Memory Is The Bottleneck

“memory is the bottleneck in AI, not compute”

memoryhbmsemiconductorsai infrastructureby theory · 19 days ago

AI accelerators are now bound by how fast data moves in and out of memory, not by raw compute flops, so the scarcity premium is shifting to DRAM, HBM and NAND makers. Micron and SK Hynix sit at the center of the HBM supply crunch feeding every major GPU cluster. Packaging and equipment names capture the picks and shovels needed to stack and test that memory, while interconnect chips like Astera Labs and Credo solve the actual bandwidth choke point between compute and memory. This basket bets on the physical constraint, not the model layer.

Holdings: 9
Largest: 30%
Pillars: 3
Revisions: 1

Connect a wallet to launch this theory as a vault on Robinhood Chain. You sign it, so the vault is yours to revise.

How it breaks down

Memory Chipmakers

67%

The companies that actually make the DRAM, HBM and NAND that AI systems are starved for.

Packaging and Equipment

23%

The physical build-out layer, advanced packaging and fab tools, that determines how fast memory supply can actually grow.

Bandwidth Interconnect

10%

Chips that move data between compute and memory faster, attacking the bottleneck from the connectivity side.

Holdings

Every position, why it is here, and its token on chain. Weights sum to exactly 100%.

Weight	Ticker	Company	Why it is here	Token
30%	MU	Micron Technology	Micron is the largest US HBM and DRAM producer and is sold out of leading edge memory capacity well into next year.	0xfF08…4afD
15%	SKYHY	SK Hynix	SK Hynix is the dominant HBM supplier to the top AI accelerator makers and sets the pricing power narrative for the whole memory market.	0x84CA…FeF8
12%	SNDK	Sandisk Corporation	Sandisk's NAND flash business benefits directly as AI data pipelines demand more storage bandwidth alongside compute.	0xB90A…6400
10%	WDC	Western Digital Corporation	Western Digital's storage and NAND exposure captures the same capacity crunch on the data lake side of AI infrastructure.	0xF525…7D3e
10%	AMKR	Amkor Technology, Inc.	Amkor does the advanced packaging that stacks HBM dies onto logic, a physical step that is now a gating factor for memory supply.	0xDd35…305B
8%	LRCX	Lam Research Corporation	Lam Research's etch and deposition tools are essential for building the 3D NAND and DRAM structures memory scarcity depends on.	0x57b0…29a4
7%	ALAB	Astera Labs, Inc.	Astera Labs makes the retimers and connectivity chips that move data between GPUs and memory, directly addressing the bandwidth bottleneck.	0x748c…A3Dc
5%	AMAT	Applied Materials	Applied Materials supplies the fab equipment memory makers need to expand HBM and DRAM capacity, a direct beneficiary of the bottleneck.	0x3604…16d0
3%	CRDO	Credo Technology Group Holding Ltd Ordinary Shares	Credo's high speed connectivity silicon reduces the data movement penalty that memory bandwidth constraints impose on AI clusters.	0x4D67…C41b

MUMicron Technology30%
Micron is the largest US HBM and DRAM producer and is sold out of leading edge memory capacity well into next year.
0xfF08…4afD
SKYHYSK Hynix15%
SK Hynix is the dominant HBM supplier to the top AI accelerator makers and sets the pricing power narrative for the whole memory market.
0x84CA…FeF8
SNDKSandisk Corporation12%
Sandisk's NAND flash business benefits directly as AI data pipelines demand more storage bandwidth alongside compute.
0xB90A…6400
WDCWestern Digital Corporation10%
Western Digital's storage and NAND exposure captures the same capacity crunch on the data lake side of AI infrastructure.
0xF525…7D3e
AMKRAmkor Technology, Inc.10%
Amkor does the advanced packaging that stacks HBM dies onto logic, a physical step that is now a gating factor for memory supply.
0xDd35…305B
LRCXLam Research Corporation8%
Lam Research's etch and deposition tools are essential for building the 3D NAND and DRAM structures memory scarcity depends on.
0x57b0…29a4
ALABAstera Labs, Inc.7%
Astera Labs makes the retimers and connectivity chips that move data between GPUs and memory, directly addressing the bandwidth bottleneck.
0x748c…A3Dc
AMATApplied Materials5%
Applied Materials supplies the fab equipment memory makers need to expand HBM and DRAM capacity, a direct beneficiary of the bottleneck.
0x3604…16d0
CRDOCredo Technology Group Holding Ltd Ordinary Shares3%
Credo's high speed connectivity silicon reduces the data movement penalty that memory bandwidth constraints impose on AI clusters.
0x4D67…C41b

What breaks this

Written by the same model that built the basket, and not edited to sound better.

01Extreme concentration in a handful of memory names means any pricing downturn or oversupply cycle hits the whole basket at once.
02Memory has historically been a brutal boom bust commodity market, and a capacity glut could crater margins even if AI demand stays strong.
03If AI progress shifts back toward compute bound architectures or new memory technologies emerge from unlisted players, this thesis loses its edge.
04SK Hynix and Sandisk exposure depends on ADR or foreign listing liquidity which can behave differently than the underlying business.
05Packaging and equipment names are cyclical capex plays that can lag or overshoot the actual memory shortage narrative.

Changelog

Every revision to this basket, who made it, and when. A theory is allowed to change its mind in public.

by the modelAug 14, 2026
Basket created from the thesis. 9 holdings.

How this basket was made

Source: Generated
Model: claude-sonnet-5
Chain: Robinhood Chain · 4663
Created: Aug 14, 2026

The model was given the full list of equities tokenized on Robinhood Chain and could only pick from it. Token addresses were attached afterwards from that same list, so a ticker the model invented would have been dropped rather than shown here. This is not investment advice and it has not been backtested.