Bargo
AI Infrastructure · Memory

The memory race: why Gavin Baker says HBM is the real AI bottleneck

The AI investor discourse has been pricing NVIDIA GPU supply as the constraint for two years. This week, one of the most respected long-only tech investors said out loud what SemiAnalysis and SK Hynix have been arguing for months: it is not GPUs. It is memory. And that reframes the winners.

Bargo · 2026-07-14

On July 13, 2026, Gavin Baker of Atreides Management said memory bandwidth, not GPU count and not power, is the single hardest bottleneck in the AI infrastructure cycle. The same day, SanDisk (SNDK) crashed 14.7% while Micron (MU) held at $920, and SK Hynix marketing dominated authority X feeds. This piece unpacks what HBM is, who wins from it, and why the NAND names got left behind. Source: BargoAI research.

SK Hynix HBM share MU HBM revenue SNDK move (Jul 13)
~50% $8B+ -14.7%
of HBM3E supply, 2026 FY2026 run rate NAND-vs-HBM divergence

What did Gavin Baker actually say about the memory race?

Gavin Baker, managing partner at Atreides Management and a 25-year Nvidia holder who has been on the AI-infrastructure trade since 2019, made the case on the All-In Podcast (episode E278) on July 13. His exact framing:

The AI race is a memory race. Most investors still don't see it. Memory bandwidth is the single hardest AI bottleneck, not GPUs, not energy. GPUs you can add. Energy you can route. HBM packaging remains supply-constrained.

Gavin Baker, Atreides Management, on The All-In Podcast E278, July 13, 2026

Two things make this quote matter more than a typical thesis update:

Within 24 hours, memory names in Bargo's tracked X feeds jumped in mention count. SK Hynix (@SKhynix) alone recorded 27 authority mentions this week, the highest single-company count in the semi complex.

Why is memory bandwidth the real bottleneck, and not GPU compute?

A GPU runs AI models by pulling weights and activations from memory, doing math on them, and writing results back. If the math is fast but the memory is slow, the GPU sits idle waiting for data. This is called being memory-bandwidth-bound. Most modern AI workloads, especially inference on large language models, are memory-bound, not compute-bound.

The specific memory type NVIDIA and AMD use inside their AI GPUs is called High Bandwidth Memory (HBM). HBM is not standard DRAM soldered on a motherboard. It is a stack of memory chips manufactured with special interconnects and mounted directly on the same silicon substrate as the GPU (using a TSMC process called CoWoS, "Chip on Wafer on Substrate"). The physical proximity is what makes it "high bandwidth."

Two things constrain HBM supply:

Baker's framing captures this cleanly: you can add another GPU factory. You can route more power to a datacenter. You cannot easily add HBM capacity because it is bottlenecked at TSMC's packaging fab, not at the memory fabs themselves.

Which companies actually make HBM, and who benefits?

The HBM market is a three-vendor oligopoly. Two are Korean, one is American.

Vendor Est. HBM share Status Ticker
SK Hynix ~50% Sole HBM3E supplier at NVIDIA for 18 months; HBM4 sampling KRX: 000660 (foreign)
Samsung Electronics ~35% HBM4 mass production began this week; catching up on qualification KRX: 005930 (foreign)
Micron Technology ~15% Third supplier; ramping HBM3E, HBM4 in 2027 NASDAQ: MU

For US-listed investors, Micron (MU) is the direct play. The company reported $41.5B in Q3 FY2026 revenue with 84.9% gross margins, per its June 24, 2026 earnings release. Management said memory tightness persists "beyond calendar 2027" and disclosed 16 multi-year Strategic Customer Agreements covering roughly half of company revenue. This is not a spot-price cyclical anymore. It is a contracted franchise.

Downstream beneficiaries are the enablers of HBM manufacturing:

Why did SanDisk crash 14.7% on the same day memory dominated the story?

SanDisk (SNDK) dropped 14.7% to $1,634 on July 13, the sharpest single-name move in the memory complex that day. The tape reason was broad-market semi selloff on Diffusion Rule uncertainty. The underlying reason is more specific: SNDK sells NAND flash, not HBM.

NAND is the storage technology inside SSDs and enterprise storage arrays. It is high-volume, low-margin, and highly cyclical. It benefits from AI datacenter buildouts (every server needs storage) but not from the specific AI-training/inference bottleneck Baker highlighted. When the market rotates capital toward memory names, it is HBM winners it wants to own, not NAND commodity suppliers.

Samsung and SK Hynix have both been publicly deprioritizing NAND capacity to chase HBM production. That is bearish for NAND spot pricing over the medium term (less supply reduction, more demand rotation to HBM). Micron shares the same dynamic, but HBM is the larger and growing part of MU's revenue mix. SNDK does not have that HBM offset. It is pure-play NAND.

The read. The July 13 tape action put a fine point on Baker's thesis. Same industry (memory), same catalyst day (Diffusion Rule uncertainty + AI trade rotation), and MU held while SNDK crashed. The market is now discriminating within memory. HBM winners get the capital. NAND names get sold.

What this means for your portfolio

If you own NVIDIA (NVDA) or the AI hyperscaler basket, Baker's memory thesis matters to you in specific ways:

This is not investment advice. All live financials, options positioning, and signals are on bargo.ai.

Sources

Related reading

Reviewed by the Bargo editorial desk. Gavin Baker quote per The All-In Podcast E278, July 13, 2026. HBM market share estimates per SemiAnalysis and industry consensus. Micron financials per Q3 FY2026 earnings release. Market signals from live BargoAI feeds. This is research and educational content, not investment advice.

Related research

HBM
HBM week: SK Hynix declares victory as Samsung enters the game

Two signals landed on the memory market inside the same week. SK Hynix ran a marketing barrage claiming HBM dominance. Samsung Electronics began HBM4 mass production, entering the market as a real third supplier. Both stories are true. The pricing implication for NVIDIA, AMD, and Micron is not what the headlines suggest.
Memory
Micron Q3 FY2026 earnings: 84.9% margins and $22B in contracts

A plain-English decode of Micron's Q3 FY2026 print, the $22 billion in customer contracts nobody flagged first, and what it means for NVIDIA, ASML, and the equipment trio.
semiconductors
Korea's Semiconductor Exports Just Hit an All Time High. The Stocks Are Tanking.

South Korea just posted $44.8 billion in June semiconductor exports. Micron reported the best quarter in its history. And yet the entire memory complex is selling off. Here's what's really going on.