Dylan Patel

A SemiAnalysis teardown of Huawei's Kirin 9030 (SMIC N+3) compares it against TSMC N6 (via a MediaTek Helio G99) and Intel's 18A, finding that SMIC matches TSMC N6-class logic density only through aggressive DUV multi-patterning and DTCO rather than EUV, at the cost of process maturity and cost efficiency. The report details die floorplans, CPU/GPU/NPU architecture changes versus the Kirin 9020, memory (Samsung and CXMT LPDDR5X) and iPoP packaging, showing Huawei's silicon closing some gaps under export controls while still trailing leading-edge Apple, Qualcomm and MediaTek chips in performance and efficiency.

SemiAnalysis founder Dylan Patel walks through lab compute, capex trajectories and revenue, arguing the world economy increasingly tracks lab economics and the compute market, with figures out to 2028.

Premium fast modes show users will pay for lower latency, so frontier labs are evaluating purpose-built inference hardware that trades batched throughput for interactivity, which is what real-time assistants and full-duplex voice need.

A SemiAnalysis technical article arguing that RL training system efficiency hinges on matching trainer (learning) and generator (inference/rollout) throughput, covering GRPO, asynchronous PipelineRL, policy staleness, and RL sandbox infrastructure challenges. It includes case studies training MoE models on agentic coding tasks using open-source RL frameworks, comparing costs against hosted solutions like Thinking Machines' Tinker.

Dylan Patel and Dwarkesh Patel on where lab economics are heading: spend rotating back toward training as recursive self-improvement nears, OpenAI and Anthropic outbidding everyone for usable FLOPs, and whether a decade of multi-trillion AI capex ends in a sovereign debt crunch.

An open-source SemiAnalysis initiative that benchmarks DeepSeek v4 Pro inference performance across hardware (GB300 NVL72, Huawei Ascend 950DT, MI355X, B200, H200) and engines (vLLM, SGLang, TensorRT-LLM, ATOM) from Day 0 through subsequent weeks, tracking how throughput and interactivity improve as vendors patch and optimize their stacks. The article documents specific bugs (e.g., a hardcoded hidden-size guard in TensorRT-LLM, broken AITER kernels on AMD ROCm) and shows AMD's MI355X achieving over 100x throughput improvement within 26 days via kernel-level fixes.

SemiAnalysis breaks down Cerebras CS-4, which doubles inference throughput on the same 5nm wafer by roughly doubling clock speed, plus the power, cooling and I/O changes that make disaggregated inference possible.

SemiAnalysis extended its building-level datacenter tracking model to China, finding over 24GW of built AI datacenter capacity, more than EMEA or the rest of Asia-Pacific, challenging assumptions that Chinese AI infrastructure is small or largely vacant.

SemiAnalysis explains Engram, a model architecture that extends token embeddings with learned multi-token lookups addressable by token ID, letting those rows prefetch from host DRAM instead of sitting in HBM, cutting HBM needs without hurting quality.

SemiAnalysis tests Trainium3 racks and finds AWS Neuron now supports native eager-mode PyTorch with torch.compile, making research training viable, while noting bugs and a large maturity gap versus CUDA. Anthropic already runs inference and RL rollouts on it.

SemiAnalysis claims over 24GW of datacenters are in advanced construction for 2027 delivery across 40 campuses, supporting a forecast of about 50GW added that year and pushing back on a 25GW skeptic estimate, while noting labor and supply constraints.