Summary of "He Predicted the AI Bubble in 2023 | Doug Clinton and Gene Munster on Why We're Still in 1996"

Summary of Main Arguments & Reports

AI demand will outstrip power capacity, extending the “AI bubble” timeline. Doug Clinton (Deep Water / Intelligent Alpha) argues that AI is fundamentally about converting electricity into “intelligence.” If demand for intelligence is effectively infinite, demand for power is also effectively infinite. This supply constraint—energy and data center buildout—is presented as a key reason markets may remain in an extended boom rather than peaking immediately. Clinton reiterates an earlier view (end-2023) that AI culminates in a bubble larger than the dot-com bubble, and suggests the timing is closer to 1995–1996 than 1998, implying more runway.
Bottlenecks are power/data centers, but utility adoption is already accelerating. The discussion shifts from “when will it peak?” to “what’s driving the trade now?” Clinton credits recent model capabilities (especially coding agents) with unlocking enterprise productivity, while the main bottlenecks remain building data centers and powering them. Gene Munster (framed as having a “1996” view in the title) asks whether models have reached mass adoption. Clinton clarifies that “mass adoption” is mainly enterprise adoption, pointing to rapid increases in enterprise spend on inference.
A coding-agent paradigm shift is central to the rally (and investor confidence). Clinton highlights a specific catalyst: Anthropic’s Opus (described as a version released in late 2023) making coding agents (e.g., “Claude”-style tooling) accessible to mainstream users and enterprises. The claim is that this moved AI from “promising” to usable: coding agents can translate natural language into code, execute tasks, and expand productivity benefits across knowledge work.
Job disruption is expected to be sharper near term, but the market adapts. Clinton predicts more acute knowledge-worker unemployment in the next ~5 years than in earlier waves (mobile/internet), but expects it to eventually “fix itself” as workers retrain and reposition—becoming “detectives,” salespeople, and other roles that generate unique data feedstock for models. The framing emphasizes that AI can supercharge individuals or make others irrelevant, especially for the marginal segment.
Model performance is improving quickly; “general intelligence” is largely semantic. The hosts discuss whether current models qualify as AGI. Clinton argues the AGI/superintelligence debate is mostly semantic, since today’s models can handle a very high percentage of user tasks with good accuracy. He also notes fast improvement cycles—rankings change frequently across model releases.
Competitiveness of model providers: GPT currently #1 in internal testing. Clinton reports a leaderboard from Intelligent Alpha’s testing harness:
- GPT (top model) is ranked first on their benchmark.
- Opus/Anthropic, Gemini/Google, and others cluster near the top depending on version.
- Open-source/CN providers are present, but Clinton says their tests show closed-source “major U.S.” models tend to outperform open-source models on a financial-direction benchmark (while stressing that leaderboards can shift over time).
How Intelligent Alpha measures models: a standardized earnings-direction benchmark. Intelligent Alpha’s product (“Intelligent Earnings Benchmark”) uses a consistent evaluation harness where multiple LLMs ingest the same prepared financial data (earnings call transcripts, street estimates, historical statements, etc.) for ~700 stocks. The models predict whether revenue/EPS move up or down, and also (in buckets) estimate how much. Clinton attributes GPT’s edge mainly to incremental improvements—being better at estimating base rates and likely outcomes.
AI investing thesis for markets: infrastructure (especially power/energy) remains the core. Clinton argues investors still underestimate energy constraints. He highlights nuclear and energy storage as key themes, suggesting data center expansion will keep accelerating due to insatiable power demand. He also points to rising hyperscaler capex growth expectations (up substantially from earlier forecasts), reinforcing that the “brain” (data centers) is expanding faster than Wall Street expects.
AI-related IPOs are viewed as a broader confidence signal. Clinton cites Cerebras’s IPO as evidence of market appetite and suggests a second tier of AI-related companies (e.g., data/coding tooling) may consider IPOs as conditions improve.
Market rotation among mega-cap hyperscalers is tied to competence and AI tooling ecosystems. The hosts debate whether the market rewards hyperscalers that spend aggressively on AI infrastructure. One core idea is that markets may re-price companies where management is judged to be competently anticipating demand. Outcomes may differ based on AI ecosystem partnerships—such as Anthropic’s correlation with AWS users versus competitors more tied to OpenAI.
Second-order opportunities: “space” as an energy and infrastructure frontier. Clinton and Munster outline early-stage thematic ideas around space:
- Orbital data centers (prototypes first; limited initial performance) as a way to bypass local permitting bottlenecks and possibly improve efficiency.
- Novel drug/product development in space, where gravity/environment become production constraints. Overall, “space” is positioned as a longshot but potentially transformative pathway for energy and productivity.

Presenters / Contributors

Doug Clinton — Deep Water Asset Management / Intelligent Alpha
Gene Munster — partner/host; referenced throughout as part of Excess Returns coverage
Jack — host; referenced when asking questions
Justin — host; referenced during model/testing and investment discussions