Dylan Patel

Dylan Patel is the founder, CEO and chief analyst of SemiAnalysis. His coauthored work connects semiconductor manufacturing, hardware and software to practical AI questions: what limits a model’s speed, why advertised accelerator performance differs from usable performance, and how much serving an application costs. His contributions include analyses of inference economics, comparisons of AMD and Nvidia systems, and SemiAnalysis’s continuous inference benchmarking project.[1][4][6][7][8]

Patel and Gerald Wong’s July 10, 2023 GPT-4 analysis argued that recurring inference costs were a more consequential scaling constraint than the expense of a training run: “The much more important issue with scaling AI, the real AI brick wall, is inference.” They connected sparse activation—using only some parameters for each token—to reduced active computation, while explaining how memory bandwidth and hardware utilization constrain economical serving. Their specific GPT-4 implementation claims were SemiAnalysis’s reporting; OpenAI’s technical report withheld architecture and model-size details.[4][15]

With Daniel Nishball, Patel framed unequal access to compute as a divide between the GPU-rich and GPU-poor in the August 28, 2023 Gemini analysis. They criticized leaderboard-oriented work and urged open efforts toward better evaluation, cleaner data, sparse architectures, speculative decoding and efficient serving of multiple fine-tuned models. The headline’s fivefold comparison was a forecast of Google’s total pretraining computation, rather than a measured fivefold improvement in model quality.[5]

In his November 2023 Latent Space interview, Patel recalled composing the Gemini piece in four hours after accumulated research into TPU orders, HBM capacity and the supply chain, then revising it after feedback. He credited specialist teammates with the deeper manufacturing work behind the publication, including Myron’s work on an extended manufacturing-process explanation prompted by Kokusai Electric’s public offering.[2][5]

Patel, Nishball and Reyk Knuhtsen’s December 22, 2024 training comparison made software availability part of the result. The team reported five months of testing MI300X, H100 and H200 systems and delayed publication to investigate fixes with AMD. It shared code and intermediate results with both vendors, separated public stable releases from nightly, custom and experimental builds, and disclosed corrections followed by retesting. The authors reported that MI300X’s theoretical advantages were not realized on public stable training software; custom development builds changed the comparison.[6]ⁿ¹

In January 2025’s DeepSeek Debates, Patel and coauthors AJ, Doug and Knuhtsen challenged the treatment of an approximately $6 million pretraining figure as the entire cost of developing DeepSeek V3. Their distinction was between GPU costs for one run and the broader work of experimentation, architecture development, data collection and cleaning, staff and infrastructure: “The pre-training cost is a very narrow portion of the total cost.”[23]

The May 23, 2025 inference report, by Kimbo Chen, Patel, Nishball and Ivan Chiam, compared throughput at a given per-user latency. Increasing concurrent requests can improve aggregate token production while making individual responses slower. The team tested specified Llama and DeepSeek models, precisions, serving frameworks and tensor-parallel configurations within a single node. It also separated ownership economics from short-term rental economics: under its historical prices, different workloads favored different vendors for owners, while Nvidia’s more competitive rental market favored renters.[7]ⁿ¹

Repeated reruns and rapidly changing serving frameworks made fixed-date comparisons difficult to keep useful. Patel coauthored the October 9, 2025 launch of InferenceMAX with Chen, Nishball, Cam Quilici and Cheang Kang Wen. The project introduced automated nightly benchmarks and a free dashboard tracing throughput against interactivity, cost and provisioned power. Its launch power comparisons used modeled all-in utility capacity, including facility overhead, rather than measured accumulated energy.[7][8]

The project later became InferenceX, with an Apache-2.0-licensed repository recording v2 in February 2026 and AgentX in August 2026. Its documentation describes a shift from nightly scheduling to rerunning configurations when software, drivers or models change. AMD and Nvidia supplied hardware access and engineering assistance; inference-framework maintainers and compute providers including Crusoe, CoreWeave, Nebius, TensorWave, Oracle and TogetherAI supported the work.[9][10]

AgentX exercises long, multi-turn coding sessions that accumulate histories, spawn child conversations, wait for results and resume. It uses synthesized text with recorded lengths, shared-prefix relationships, subagent topology and timing to test scheduling and cache reuse. Concurrency counts live session trees, so child conversations can produce more simultaneous requests than that count. Documentation credits SemiAnalysis with proposing the benchmark, AIPerf’s team with implementation and documentation, and Callan Fox and Weka with the underlying trace work.[10][12]

The AIPerf tutorial identifies AgentX as a work-in-progress MVP. Corpus variants, context filtering and trajectory-start settings can change the workload; globally idle periods may be compressed. Its validity flag checks only some conditions and does not certify full comparability.[11][12]

In a Sequoia interview, Patel argued that accelerator suitability depends on co-designing model architecture, systems software and hardware. Memory, networking topology and model sparsity interact, so an accelerator cannot be judged in isolation from the model it serves. He distinguished small teams’ reliance on existing open software from large labs’ ability to build custom infrastructure around their chosen hardware.[22]

In the February 28, 2026 In-Context Cooking interview, Patel argued that hyperscalers would keep spending heavily because infrastructure-building capacity was their advantage and AI threatened their existing businesses. He predicted that Google would spend its cash generation on AI infrastructure and models in 2027, explicitly framing the claim in terms of cash flow. He also described expanding SemiAnalysis’s coverage into model usage, running costs and token economics.[3]

In September 2026, Patel added an investment and board role at Positron. Its September 10 financing announcement identified him as founder and CEO of SemiAnalysis Capital, a co-lead of the $375 million Series C. Patel said he was investing and joining the board because Positron’s forthcoming architecture addressed memory constraints without depending on HBM or advanced packaging, and because its first generation had deployment experience at Oracle. This was his investor assessment; the next-generation tapeout and production remained planned milestones.[16][17]

Forbes reported in September 2026 that Patel held a stake in Fluidstack. It also reported former employee Wei Zhou’s allegation that Patel pressed him to include sensitive Fluidstack TPU data in SemiAnalysis reports. SemiAnalysis filed its own suit alleging breach of contract and trade-secret misappropriation; Patel publicly denied the former employee’s claims on April 4. Forbes reported that the cases had moved to arbitration. The allegations remain contested.[19][20][21]

Before these AI analyses and investment roles, Patel developed his hardware interests through repair experiments and online forums. In the Sequoia interview, he recalled an Xbox’s red-ring failure leading him to open the console and try several fixes, followed by participation in hardware and smartphone communities. He later combined consulting during a national-park trip with textbooks and audiobooks about semiconductors and AI, and described learning progressively more through repeated visits to specialist conferences.[22]

Patel recalled that exposure of his anonymous identity briefly interrupted his posting before he began more deliberate publication under his name. SemiAnalysis describes its business-building from a solo venture since 2020; his April 5, 2021 welcome post invited subscriptions before the first issue. In the February 2026 cooking interview, he credited Doug with repeatedly encouraging named publication on Substack and charging for it. Patel initially resisted charging, eventually adopted it, and recalled Doug joining the company years later.[22][1][18][3]