Semianalysis

SemiAnalysis is a semiconductor and AI research company whose work follows the chain from chip fabrication to the cost of running models. It publishes research, sells institutional data models and provides consulting. Its coverage spans fabrication, packaging, chip design, networking, datacenters, software, training and inference, connecting accelerator supply with the performance and economics of deployed AI systems.[1][4]

As of September 2026, Dylan Patel is its founder, CEO and chief analyst, and Doug O’Laughlin is president. The business combines a retail newsletter with institutional research and consulting; its institutional catalog includes infrastructure models, Core Research and the monthly semiconductor tracker ChipBook.[2][3][1][4]

SemiAnalysis’s Accelerator & HBM Model connects accelerator shipments and installed capacity to wafer starts, packaging volumes and yields, and high-bandwidth memory requirements. It tracks these dependencies by chip, supplier and customer, making fabrication, packaging and memory supply part of the forecast of accelerator availability.[4]

Its AI Cloud TCO model estimates GPU-server cost per hour, training cost per unit of computation and inference cost per million tokens. TCO means total cost of ownership: the model includes throughput, parallelism, depreciation, financing, power and colocation. Comparing GPU services therefore involves both their operating expense and the useful output each configuration can serve. The Tokenomics Model extends the calculation from hardware and workload throughput to usage, revenue and forecasts of future accelerator demand.[4]

The company describes a research method combining supply-chain tracking, fieldwork and physical records. Its Datacenter Industry Model draws on permits, property records, freedom-of-information requests, power data and satellite imagery assisted by computer vision. Its foundry model combines equipment-based estimates with filings, customs data and primary research. SemiAnalysis says fieldwork sharpens its models, while the models direct further fieldwork.[1][4]

The present mix of publishing, models and benchmarking developed from a smaller operation. Patel’s biography traces SemiAnalysis’s evolution since 2020 from a solo venture into a team. Its early publishing and podcast decisions provide context for the broader research business.[2][5]ⁿ¹

On January 3, 2022, Patel announced Transistor Radio with O’Laughlin, then associated with Fabricated Knowledge. Patel explained that long articles left little room for news tidbits, informal market projections and weekly observations. Three days later, he said they had not liked the rapid-fire webinar enough to declare it their first episode and changed the planned format: “We will do more in depth single topic episodes instead.”[5][6]

On May 4, 2023, SemiAnalysis published the Google “We Have No Moat” memo. The document argued that open development, smaller models and rapid iteration threatened the advantages of closed large-model development. In their introduction, Patel and Afzal Ahmad described it as one Google employee’s opinion, said they had verified its authenticity, and explicitly rejected its argument: “We do not agree with what is written below”. They published the employee’s memo rather than presenting it as their own analysis or Google’s corporate position.[7]

On August 28, 2023, Patel and Daniel Nishball’s Gemini article framed unequal access to accelerators through GPU-rich and GPU-poor organizations. Their analysis considered both total compute capacity and capacity available per researcher, arguing that this access constrained experimentation and commercial ambitions. The headline’s fivefold comparison concerned projected total pretraining FLOPS, rather than a measured fivefold advantage in model quality.[8]

A November 2023 Latent Space episode with Patel discussed the framework and its implications for training and inference. Patel distinguished model FLOPS utilization in training from memory-bandwidth utilization in low-batch inference: serving a model quickly can depend on how fast its parameters can be read, even when much of an accelerator’s arithmetic capacity remains unused.[9]

SemiAnalysis launched the first version of its open inference benchmark in October 2025 under the name InferenceMAX; the project is now InferenceX. It addresses the rapid evolution of serving software: changes to kernels, scheduling and distributed inference can make a fixed benchmark obsolete before the hardware changes. The Apache 2.0 project publishes a free dashboard of performance results.[10][13]

The dashboard compares throughput against latency while sweeping tensor-parallel sizes—how model computation is divided across accelerators—and concurrent request counts. This exposes the tradeoff between serving more work and responding quickly. Benchmarks originally ran nightly; the current methodology says the growing number of combinations led the project to rerun tests when configurations change, such as after a software release, driver update or model addition.[11][12]

In August 2026, the project added AgentX, a long-context, multi-turn coding-agent workload. Its methodology explains: “The prompt text is synthesized, but the workload shape comes from recorded sessions”. The replay preserves token lengths, shared prefixes, per-turn timing and, in supporting corpora, subagent fan-out. Repeated context and overlapping conversations exercise serving, routing and cache behavior differently from isolated prompts; the workload measures infrastructure behavior rather than whether a model writes good code.[10][12]

AgentX is implemented with AIPerf. SemiAnalysis credits NVIDIA’s AIPerf team for the scenario’s implementation and documentation, and Callan Fox and Weka for the underlying agentic-coding trace work.[12]

InferenceX links chart results to public GitHub Actions runs, with recipes, logs and artifacts documenting the configurations tested. Its acknowledgements credit AMD and NVIDIA for hardware access and engineering support, the SGLang, vLLM and TensorRT-LLM maintainers for serving software, and cloud providers including Crusoe, CoreWeave, Nebius, TensorWave, Oracle and TogetherAI for compute resources.[12][10]

ClusterMAX evaluates managed GPU cluster services for teams running AI workloads. Its ten criteria cover performance, reliability, security and operational usability, including managed Slurm and Kubernetes. The rankings focus on the managed service layer; they do not rate a provider’s data center construction, bare-metal clusters or inference endpoints. The team described the project’s evolution in a Latent Space discussion of ClusterMAX 3.0.[14]

The researchers said they were tracking 323 providers by the third edition and had moved from running benchmarks largely by hand to a more automated testing suite, including coding agents that run tests and help fix issues. The rankings provide a broad first pass; SemiAnalysis also offers more tailored consulting for a customer’s specific requirements.[14]

Security checks include software versions, known vulnerabilities and network configuration. SemiAnalysis released the CMAX command-line tool to help users check versions and locate upgrades. The team said some providers had exposed other customers’ jobs or data during its assessments; those findings are the researchers’ reported observations.[14]