A captain's watchlist, not a public recommendation. After Micron's record print and ~$50B Q4 guide confirmed the memory super-cycle, I'm concentrating on the picks-and-shovels of the AI build-out — compute, custom silicon, memory, foundry and lithography — companies whose order books are bound to physics, capex and multi-year customer agreements, not to a perfect interest-rate path. SK Hynix's Jul 10 Nasdaq ADR debut was the newest data point on that same thesis.
My read (Jun '26): still the core lock. Data-center demand is intact, the CUDA + networking moat holds, and BioNeMo / Halos push the platform beyond GPUs. The trade is crowded and headline-sensitive — own it, don't chase green days into the Fed.
core lockMy read: the cleanest structural counterweight to Nvidia. A 10GW OpenAI custom-ASIC deal, ~$73B backlog and ~60% projected AI-compute ASIC share by 2027 with Microsoft, Google, Amazon and Meta anchoring. Hyperscaler capex flows straight through it.
custom-silicon kingMy read: the super-cycle is now confirmed, not hoped-for. Record FQ3 ($41.5B rev, 85% gross margin), a ~$50B Q4 guide and multi-year Strategic Customer Agreements that lock pricing and make earnings durable. HBM4 is shipping in volume. Samsung's Jul 7 Q2 guide — op. profit ~₩89.4T, ~19x YoY, on 40–60% memory price gains and its first quarter of commercial HBM4 shipments — is independent, cross-vendor confirmation of the same cycle, not just a Micron story. Highest torque in the stack — expect sharp two-way swings.
memory super-cycleMy read (Jul '26): the HBM chokepoint becomes directly ownable. SK Hynix commands ~50–70% of global HBM share — feeding NVIDIA's GPUs — and its Jul 10 Nasdaq ADR debut (ticker SKHY, 10:1 vs KRX:000660) ends the old "Korea discount." Late-Jul SOX index inclusion forces passive buying; the ~30% pre-listing pullback and ~2.44% dilution are the entry cost. Adding it, not replacing Micron — different balance sheet, different FX/ADR-structure risk. Swapped in for Marvell this cycle: more differentiated exposure, a real re-rating catalyst.
hbm gatekeeperMy read: the company that physically makes the AI era. Every GPU and custom ASIC routes through its leading-edge nodes, and CoWoS advanced packaging is now a tighter constraint than fabrication itself. The lowest-drama way to own the whole build-out.
the chokepointMy read: the one true monopoly in the stack. No high-NA EUV, no denser chips — and its ~$400M machine is seen powering chipmaking for a decade. Lumpy orders and China headlines create the volatility; the moat is unmatched. Accumulate on capex-fear dips.
monopoly moatScroll to let the divide draw itself: the US bloc settles left in blue, the Chinese bloc right in red, and the suppliers both sides depend on sit on the line. Tap or click a logo to inspect its products and verified relationships; desktop users can also pan or reposition nodes. Zoom is intentionally disabled.
SCROLL TO UNFURL ↓Both blocs light up at once. Claude Opus 5 launched 24 July 2026 and tops the independent Artificial Analysis Intelligence Index at 61, ahead of Fable 5 at 60 and GPT-5.6 Sol at 59 — while Moonshot's Kimi K3 sits within four points of the leader. The capability spread that ran past 30 points in 2023 is now measured in single digits.
US private AI investment runs roughly 23 times China's, partly offset by state-guided funds on the other side. Alibaba's disclosed Moonshot stake, cloud-linked frontier investments and the 25-member open-weight coalition show how capital, policy and distribution now move as one bloc-level decision.
Cloud capacity flows through accelerators and custom silicon into HBM, DRAM and advanced packaging — and the lines that cross the meridian are the dependency. Advanced AI-processor die capacity still runs about 35–38× in the US bloc's favour on a quality-adjusted basis, which is why capability convergence has not closed the compute gap.
The full network is now visible. Select any branded node to trace primary products and source-backed investment, platform and compute-supply relationships across the divide — then read the measured comparison below.
Frontier models from both blocs on one normalized radar, then the full grid behind it: intelligence, professional and agentic tasks, operational efficiency and standardized API economics.
Ten US-listed and ten China-listed names mapped to the layer of the stack they own. Desk weights are the operator's own conviction — this is a captain's log, not a recommendation, and no position here is advice.
Four axes on a 1–100 scale with the method and evidence printed next to every score. Weights are an editorial judgement about what is hardest to substitute, not a published model.