Is there too much hype for AI and the hyperscalers?

Episode Summary

Executive Summary: The episode debates whether DeepSeek marked a turning point for AI capex and the U.S. hyperscalers. Brad Rogoff and Ross Sandler agree DeepSeek exposed how efficient models can be built with less compute, but Ross argues AI demand is still early and shifting toward reasoning and agentic AI will likely sustain, if not increase, long-term compute needs—especially inference. The key question is less bubble/no bubble and more how much capex is truly required, and which layer of the stack benefits most.

Main Topics: DeepSeek as an AI efficiency shock (Priority: 5/5): DeepSeek is presented as a major market and narrative shock because it demonstrated strong model performance using more efficient engineering, open-source tools, and less dependence on frontier chips. It challenged assumptions that ever-larger compute budgets are required for competitive AI. Hyperscaler capex and the risk of overbuilding (Priority: 5/5): The discussion centers on whether Microsoft, Amazon, Alphabet, Meta and peers have overbuilt data centers and chip budgets. The speakers frame the issue as potentially being both bubble-like enthusiasm and a timing mismatch between capex and monetization. Open source, China, and geopolitical context (Priority: 4/5): DeepSeek’s open sourcing, plus its Chinese origin during U.S.-China trade tensions, amplified market attention. Ross argues China is strong in software and AI engineering even if it lacks cutting-edge semiconductor access. Shift from training to inference and reasoning models (Priority: 5/5): A major theme is the transition from dense, general-purpose training models to sparse reasoning models that can act more autonomously. This shift changes where compute is spent and raises the importance of custom inference chips. Agentic AI and the next use cases (Priority: 4/5): The conversation argues that AI’s next phase will be agentic systems that can complete tasks end-to-end, such as ordering groceries, booking travel, and handling enterprise workflows, which may materially increase adoption and compute demand. Investment winners and losers in the AI stack (Priority: 4/5): Ross suggests that if capex remains high, semiconductor, networking, and data center suppliers benefit most; if model efficiency improves faster, hyperscalers may regain free-cash-flow leverage and software/application companies with strong AI usage could outperform.

Key Arguments: DeepSeek did not invent new AI; it improved engineering and efficiency using open-source foundations, showing that strong results do not always require the most expensive chips. The DeepSeek moment likely exposed overbuilt AI infrastructure assumptions and put greater scrutiny on hyperscaler capex plans. The right framing is not simply 'bubble or not' but whether current and expected AI capex is calibrated to future demand. AI adoption is accelerating, so a slowdown in training intensity may be offset by much larger inference demand and broader usage. Reasoning models and agentic AI can expand AI from text-completion tools into task-completion systems, which should increase real-world utility and consumption. Even if DeepSeek lowers the cost curve, overall AI compute demand may still rise sharply because new use cases are emerging quickly. If chinchilla scaling continues, chip and infrastructure suppliers remain the main beneficiaries; if it stalls, hyperscalers could see stronger free cash flow and software/application names may gain share.

Data Points: NVIDIA one-day loss: Largest on record - Ross described the January DeepSeek selloff as a historic hit to AI hardware stocks. Market move comparison: Semiconductor names relative to telecom names moved the most since the dot-com bust in 2000 - Used to characterize the severity of the AI hardware selloff after DeepSeek. Microsoft forward P/E: About 30x - Brad cited this as evidence that sentiment toward hyperscalers remained relatively robust. Hyperscaler total capex consensus: A little over $300 billion - Ross used this to frame current expected spending by major cloud providers. AI semiconductor portion of capex: About $100 billion - Ross said roughly one-third of total hyperscaler capex is going to AI semiconductors. Compute gap to 2028: Around 250 billion exaflops - Estimated gap between current compute capacity and what may be needed by 2028. Exaflop definition: A billion billion calculations per second - Ross defined the unit when explaining next-generation compute scale. Inference compute growth: Could quadruple between 2025 and 2028 - Ross argued inference demand may need to rise substantially as reasoning models scale. Inference share of industry compute: A little over half - Projected inference share of total industry compute at the end of the forecast period. ChatGPT users: 400 million - Ross said ChatGPT crossed this user milestone earlier in the year. Fast user growth: Additional 100 million users in about two months - Used to show adoption is accelerating. Earlier user growth pace: First 100 million users took about a year - Compared with the recent acceleration in user adoption.

Pivotal Quotes: "DeepSeek didn't really invent anything new in AI. They just engineered everything much better than what was done before." — Ross Sandler: Ross explains why DeepSeek mattered: efficiency and execution, not a new AI paradigm. "The question is not really answer the bubble or no bubble question. But really, what's the right level of AI CapEx that these hyperscalers should be underwriting?" — Ross Sandler: He reframes the debate from a binary bubble call to a capex-sizing question. "The same thing is about to happen in AI where the capability gets much better." — Ross Sandler: Ross compares AI’s adoption trajectory to the evolution from early mobile apps to indispensable native apps.

Implications: AI investment is shifting from raw model training toward efficient inference, reasoning, and agents. Winners may rotate between chip/infrastructure vendors, hyperscalers, and AI software leaders depending on how quickly costs fall versus usage grows.