Founder, CEO & Chief Analyst, SemiAnalysis · semiconductor & AI-infrastructure research · Expert Watch · weekly scan · our rolling analysis
Why we watch him. Dylan Patel is among the most granular public trackers of the AI-infrastructure buildout, following data centres by satellite imagery and mapping hundreds of billions in capital flows across chips, memory, power and cloud. That cost curve sets what every AI capability under Dominic's six companies can afford. Portfolio relevance: defence and intelligence data (chip supply chain, US-China export controls) and AI-bot monetisation (token and inference-cost economics); his "AI spend could exceed salary budgets" framing touches the knowledge-work value question. Weekly output makes him a live tripwire on the compute-economics debate.
Current headline view: the binding constraint on AI has moved from GPUs to the deeper supply chain (memory, EUV logic, power), while inference demand stays near-unbounded, so the buildout compounds and the real gains come from hardware-software co-design, not the chip alone.
Latest 1-page summary (as at 10 Jul 2026)
The eight core theses we track him against:
Three bottlenecks to scaling compute: logic, memory, power. EUV litho (ASML, ~70-100 tools/yr) is the #1 constraint to 2030.
Memory is the emerging crunch. Reasoning models need huge KV caches, DRAM/NAND toward tripling, ~30% of Big Tech 2026 capex going to memory.
GPUs are holding or gaining value. An H100 is worth more than three years ago (scarcity plus efficiency), countering the depreciation-bear case.
Inference demand is near-unbounded. Only the frontier model is wanted, willingness-to-pay near-limitless, inference could be "bigger than oil".
The real 100x is hardware-software co-design. Model plus kernels plus silicon, not any single faster chip; the "CUDA moat" was never really about CUDA.
Cloud providers face a Pascal's wager. Forced to sacrifice profitability through 2027 on AI capex; Google possibly no profit in 2027.
US-China chip rivalry and export controls are central. The biggest AI risk may be voters, not China.
Nvidia and Jensen back "neocloud" providers. They do this to keep compute a multipolar market.
Method note. This v1 baseline is distilled from Dylan's SemiAnalysis research and his recent guest interviews (Dwarkesh, Invest Like the Best, Latent Space, Sequoia). It is not yet a full X post-by-post read. The weekly scan refines it against X and SemiAnalysis from here, and flags any deviation.
Analysis backlog (newest first)
Deep source docs are held locally in the workspace at expert-watch/, not on this page.
10 Jul 2026 · v1 profile built: identity verified (Dylan Patel, SemiAnalysis), surfaces, eight core theses, deviation baseline. Themes from his SemiAnalysis research plus recent guest interviews (Dwarkesh, Invest Like the Best, Latent Space, Sequoia).
Notable reactions (operators & researchers)
Up to three recent, high-signal reactions from operators or researchers (not commentators) to a specific new argument of his. Challenges first. As at 10 Jul 2026.
No high-signal reactions logged this cycle (v1 baseline). The first weekly scan will populate this.
Recent media
Recent YouTube videos and podcast episodes, most recent first. As at 10 Jul 2026. He appears as a guest only.