SemiAnalysis's STEEL teardown lab found that Intel's Panther Lake, the first commercial chip to use backside power delivery (PowerVia) and Intel's first gate-all-around RibbonFET transistors on the 18…
Mixture of Experts has changed the structure of AI inference serving by altering which tensors are active per token, what must remain close together, which transfers need strong local bandwidth, and h…
SemiAnalysis reported that DeepSeek's Engram architecture, which extends standard token embeddings with learned multi-token lookups, can lower required HBM capacity for models at the same quality by o…
SemiAnalysis estimates that only about 2.3 GW of planned US datacenter capacity is genuinely delayed by local moratoriums and New York's executive order, despite four states acting in under two months…
SemiAnalysis reported that Nvidia's Vera Rubin NVL72 platform delivered up to 7x better token throughput per megawatt than Blackwell on its AgentX agentic inference benchmark, even on early pre-releas…
Nvidia's next-generation Rubin Ultra accelerator will ship with 192GB of HBM per GPU, down from 288GB in standard Rubin and B300, according to SemiAnalysis, which first reported the change. SemiAnalys…
Physical AI is moving AI beyond screens into robots, but the field is split on where inference should run: on-device versus in datacenters. Figure runs its Helix model entirely onboard, while Physical…
SemiAnalysis published the first third-party inference results for Google's TPUv7 Ironwood, showing up to 50% better performance per dollar than NVIDIA's B200/B300 in apples-to-apples comparisons. The…
South Korea's government launched the 'Independent AI Foundation Model' project in June 2025, a tournament-style initiative to develop a domestic frontier AI model, with five consortiums selected in A…
SemiAnalysis's ClusterMAX 3.0 testing found that most neoclouds have serious security vulnerabilities, with five frightening patterns identified. The company urges neocloud operators and users to upda…
OpenAI's new inference chip, Jalapeño, outperforms Nvidia's Blackwell and other competitors in performance per watt across almost all scenarios, according to benchmarks run by SemiAnalysis with OpenAI…
SemiAnalysis announced AgentX 1.0, the world's first fully open source multi-turn agentic coding inference benchmark at 1 million context, released under Apache 2.0, after spending more than $3M build…
Open-source AI models are catching up to closed frontier models twice as fast with each generation, according to a SemiAnalysis analysis that found open models now match closed models on coding and ag…
Cerebras Systems unveiled its fourth-generation CS-4 rack, doubling the performance of the CS-3 by increasing clock speeds and power delivery on the same 5nm wafer-scale engine (WSE-3), enabling doubl…
TileRT's persistent engine on NVIDIA GPUs achieves up to 500 tokens/s/user on the InferenceX GLM5 FP8 744B benchmark on a single B200 decode server, approximately 3× faster than GB300 NVL72 running tr…
SpaceX aims to build and deliver 6-8GW of AI datacenter capacity in 2027, potentially exceeding 10GW, which could generate $300-500B in capex and $300B in annual recurring revenue for SpaceX, accordin…
Google announced on August 5th a complete overhaul of DeepMind leadership, with co-founder Demis Hassabis stepping back from day-to-day operations and former Google Chief Scientist Jeff Dean leaving t…
Moonshot AI released Kimi K3, an open frontier model that swept leaderboards, featuring a hybrid attention mechanism with Kimi Delta Attention (KDA), a linear attention layer derived from DeltaNet and…
Modular construction has become the default playbook for building datacenters, with SemiAnalysis tracking over 61GW of modular capacity across 1,000+ sites and estimating modular penetration will reac…
AMD has a great chance of breaking Nvidia's CUDA software moat in AI accelerators, according to SemiAnalysis, which upgraded its outlook from zero to non-zero to now a strong probability of success, c…