Compute Substrate — AGI Track

Investment thesis for the physical/manufacturing/cooling stack that gates FLOPs/$/W. Dated 2026-05-26.

1. Thesis

Cognition is becoming abundant. Once recursive self-improvement starts compounding inside the labs, the binding constraint shifts to the physical machine that runs the cognition. The labs are not capital-constrained or model-constrained — they are substrate-constrained. NVIDIA shipping next-gen XPUs, Anthropic and OpenAI training frontier models, and the US/China governments each subsidizing domestic fabs all converge on a single physical bottleneck: advanced logic wafers + HBM + advanced packaging + the optics that connect them, all of which have multi-year capex cycles, single-digit-supplier concentration, and >50-week lead times today.

The investable insight is that the constraint is not the GPU vendor. It is the suppliers upstream of the GPU vendor. TSMC's CoWoS lines are sold out through 2027 (Silicon Analysts Q1'26); HBM3E/HBM4 capacity is sold out for 2026 across all three suppliers; ASML is the sole EUV vendor; Lumentum/Coherent/Fabrinet are sold what they can build. The market has partially noticed (LRCX +318% in 12 months, Lumentum +1,455%, Coherent +386%) but the path-of-least-resistance bid is still flowing into NVIDIA. Where supply cannot expand fast — fabs (4-year build), packaging (2-year build), photoresist (chemistry), specialty optics (single-supplier reticle steppers) — pricing power is being underestimated.

2. The bottleneck hierarchy (as of May 2026)

Ranking is by current scarcity and time-to-relieve:

RankBottleneckQuantified scarcityTime to relieveWho owns it
1CoWoS advanced packaging52–78 week lead times across all 3 TSMC AP fabs; sold out through 2027; NVIDIA ~60–70% of allocation2–3 years (new AP fab build)TSMC (dominant), Amkor & ASE overflow at 26–39 wk lead times
2HBM (3E 8-Hi/12-Hi, HBM4)All 3 suppliers' 2026 capacity sold out; shortage forecast through 2028; HBM4 per-stack pricing ~$500 vs $300 for HBM3E 12-Hi (+67% generational uplift)~24 mo (capex + DRAM-to-HBM line conversion)SK Hynix ~50%, Samsung ~28%, Micron ~22% bit share in 2026
3Leading-edge logic (N3, N2, A16)TSMC N3 fully booked; 2nm/A16 70% CAGR 2026–2028 planned; 9 phases of fab build started in 20264–5 years per fabTSMC near-monopoly at leading edge; Samsung distant, Intel struggling
4EUV lithography systemsSingle supplier; High-NA EUV shipping in low units; AI customer capex on equipment +18.6% in 2026Permanent constraint — no substituteASML (100%)
5Optical interconnect (800G/1.6T)Lumentum revenue +69% TTM; Fabrinet revenue +30% TTM; rack-to-rack and pod-to-pod optical link demand outrunning supply12–18 moCoherent, Lumentum, Fabrinet (manufacturing); Marvell + Broadcom (silicon)
6EUV mask inspection & specialty WFELasertec near-monopoly in actinic mask blank inspection; ¥3.7T market cap; ¥213B FY24 revenue3–5 years (no near-term challenger)Lasertec (actinic), KLA (broader inspection), Disco (dicing/grinding)
7Liquid cooling at hyperscaleVertiv 30% organic Q1; Modine just signed $4B hyperscaler deal (2027–2029), 6x its current data center revenue12–24 mo for capacity; tech is the gating factorVertiv, Modine (Airedale), nVent, Asetek
8Specialty materials & gasesEUV photoresist controlled by 3 Japanese firms (JSR, TOK, Shin-Etsu Chem); electronic-grade neon & helium niche but recoverable2–4 years for new chemistry qualJSR, TOK, Shin-Etsu Chem, Entegris (delivery systems)
9Silicon wafers (300mm)Demand recovering after 2024 inventory destock; 4-firm oligopoly3–4 years for new ingot capacityShin-Etsu Handotai, Sumco, GlobalWafers, Siltronic
10EDA / IPCadence + Synopsys + Siemens EDA = effective triopoly with deep customer lock-in; chip-design productivity is the actual bottleneck for custom siliconPermanent moatSynopsys, Cadence, Siemens EDA
Cascade dynamic: Per Silicon Analysts Q1'26, "even if a customer secures N3 or N5 wafer slots, the chip cannot ship without CoWoS packaging." When CoWoS doubles in 2027, the gating constraint moves up the stack to HBM (HBM4 ramp) and to optical interconnect (rack-scale 1.6T). Plan for the bid to migrate.

3. Supply/demand gap model

Setting up demand-side and supply-side numbers:

Demand side (cited)

  • TSMC's own forecast: AI accelerator wafer demand growing 11x from 2022 to 2026 (TrendForce). FACT
  • CoWoS capacity: 80%+ CAGR 2022–2027 (TSMC). FACT
  • HBM TAM: ~$30B in 2024 → $80B+ in 2026e per TrendForce composite. FACT
  • Semiconductor capex of top customers (TSMC/Samsung/Intel/Micron/SK): +19.6% in 2025, +18.6% in 2026 (ASML reference). FACT
  • Global chip industry sales 2026: ~$975B target (+~25% YoY). FACT

Demand side (my estimate, AGI-pilled)

The consensus 2026–2028 capex curves assume a "normal" AI investment cycle. Reasoning from "given AGI is happening," I expect the labs to attempt to double effective compute every 6–9 months through 2028 — not the ~12–18 month cadence implied by current capex. The two delta sources will be: (1) frontier-lab self-funded GPU buildouts (xAI, Anthropic, OpenAI, Meta hitting $200B+ combined run-rate by 2027) and (2) sovereign-AI buildouts (US/EU/UAE/Saudi) accelerating from low base. ESTIMATE Net: 2026–2028 demand at the wafer level is probably 30–50% above current TSMC/ASML guidance.

Supply side (cited)

  • TSMC capex 2025: $42B; 2026e ~$45–50B; 9 fab phases starting in 2026. FACT
  • Fab build time: ~4 years shell-to-production for leading-edge. DERIVED
  • HBM line conversion: 18–24 months from greenlight to qualified bits. DERIVED
  • ASML High-NA EUV: ramping at ~5–10 systems/yr through 2027 — that is the literal upper bound on how fast the world can install new leading-edge capacity. ESTIMATE

Gap conclusion

The gap is real and structural for the next ~3 years. Supply cannot physically respond to a 30%+ demand surprise inside 2 years. Pricing power will accrue to whoever sits in the narrowest part of the funnel. By severity of "can't expand fast" in 2026–2028: EUV mask inspection > CoWoS > HBM > leading-edge logic > EUV machines > photoresist > wafers. NVIDIA is downstream and can theoretically dual-source; the chokepoints above cannot be dual-sourced.

4. Investable public companies

Prices and market caps as of close 2026-05-26 unless noted. "Already-priced-in" is my honest assessment of how much of the AGI/AI supply-constraint bid the stock has already absorbed; lower score = more headroom.

TickerMkt CapPriceValuationOwns this bottleneckKey risksPriced-in (1-10)
TSM$1.87T$412~30x FY26 P/ESole leading-edge logic + dominant CoWoSTaiwan geopolitics; Intel 18A; capex inflation6
ASML$621B$1,632~36x P/EEUV monopoly (incl. High-NA); structurally un-substitutableChina 42% of 2025 revenue, falling; export controls; long order cycle6
AVGO$2.00T$422~40x FY26 P/E (est)Custom XPU (Google/Meta) + AI networking silicon (Tomahawk/Jericho)Custom-silicon customer concentration; software (VMware) drag8
SK Hynix (000660.KS)₩1,453T (~$1.07T)₩2,333,000~13x P/E~50% HBM bit share; first to HBM4; ~70% of NVDA Rubin HBM4 allocationCyclicality risk if AI capex pauses; Samsung HBM4 finally qualified5
MUest ~$1T$895.88~15x FY26 P/EHBM3E ramp + HBM4 entry; only US-listed pure-play HBM proxyDRAM cycle; HBM yield ramp risk; only ~10% of Rubin allocation7
LRCX$404B$322.68~30x P/EEtch/deposition leader; HBM TSV etch is core to memory stack scalingWFE cyclicality; China exposure7
AMAT$361B$454.89~28x P/EBroadest WFE portfolio; advanced packaging tools; epi for GAALess HBM-levered than LRCX; China exposure6
KLAC$263B$2,011~36x P/EProcess control monopoly; every fab needs them; very high incremental marginCyclical; some Lasertec overlap7
Lasertec (6920.T)¥3.74T (~$24B)¥41,680~43x P/EActinic EUV mask blank inspection — single-source for ASML/TSMC/SamsungOne-product-line concentration; capex lumpy; ADR liquidity (LSRCY) thin4
Tokyo Electron (8035.T)¥23.2T (~$148B)¥54,500~41x P/ECoater/developer monopoly (EUV track tool); advanced etchWFE cycle; Japan-only listing for many investors5
Disco Corp (6146.T)est ~¥6T (~$38B)~40x P/E (est)De-facto monopoly in dicing/grinding for HBM TSV and CoWoSEquipment cycle; Japan listing4
Amkor (AMKR)est ~$10B~22x P/E#2 OSAT (advanced packaging overflow from TSMC); Arizona AP fab buildLower margins than TSMC; capex headwind; small vs. TSMC3
ASE Tech (ASX)est ~$30B~14x P/ELargest pure-play OSAT; FOPLP and advanced packaging investmentsCommodity packaging drag; pricing power weaker than TSMC4
Shin-Etsu Chem (4063.T)est ¥10T (~$64B)~18x P/ESilicon wafers (Shin-Etsu Handotai) + EUV photoresist + chemicalsBig conglomerate dilutes pure exposure4
SUMCO (3436.T)est $5B~18x P/E#2 silicon wafer makerSlowest part of the stack to bid up; wafer pricing weak2
JSR (4185.T)delisted — JIC tender 2024n/an/aEUV photoresist leadership — no longer investable publiclyNot investablen/a
SNPS$102B$534.56~35x fwd P/EEDA + IP; Ansys deal closed; AI design copilotsEDA growth maturing; Ansys integration risk6
CDNS$105B$381.75~47x fwd P/EEDA #2 with strong IP and verification franchisePremium multiple; competition with Synopsys6
COHR$74.6B$381~38x fwd P/E (est)800G/1.6T datacom optics; vertically integrated InP laserAlready +386% in 12mo; Aschenbrenner just sold8
LITE$70.9B$910.81~57x fwd P/EEML lasers for hyperscaler optics; CW lasers for co-packaged optics+1,455% in 12mo — bid is in price9
FN$24.4B$680.77~32x P/EContract optical/photonic mfg — silent benefactor of every optics rampCustomer concentration (NVIDIA, Cisco)7
MRVLest ~$130B~35x fwd P/ECustom XPU (one US hyperscaler, second engagement ramping); coherent DSPs; AECConcentration on one hyperscaler; NVDA/AVGO competition6
ALAB$54.6B$318.72~60x+ fwd P/EPCIe retimers + Scorpio scale-up fabric chip — only credible alt to NVLinkExtreme multiple; product-cycle risk; reliance on PCIe 6.0 adoption8
VRT$124B$323.91~50x fwd P/EData-center thermal+power leader; FY26 guide 29–31% organic growthStopped reporting backlog Q4'25 (lower transparency); priced for perfection7
MOD$14B$296.93~41x fwd P/EAiredale liquid cooling; just signed $4B hyperscaler deal (2027–29)Single-customer concentration on the new deal; non-DC business is legacy5
nVent (NVT)est ~$15B~22x P/ELiquid cooling acquisitions (CIS, ECC); rack-level thermalLess pure-play than Vertiv/Modine5
ENTG$21.7B$142.12~37x fwd P/EFluid handling, filtration, materials delivery for fabs — quiet picks-and-shovelsRevenue still flat YoY (recovery, not boom)4
Linde (LIN)est ~$220B~30x P/ELargest electronic-grade industrial gases supplier (incl. neon, helium)Industrial-gas pricing is regulated; only partial AI exposure4
Air Products (APD)est ~$70B~22x P/EElectronic gases + on-site supplyHydrogen capex overhang; less AI exposure than LIN3

5. Pre-IPO / private to watch

CompanyWhat it isStatusWhy watch
Cerebras SystemsWafer-scale AI accelerator (CS-3)IPO filed; $510M revenue, $10B OpenAI commitment, $23B valuationOnly credible alternative to NVIDIA at scale for inference; will trade like NVDA-beta
LightmatterSilicon photonics interconnect + photonic compute~$4B+ valuation; Series DIf photonic interposer adoption hits, could disintermediate copper SerDes
AYAR LabsIn-package optical I/O (TeraPHY)Late-stage private; Intel/Nvidia investorsCo-packaged optics is the most likely "next bottleneck" past copper limits at 224G
Celestial AIPhotonic Fabric memory-compute interconnectSeries C doneMemory wall is the constraint AGI labs care most about
Rain AI / EnCharge AI / MythicAnalog in-memory computeEarly; commercial in 2026–2027If energy/W is the binding constraint post-AGI, analog wins niches
TenstorrentRISC-V + AI accelerator (Wormhole/Blackhole)~$2B valuation; Korea/SK Hynix backedOpen RISC-V plus Jim Keller credibility
SiPearl / RISC-V sovereign playsEU/India/Japan/Korea sovereign AI siliconGovt-fundedSovereign-AI capex is real and will buy from non-NVDA where allowed
Groq / SambaNova / EtchedInference-specialized siliconPrivate, mid-stageInference TAM dwarfs training in the AGI deployment phase

6. Commodity / physical resource / ETF plays

  • SMH (VanEck Semiconductor): Most-liquid US semis ETF; ~25% NVDA, ~14% TSM, ~10% AVGO. Beta to the whole stack. Best "one-trade" exposure if you don't want to pick.
  • SOXX (iShares Semiconductor): Similar holdings, more equal-weighted at top. Slightly less concentrated.
  • SOXQ (Invesco): Lower fees, similar exposure.
  • FTXL (First Trust Nasdaq Semis): Adds VRT and equipment/materials weighting.
  • Materials & specialty chemicals: No clean US ETF. Closest is direct buys: ENTG, SHECY (Shin-Etsu Chem ADR), Linde (LIN). Tokyo Ohka (4186.T) and JSR (delisted) round out photoresist exposure.
  • Neon & helium: No pure-play. Linde + Air Products + Air Liquide carry it; Ukraine supply has largely recovered post-2022 dislocation. Not a current bottleneck.
  • Quartz (high-purity for crucibles): Sibelco (private) + Mineração Curimbaba (private). Effectively un-investable publicly; tracked via Shin-Etsu Chem and Sumco indirectly.
  • Cooling fluids (PFAS-free): 3M exiting PFAS; Chemours, Solvay (now Syensqo) emerging. Speculative.

7. People & talent concentration

Critical know-how concentration matters because hiring is the most leading indicator of capability transfer:

  • Advanced packaging engineers (CoWoS, FOWLP, HBM TSV): Concentrated at TSMC, Amkor, ASE, Disco. Signals: any major hire-out by Samsung Foundry or Intel Foundry indicates closing the gap. Watch LinkedIn movement at TSMC Taoyuan and TSMC Arizona.
  • EUV photonics & optics: ASML (Veldhoven) + Zeiss SMT + Cymer (acquired). No real out-flow possible — this is the single rarest engineering pool in the world.
  • HBM design teams: SK Hynix has 5-year lead from sustained investment; Samsung's HBM org was reorganized 2024–2025 after repeated qualification failures (signal: read Korean financial press on Samsung memory leadership changes).
  • Chip-architecture leadership: Jim Keller (Tenstorrent), Raja Koduri (Mihira AI), key NVIDIA alumni at xAI/Anthropic/OpenAI. Any of these landing at a public company is a 5–10% pop signal.
  • Process integration (3nm/2nm): TSMC vs. Intel Foundry attrition is the most-watched proxy for whether Intel 18A actually ramps.

Acquihire signals to watch: any large public buyer paying premium for a 30–50 person photonics, in-memory compute, or HBM controller startup is a tell that the listed players see a gap.

8. Top picks ranked

#1 — SK Hynix (000660.KS) HIGH CONVICTION

Thesis: ~50% bit share in HBM in a market that grew from $30B (2024) to $80B+ (2026e) and is forecast in shortage through 2028. ~70% of NVIDIA Rubin HBM4 allocation. HBM4 per-stack pricing is +67% generational uplift over HBM3E 12-Hi, which is unheard of in DRAM. P/E of ~13x is materially below TSM (~30x), Lasertec (~43x), and the WFE names (~30–36x). Despite a +1,043% market-cap increase over the year, valuation is still cyclical-discount because the market treats it as DRAM. The whole point of the AGI thesis is that HBM is no longer cyclical — it's structurally undersupplied.

Why above #2: Cheapest valuation in the whole substrate stack, sitting on the most acute non-CoWoS bottleneck, with Korean listing access friction depressing the bid relative to fundamentals. ADR (HXSCL) exists for US holders.

#2 — TSMC (TSM) HIGH CONVICTION

Thesis: Sole supplier of leading-edge logic AND dominant advanced packager. Already crystallized as a near-monopoly with 30x earnings. The only reason it isn't #1 is that some of the bid is in price. But $1.87T market cap on a company that literally cannot lose the next 3 years of AI capex is still cheap on a sum-of-parts basis. CoWoS sold out through 2027; 80% CAGR through 2027; 9 fab phases breaking ground in 2026.

Why above #3: Monopoly + monopsony-like pricing power with NVIDIA. Geopolitical discount remains visible vs. ASML.

#3 — Lasertec (6920.T) MEDIUM-HIGH CONVICTION

Thesis: Effective monopoly on actinic EUV mask blank inspection — without their tools, every High-NA EUV mask is shipped blind. ~$24B market cap is small relative to the gating role. The risk-adjusted upside in a High-NA EUV ramp scenario is asymmetric. ADR (LSRCY) is illiquid; Tokyo direct preferred.

Why above #4: Smallest mega-cap exposure to the deepest single-supplier moat in the stack. The market under-prices Japanese specialty equipment makers due to listing friction.

#4 — ASML MEDIUM CONVICTION

Thesis: EUV monopoly is permanent. Trading at ~36x P/E with TTM FCF margin ~28%. China is a real overhang (42% of 2025 sales, falling sharply). 2026 base case is only ~5% revenue growth because China decline offsets core customer +18.6%. But High-NA EUV upgrade cycle through 2028 is uncorrelated with cycle worry. Below DCF fair value of ~$1,233 vs $1,632 spot — wait, that's actually above fair value. The Trading News piece argues upside but the multiple is rich.

Why not higher: Multiple already prices in monopoly. China downside risk is real and not in numbers. Slot 4, not slot 1.

#5 — Vertiv (VRT) MEDIUM CONVICTION

Thesis: Best public exposure to the cooling/power bottleneck. FY26 organic guide 29–31%. Liquid cooling is mandatory for B200/B300/Rubin density. 50x fwd P/E is rich but the data-center thermal market is in early innings of a structural reset. Concerning: management stopped reporting quarterly backlog Q4'25, citing "lumpiness" — that's a transparency downgrade right when investors most want it.

Honest note: Modine (MOD) is the higher-beta version of the same thesis — $4B deal at 6x current DC revenue is mathematically more meaningful than any single VRT print. But VRT is the steadier compounder. Pick MOD if you want torque, VRT if you want quality.

Honorable mentions (not in top 5 but worth tracking)

  • Amkor (AMKR) — small, undervalued OSAT #2 with Arizona AP fab build. If TSMC AP allocation gets rationed, overflow goes here. Best risk/reward in OSAT.
  • Disco Corp (6146.T) — quiet monopoly in HBM/CoWoS dicing & grinding. Japanese listing friction = persistent discount.
  • Entegris (ENTG) — quiet materials/filtration play; not yet bid up like equipment. Highest-quality compounder of the materials names.

9. What would change my mind

  • Frontier-lab capex pause >6 months. If Anthropic/OpenAI/Meta/xAI collectively pull back GPU orders, CoWoS lead times compress fast and the bid evaporates. Watch hyperscaler Q4 capex commentary.
  • Real algorithmic efficiency breakthrough. A model that achieves GPT-5-class capability on 1/10 the compute kills the supply/demand gap. Possible but historically Moore's-Law-for-models has been "more compute = more capability" not "less."
  • HBM yield breakthrough at Samsung. If Samsung qualifies HBM4 at NVDA in volume (already qualified in Q1'26 per cited source), price discipline cracks and SK Hynix's pricing power erodes.
  • Intel 18A actual ramp. If Intel Foundry hits 18A volume with credible yields by H2 2026, TSMC pricing power compresses. Currently no evidence this happens — but watch for big-customer commitments (Microsoft, Apple).
  • Taiwan strait incident. Tail risk; not investable as base case but warrants partial geo-diversification (ASML, Samsung, Intel Foundry hedges).
  • China EUV breakthrough. SMEE/Huawei prototypes claimed at ~2028. If credible by 2027, ASML's China discount cap goes from "decline" to "structural loss." Currently 10–15 years behind on light source per cited assessment.
  • Co-packaged optics actually hits. If silicon photonics CPO ramps in 2027 as some Marvell/Broadcom slides claim, Lumentum/Coherent transceiver demand could compress fast. Watch hyperscaler CPO RFPs.

10. Sources & date

Analysis date: 2026-05-26. All prices and market caps are as of market close that date except where noted.

Disclaimer: Analysis prepared for internal investment thinking. Not investment advice. Estimates labeled where used. Multiple data points are derived from secondary sources; primary 10-Q/10-K verification recommended before sizing positions.