The structurally important development this cycle is that the chips-compute value chain has bifurcated clearly enough to act on: a durable-rent tier sitting upstream of NVIDIA, and a contestable tier that includes NVIDIA itself. That bifurcation is sharpened by the single most novel signal of this scan — a US-government directive on 12 June 2026 suspending all access to Anthropic's Fable 5 and Mythos 5 model variants (https://www.anthropic.com/news) — which, if it establishes a domestic-model-variant export-control category, converts what has been a market-pricing question about platform-layer scarcity into a regulatory-permission question.
The revenue gap has not closed
The Cahn arithmetic underlying the AI capex ceiling thesis remains unchanged: roughly $840 billion per year implied against less than $50 billion observed, a 16× gap. NVIDIA's fiscal Q3 FY26 revenue of $57 billion — approximately 90% data-centre, with four hyperscaler customers (MSFT, META, AMZN, GOOGL) each exceeding 10% of FY26 revenue and two of them accounting for 36% combined — is impressive in isolation and nowhere near closing that arithmetic. Anthropic's Series H at $65 billion, valuing the company at $965 billion post-money and followed by a confidential S-1 submission to the SEC on 1 June 2026 (https://www.anthropic.com/news), is being priced against a forward revenue trajectory no historical software cohort has demonstrated at this scale. When the S-1 becomes public — expected Q3 to Q4 2026 — it is the cleanest single falsification test for the $200 billion AI-attributable-revenue threshold.
The commoditisation leg of the thesis has simultaneously strengthened. Open-source enterprise deployment extended from 23% to 67% across 2025. DeepSeek-R1 was trained at a verified cost below $6 million, open-source, runnable on Huawei Ascend hardware [arXiv:2501.12948]; Chinese AI's global use share moved from under 1% to roughly 15% over the same window. More structurally, Anthropic's distribution configuration now routes Claude through three competing hyperscaler channels — GCP Vertex AI, Microsoft Azure Foundry, and AWS — plus new enterprise-SI partnerships with TCS (12 June 2026) and DXC (11 June 2026) (https://www.anthropic.com/news). The model lab has settled into the role of content layer inside someone else's distribution stack. The early-2000s cable-vs-content analogy is exact: HBO earned the high gross margin, Comcast earned the volume and the relationship. The 60%-plus AI-service gross margin at two or more hyperscalers for three consecutive years that would refute commoditisation moves further out of reach under this arrangement.
Where rent durably sits
The investable implication is to migrate exposure away from the contestable merchant-GPU sub-tier toward the upstream-monopoly and picks-and-shovels layers where rent compresses least.
At the apex sits ASML (ASML), the sole supplier of EUV lithography globally — its Twinscan NXE:3800E and NXE:5000 High-NA units gate every 2-nm-class node at TSMC, Samsung, and Intel Foundry. Substituting ASML requires a multi-decade national-industrial-policy commitment; Japan's national EUV consortium and China's SMEE programme are both at engineering-development stage, not production. One layer inside that monopoly, Lasertec (6920.T) holds the only EUV-actinic mask-blank inspection franchise — a second-order monopoly inside an already irreplaceable stack. The wafer-fab-equipment oligopolists — Tokyo Electron (8035.T), Applied Materials (AMAT), Lam Research (LRCX), KLA Corp (KLAC) — carry 3-to-5-year requalification cycles that insulate them from displacement.
The binding sub-bottleneck for 2026 to 2028 is not transistor density at N3/N3P — wafer-out is sufficient for Blackwell and Rubin demand — it is TSMC's (TSM) CoWoS-S/L advanced-packaging throughput. CoWoS interposer area, redistribution-layer yield, and hybrid-bond pitch constrain finished-accelerator exits regardless of die availability; the constraint relaxes only after 18 to 36 months of capex and qualification. The picks-and-shovels into that bottleneck are BESI (BESI.AS) on hybrid-bonding equipment and ASM Pacific (0522.HK) on bonding-and-handling — neither features in the typical NVIDIA-centric AI conversation. In HBM memory, SK Hynix (000660.KS), Micron (MU), and Samsung (005930.KS) form a three-supplier oligopoly in tight allocation against demand for HBM4 — 12-Hi stacks, 1024-bit interface, approximately 1.5 TB/s bandwidth — the qualifying memory for Blackwell Ultra and Rubin. Micron's HBM revenue is tracking toward roughly $8 to $10 billion annualised by end of FY2026.
At the accelerator silicon layer the relevant distinction is between NVIDIA (NVDA) — an avoid at cap-weighted peak consistent with both the capex-ceiling and value-migration theses — and the custom-ASIC design houses: Broadcom (AVGO), designing Google's TPU v5p/v6e, Meta's MTIA, and ByteDance silicon; Marvell (MRVL), designing AWS Trainium and Inferentia. Hyperscaler custom-ASIC decisions are 18-to-36-month commitments that NVIDIA cannot displace inside one product cycle; AVGO's AI revenue FY2026 run-rate is tracking toward $25 to $30 billion. AMD (AMD) provides merchant-GPU second-source optionality. The commissioned design-house sub-layer — Alchip Technologies (3661.TW) and Global Unichip (3443.TW) — sits one tier below AVGO and MRVL. System integrators (Super Micro SMCI, Dell DELL, HPE HPE) and the neocloud cluster (CoreWeave CRWV, Nebius NBIS, Applied Digital APLD, Iris Energy IREN) occupy the most commoditised downstream layer and are the transmission channel through which GPU-collateralised debt — funded partly by hyperscaler issuance of $79 billion from MSFT and $25 billion from META in 2025 alone — routes into private-credit and insurance balance-sheet exposure.
What would change my mind
Three concrete observations would falsify the positioning. First: TSMC CoWoS capacity additions accelerate materially beyond the current +60% year-on-year 2026 expansion plan, signalling supply catching demand and compressing advanced-packaging rent durability. Second: NVIDIA prints a Q3 FY27 gross margin below 70%, ahead of the commodity-compression timeline the thesis assumes. Third: Anthropic's S-1 reveals an annualised revenue run-rate below $30 billion — breaking the $965 billion private mark and triggering a re-rate across the frontier-lab tier that pulls hyperscaler capex guidance lower and closes the Cahn gap via commitment cuts rather than revenue compounding.
What I'm watching next
The 12 June 2026 US-government directive against Anthropic's Fable 5 and Mythos 5 (https://www.anthropic.com/news) is the load-bearing new datapoint of this cycle: a domestic-model-variant restriction is categorically distinct from the BIS export-control regime applied to China since 2022. I am monitoring Commerce Department BIS docket filings for July through September 2026; if this resolves as a Section 1758 Export Control Reform Act action against a domestic commercial product, the regulatory-moat dynamic enters operational reality and the platform layer's revenue ceiling caps lower. Alongside that: NVDA Q1 FY27 10-Q (SEC EDGAR returned 403 this cycle), TSMC May and June 2026 monthly revenue, SK Hynix HBM4 mass-production milestone, and ASML Q1 2026 High-NA EUV bookings are all deferred primaries for the next cycle. The Anthropic S-1 remains the cleanest single falsification observable in the pipeline.