NewMoneyMoves
PerspectivesResearchTerminalsWeekend RWABoard
VIEW
InsightEditorialAdd a viewContextThesisReadingOutlook
››Thematic Deep Dive
EDITORIAL VIEW

The Structure of the Inference Bottleneck — Why Now Is the Turning Point

Memory Wall x Inference Cost: Why the HBM Supercycle Is a Structural Shift, Not a Cyclical Peak

HHaelangdal·Founder AnalystMarch 11, 202618 min readThematic Deep Dive
NewMoneyMoves

Global Investment through Themes

Content

ThemesIdeasReportsEarnings

Explore

GuideETFsPerspectivesTelegramNVIDIA PortfolioWeekly ReviewCredit & Leverage

Community

Daily News

Legal

Privacy PolicyTerms of ServiceDisclaimer

© 2026 NewMoneyMoves. All rights reserved.

Bottom Line

The real bottleneck in AI inference is memory bandwidth — not network latency — and HBM demand has entered a structural, not cyclical, growth trajectory driven by the inference demand explosion.

Reader's Brief — 30-second TL;DR

Advanced
Why Now

Oracle's earnings call confirmed memory bandwidth as the AI inference constraint, while NVIDIA's $20B Groq acquisition strategically reinforced this thesis.

Winners ?? Losers

Beneficiaries — HBM suppliers (SK Hynix, Samsung, Micron), HBM-SRAM tiered architecture designers, inference-specialized chip developers. Headwinds — investors expecting KV cache compression technologies (like TurboQuant) to substitute for memory demand, legacy CPU-based inference infrastructure companies.

Watch For

Quarterly capacity expansion announcements from major LLM service providers (OpenAI, Anthropic, Google) — leading indicator for memory bandwidth demand trajectory.

Reading depth
  1. 01Ch 1. The Physics of the Bottleneck: Why Memory Is the LimitInference generates tokens sequentially, loading all weights each time, so it is memory-bandwidth-bound. A GPT-4-class 3.6TB does not fit one chip, making memory the limit.Jump to section
  2. 02Ch 2. KV Cache — The Hidden Memory Monster Dominating InferenceThe KV cache explodes with context, can exceed the weights, and caps concurrent requests, a direct ceiling on monetization. NVIDIA's ICMS improves it 5x.Jump to section
  3. 03Ch 3. HBM Supply Chain — The Physical Reality of the BottleneckHBM is effectively sold out for 2025-2026, with HBM4 at a 40-50% premium over HBM3E. The TSV density penalty blocks quick expansion, cutting general DRAM in a zero-sum trade.Jump to section
  4. 04Ch 4. SRAM vs HBM — The Essence of the Architecture WarGroq's SRAM offers 80 TB/s on-chip bandwidth and ultra-low latency but small capacity, hurting large-model cost-efficiency. NVIDIA's $20B buy signals the SRAM-HBM war.Jump to section
  5. 05Ch 5. Next-Gen Memory Hierarchy Map — The Rise of HBF and PIMThe next-gen memory hierarchy reshapes. SK Hynix's HBF at 512GB/stack offloads static weights to Flash, and PIM cuts energy ~70% by computing inside memory.Jump to section
  6. 06Ch 6. Investment Implications — Who Profits from This BottleneckThe top beneficiaries are direct HBM suppliers. SK Hynix maximizes pricing power with share and sellout, and Micron overtook Samsung for a solid No. 2.Jump to section

Ch 1. The Physics of the Bottleneck: Why Memory Is the Limit

We need to start by unpacking the physical basis behind the statement by Oracle co-CEO Clay Magouyrk during the Q3 FY2026 earnings call (2026.3.10) that'the real bottleneck is hardware architecture.'

The computational structure of LLM inference is fundamentally different from training. Training processes thousands of batches in parallel against the same weights, making GPU's massively parallel cores optimal. In contrast,inference generates tokens sequentially, one at a time. With each token generated, the model's entire weights must be loaded from memory.

Loading a GPT-4 class model (~1.8 trillion parameters estimated) at BF16 precision requires approximately3.6TB of memory. The current NVIDIA H200's HBM3E capacity is 141GB. Therefore, single-chip deployment is impossible, and distributed inference across dozens of GPUs connected via NVLink is required.

Looking at the key metric, Arithmetic Intensity (FLOP per byte):

  • This value is very low in LLM inference
  • The model reads weights from memory, performs very few computations, and discards them
  • Such workloads are calledMemory-Bound, where performance is determined bymemory bandwidth (GB/s), not FLOPS

A co-authored paper by Google DeepMind's David Patterson (Turing Award laureate) and Xiaoyu Ma (arXiv 2601.05047, forthcoming in IEEE Computer 2026) quantitatively proves this:NVIDIA GPU 64-bit FLOPS grew 80x from 2012-2022, but memory bandwidth grew only 17x over the same period. This gap is the numerical definition of the ' Wall.'

HBM Specs by Generation: -H100: HBM3, 80GB, 3.35 TB/s -H200: HBM3E, 141GB, 4.8 TB/s -Rubin R100: HBM4, 288GB,22 TB/s

The Rubin GPU is the first to incorporate HBM4, achieving 22 TB/s bandwidth per GPU — approximately2.8x improvement over Blackwell. This is not a simple performance upgrade but an attempt to break through the memory itself.

However, there's a paradox: every time HBM capacity and bandwidth increase, model developers immediately fill that space with larger parameters, longer contexts, and bigger caches. MoE models like DeepSeek-V3 with 256 experts cause memory footprint to explode.When better HBM is supplied, models grow to match, so memory always remains the bottleneck for the next generation.

CharacteristicTraining (Training) (Inference)
computation typethousands of batches processed in parallel on same weightsTokens generated sequentially one at a time
Memory usageweights + gradients + optimizer Statusweights + KV cache (context proportional growth)
bottleneckGPU parallel cores (Compute-Bound)Memory Bandwidth (Memory-Bound)
Key MetricFLOPS (computation Throughput)GB/s (Memory Bandwidth)
arithmetic intensityHigh (compute-intensive)very Low (read-intensive)
Characteristic
computation type
Training (Training)
thousands of batches processed in parallel on same weights
Inference (Inference)
Tokens generated sequentially one at a time
Characteristic
Memory usage
Training (Training)
weights + gradients + optimizer Status
Inference (Inference)
weights + KV cache (context proportional growth)
Characteristic
bottleneck
Training (Training)
GPU parallel cores (Compute-Bound)
Inference (Inference)
Memory Bandwidth (Memory-Bound)
Characteristic
Key Metric
Training (Training)
FLOPS (computation Throughput)
Inference (Inference)
GB/s (Memory Bandwidth)
Characteristic
arithmetic intensity
Training (Training)
High (compute-intensive)
Inference (Inference)
very Low (read-intensive)
Takeaway

Inference generates tokens sequentially, loading all weights each time, so it is memory-bandwidth-bound. A GPT-4-class 3.6TB does not fit one chip, making memory the limit.

Full access requires 🥉 Bronze tier

Sign in with Google — your tier will be checked automatically and access granted if eligible.

Sign in with Google

This site runs on ads — the tier system rewards community contributions.

Comments

This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.

All Reports
Home
Perspectives
Company
Macro
Theme