AI Infrastructure

1 call tracked · 1 active

1
Total Calls
1
Active
—
Accuracy
Direction Mix
Bull 100%

All Calls

LONGActive

Parameter Scaling Is Dead. Multi-Dimensional Scaling Is the New Paradigm.

AI scaling has moved from raw parameter count to four memory-hungry dimensions — inference-time compute, MoE, data curation, architecture — making memory bandwidth the universal bottleneck.

Parameter scaling hit diminishing returns — GPT-4 to GPT-5 gap smaller than GPT-3 to GPT-4

Inference-time compute (o3) generates 10-100x more tokens per query — massive KV cache pressure

ai-infrastructurememory-over-computescaling
High1YAI Infrastructure — Scaling StrategyUpdated 4d ago13 Apr
LONGActive

Inference Costs Will Fall 90% by 2028 — The Memory Stack Is the Mechanism

Four independent memory-stack improvements — HBM4, KV-cache compression, optical interconnect, and architecture — could compound to a ~10x fall in per-token inference cost by 2028.

Four independent memory stack improvements compound: HBM4 (2x) × KV compression (4-6x) × optical (70% energy) × architecture (2-10x)

Even three of four layers delivering = 5-8x cost reduction by 2028

ai-infrastructurememory-over-computeinference
Med→ Med3YAI InfrastructureUpdated 4d ago13 Apr
LONGActive

BTC Miners Pivoting to AI Infrastructure: The Power Arbitrage

Bitcoin miners already hold the power, cooling, and space AI needs — mining revenue is expected to fall from 85% to under 20% of total as they convert to AI hosting.

BTC miners own power-ready facilities — the scarcest asset in AI infrastructure

Mining revenue dropping from 85% to under 20% of total by late 2026 — pivot is underway

ai-infrastructurememory-over-computeenergy
High▲ High1YAI Infrastructure — Data CentersUpdated 4d ago13 Apr
WATCHActive

Transformer Successors: State-Space Models and Linear Attention Are Worth Watching

Transformers' quadratic memory scaling is the KV-cache problem. State-space and linear-attention models scale linearly — hybrids blending both look like the likely path.

Transformer attention scales O(n²) in memory — the fundamental cause of KV cache and HBM bottlenecks

State-space models (Mamba, RWKV) offer O(n) linear memory scaling — potentially 10x more efficient

ai-infrastructurememory-over-computememory
Med→ Med3YAI Infrastructure — Model ArchitectureUpdated 4d ago13 Apr
LONGActive

KV Cache Compression Will Cut Inference Costs 50% Before HBM4 Ships at Scale

KV caches are AI inference's biggest bottleneck. Software compression (6x cited) is shipping faster than HBM4 hardware — and will drive the cost collapse of 2026-2027.

KV cache is THE bottleneck: 128K prompt on Llama 3.1-70B = 40GB HBM just for key-value storage

TurboQuant: 6x compression at inference time, no retraining required, minimal quality loss

ai-infrastructurememory-over-computememory
Med→ Med1YAI Infrastructure — Inference SoftwareUpdated 4d ago13 Apr
← View all market calls