Dylan Patel — AI Compute Bottlenecks Investment Thesis

Source: Dwarkesh Podcast — "Deep dive on the 3 big bottlenecks to scaling AI compute", Dylan Patel / SemiAnalysis, March 13, 2026.

The Framework: The "X-1 Supply Chain" Problem

Dylan's organizing model: every supplier in the AI compute stack builds X-1 of what AI actually needs, because nobody outside the labs is fully AGI-pilled. This creates perpetual scarcity at the most constrained link — and the most constrained link earns the highest margin.

Constraint Layer	Timeline	Most Constrained Supplier	Investable Implication
Fab space / cleanrooms	Now–2027	TSMC, Samsung, Micron	Anything that speeds fab construction
Logic wafers (TSMC N3/N2)	2026–2028	TSMC (NVIDIA has 70%+ allocation)	TSMC; EDA tools for chip design velocity
Memory (HBM)	Now–2028+	SK Hynix, Micron, Samsung	HBM vendors; consumer electronics (hurt)
EUV lithography tools	2028–2030	ASML (sole supplier)	ASML; fab tool ecosystem (AMAT, LRCX)
Power / energy	Now (near-term)	Gas turbines, fuel cells	GE Vernova; Bloom Energy; not terminal constraint

Investment Thesis #1: ASML Is the Most Underpriced Monopoly in Technology

"By 2028 or 2029, the bottleneck falls to the lowest rung on the supply chain, which is ASML... Currently they can make about 70 [EUV tools]. Next year 80. Even under very aggressive supply chain expansion, they only get to a little bit over 100 by the end of the decade."

The argument: 3.5 EUV tools = 1 GW of AI compute. 1 GW of compute = ~$50B in data center CapEx. The entire $50B rests on $1.2B of tooling ASML produces at an un-expandable rate. ASML has 700 tools in cumulative fleet by 2030 → 200 GW ceiling on global AI compute, if all allocated to AI. Sam Altman wants 52 GW/year alone — that's 25% of the ceiling.

The margin anomaly: Unlike Nvidia (70%+ margins) and memory vendors (tripling price), ASML has never raised prices faster than capability gains. This chronic underpricing is a gift to compute buyers that will not last indefinitely.

The market debates Nvidia vs. AMD. Nobody is pricing ASML as the true choke-point. When the constraint cascades from fab space to EUV tools in 2028, ASML's supply/demand math becomes undeniable.

Trigger: ASML quarterly unit shipments vs. demand guidance; any analyst revision acknowledging EUV as compute ceiling.

Names: ASML (ASML) — sole EUV supplier; Applied Materials (AMAT) and Lam Research (LRCX) — the second-tier fab tooling bottleneck (CVD, etch chambers). ASML's optics supplier Carl Zeiss (private).

Investment Thesis #2: GPU Depreciation Is Inverted — Long-Term Contract Holders Win

"An H100 is worth more today than it was three years ago."

The argument: GPT-5.4 is both more capable AND cheaper per token than GPT-4 — meaning each H100 now serves a model with a $100B+ token TAM vs. a few billion for GPT-4. H100 TCO is $1.40/hr over 5-year depreciation. Current market is signing 2–3 year deals at $2.40/hr. The company that committed early (long-term 5-year contracts at $1.40 cost) is generating 70%+ gross margins.

Michael Burry argued for 3-year GPU depreciation. Patel's data shows the opposite — as models improve, chips that run them become more valuable, not less. The whole bear case on GPU clouds is built on the wrong depreciation assumption.

Trigger: Any H100/Blackwell spot pricing above $2/hr; CoreWeave gross margins in earnings; GPU contract renewal rates.

Names: CoreWeave (CRWV) — 98%+ of compute on 3+ year long-term contracts locked in at historical cost; the pure-play expression of this thesis.

Investment Thesis #3: The Memory Crunch Is the Most Actionable Near-Term Trade

"Smartphone volumes are going to go from 1.4 billion to potentially 500 or 600 million next year. 30% of Big Tech's CapEx in 2026 is going towards memory."

The argument: HBM uses 4x more wafer area per bit than standard DRAM. AI is bidding up HBM prices, destroying DRAM supply for consumer devices. Smartphone volumes collapsing 40%+. iPhone BOM increases ~$150–250. Memory vendors didn't build new fabs 2021–2024 (losses in 2023). New fabs take 2 years — meaningful relief not until late 2027.

The market treats memory as a commoditized cyclical. Patel identified this crunch 1.5–2 years before it hit pricing, based on reasoning + long context = large KV cache = massive memory demand. The scarcity is structural, not cyclical.

Trigger: SK Hynix / Micron earnings — HBM ASP guidance; quarterly IDC smartphone shipment data tracking toward 500-600M; Apple BOM press reports citing memory cost increases.

Names: SK Hynix (000660.KS / HXSCF) — #1 HBM supplier, locked-in Nvidia relationship; Micron (MU) — acquiring Taiwan fab, catching up; Samsung (SSNLF) — secondary HBM beneficiary.

Hurt: Apple (AAPL) — ~$150–250 BOM increase per iPhone; Xiaomi / Oppo — cutting low/mid-range volumes by half.

Investment Thesis #4: Power Is Solved by Capitalism — Bloom Energy Is the Contrarian Bet

"Any of [the power sources] individually will do tens of gigawatts, and as a whole they will do hundreds of gigawatts... They're just way simpler than chips."

The argument: Patel explicitly pushes back on the "energy is the terminal AI constraint" thesis. He identifies 16+ power generation manufacturers across combined-cycle gas, aeroderivatives, ship engines, reciprocating engines (diesel→gas), and fuel cells — with Bloom Energy named specifically as having "fast production ramp and quick payback period." Power constraints are real but solvable through capitalism; chip manufacturing cannot be solved the same way.

The consensus AI energy trade is nuclear (Constellation, Vistra) and combined-cycle (GE Vernova). Bloom Energy is the under-owned alternative. Patel has been "very positive on them for a year and a half" — a rare Tier 1 endorsement for a non-chip company.

Trigger: Bloom Energy data center customer wins; GE Vernova order backlog and turbine lead times; behind-the-meter power announcements from hyperscalers.

Names: Bloom Energy (BE) — fuel cells with fast production scale-up, explicitly endorsed; GE Vernova (GEV) — combined-cycle turbines (sold out through 2030; the natural scarcity play on the gas turbine oligopoly).

Investment Thesis #5: China Timeline — Export Controls Are the Single Decisive Variable

"It's fast timelines, the US wins; long timelines, China wins."

The argument: At current trajectory (Anthropic/OpenAI at 10 GW each by end of 2027), Chinese labs cannot close the compute gap. US model distillation becomes harder as models stop showing reasoning chains. But if AGI-level timelines slip to 2035, China will have fully indigenized DUV (by 2030) and working EUV. Their manufacturing volume will exceed the West.

Most US investors assume US wins regardless of timeline. Patel's thesis is explicitly conditional: fast takeoff = US; slow takeoff = China. Huawei without TSMC would already be the dominant AI chip company. Export controls on ASML are the single most important policy lever in the race.

Trigger: ASML export control policy changes; China domestic EUV tool announcements; Huawei Ascend performance benchmarks vs. Nvidia.

Names: Hold ASML (ASML) as both an investment and a geopolitical hedge. Any loosening of EUV export restrictions is immediate downside for US compute supremacy; any tightening is upside for ASML's domestic market.

Ecosystem Map

Chokepoint monopolies (highest margin capture):

ASML — EUV tools; sole supplier; chronically underprices vs. value enabled
Nvidia (NVDA) — GPU platform; 70%+ gross margins; locked up TSMC N3 allocation years early

Memory infrastructure (near-term scarcity):

SK Hynix (HXSCF) — #1 HBM; Nvidia-preferred supplier; no fast substitute
Micron (MU) — catching up; domestic US HBM ramp; fab acquisition signal
Samsung (SSNLF) — broad memory exposure; 3D DRAM R&D bet

Fab tooling (second-tier bottleneck):

Applied Materials (AMAT) — CVD, etch, deposition; expanding capacity
Lam Research (LRCX) — etch chambers; direct fab bottleneck downstream from ASML
KLA Corporation (KLAC) — metrology; every fab needs inspection tools

Power (near-term bottleneck, more solvable):

Bloom Energy (BE) — fuel cells; fast scale-up; Patel's specific endorsement
GE Vernova (GEV) — gas turbines; sold out through 2030; pure scarcity pricing

Long-term compute holders:

CoreWeave (CRWV) — 98%+ on 3+ year GPU contracts at historical cost

Key Risks

Taiwan risk is existential: TSMC disruption collapses global AI capacity from 200+ GW/year to ~20 GW across Intel and Samsung.
3D DRAM arrives early: if in mass production before 2030, bits-per-EUV-pass improves dramatically, relieving memory constraints faster than projected.
AI revenue inflection stalls: if Anthropic's $20B ARR doesn't compound, demand for HBM and EUV moderates.
Alternative lithography (X-ray / synchrotron): low probability this decade but would disrupt ASML's monopoly position.
ASML and TSMC never take aggressive pricing: their chronic underpricing relative to value enabled could persist indefinitely.

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	ASML (ASML)	Sole EUV supplier; 3.5 tools = 1 GW; 200 GW ceiling by 2030	Named as #1 constraint for AI compute scaling by 2030
1	SK Hynix (HXSCF)	#1 HBM; 4x wafer area per bit vs. DRAM; Nvidia locked-in	30% of Big Tech CapEx going to memory; smartphone volumes halving
1	NVIDIA (NVDA)	GPU platform + 70%+ N3 wafer allocation + $90B LT contracts	H100s worth more today; GPU depreciation thesis is inverted
1	CoreWeave (CRWV)	H100 TCO $1.40; current market pricing $2.40; 98%+ LT contracts	Long-term contract margin thesis confirmed by spot pricing
2	Applied Materials (AMAT)	CVD/etch fab tooling; secondary bottleneck downstream ASML	Expanding capacity; every new fab requires their tools
2	Lam Research (LRCX)	Etch equipment; direct fab production bottleneck	Same thesis as AMAT; dual-exposure to logic and memory fabs
2	GE Vernova (GEV)	Gas turbines sold out through 2030; oligopoly pricing power	Hundreds of GW of data center power orders confirmed
2	Micron (MU)	HBM ramp; Taiwan fab acquisition; catching up to SK Hynix	January 2026 fab acquisition = committed to HBM buildout
3	Bloom Energy (BE)	Fuel cells for behind-the-meter data center power; fast ramp	Named specifically by Patel as "very positive for 1.5 years"
3	Samsung (SSNLF)	HBM + DRAM + NAND + 3D DRAM R&D; broad memory exposure	Memory price triple; 3D DRAM roadmap option
4	ASML EUV capacity option	Theoretical: lock up forward EUV tool contracts before demand crisis	Patel explicitly described this arbitrage as rational but unexecuted

Monitoring Checklist

Bottom Line

ASML at 3.5 tools per GW is the most concentrated chokepoint in technology. $1.2B of tooling enables $50B of compute. The market prices ASML as a semiconductor equipment company; it should be priced as the toll booth on the singularity.
Michael Burry's GPU depreciation thesis is empirically wrong. H100s signing at $2.40/hr when TCO is $1.40. The value of a GPU increases as better models run on it. Buy the long-term contract holders (CoreWeave); short the depreciation narrative.
Bloom Energy is the most under-owned data center power play. Patel has been explicitly positive for 1.5 years; fast production scale-up, quick payback, no dependency on turbine blade backlog. Not on most AI infrastructure watchlists.
SK Hynix is the cleanest memory trade. The scarcity is structural (HBM = 4x wafer area per bit), the pricing power is confirmed (memory tripling), and the customer lock-in is established (Nvidia preferred supplier). Smartphone volume collapse releases no relief to HBM.
The whole supply chain builds X-1. Buy the next constraint 12–18 months early. Cleanroom/fab space is acute now. EUV tool ceiling arrives 2028. The pattern is fully predictable; the market only prices it after it's acute.