How semiconductor physics will constrain AI compute from 2026 to 2040 — and which companies will profit from clearing each bottleneck.
For the bottleneck framework, the current pass keeps the original supply-constrained thesis but shifts near-term emphasis toward power interconnection, HBM, advanced packaging, optics/CPO, and verification/test capacity. The active ranking pass cross-checks the original thesis against current market prices, market cap, YTD rerating, AI 2027 compute/security assumptions, hyperscaler capex pressure, power constraints, HBM and advanced-packaging evidence, and physical-AI adoption signals.
Key metrics from the AI compute supply chain, cross-checked against the latest ASML, TSMC, IEA, and market-data snapshots used on this page.
The binding constraint shifts every 2-3 years as earlier bottlenecks get cleared and new ones emerge.
ASML now expects to ship 60 low-NA EUV tools in 2026 and 80 in 2027. At 3.5 tools per GW, that's only 12.0-18.3 GW/year of new AI compute depending on allocation.
HBM requires 4x more wafer area per bit than commodity DRAM. Prices have quadrupled. Consumer electronics are getting squeezed out.
Sources: SK hynix Q1 2026 results, TrendForce Apr. 21 HBM bulletin, Micron earnings. Values in $B.
30% of Big Tech CapEx goes to memory alone. Source: SemiAnalysis estimates.
CXMT can produce 2.2M stacks in 2026 — enough for only 250K–400K Ascend 910C packages. Global HBM supply remains effectively sold out through 2026.
Smartphone volumes projected to decline 40% as memory is reallocated. HBM margins (50–70%) vs DRAM (20–30%) make conversion irresistible.
HBM4: 2.5 TB/s per stack. DDR5: 64–128 GB/s same shoreline. A 20x gap that makes commodity DRAM unsuitable for AI.
Strategic dynamics between labs, chip makers, enterprise buyers, and nations. Hover over cells to explore payoffs.
Anthropic vs OpenAI: Who commits to compute early?
| OpenAI | |||
|---|---|---|---|
| Commit Early | Be Conservative | ||
Anthropic | Commit Early | 8,8 Nash | 10,5 |
| Be Conservative | 5,10 Actual | 6,6 | |
Strategic InsightOpenAI committed early while Anthropic was conservative. Now Anthropic pays premium for spot capacity. In supply-constrained markets with increasing returns, early commitment dominates.
10 bottleneck categories. Public names now show a current price plus USD market cap snapshot, while private, acquired, and subsidiary names stay in the mix for the same bottleneck.
Peak bottleneck period: 2028-2032
Prices use the local listing or ADR quote. Market caps are normalized to USD. Refreshed July 9, 2026.
Sole EUV manufacturer. Projects $71B revenue by 2030.
Leading metrology/inspection. ~25% mask inspection market.
ONLY company making EUV mask inspection tools. Monopoly.
Largest semi equipment maker. Deposition, etch, inspection.
Major etch/deposition. Benefits from increasing EUV layers.
Photodetectors and sensors for lithography metrology.
Light sources for DUV lithography and mask inspection.
Sole supplier of ALL EUV optics. 18 mirror sets per tool. Irreplaceable.
Sole supplier of CO2 laser for EUV sources. Single-source dependency.
Produces all EUV light sources. Tin droplet laser system.
A $400M EUV tool enables $14.3B of downstream economic value. ASML captures less than 3% of the value it creates.
If ASML raised prices by just 50% (to $600M/tool), their earnings would roughly double — and the tool would still represent less than 5% of downstream value. As the EUV bottleneck tightens from 2028-2032, pricing power increases dramatically.
Fast AI timelines favor the US (infrastructure lead). Slow timelines favor China (vertical supply chain). Drag the slider to explore.
Knife’s edge. China achieves working EUV in lab. Both sides have significant capacity.
“Fast timelines, the US wins; long timelines, China wins.”
Dylan Patel, SemiAnalysis
The binding constraint shifts every 2-3 years. Hover or click to see which stocks to position in at each phase.
HBM4 entering volume (Samsung Feb 2026), but supply lags demand. The 2026 HBM market is now closer to $54.6B, and CoWoS remains the throughput gate with TSMC near 130K wafers/month against ~940K wafers of implied Nvidia need alone. Micron new HBM packaging comes online 2027. Amkor Arizona not until 2028. (Sources: Samsung Q4-2025, TrendForce Apr. 21 HBM bulletin, TSMC Q1 2026 results, Amkor press release)
62% HBM share. HBM4 mass production. Nvidia preferred partner. (SK hynix 12-layer HBM4 announcement)
Dominates CoWoS. A16 volume 2H 2026. $52-56B capex. (TSMC 2024 Annual Report)
HBM4 mass production Feb 2026 at 3.3 TB/s. Tripling HBM sales. (Samsung earnings)
#2 OSAT. Arizona campus first production 2028. Up to $7B investment. (Amkor Arizona announcement)
Singapore HBM packaging 2026, meaningful expansion 2027. $200B US plan. (Micron press releases)
Largest semi equipment maker. HBM and packaging equipment supplier. (AMAT earnings)
Some companies span multiple bottleneck phases. Darker = higher conviction in that period.
| Company | Phase 1 2026-2027 | Phase 2 2028 | Phase 3 2029-2030 | Phase 4 2031-2032 | Phase 5 2033-2035 | Phase 6 2036-2040 |
|---|---|---|---|---|---|---|
ASML ASML | High | Med | ||||
TSMC TSM | High | High | Med | Low | ||
SK Hynix 000660 | High | |||||
GE Vernova GEV | High | High | Low | |||
Eaton ETN | High | Med | Low | |||
Constellation CEG | Med | High | Med | |||
Intel INTC | Low | High | High | |||
Lam Research LRCX | Low | High | ||||
Amkor AMKR | High | Med | ||||
Coherent COHR | High | Low |
These companies have absolute monopolies with no alternative supplier. They represent the highest-conviction plays across ALL phases.
Sole EUV tool maker
Sole EUV optics/mirrors
Sole EUV laser source
Sole ABF film manufacturer
Sole EUV mask inspection
100% of HBM production
Fifteen chokepoints that sit underneath the AI buildout, each mapped to five US-based and five non-US-based public equities. The goal is not purity for its own sake. The goal is to identify who gets paid when the shortage moves from theory into capex.
US bucket means the company is based in the United States. Non-US bucket means the business is based outside the United States, even if an ADR exists.
Direct means the company makes the constrained part. Tooling means it sells the equipment or process control behind the constrained part. Adjacent means it benefits when the bottleneck forces more spend elsewhere in the stack.
Some sectors, especially MLCCs, inductors, glass-core substrates, and ABF substrates, have very few pure-play US-listed options. Those baskets intentionally widen to adjacent public beneficiaries rather than forcing bad fake precision.
These are research baskets, not price targets. The section is designed to extend the page's bottleneck logic into a company map that is easier to monitor quarter by quarter.
Lasers and silicon photonics determine how fast clusters can talk.
HBM, hybrid bonding, ABF, and advanced packaging determine how much compute you can actually ship.
800V distribution, busbars, transformers, and turbines determine whether the campus can be energized.
MLCCs, inductors, cooling, and PCBs decide whether the system stays stable after it is built.
Bottleneck scores measure how likely the shortage is to gate AI system throughput. Pricing-power scores estimate how much the shortage can translate into economics. US-depth scores show how many usable US public equities exist in the lane.
| Sector | Bottleneck | Pricing power | US depth |
|---|---|---|---|
HBM and advanced memory Packaging | Bottleneck5/5 | Pricing power5/5 | US depth3/5 |
Hybrid bonding Packaging | Bottleneck5/5 | Pricing power5/5 | US depth4/5 |
Gas turbines Power Delivery | Bottleneck5/5 | Pricing power5/5 | US depth3/5 |
Silicon photonics Photonics | Bottleneck5/5 | Pricing power4/5 | US depth3/5 |
800V power delivery Power Delivery | Bottleneck5/5 | Pricing power4/5 | US depth4/5 |
Advanced packaging Packaging | Bottleneck5/5 | Pricing power4/5 | US depth4/5 |
Grid transformers Power Delivery | Bottleneck5/5 | Pricing power4/5 | US depth3/5 |
ABF substrates Packaging | Bottleneck5/5 | Pricing power4/5 | US depth1/5 |
Thermal management and liquid cooling Thermal | Bottleneck5/5 | Pricing power4/5 | US depth4/5 |
High-power, narrow-linewidth lasers Photonics | Bottleneck4/5 | Pricing power4/5 | US depth2/5 |
MLCCs Passives | Bottleneck4/5 | Pricing power4/5 | US depth1/5 |
Glass core substrates Packaging | Bottleneck4/5 | Pricing power3/5 | US depth1/5 |
Inductors Passives | Bottleneck4/5 | Pricing power3/5 | US depth1/5 |
Power shelves and busbars Power Delivery | Bottleneck4/5 | Pricing power3/5 | US depth4/5 |
Advanced PCBs Packaging | Bottleneck4/5 | Pricing power3/5 | US depth3/5 |
AI clusters need cleaner optical links at higher data rates, which raises the value of narrow-linewidth CW, EML, and external laser source platforms.
Laser quality, yield, and qualification time constrain the optical engines that feed 800G, 1.6T, and CPO ramps.
Hyperscaler optical demand has shifted from generic telecom optics to AI-centric interconnects, where spectral purity, thermal control, and packaging matter more.
This is a concentrated basket. The best public names are component and photonics-platform companies rather than pure single-product laser foundries.
Owns the laser, detector, and transceiver stack that rides 400G to 1.6T upgrades.
Datacom cycles can whipsaw margins if cloud capex pauses.
Its UHP laser roadmap is tied directly to SiPh transceivers and co-packaged optics.
Customer concentration is meaningful in high-end optics.
The photonics portfolio gives it leverage to advanced laser manufacturing and optical assembly.
Semicap cyclicality and integration execution can dilute the optics thesis.
It is one of the few liquid public laser platforms with deep manufacturing know-how and scale.
Industrial mix remains larger than AI interconnect exposure.
The company offers a smaller-cap way to express high-power laser process and packaging know-how.
Defense and industrial exposure dominate current revenue mix.
Jenoptik benefits where precision optics and photonics tooling get pulled into higher-end optical stacks.
More optics-platform exposure than direct AI datacom laser exposure.
Management is explicitly positioning the business around AI data-center interconnect photonics.
Balance sheet repair remains part of the equity story.
Hamamatsu gives broad photonics exposure across detectors, emitters, and sensing infrastructure.
AI optical networking is only one piece of the business mix.
It is one of the few listed non-US laser specialists left on the board.
Liquidity and scale are much lower than the US leaders.
The company benefits as next-generation optics require tighter test and characterization loops.
A narrower portfolio means higher program-level volatility.