
- •Altman’s self-assessment: limited operating experience early on
- •Investor vs. CEO mindset: deal-making vs. long-term operationalization
- •Running OpenAI revealed the depth of organizational dynamics required
- •Partnership structures improved as operational understanding increased
Partnering with potential competitors to scale infrastructure end-to-end
OpenAI’s recent partnerships (AMD, Oracle, NVIDIA) illustrate a strategy of collaborating broadly to make an aggressive infrastructure bet. Altman argues the build requires coordinated support across the entire supply chain—from electricity to distribution—and that limits are far from current scale if capability keeps rising.
- •Rationale: aggressive infrastructure bet backed by research confidence
- •Need ecosystem support: ‘from electrons to model distribution’
- •Collaboration can include firms that may compete in some layers
- •Scale bounded by GDP/knowledge work, but far beyond today if roadmap holds
- •GPU allocation: research typically prioritized over product demand spikes
Measuring progress: beyond benchmarks to revenue and real-world science
Altman downplays static benchmark evals due to saturation and gaming, and points to harder-to-fake indicators like scientific discovery and even revenue as meaningful measures. The chapter also touches on shifting “AGI vibes” and how perception can lag reality.
- •Static benchmark evals are increasingly gamed and less informative
- •Scientific discovery is a durable, meaningful eval target
- •Revenue/usefulness as another practical capability signal
- •Online sentiment (‘AGI-pilled’ vs. not) can be misleading about progress
AGI won’t feel like a Big Bang: continuity, adaptation, and safety focus
Altman and Horowitz argue AGI may ‘whoosh by’ and feel more continuous than apocalyptic, because society adapts quickly. On safety and regulation, Altman prefers minimal broad regulation, with targeted stringent testing only for extremely superhuman frontier models to avoid stifling beneficial uses.
- •Expectation: AGI feels continuous; society adapts faster than expected
- •Still anticipates ‘strange or scary moments’ and real harms
- •Regulatory stance: most regulation has downsides; focus on frontier superhuman models
- •Avoid broad clampdowns that would block beneficial, less capable models
- •Geopolitical concern: overregulation could disadvantage the US vs. China
Copyright, likeness, and the emerging economics of IP-enabled generation
Altman predicts training may be deemed fair use, while generation ‘in the style of’ or with explicit IP may require new licensing models. They explore a surprising dynamic: some rights holders may want more inclusion (with safeguards) because interactive AI can increase franchise value.
- •Prediction: training likely treated as fair use; generation may need new rights frameworks
- •Analogy: humans can be inspired by works but can’t reproduce them verbatim
- •Rights-holder split: some want restrictions; others want more character usage
- •Name/likeness and character safety guardrails are central
- •Creative industries may behave irrationally due to intermediaries and incentives
Open source strategy and the risk of ceding default models to China
Altman states he’s pro–open source and is pleased users value OpenAI’s open model release. Horowitz raises strategic concerns that Chinese-origin open models could dominate academia and shape defaults, including potential hidden influence in weights and alignment choices.
- •Altman: open source is broadly good; positive response to OpenAI’s open release
- •Concern: dominance of Chinese open models in universities
- •Risk framing: ceding ‘interpretation’ and defaults to potentially state-influenced systems
- •Open weights enable unknown modifications and downstream behaviors
Energy as AI’s limiting factor: gas now, solar+storage and nuclear later
Altman explains energy as a foundational driver of human prosperity and now deeply entangled with AI scaling. He expects natural gas to supply much near-term incremental load, while long-term dominance comes from solar+storage and advanced nuclear—if nuclear becomes decisively cost-competitive and policy allows rapid buildout.
- •Energy abundance historically correlates with quality-of-life improvements
- •Near-term expectation: natural gas provides much net-new baseload
- •Long-term expectation: solar+storage plus nuclear dominate
- •Advanced nuclear includes SMRs and fusion; speed depends on economic advantage
- •Policy and permitting (e.g., NRC) are pivotal; cheap energy can shift politics quickly
Monetization, trust, ads, and new user behaviors in AI-generated media
Altman says real usage often diverges from expectations—Sora is being used heavily for meme-like social sharing, which pressures pricing and packaging since video generation is costly. He’s open to ads but warns that ChatGPT’s trust relationship makes pay-for-placement recommendations dangerous, and notes emerging manipulation like AI-targeted review spam.
- •Sora users create high volumes of social content; cost forces new pricing models
- •Likely need per-generation or usage-based monetization for expensive modalities
- •Ads: acceptable in principle, but must not compromise user trust
- •Paid-placement recommendations would destroy credibility in assistants
- •New threat: ecosystem manipulation (fake reviews/content optimized for AI consumption)
- •Broader question: preserving incentives to create content on the internet
Talent wars, personal arc, and founder advice: curiosity, trenches, humility
Altman reflects on the post-ChatGPT era as exhausting and chaotic compared to the earlier joy of running a research lab. He describes his broader investing as funding what he believes in (energy, longevity), and closes with advice: predicting the next trillion-dollar opportunity requires hands-on exploration—humility beats armchair forecasts, and following curiosity keeps you close to real inflection points.
- •Post-ChatGPT: operating intensity rose dramatically; adaptation over time
- •No ‘master plan’ across investments—just funding high-conviction areas
- •AI has been a lifelong interest; early AI periods felt like ‘not working’
- •Founders/investors must explore in the trenches; forecasts are usually wrong
- •Investors often pattern-match; breakthroughs create opportunities that look different from prior winners
