Dylan Patel

SemiAnalysis examines how Huawei's Ascend 950PR hardware and DeepSeek's kernel software (TileLang, DeepGEMM, FlashMLA, DeepEP) now interlock, arguing DeepSeek will increasingly run on a fully domestic Chinese stack rather than Nvidia compute.

Hands-on testing across GPU neoclouds for ClusterMAX 3.0 turned up five recurring security failure patterns in the providers that AI labs increasingly rent from, alongside an open CLI that audits your own cluster or standalone GPU box.

SemiAnalysis publishes the first third-party benchmarks of Google's TPUv7 Ironwood, finding up to 50% better performance-per-dollar than Nvidia B200/B300 on inference. It compares Google's internal TCO against what external customers actually pay as Anthropic scales toward over a million TPUs.

SemiAnalysis examines why robotics inverts the LLM design pattern: real-time control loops and per-unit cost constraints force robot inference onto the device rather than the datacenter.

An argument for why governments are funding sovereign AI: frontier model access sits with a handful of US labs and their regulator, releases have been delayed on safety and policy grounds, and open weight licenses now carry revenue thresholds. Korea's buildout is read as an Nvidia win.

SemiAnalysis publishes ClusterMAX 3.0, a hands-on tested ranking of 77 GPU cloud providers across compute, networking, storage, orchestration, and security, drawn from direct probing of provider infrastructure.

A SemiAnalysis report arguing Meta's datacenter and compute procurement is accelerating rather than slowing, with over 5GW contracted in just the first half of the year across cloud and colocation deals. It outlines four ways Meta could monetize its massive compute buildout beyond training its own frontier models: scaling ads recommendation systems 10x, striking a Bedrock-style deal to resell private Anthropic Claude access, pursuing short-term high-margin 'SpaceX-style' compute deals like Elon Musk's Anthropic and Google contracts, and building vertical AI-powered sales/marketing products atop its ad platform.

A SemiAnalysis technical deep-dive into Kimi K3's architecture, tracing Kimi Delta Attention (KDA) from linear attention through DeltaNet and Gated DeltaNet, and detailing Moonshot's FlashKDA kernel implementation with FLOP and memory-traffic complexity analysis. The piece also examines Kimi K3's hybrid KDA/MLA attention design, KV cache throughput and residency across memory tiers, and the challenges linear attention poses for prefix caching in inference engines like vLLM.

An analysis of whether open weights have caught the closed frontier on coding and agentic work, with figures on inference volumes, provider competition and what capable open models do to token pricing.

SemiAnalysis argues the industry is pivoting from taller 12-hi HBM stacks toward 4-hi and 8-hi configurations, pointing to Nvidia's Rubin Ultra capacity cut to 192GB, and explains why lower stacks offer better bandwidth-per-dollar for inference workloads.

A SemiAnalysis deep-dive argues Meta's Superintelligence Labs (MSL) is uniquely positioned among AI labs to become world-class in data, talent, and compute simultaneously, despite a rocky Llama 4/Muse Spark start. It details Meta's aggressive RL-environment data strategy (including employee screen-recording programs and a new 3,000-engineer applied AI org), massive multi-gigawatt 'titan' datacenter buildouts across five sites, and compute projections showing Meta potentially surpassing OpenAI and Anthropic in total AI compute by year-end.

SemiAnalysis publishes the first verified agentic-inference benchmarks for Nvidia's Rubin NVL72 platform, finding up to 7x better token throughput per megawatt than Blackwell on early pre-release software.

SemiAnalysis argues the narrative that state and local datacenter moratoriums are halting the US AI buildout is overstated, following its earlier debunking of water-use and cancelled-capacity myths.

SemiAnalysis revisits its AMD call, from giving it a 0% chance of closing the gap with NVIDIA to a materially better one, and assesses whether AMD can break the CUDA moat.

Reports say enterprise token consumption has hit a budgeting wall. SemiAnalysis talked to more than 50 customers and argues the widely reported clampdowns at Meta and Uber are overstated and rooted in poor incentives.

SemiAnalysis's Dylan Patel breaks down the semiconductor and power bottlenecks limiting AI compute, how hyperscaler capex converts to gigawatts online, and what lab fundraises cover in rental compute cost.

SemiAnalysis argues SpaceX's 6-10GW datacenter target for 2027 is achievable, walking the suitable sites and gas-generation equipment that make it possible, then prices scarce near-term capacity at up to $50B per GW per year against the $100B-plus per GW that labs can earn selling API inference.

A SemiAnalysis technical report summarizing ECTC 2026 disclosures on AI accelerator packaging, covering Intel's EMIB-T bridge technology (TSVs, MIM capacitors, sub-25μm bump pitch scaling), Marvell's custom HBM4E base-die architecture, Samsung's silicon interposer and hybrid copper bonding thermal work, and emerging microfluidic cooling and photonic interconnect approaches from TSMC, Microsoft, and Lightmatter. It argues that packaging (not transistor scaling) is now the primary bottleneck for AI accelerators, detailing specific bandwidth, power-delivery, and thermal-resistance tradeoffs across vendors.

SemiAnalysis breaks down how Anthropic and OpenAI subscription plans convert monthly payments into credits per model and token type. It estimates that subscriptions use over 40% of Anthropic inference compute and cut blended revenue per MW.

SemiAnalysis tracks 75GW of binding behind-the-meter power orders for AI datacenters, sharply up in Q2 2026, and lays out the execution risk between a signed power order and an operating facility.

SemiAnalysis argues robotics inverts typical LLM economics: real-time control loops and per-unit hardware costs force robot AI models to be designed around fixed, affordable on-device compute rather than scaled freely like cloud models.

SemiAnalysis rebuilt PJM's black-box Reserve Requirement Study and argues its methodology ignores winter efficiency gains and post-Elliott resilience, understating capacity by ~4GW. The knock-on: an estimated $12B in avoidable ratepayer cost and a capacity market structurally hostile to load growth.

SemiAnalysis on Vera Rubin NVL72, the second generation of NVIDIA's rack-scale Oberon architecture, where engineering samples running DeepSeek R1 show 5.4x performance per megawatt and 5x per dollar over GB200 NVL72.

SemiAnalysis examines why sparse attention does not shrink memory capacity needs, since top-k selection wants the whole context in HBM, and how SGLang's HiSparse offloads KV cache to host DRAM with LRU caching and layer-wise overlap.