State of the AI industry — the OpenAI Podcast Ep. 12

AI’s next phase hinges less on whether models get smarter and more on whether people and companies learn to use that capability to complete real tasks—trip planning, contract review, clinical triage support—because compute availability is the main bottleneck today. In 2025, “vibe coding” and agent-like experimentation grabbed attention; in 2026, multi-agent systems are expected to mature enough to deliver visible outcomes, first inside enterprises for end-to-end workflows and later for consumers in messy, multi-step situations.

The discussion frames the industry as moving along a “capability gap” continuum. Most users still operate AI as a chat interface for questions, even though the real value comes from turning prompts into outcomes. On the consumer side, the goal is shifting from “ask and answer” to “task worker” behavior—booking travel across calendars, reservations, and flight schedules; generating meal plans for specific health needs; or helping people prepare for medical conversations. On the enterprise side, the same continuum applies, but with vertically specialized deployments that can transform core business processes. Examples range from healthcare workflows like drug discovery and hospital operations (e.g., admission and discharge timing) to retail outcomes such as higher conversion rates and larger basket sizes.

A key adoption argument is that usage growth will outpace capability growth because learning curves lag behind model improvements. Only a small fraction of people are believed to be using a meaningful portion of available AI capability, so the number of weekly users—cited as 800 million ChatGPT users—should expand further, with a long runway (described as roughly a decade) for broader, deeper adoption. The “bubble” debate is addressed directly: demand is constrained mainly by compute, not by lack of interest. The analogy used is electricity—people can’t use the full range of what’s possible until they learn what to do with it.

Health is treated as both the clearest marker of acceleration and the hardest regulatory domain. Usage signals are already large: 230 million people ask ChatGPT health questions weekly, and 66% of U.S. physicians report using it in daily work. Yet diagnosing and prescription-writing face legal and regulatory constraints; no AI medical device is approved yet, and institutions like the FDA and the American Medical Association limit what systems can do. Even so, the practical near-term impact is framed as augmentation: helping doctors stay current, supporting triage and differential diagnosis, and giving consumers better-prepared conversations and second opinions.

On compute and investment, the CFO perspective ties spending to revenue and future capacity. Compute capacity is described as rising from 200 megawatts to 600 megawatts and then to 2 gigawatts, alongside ARR growth from $2B to $6B and then to over $20B. Decisions must be made years ahead because data centers and orders can’t be created instantly. The broader market is also investing heavily, with global hardware and chip spending rising by hundreds of billions.

Finally, the conversation argues that “bubble” metrics based on stock prices miss the real signal. Vinod Khosla proposes measuring AI demand by operational usage—API calls or internet traffic analogs—because those reflect real utility rather than investor sentiment. The enterprise and startup implications are clear: agents are still early (only a minority of companies use them today), but there’s room for startups to build moats around proprietary data, complex workflows, permissions, and identity—especially where general models can’t safely or efficiently handle governance-heavy tasks like procurement approvals. The outlook extends beyond software into robotics and a future where labor and expertise become cheaper, raising the question of what societies will do when more services become effectively free—housing remains a major unresolved challenge.