SemiAnalysis

22 articles in the TruthFoundry index. Each links out to the original.

• Anthropic AI Subscriptions Deliver Over 5x More Value Than OpenAI Plans · SemiAnalysis analysis finds Anthropic subscriptions offer higher value than OpenAI equivalents.
• GLM 5.3 Sparse Attention Memory Impacts And Inference Benchmarks · Technical analysis evaluates sparse attention memory behavior and GLM5.3 inference serving performance.
• Intel Panther Lake Debuts First Commercial Backside Power Delivery and GAA Transistors · Intel's Panther Lake chip introduces the first commercial backside power delivery network and gate-all-around transistors.
• China's AI Datacenter Capacity Surpasses Europe Amid Hyperscale Boom · China now leads global datacenter capacity with over 24GW, driven by massive hyperscale investment.
• ClusterMAX 3.0 Ranks Neoclouds: Nebius Platinum, CoreWeave Bar Setter · SemiAnalysis releases ClusterMAX 3.0, ranking 77 neoclouds with Nebius in Platinum and CoreWeave leading.
• Mixture of Experts Transforms AI Inference Architecture and Economics · Mixture of Experts models have fundamentally altered AI inference system design.
• Engram Architecture Optimizes DRAM/SSD Offloading for AI Models · Engram model architecture uses learned multi-token lookups to reduce HBM demand and enable efficient offloading to DRAM and SSDs.
• Analysis Debunks Claim That US Datacenter Moratoriums Are Halting Buildout · Semianalysis maps 300+ moratoriums to find only 2.3 GW of genuine datacenter capacity delays.
• NVIDIA Rubin NVL72 delivers 67x better performance per dollar on agentic inference · Benchmark testing shows NVIDIA Rubin NVL72 far outperforms prior AI accelerators.
• Robotics Inference Shifts to Datacenters Due to Latency and Cost Constraints · Generalist robots increasingly rely on off-robot datacenter compute to overcome hardware limitations.
• 4-hi HBM Emerges As Optimal AI Memory Amid Supply Shortages · SemiAnalysis explains why shorter 4-hi HBM stacks deliver better cost efficiency for AI inference.
• Nvidia Expands $530B Off-Balance Sheet Guarantees to Back AI Buildout · Nvidia disclosed $530B in off-balance sheet guarantees to support AI infrastructure and non-investment-grade operators.
• AI Hyperscalers Adopt Behind-The-Meter Power Despite Execution Risks · Major AI labs and hyperscalers are rapidly deploying behind-the-meter power solutions to overcome grid constraints.
• TPUv7 Ironwood Delivers 50% Better Performance Per Dollar Than NVIDIA GPUs · Third-party benchmarks show Google's TPUv7 Ironwood outperforms NVIDIA B200/B300 on cost-efficiency for AI inference.
• South Korea Launches Sovereign AI Project, Nvidia Benefits Over Hynix · South Korea runs a national AI model tournament with major industry implications.
• Neocloud Security Testing Reveals Flawed AI Infrastructure · ClusterMAX 3.0 testing exposes significant security gaps in major AI neocloud providers.
• https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia
• InferenceX Releases AgentX 1.0 Benchmark for Agentic Coding · InferenceX launches AgentX 1.0, the first open-source benchmark for multi-turn agentic coding inference.
• Open Source AI Models Rapidly Closing Gap With Closed Frontier · Analysis shows open-source AI models are catching up to closed models in half the time of previous eras.
• Cerebras Unveils CS-4: Double Performance, Modular Rack Design · Cerebras launches CS-4, doubling token throughput via higher clock speeds and modular rack architecture.
• SemiAnalysis says PJM's modeling errors cost ratepayers $12B; emergency auction may repeat the waste · SemiAnalysis estimates PJM's modeling errors wasted $12B of ratepayer funds and its emergency auction repeats the risk.
• TileRT Boosts NVIDIA GPU Inference Speed for Interactive AI · TileRT software achieves up to 3x faster token generation on NVIDIA B200 GPUs for ultra-low latency AI workloads.