OpenAI board chair and Sierra CEO Bret Taylor says the industry is in an AI bubble — and that’s OK. In a recent interview, Taylor echoed Sam Altman’s warning that “someone is going to lose a phenomenal amount of money” but insisted the underlying technology will continue to accrue outsized long-term value. It is a position that recalls the late-’90s internet cycle: exuberance first, enduring platforms second.
Why does everyone call it a bubble?
Look around: capital, compute and hype are all growing faster than proven returns. CB Insights and PitchBook have tracked tens of billions pouring into generative AI startups, minting dozens of unicorns with scant history of revenue. NVIDIA market cap swelled to the multi-trillion range on demand data center demand, and hyperscalers flagged major infrastructure outlays to feed model training and inference. The International Energy Agency estimates that data center electricity consumption could double in a few years, driven partly by AI.
Taylor’s point is straightforward: froth and fundamentals can happily coexist. We can be overpaying for so many companies at the same time that the foundational shift is genuine. That paradox was the hallmark of the dot-com era, and it seems certain to define today’s model- and agent-led wave of A.I.
Dot-com déjà vu — with a twist
The internet’s core thesis wasn’t wrong; many startups from that era disappeared when capital tightened in the late ’90s, but there was always a hard core of successful survivors. The survivors — the ones who built infrastructure, trusted brands and genuine distribution — compounded for decades. And there’s a parallel to that in Taylor’s analogy: A shakeout that’s due, or overdue, the outcome of which is not just leaving noise behind but also some platforms and tooling and norms by which the next stage gets powered.
This cycle has new dynamics. It is delivered faster, often through APIs and cloud marketplaces. The barrier to differentiation at the model layer is being squeezed by eroding margins from open models (such as those that Meta and Mistral have open sourced) given all of the compute supply remains under control with a handful of providers and cloud vendors. Industry is as much setting the pace of progress in research terms power, latency and memory constraints.
What a healthy correction might look like
Anticipate consolidation: teams and IP taken on by bigger players, as when the founders and team at a hot agent startup joined a cloud giant even though that company licensed its tech. Commoditization at the base model layer pushes value to custom orchestration, proprietary data and workflow integration. Buyers will want auditability, cost predictability and security — where many early-stage products remain thin.
Unit economics will decide winners. Inference is the new hosting bill, and costs can escalate if you’re not engineering conscientiously. Approaches such as retrieval-augmented generation, model distillation, and dynamic routing to smaller task-specific models are transitioning from the research to production running foot for restraining per-query costs. Companies that can prove out a durable ROI per seat or per transaction — not just show an impressive demo — will exist when capital tightens up.
Where Taylor finds durable value
Taylor runs Sierra, an AI agent startup that collaborating on enterprise workflows — and his optimism is based on concrete productivity gains.
This direction is also supported by independent research: a highly-cited study by Stanford and MIT; for example, found that access to an AI assistant increased customer support productivity by roughly 14%, with the greatest gains among less experienced workers. Those are some of the kinds of impacts CFOs can also underwrite.
In customer service, agents that can intelligently triage, draft and resolve tasks on their own are already reducing handle times and boosting consistency. In the world of software, code assistants are speeding up evaluation and resolution. In medicine, natural language generation/processing tools are being used for ambient documentation of clinical care. According to McKinsey, generative artificial intelligence could unlock an additional $2.6 trillion to $4.4 trillion in annual value among functions such as sales, marketing, customer service and R&D — if companies can go from pilots to governed production.
The enterprise stack is finally coming together: vector databases for retrieval, evaluation harnesses for quality and safety, policy engines to enforce privacy, IP and compliance. The kind of scaffolding that turns a viral demo into a reliable system — exactly the sort of infrastructure that has historically been left in place after an air pocket deflates.
Risks and governance realities
Regulation is catching up. The EU’s AI Act established a risk-based set of obligations for model providers and deployers, while U.S. agencies have indicated they’ll use consumer protection, competition and employment laws on the books to police AI claims and harms. “The [NIST AI Risk Management Framework] is emerging as a de facto standard for assessments, logging and responses to issues.”
Buyers are increasingly demanding meticulous evaluations, red-teaming of results, provenance controls and transparent fine-tuning data. The risk of hallucinations, the need to inject promptly, data leakage and copyright exposure are now board-level risks. “Teams that are actually building with security-by-design and robust monitoring will find budget, even as vanity pilots dry up.”
The bottom line
Taylor’s take is not an inconsistency. Markets may over-shoot, but what is happening under the hood — software that speaks and writes across modalities, and agents that do stuff — has legs. The shakeout is likely to separate the bulk from those wrapped in thin paper and speculative bets, rewarding disciplined engineering, reliable governance and transparent unit economics. In Taylor’s estimation, that, very likely, is a good thing: when the next decade takes off running, it will do so on the rails that the bubble’s whatever-it-takes-to-prevent-crimson-donut-pavement excess poured. If history is any guide (and with dot-com it was), then yes- we-wanna-bubble like Kavanaugh wants beer but: “Prices just went up for subways” – Wall Street Journal)((( ENOUGH!