Loading activity…
@semianalysis_
AI
The additions expand Google's eighth-generation TPU lineup to four systems, with AMD host processors supporting TPU 8tx.
Technology
Galaxy reportedly delivered its first 120MW phase in August 2026, within a year; it also hosts an 800V DC pilot.
AI
The reported setup used the default SGLang recipe, with roughly 60 tokens per second as the K3 comparison.
AI
Together says its compiler serves trillions of tokens daily and adapts to traffic shifts within hours of recompilation.
Technology
A research firm estimates 3.6GW under construction across multiple sites, not capacity already operating.
AI
An estimated 3.6 GW of capacity is under construction across existing and new sites.
Technology
SemiAnalysis puts the top model score at 12% and warns that safety classifiers can depress some results.
AI
The v1.0 benchmark spans 11 domains and 206 tasks, with a score that averages performance across domains.
AI
AMD's Anush Elangovan and his team reportedly got there through software optimizations including AITER kernel tuning and MonoKernel decode.
AI
SemiAnalysis says it can reach over 850 tokens per second per user, even at 2.8 trillion parameters.
AI
A post citing Anthropic's S-1 lists $974M in Q2 operating income after a $1.7B operating loss in Q1.
AI
Anthropic says its registration statement is not yet effective, so securities cannot be sold or offers to buy accepted.
World & Business
One post says roughly $72B of the reported loss came from non-cash charges, putting the underlying loss near $9B.
Technology
Kalshi says the filing is due at 9 a.m. EST; a separate post also expects it on October 26.
AI
SGLang cites results on B300 and GB300 hardware; vLLM's decode claim covers batch sizes 8 to 256.
AI
SemiAnalysis reports roughly 20% less monthly API-equivalent value but roughly 20% higher five-hour usage limits for some accounts.
AI
SemiAnalysis says Nvidia increased Rubin interactivity by close to 2x in less than eight days using low-latency specialized kernels.
AI
The contest page lays out GPU-kernel and SGLang serving work, with different tests for each.
AI
The SGLang project says early reinforcement-learning training on TPUs with Miles and TorchTitan produced promising results.
AI
In a post, kushal1t claims the framework boosts GPU utilization by packing independent workloads into one kernel.
