Dylan Patel — AI Compute Bottlenecks Investment Thesis

Source: Dwarkesh Podcast – "Deep dive on the 3 big bottlenecks to scaling AI compute"
Speaker: Dylan Patel, founder & CEO, SemiAnalysis
Host: Dwarkesh Patel
Date: March 13, 2026
Summary Generated: May 2, 2026

Why This Interview Matters

Dylan Patel runs SemiAnalysis, the most cited supply chain intelligence firm covering the AI compute stack. His data subscribers include every major hyperscaler, AI lab, neocloud, and semiconductor company, plus ~40% from hedge funds. This is not commentary — it is on-the-ground semiconductor supply chain tracking, updated daily, across every fab, data center, wafer order, and tooling contract globally.

The central, non-consensus thesis of this interview: the dominant bottleneck to AI compute scaling is not power, not data centers, not software — it is the physical semiconductor supply chain, ultimately bottlenecking at ASML's ability to produce EUV lithography machines.

This conversation provides an investment framework that maps exactly which entities will extract the most margin dollars as AI compute scales from ~20 GW today to a projected 200 GW by end of decade.

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

The Core Claim

"By 2028 or 2029, the bottleneck falls to the lowest rung on the supply chain, which is ASML. ASML makes the world's most complicated machine: an EUV tool... Currently, they can make about 70. Next year, they'll get to 80. Even under very aggressive supply chain expansion, they only get to a little bit over 100 by the end of the decade."

Why This Is Investable Alpha

The math is extraordinary:

3.5 EUV tools = 1 GW of AI compute capacity
1 GW of compute = ~$50 billion in data center CapEx
3.5 EUV tools costs = $1.2 billion
The ratio: $50B of economic value resting on $1.2B of tooling

The supply ceiling:

ASML has ~250–300 installed EUV tools today
Adding 70/year → 80/year → max ~100/year by 2030
Total fleet by 2030: ~700 EUV tools
700 ÷ 3.5 = 200 GW maximum global AI compute capacity by end of decade

The demand gap:

Sam Altman wants 52 GW/year by 2030
Anthropic, Google, Meta, Microsoft each want similar scale
Total demand from major players alone would exceed the 200 GW ceiling if all allocated to AI

ASML's pricing anomaly:

"ASML has never raised the price more than they've increased the capability of the tool."

EUV tools went from $150M to $400M, but throughput and overlay accuracy have more than doubled. ASML is chronically undercharging relative to the value it enables. Nvidia takes 70%+ gross margins. Memory vendors are doubling/tripling price. ASML has held pricing discipline that no rational monopoly should maintain.

"They haven't taken price and margins up like crazy... you can take the margin. Nvidia takes the margin. Memory players are taking the margin. But ASML has never raised the price more than they've increased the capability of the tool."

Why the supply chain can't scale fast:

4 core components: Zeiss optics (Germany), Cymer source (San Diego), reticle stage (Wilmington CT), wafer stage (Eindhoven)
10,000+ suppliers in ASML's supply chain
The EUV source fires tin droplets hit by three successive lasers at a precise shape to emit 13.5nm light
Zeiss lenses require 18 multilayer mirror optics per tool, each with sub-nanometer accuracy
The machine moves at 9Gs, requires sub-1nm alignment across physical parts, and takes months to install at customer sites
Cannot be meaningfully expanded on short timescales — "too artisanal"

The AGI-pilled gap: ASML management does not believe demand projections from the AI labs. Patel says the entire supply chain keeps telling him "your numbers are too high" — until 6–12 months later when they're exactly right. ASML does not see demand for 200 GW/year of AI chips, so they're not expanding aggressively.

Investment implication:
ASML ($ASML) is the ultimate "picks and shovels" monopoly with zero competition, demonstrably underpricing its products, and sitting at the chokepoint of a supply-constrained 10x growth market. The market is only beginning to price in the ASML ceiling. Applied Materials ($AMAT) and Lam Research ($LRCX) are secondary bottlenecks in the same constraint chain.

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

The Core Claim

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

This is the most counterintuitive claim in the entire interview — and it has massive financial implications.

Why This Is Investable Alpha

The standard bear case: Nvidia releases faster chips every 18 months; H100s depreciate rapidly because Blackwell is 3–5x more performant; therefore cloud providers have been overbuilding.

Why the bear is wrong:

GPT-5.4 (released March 2026) is far more capable AND cheaper to run per token than GPT-4
An H100 can serve more tokens per GPU of GPT-5.4 than of GPT-4
The TAM for GPT-5.4 tokens: "probably north of a hundred billion dollars"
The TAM for GPT-4 tokens: "maybe a few billion, maybe tens of billions"

"If improvements stopped here, the value of an H100 is now predicated on the value that GPT-5.4 can get out of it instead of the value that GPT-4 can get out of it."

The long-term contract advantage:

H100 TCO: $1.40/hour over 5-year depreciation at volume
Current market H100 deals: $2.40/hour for 2–3 year contracts signed in 2026
Companies that locked in compute at $1.40 cost, now renting at $2.40, are generating 70%+ gross margins
CoreWeave (CRWV): 98%+ of compute on 3+ year long-term contracts

"The person who committed early has better margins in general. The percentage of the market that is in long-term contracts is much larger than the percentage of the market in short-term contracts."

The Anthropic vs. OpenAI lesson:

Anthropic was deliberately conservative on compute commitments ("principled, don't want to go bankrupt")
OpenAI signed "crazy deals" with Microsoft, Google, Amazon, CoreWeave, Oracle, NScale, SoftBank Energy
Result: OpenAI will end 2026 with significantly more compute than Anthropic
Anthropic at $20B ARR must now pay premiums for last-minute spot compute or route through Bedrock/Vertex/Foundry revenue-share arrangements

"Anthropic was a lot more conservative. They were like, 'We'll sign contracts, but we'll be principled.' In some sense, Dario has screwed the pooch compared to OpenAI."

Investment implication:
CoreWeave (CRWV) is the purest expression of the long-term contract thesis — locked in H100 capacity at $1.40 TCO, now renting at $2.40+. If H100 value continues to appreciate as models improve, their position strengthens further. The Alchian-Allen effect (as fixed GPU costs rise, customers prefer paying marginal premium for the best model) also structurally benefits premium cloud/inference providers.

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

The Core Claim

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

Why This Is Investable Alpha

The math:

Big Tech combined CapEx $600B this year; 30% = **$180B going to memory**
HBM uses 3–4x more wafer area per bit than standard DRAM
AI wants HBM; AI will destroy consumer DRAM demand to get it

The cascade:

Memory vendors didn't build new fabs 2021–2024 because prices were low (losing money in 2023)
"Reasoning models → long context windows → large KV cache → enormous memory demand" — Patel flagged this 1.5–2 years before it showed in prices
New fabs take 2 years to build; meaningful new capacity won't arrive until late 2027 or 2028
HBM bandwidth vs. DRAM: 2.5 TB/sec per HBM4 stack vs. ~128 GB/sec for equivalent DDR5 shoreline — order of magnitude difference; there is no DDR substitute

Consumer device impact:

iPhone memory BOM: $50 → $150 (DRAM alone, tripling prices); total memory/storage increase: ~$150–250 per iPhone
Xiaomi and Oppo cutting low-end/mid-range smartphone volumes by half
Smartphone volumes: 1.4B → potentially 500–600M (SemiAnalysis projection)
NAND prices also rising, further squeezing consumer electronics

Who benefits:

SK Hynix (HXSCF/000660.KS): #1 HBM supplier; Nvidia locked them up on long-term deals
Samsung (SSNLF): secondary HBM beneficiary; also moving faster on 3D DRAM R&D
Micron (MU): acquiring Taiwan fab; late to HBM ramp but catching up

Who gets hurt:

Apple (AAPL): becoming a smaller percentage of TSMC revenue; facing massive BOM increases; customers feeling squeeze on iPhone pricing
Consumer electronics broadly: PC gaming GPU prices already causing memes — "it's going to be even worse when memory prices double again"

"People are going to hate AI even more. Today, you already see all the memes on PC subreddits and gaming PC Twitter... It's going to be even worse when memory prices double again."

Key forward-looking signal from Patel:
Micron's acquisition of a Taiwan fab (January 2026) signals the memory vendors are finally acting. But structural relief isn't coming before 2027–2028.

Investment implication:
SK Hynix is the highest-conviction memory play given their established HBM supply relationship with Nvidia. Samsung is a broader bet. Micron is a catch-up play with more execution risk. All three benefit from the same structural dynamic.

Investment Thesis #4: Power Is a Solved Problem — Not the Long-Term AI Bottleneck

The Core Claim

"The supply chains [for power] are just way simpler than chips... at the end of the day, it's still simple enough that we will be able to solve it through capitalism and human ingenuity on the timescales required."

Why This Is Investable Alpha (Contrarian)

The dominant narrative in 2024–2025 was "energy is the scarce resource for AI." Patel explicitly pushes back, naming a much longer list of power generation sources than the market is pricing:

Source	Notes
Combined-cycle gas turbines	Only 3 major manufacturers; long lead times on blades
Aeroderivatives (jet engines repurposed)	10+ manufacturers; Boom Supersonic + Crusoe model
Reciprocating engines (diesel → gas)	10+ manufacturers; auto demand decline = freed capacity
Ship engines	Already deployed for Microsoft data center in NJ
Bloom Energy fuel cells	Named specifically as fast scale-up, quick payback
Solar + battery	Cost curves continuing to fall
Utility-scale batteries	Can unlock 20% of US grid's idle peak-reserve capacity

"There are 16 different manufacturers of power-generating things just from gas alone... Any of these individually will do tens of gigawatts, and as a whole, they will do hundreds of gigawatts."

The key insight on cost tolerance:

Even if power cost doubles, H100 TCO goes from $1.40 to $1.50/hour
This 10-cent increase is irrelevant when models are improving at the current rate
The value per GPU is going up faster than any cost increase from energy

The grid battery unlock:

US grid = ~1 terawatt
20% sits idle as peak reserve, only needed a few hours per year
Utility-scale batteries absorb peak load → that 200 GW unlocked for data centers
Regulatory mechanism is the hard part, not the physical infrastructure

Bloom Energy specifically:

"Bloom Energy fuel cells... we've been very positive on them for a year and a half now because they have such a capability to increase their production. Their payback period for a production increase is very fast, even if the cost is a little bit higher than combined-cycle."

Investment implication:
Power is NOT the terminal constraint — chips are. This means:

The "pure energy plays" thesis (nuclear, grid infrastructure) may be somewhat overbought as terminal AI bottleneck plays
Bloom Energy (BE) is a specific contrarian bet on a fuel cell provider with fast capacity ramp that Patel has flagged repeatedly
GE Vernova, Vistra, Constellation, and other "AI energy" plays remain valid but should be sized as infrastructure support plays, not choke-point monopolies

Investment Thesis #5: China Timeline — Fast Timelines = US Wins; Long Timelines = China Wins

The Core Claim

"If timelines are slow enough [to AGI], I don't see why China wouldn't be able to catch up drastically."
"We're in a fast takeoff. The revenue is compounding at such a rate that it does affect economic growth. China hasn't done that yet."

Why This Is Investable Alpha

Dylan's framework:

Fast timeline (AGI-like capabilities 2026–2030): US wins. Anthropic and OpenAI will each be at 10 GW by end of 2027. Chinese labs cannot match this compute scale. As US models become opaque (stop showing reasoning chains), distillation into Chinese models becomes much harder. US economic growth compounds from AI revenue.
Slow timeline (2035+): China wins. They will have fully indigenized DUV by 2030, working EUV by 2030. They can outscale the West on sheer manufacturing volume + vertical integration.

China's current status:

DeepSeek and other Chinese models "as competitive as they've ever been" but Opus 4.6 and GPT-5.4 have "pulled away a little bit"
China uses ASML DUV tools for their 7nm and 14nm capacity — still dependent on Western tooling
By 2030: China likely has ~100 DUV tools/year from domestic production
By 2030: China has working EUV but not in mass production (production hell takes 5–7 years, as ASML experienced)

The most critical inflection:

Huawei's Ascend 910C/D is competitive — 60% of H100 inference performance per Dylan's data
"If Huawei had TSMC, they would have a better accelerator than Rubin." (Huawei has better networking, AI research talent, and end-market than Nvidia in several dimensions)
The US export control of TSMC access to Huawei in 2020 is the single most important policy decision in the AI race

Investment implication:
The US-China AI divergence is primarily a function of compute access, not model architecture. Export controls on TSMC access, ASML tools, and HBM are the key levers. Companies that benefit from a fast US AI takeoff (inference providers, model labs, compute infrastructure) are positioned well. Taiwan risk (TSMC disruption) is existential for global AI and should be priced as a fat-tail risk in any compute-heavy portfolio.

Investment Thesis #6: The "Everyone Is X-1" Supply Chain Underestimation Creates a Perpetual Mismatch

The Core Claim

"OpenAI and Anthropic know they need X. Nvidia is not quite as AGI-pilled. They're building X - 1. You go down the supply chain, everyone's doing X - 1. In some cases, they're doing X ÷ 2."

Why This Is Investable Alpha

The semiconductor supply chain repeatedly tells Patel his projections are too high — and is repeatedly proven wrong 6–12 months later. This creates a systematic, exploitable inefficiency:

The leopard-spots pattern:

Labs identify demand for X compute
Nvidia ships X-1 (not fully AGI-pilled)
TSMC allocates for X-2 (wants stable CPU/mobile business)
Memory vendors build for X-3 (lived through 2022–2023 price crash, traumatized)
ASML targets X-4 (doesn't believe AI is permanent demand surge)

The compound undershoot:

"'We're told our numbers are way too high, and then when they're right, they're like, Oh, yeah, but your next year's numbers are still too high.'"

What this means in practice:

The "memory crunch" was predictable 1.5–2 years in advance but nobody priced it
The "fab space crunch" (cleanroom capacity) is predictable now but nobody is pricing it
The "EUV tool ceiling" for 2028–2030 is predictable now but nobody is pricing ASML at its correct value

This year's undershoot: Anthropic was so conservative on compute commitments that they are now compute-constrained. Their reliability is low because they can't serve demand. The $20B ARR business is being throttled by insufficient compute.

Investment implication:
The systematic underestimation creates repeatable trades: buy the supply chain constraint 12–18 months before it becomes acute. Patel's current forward warnings:

Cleanroom/fab space — bottleneck THIS year and next
EUV tool capacity — bottleneck by 2028–2030
3D DRAM transition — could either relieve or tighten memory constraints depending on retooling timeline

Key Company-Level Investment Summary

Company	Ticker	Investment Case	Conviction
ASML	ASML	EUV monopoly; 3.5 tools = 1 GW; chronically underpricing its product; ultimate AI choke-point	Tier 1
SK Hynix	000660.KS / HXSCF	#1 HBM supplier; locked-in Nvidia relationship; smartphone volume collapse frees capacity for AI	Tier 1
CoreWeave	CRWV	Long-term H100 contracts at $1.40 TCO now worth $2.40+; 98%+ on 3+ year deals	Tier 1
Applied Materials	AMAT	Fab tooling (CVD, etch) — second-tier ASML constraint; expanding capacity aggressively	Tier 2
Lam Research	LRCX	Same category as AMAT; fab process equipment bottleneck downstream from ASML	Tier 2
Micron	MU	Catching up on HBM; acquiring Taiwan fab; further from pole position than SK Hynix	Tier 2
NVIDIA	NVDA	The central GPU platform; locked up TSMC N3 allocation years early; 70%+ gross margins	Tier 1
Bloom Energy	BE	Named specifically by Patel; fuel cells with fastest production ramp and quick payback period	Tier 3
GE Vernova	GEV	Combined-cycle turbine manufacturer; demand for data center power; long turbine lead times	Tier 2

Key Figures and Data Points

Metric	Value
Big Tech combined CapEx 2026	~$600B (hyperscalers alone); ~$1T across full supply chain
Memory share of Big Tech CapEx	~~30% (~~$180B)
AI compute deployed this year	~20 GW (US), adding 20 GW incremental
H100 TCO (5-yr, volume)	$1.40/hour
H100 current long-term contract price	$2.40/hour (2–3 year deals)
EUV tools produced/year (today)	~70; max ~100 by 2030
EUV tools per GW of AI compute	3.5
EUV cost per tool	$300–400M
Effective AI CapEx per GW	~$50B
EUV-enabled AI capacity by 2030	~200 GW (if all allocated to AI)
Anthropic ARR (early 2026)	~$20B
Anthropic gross margins	Sub-50% (as reported by The Information)
Anthropic compute spend implied	~$13–14B rental cost; ~$50B CapEx underlay
Anthropic target GW by end 2026	~5–6 GW (up from initial conservative plans)
OpenAI GW by end 2026	~6+ GW (ahead of Anthropic)
H100 inference perf vs. Blackwell	Blackwell 20x faster for DeepSeek/Kimi K2.5 at 100 tokens/sec
Smartphone volumes (projection)	1.4B → 500–600M (SemiAnalysis)
iPhone BOM memory/storage increase	~$150–250 per device
Fab construction timeline	2 years (vs. 8 months for a data center)
GPU RMA rate (Blackwell)	~15% on initial deployment
CoreWeave long-term contract %	98%+ on 3+ year contracts

Contrarian and Non-Consensus Views

H100s appreciate in value over time — the entire market assumes GPU depreciation; Patel's data shows the opposite is occurring as model utility and token TAM expand
Power is NOT the critical long-term bottleneck — contradicts the dominant narrative from Jensen Huang, data center operators, and most energy investment theses; chips and EUV are the real ceiling
Smaller model + more RL > larger model + less RL — the research feedback loop favors smaller, faster models; this is why OpenAI and Anthropic build smaller than Google; Google's advantage is unipolar TPU fleet that can RL at scale
China wins on slow timelines — almost nobody says this publicly; most US investors assume US wins regardless of timeline; Patel explicitly says if AGI takes until 2035, China has the manufacturing advantage
Anthropic's conservative capital approach was a strategic error — the market treated Dario's capital discipline as responsible; Patel calls it "screwing the pooch" — OpenAI's "crazy deals" strategy was correct
ASML has been chronically undercharging — one of the most powerful monopolies in the world has never taken the margin Nvidia and memory vendors take; this should be viewed as a structural gift to compute buyers that won't last forever

Risks and Caveats

Taiwan risk is real and existential: TSMC disruption would collapse global AI capacity from hundreds of GW/year to ~20 GW across Intel and Samsung. This is the primary tail risk in every position in this space.
China closing the gap faster than expected: Patel says China likely has indigenized DUV by 2030 and working EUV. If production hell is faster than expected, the US compute advantage window is shorter.
AI revenue inflection stalls: The entire thesis is predicated on AI revenue growing fast enough to justify the CapEx. If Anthropic's $20B ARR doesn't compound to $200B, demand for H100s, HBM, and EUV tools moderates.
3D DRAM arrives early: If 3D DRAM enters mass production before 2030, the bits-per-EUV-pass equation changes dramatically, relieving memory constraints faster than projected.
Alternative lithography disruption: Low probability but Patel acknowledges X-ray lithography (particle accelerator-based) companies could disrupt ASML's monopoly "beyond EUV." Assigns very low probability for this decade.