AI

AI infrastructure capex is the largest concentrated build-out in technology history. The durable economic value forms in layers, and each layer has a different risk and ownership profile. Foundational chokepoints (EUV lithography, leading-edge foundry, GPU compute, high-bandwidth memory, datacentre power and cooling) are the deepest moats and where we anchor public-equity exposure. The platform layer (NVIDIA's CUDA stack, hyperscaler custom silicon, networking) is where competitive dynamics actually decide. The frontier-lab layer (OpenAI, Anthropic, xAI, Mistral, SSI, DeepSeek) is where most of the demand originates but where almost nothing is publicly investable. The application layer is the spillover thesis and the reason this worldview cross-references Defense, Biotech, Fintech, and SpaceTech.

NVIDIA is the most concentrated single bet in AI infrastructure, and the platform-layer call we have to make explicitly. The CUDA software stack creates switching costs that compounded over fifteen years; every framework, every researcher, every inference runtime is calibrated to it. The Mellanox acquisition extended the moat to multi-rack training clusters via InfiniBand and Spectrum-X. The bull thesis is durable platform pricing power on the demand explosion ahead. The bear thesis is customer concentration (three or four hyperscalers absorb the majority of revenue), the rise of hyperscaler-internal custom silicon (Broadcom-fabricated AWS Trainium, Google TPU, Microsoft Maia), and AMD's MI accelerators taking share where workloads tolerate switching. Blackwell shipments versus HBM3e availability is the gating ratio we monitor for the next twelve months of the call.

Frontier labs are the demand engine but mostly not publicly investable. OpenAI's $40B-plus raise, Anthropic's multi-round $20B+ stack, xAI's $6B round, and the next wave (SSI, DeepSeek, Mistral, plus China incumbents) determine the capex landing at NVIDIA, TSMC, and ASML. We watch them, we cannot hold them. Public exposure to the frontier-lab thesis lives through proxies: Microsoft's OpenAI position and exclusive Azure compute relationship, Alphabet's Anthropic stake plus internal Gemini, Amazon's Anthropic investment plus Trainium build-out, Meta's Llama open-weights strategy and internal capex draw. Each funding round and capability release is a leading indicator for the chokepoint thesis below it, and a stress test of whether public-side multiples are sustainable.

Edge sources we lean on: (1) hyperscaler capex guidance and the capex-to-cloud-revenue ratio as the bull/bear pivot for the entire stack; (2) ASML book-to-bill and high-NA EUV tool shipment cadence; (3) TSMC N3 and N2 capacity utilisation and pricing discipline; (4) NVIDIA customer concentration disclosures and Blackwell-vs-HBM shipment commentary; (5) HBM availability and pricing (SK Hynix is the real bottleneck more than any GPU); (6) frontier-lab funding rounds and capability releases as demand-side leading indicators; (7) Chinese substitution velocity (Huawei Ascend, SMIC yield) as the tail risk; (8) 13F flow into Coatue, Tiger Global, ARK on the AI stack.