Name: FULL INTERVIEW: Dylan Patel Says We’re Still Underestimating AI
Uploaded: 2026-02-03T23:08:03+00:00
Duration: 2623 s

Speakers

Dylan Patel — Founder, CEO, and Chief Analyst at SemiAnalysis

Summary

Dylan Patel discusses AI infrastructure constraints, data centers in space, and the shifting AI chip landscape. He argues AI demand and revenue are underestimated, power constraints are easing, and semiconductors, especially leading-edge fabs and memory, could become the key bottleneck again by 2027. He also highlights Meta as an AI winner, Oracle as oversold, Cerebras for fast inference, and robotics as a large long-term opportunity, while being skeptical on space data centers and Adobe's AI positioning.

Dylan Patel, SemiAnalysis founder and CEO, joins TBPN to discuss AI infrastructure, chips, power, space data centers, and geopolitics.
He argues AI revenue and compute demand are underestimated, supporting continued AI infrastructure growth.
He sees power constraints easing, but expects leading-edge fabs and memory to become the bottleneck again by 2027.
He discusses Nvidia's expanding AI chip portfolio, Google's TPU roadmap with Broadcom and MediaTek, and Cerebras's fast-inference niche.
He views Oracle as fine despite bad communications and Meta as a major AI winner via ads and wearables.
He is bullish on robotics for labor-heavy tasks and skeptical of Adobe as an AI company.
He takes the under on space data centers reaching 1% of global data center capacity by 2028.
He debates China export controls, arguing for restricting equipment while allowing model access.

Ideas

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 5:39

Nvidia broadens AI chip portfolio.

Nvidia is shifting from a one-size-fits-all GPU strategy to a multi-product portfolio with standard GPUs, CPX for prefill/video, and Groq chips, because AI model architectures and workload needs are uncertain. It is covering multiple points on the Pareto curve rather than betting on one design.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 7:20

Cerebras fills fast inference demand.

Speed/latency of inference matters more than data-center location for agentic workloads, and price-insensitive users will pay 10x for 10x faster completion. OpenAI's 750 MW Cerebras deal shows demand for cutting-edge fast inference, so Cerebras makes sense in that niche.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 12:26

Google TPU roadmap diversifies suppliers.

Google is diversifying its TPU roadmap: Broadcom makes one TPU and MediaTek makes another next year, with a third TPU project, to cover different compute, memory, and stacking tradeoffs. This creates differentiated TPU-supplier exposure for Broadcom and MediaTek.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 13:36

Google leads distributed AI training.

Google remains ahead in cross-data-center training. It built regional clusters roughly 40 miles apart, and the shift to RL and agentic training reduces synchronization bandwidth needs, making multi-data-center training more feasible. Google already has the infrastructure cards.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 19:13

Semis tighten again by 2027.

Power constraints are easing because turbines and medium-speed reciprocating engines can be brokered or added, but leading-edge fabs and memory cannot be quickly expanded. Memory makers have not built new fabs since 2022, and new capacity takes years; TSMC's leading-edge expansion is also limited. The bottleneck shifts back from power to semiconductors by 2027, favoring TSMC and memory suppliers.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 21:59

Oracle selloff is overdone.

Oracle's stock and OpenAI-linked vendors sold off and the company's comms looked panicked, but Oracle's data-center financings are secured at market-standard rates and consistent with investment-grade deals, and the OpenAI relationship remains intact. The market is overly negative on OpenAI because Anthropic is outperforming; Oracle is fine.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 26:24

Robotics has huge labor markets.

He is super bullish on robots, especially for labor-heavy markets such as laundry, dishwashing, and construction, which are far larger in worker count and economic value than fabs. Fab robotics is less relevant because fabs employ relatively few people despite huge output.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 35:00

Adobe is not an AI winner.

Adobe was perceived as an AI winner, and AI feature launches temporarily lifted the stock, but the market is now realizing Adobe is not actually an AI company, causing shares to fall again. This makes Adobe unattractive as an AI exposure.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 35:20

AI compute demand is underestimated.

The market is underestimating AI monetization and compute buildout: AI startup revenue should exceed $100B by year-end, Anthropic's $300B end-of-decade target is too low, and OpenAI's 16-18 GW by 2028 is fundable because the economic value created will be enormous. This supports continued strong demand for AI infrastructure.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 37:20

Meta is a major AI winner.

Meta's ad algorithm improved by double digits, driving CPMs up 9% despite weak consumer spending, showing AI is already monetizing. More AI-personalized content should increase time on Meta apps, and Meta is positioned to execute on AI wearables and platform distribution, while Apple may need to rely on others for AI.

Dylan Patel Founder, CEO, and Chief Analyst at SemiAnalysis 38:13

Apple AI wearables look weak.

Apple won't be able to put good AI on its wearables or will have to rely on Microsoft or Google for AI, making it vulnerable as AI moves into wearable assistants.

FULL INTERVIEW: Dylan Patel Says We’re Still Underestimating AI

Summary

Ideas

Up Next