The money is structural, not a sidebar. Four views of the same map: the names that define it, every ticker by layer, where the single points of failure live, and what ships when.
Data as of 13 September 2026 Open in the 3D atlas
The names that structurally define the map — chokepoint monopolies, scale players and pure-plays — and exactly why each one matters.
Single points of failure — if any one stalls, the whole stack stalls. Highest barriers, near-impossible to displace.
The deepest revenue pools — where the AI build-out shows up first and largest in the P&L.
Concentrated bets — smaller floats, high leverage to one specific part of the build-out.
Grouped by where each company sits in the chain. The dot is the worst chokepoint severity of any layer it touches.
Where each critical step physically happens, ordered by single-point-of-failure risk.
Logic die (TSMC N3/N2) + CoWoS advanced packaging — ~90% of leading-edge logic and near-100% of CoWoS. A disruption here halts the entire AI hardware stack within months.
HBM manufacturing — SK Hynix + Samsung, ~80% of global HBM (Micron is the rest), made in two cities. The memory every AI GPU is built around.
EUV lithography — ASML, a 100% monopoly. No sub-7nm chip (logic or HBM base die) exists without it. Plus Besi hybrid bonding.
The silent materials chokepoint — photoresist, silicon wafers, ABF film (Ajinomoto ~95%) & substrates (Ibiden), and InP lasers (Sumitomo). Low visibility, huge leverage.
Design & IP — NVIDIA, AMD, Marvell, Astera, Broadcom — plus Micron HBM, Lumentum/Coherent lasers, Lam/KLA tools and EDA. Owns the value; outsources the fabrication.
Optical-module volume (InnoLight + Eoptolink, ~60% of NVIDIA's 800G) and a rising commodity-DRAM supplier (CXMT ~10% of global revenue) — but no EUV, no HBM3E before 2027, no domestic DSP and <15% of its high-speed lasers. Holds the gallium/germanium/rare-earth inputs the West's lasers need, and has banned purchases of NVIDIA's China-specific chips. The chokepoint that cuts both ways.
Precision optical assembly — Fabrinet builds for Coherent, Lumentum, Ciena, and InnoLight/Eoptolink now build US-bound 1.6T volume here to sidestep tariffs. A quiet single-country concentration in optics, and a "substantial transformation" audit risk.
A static snapshot, not a feed — so you can see what the market already prices into each layer. Market cap and price were confirmed from two sources; 12-month return is shown only where two sources agreed. Forward P/E is consensus as quoted by stockanalysis.com.
| Company | Ticker | Price | Mkt cap | 12-mo | Fwd P/E |
|---|---|---|---|---|---|
| NVIDIA | NVDA | $218.29 | $5.27T | +22% | 18.1× |
| TSMC | TSM | $433.24 | $1.98T | n/a | 19.3× |
| Broadcom | AVGO | $361.99 | $1.73T | -2% | 20.9× |
| Samsung | 005930.KS | ₩259,500 | $1.19T | n/a | 4.2× |
| Micron | MU | $975.26 | $1.10T | n/a | 6.8× |
| SK Hynix | 000660.KS | ₩1,812,000 | $953B | +492% | 4.4× |
| AMD | AMD | $516.13 | $843B | +224% | 46.7× |
| ASML | ASML | $1,698.30 | $659B | n/a | 30.3× |
| Intel | INTC | $102.94 | $541B | n/a | 61.7× |
| Lam Research | LRCX | $298.22 | $373B | +178% | 31.4× |
| Applied Materials | AMAT | $456.49 | $362B | +216% | 25.9× |
| KLA | KLAC | $180.64 | $236B | +94% | 33.1× |
| Marvell | MRVL | $236.10 | $207B | +252% | 43.3× |
| InnoLight | 300308.SZ | ¥926.00 | $150B | n/a | 18.0× |
| Corning | GLW | $166.40 | $143B | n/a | 44.9× |
| Eoptolink | 300502.SZ | ¥423.00 | $88B | n/a | 20.3× |
| ASE Technology | ASX | $39.47 | $86B | +249% | 26.5× |
| Lumentum | LITE | $927.03 | $83B | n/a | 42.8× |
| Shin-Etsu | 4063.T | ¥5,738 | $70B | +41% | 18.4× |
| Coherent | COHR | $305.37 | $60B | +195% | 32.4× |
| Astera Labs | ALAB | $291.22 | $51B | +26% | 51.9× |
| Ciena | CIEN | $349.54 | $50B | +169% | 33.8× |
| Sumitomo Electric | 5802.T | ¥2,077 | $42B | n/a | 17.6× |
| Microchip | MCHP | $74.20 | $40B | +15% | 18.8× |
| Disco | 6146.T | ¥51,980 | $37B | +33% | 27.1× |
| Ibiden | 4062.T | ¥19,560 | $36B | +400% | 58.1× |
| Montage Tech | 688008.SH | ¥191.96 | $35B | +68% | 49.6× |
| Ajinomoto | 2802.T | ¥4,974 | $31B | +57% | 32.3× |
| BE Semiconductor | BESI | €189.25 | $18B | +158% | 34.2× |
| Fabrinet | FN | $414.58 | $15B | +15% | 22.8× |
| Amkor | AMKR | $51.73 | $13B | +109% | 19.0× |
| Rambus | RMBS | $86.97 | $9B | +13% | 25.6× |
Snapshot 11–12 Sep 2026 · stockanalysis.com, cross-checked against Google Finance / Yahoo Finance · FX USD/KRW 1,386 · USD/JPY 153.6 · USD/CNY 6.71 · EUR/USD 1.16. n/a = could not be verified from two sources, so it is left blank rather than guessed.
What happens when, and who benefits. Entries whose date has passed carry a verdict: the atlas grades its own forecasts.
Azure ships CXL-attached memory (Astera Leo) — the first in a commercial cloud. NVIDIA Quantum-X Photonics and Broadcom's Tomahawk 6 Davisson (102.4 Tb/s CPO) begin sampling to tier-1 hyperscalers.
What actually happened — Leo CXL went live on Azure M-series on 18 Nov 2025 — as a preview, not GA. Davisson is shipping as the first 102.4T CPO Ethernet switch, but Quantum-X Photonics only went commercial in early 2026, a quarter or two late. Direction right, photonics timing slipped. (graded Sept 2026)
SK Hynix and Samsung deliver HBM4 — the next bandwidth leap (a 2,048-bit bus). TSMC races CoWoS toward ~130k wafers/month. Whoever qualifies first wins the allocation.
What actually happened — All three suppliers qualified and in production, confirmed by Jensen Huang on 5 Jun 2026 — Samsung began mass production in February, SK Hynix qualified first and took ~60–70% of the allocation. CoWoS is tracking to ~130k wafers/month by year-end. (graded Sept 2026)
1.6T transceivers ramp (InnoLight volume, Marvell DSP inside). EML lasers sell out through 2027 after NVIDIA's ~$4B Lumentum/Coherent lockup. ABF substrate back in shortage (Ibiden). Scarcity = pricing power.
What actually happened — InnoLight's Q1 revenue rose 207% on 1.6T; every EML maker's capacity is under long-term agreements through 2027 after NVIDIA's $2B-each investments in Coherent and Lumentum (Mar 2026); ABF supply gap ~10% in H2 2026 with Ajinomoto pricing up to 30% higher. What we missed: conventional DRAM going into shortage too. (graded Sept 2026)
Rubin (288 GB HBM4, 22 TB/s) samples in Q4 with 16-Hi HBM4 stacks; AMD MI400 carries the first UALink silicon; Spectrum-X Photonics ships. DRAM contract prices keep rising as HBM takes ~22% of wafers. Two scheduled policy shocks: China's gallium/germanium reprieve to the US expires 27 Nov, and BIS reviews the HBM export rule (3A090.c) in December.
Rubin in volume from Q1 on HBM4 (TSMC base dies). Intel Diamond Rapids finally lands, a year late; UALink products ship broadly; Astera's NVLink Fusion hybrid racks and Scorpio X scale. SK Hynix Yongin (May) and Samsung P5 (late 2027) open — the first real DRAM relief — while CXMT targets HBM3E. Ibiden's substrate capacity starts. The tightest supply year on record, then the turn.
Co-packaged optics reach mass deployment; optical I/O chiplets ship in products (Ayar, Lightmatter); photonic memory disaggregation arrives (Marvell/Celestial). HBM4E pushes ~4 TB/s per stack. The rack becomes one machine.