TickerTain
$SPCX····$NVDA····$MSFT····$GOOGL····$ANTHROPIC·$TSLA····$CRWV····$NBIS····$HIMS····$TSM····$ASML····$FIGURE·$LYFT····$UBER····$SNOW····$META····$ELEVENLABS·$CURSOR·$COGNITION·$OPENAI·$AAPL····$RKLB····$XAI·$AMZN····$AMD····$000660.KS····$INTC····
← newsroom
AI Infrastructure · Token demand growing exponentially with 72%+ margins at frontier labs
now playing · AI Infrastructure
AI Infrastructuretailwindscore 9/10dylan patel
Token demand growing exponentially with 72%+ margins at frontier labs
Frontier model labs (Anthropic, OpenAI) are seeing revenue grow 4-5x while compute scales slower, driving gross margins from 30% to 72%+; enterprise customers are rate-limited and willing t…
AI Infrastructuretailwindscore 9/10dylan patel
Memory, equipment, and materials bottlenecks to persist until 2028+
DRAM capacity growth capped at 20-30%/year with true incremental supply not arriving until 2028, forcing prices to double/triple; TSMC's $100B capex trajectory creates equipment tail whip (…
AI Infrastructuremixedscore 8/10dylan patel
Cost per capability dropping 100-1000x but frontier demand absorbs all gains
Research compute efficiency improves dramatically (GPT-4 class now 1/600th cost), but frontier models (Mythos, Opus 4.7) consume all available compute; release cadence compressed from 6 mon…
Robotics & Physical AItailwindscore 8/10dylan patel
Few-shot robotics breakthroughs expected in 6-18 months, unlocking physical deflation
As software implementation costs collapse, robotics will see pre-trained models enabling few-shot learning (show a robot a few examples, it learns the task); this shifts AI from software-on…
AI Economics & Business Modelsriskscore 8/10dylan patel
Best models restricted to few enterprises — risk of permanent underclass
Frontier models (Mythos) released only to select enterprises (top banks for cyber); those with capital and early access compound advantages; non-adopters face 'permanent underclass' — three…
AI Infrastructuretailwindscore 7/10dylan patel
H100 useful life extending to 7-8 years, expanding cloud margins
Contrary to 5-year depreciation assumptions, Hopper and Ampere clusters are re-signing for 3-4 additional years; this extends revenue visibility for cloud providers and increases ROI on GPU…
Semiconductorstailwindscore 7/10dylan patel
CPUs and FPGAs emerge as parallel compute bottlenecks to GPUs
Reinforcement learning environments and AI deployment stacks run on CPUs (not GPUs), creating structural CPU shortage; 120 FPGAs per next-gen AI rack for control plane functions — diversifi…
Crypto & Digital Assetsriskscore 7/10dylan patel
Real economic value of tokens invisible to GDP metrics — 'phantom GDP' problem
Tokens create massive value (better decisions, faster innovation, deflationary services) but GDP shrinks as costs collapse; no statistical framework captures this knock-on value — the 'phan…