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
$AMZN · Amazon Trainium 3 with 144GB HBM3E sees massive deployment; Anthropic and OpenAI adoption validates large-scale ASIC viability
now playing · $AMZN
$QCOM···neutral· mediumdylan patel
Qualcomm AI 200 targets inference with LPDDR5X but memory costs rising; AI 250 compute-near-memory could be worthwhile
AI 200 uses 768GB LPDDR5X on TSMC N3E for inference, but LPDDR5X prices are skyrocketing; the successor AI 250 promises 10x effective memory bandwidth with compute-near-memory arc…
$AMD···bullish· highdylan patel
AMD MI455X with 432GB HBM4 and Helios Rack could make 2026 a turning point against Nvidia if delivered on time
MI455X uses CDNA 5 architecture with 320B transistors across 12 chiplets on 2nm/3nm via 3.5D packaging, delivering 432GB HBM4 at ~20 TB/s bandwidth — a memory advantage over Nvidi…
$GOOGL···bullish· mediumdylan patel
Google Ironwood TPUv7 with optical circuit switches enables 9,216-chip super pods; external customer adoption key for inference TCO leadership
TPUv7 on TSMC N3E with >100B transistors and 192GB HBM3E uses optical circuit switches to connect up to 9,216 TPUs in super pods, offering a differentiated TCO for inference — if…
$CBRS···neutral· lowdylan patel
Cerebras WSE3's 44GB SRAM and 21 PB/s bandwidth unique but memory capacity falling behind 2026 needs; WSE4 announcement anticipated
WSE3's wafer-scale design with 4T transistors on N4P delivers extraordinary SRAM bandwidth, but 44GB capacity is insufficient for 2026 workloads; the concept remains competitive f…
$GROQbullish· mediumdylan patel
Nvidia acquired Groq for ~$20B; deterministic LPU architecture eliminates latency but requires massive scale for model size; 2nd gen on Samsung 4nm awaited
Groq's LPU uses deterministic execution with 230MB on-chip SRAM on 14nm to eliminate GPU latency, but limited memory requires connecting many chips; Nvidia's ~$20B acquisition sig…
$NVDA···bullish· highdylan patel
Nvidia Vera Rubin VR200 with 35 petaflops FP4 and NVL72 rack will likely top 2026 charts; Rubin Ultra with 1TB HBM4 to answer AMD memory advantage
VR200 on TSMC N3B with 288GB HBM4 at 22 TB/s delivers 35 petaflops FP4 per package; 72-chip NVL72 rack with NVLink scale-up network makes it the most anticipated 2026 release, wit…
$META···bullish· mediumdylan patel
Meta MTIA v3 on TSMC N3P with HBM targets recommendation model inference; custom silicon for margins, external GPUs for training
MTIA v3 moves to HBM on TSMC N3P (>100B transistors) for internal inference workloads including Facebook/Instagram recommendation models, delivering margin advantage; Meta uses ex…
$AMZN···bullish· highdylan patel
Amazon Trainium 3 with 144GB HBM3E sees massive deployment; Anthropic and OpenAI adoption validates large-scale ASIC viability
Hundreds of thousands of Trainium 2 already deployed in AWS data centers; Trainium 3 on TSMC N3P with 125B transistors and 144GB HBM3E unifies training/inference, with Anthropic's…
$MSFT···bullish· mediumdylan patel
Microsoft Maia 200 with 216GB HBM3E targets FP8/FP4 inference for in-house and OpenAI models; custom silicon gaining inference share from Nvidia
Maia 200 at 825mm² with 140B transistors on TSMC N3P delivers 5/10 petaflops FP8/FP4 with 216GB HBM3E; Microsoft will use it for internal models and future ChatGPT models, exempli…
$INTC···bearish· mediumdylan patel
Intel Jaguar Shores GPU on 18A with 288GB HBM4 competitive on paper but 2027 timeline; must prove manufacturability and software after repeated failures
Jaguar Shores specs (18A, 175B transistors, 288GB HBM4) look competitive for 2027, but Intel's history of failed AI GPU attempts means it must demonstrate both manufacturing execu…