AI Wiki

Dylan Patel

AI CompaniesAI HardwarePeople
15 min read
Updated
Suggest editHistoryTalk32 citationsFact-checked
RawGraph

Dylan Patel is a semiconductor and artificial intelligence analyst, and the founder, chief executive officer, and chief analyst of SemiAnalysis, a research firm and paid newsletter covering chips, AI infrastructure, and data centers [1]. He began publishing in 2020 as a solo writer and built the operation into a subscription research business whose reports on GPT-4, DeepSeek, AMD software quality, and Chinese chip production have repeatedly set the terms of industry debate [2][3][4]. By 2026 his work was being cited from the keynote stage at Nvidia GTC by chief executive Jensen Huang, while the firm's combination of journalism, paid consulting, and personal investing had attracted criticism and litigation [5][6].

Background

Patel has said relatively little in public about his life before SemiAnalysis, and the account below is limited to statements he has made himself or facts reported by established outlets.

In a February 2026 interview on the Latent Space podcast, Patel said he is "from rural Georgia," that he "did live in Minnesota briefly after college," and that he kept bees "for like a, a year and a half basically" [7]. In the same conversation he described running "an anonymous blog on the internet for many years" and posting under an anonymous handle in what he called "silicon Twitter" before writing under his own name. He credited Doug O'Laughlin, author of the Fabricated Knowledge newsletter and later a SemiAnalysis colleague, with telling him to stop publishing anonymously on WordPress, move to Substack, and start charging for the research [7].

Patel's public persona grew out of that anonymous-poster phase. Business Insider, profiling him in August 2024, noted that OpenAI co-founder Sam Altman had referred to him only as "that semianalysis guy" in a 2023 post on X, by which point the newsletter had already become what the article called a central source of intelligence for people tracking Nvidia [2].

SemiAnalysis

SemiAnalysis dates its founding to 2020. The company's own biography page for Patel states that "Since 2020, SemiAnalysis has transformed its business from a solo venture into a cohesive and focused team to provide breaking news and in-depth analysis for the most strategic, complex, and escalating challenges in the semiconductor industry" [1].

Growth was rapid but staged. Business Insider reported in August 2024 that the firm had gone from three analysts in two countries to twelve subject matter experts spread across the United States, Japan, Singapore, Taiwan, and France within about two years, hiring former employees of ASML, Microsoft, and Nvidia [2]. Patel told the publication, "We have the entire view of the supply chain, from manufacturing, up to models" [2]. By the February 2026 Latent Space interview he put headcount at roughly sixty people working across eight to ten countries [7]. The newsletter's Substack page listed more than 305,000 subscribers as of mid-2026 [8].

The business is not only a newsletter. SemiAnalysis says its differentiating method is to start with the technology rather than with company financials, to attend more than forty industry conferences a year, and to connect layers of the supply chain that are usually analysed separately [9]. Alongside the subscription it sells institutional data products, which are priced and licensed separately from the newsletter [10].

ProductWhat it covers
Accelerator and HBM ModelHistorical and forecast AI accelerator production by company and type, plus high bandwidth memory supply [10]
Datacenter Industry ModelCurrent and forecast critical IT power capacity for colocation and hyperscaler data center build-out [10]
AI Cloud TCO ModelOwnership economics for operators that buy accelerators and resell bare metal or cloud GPU compute [10]
ClusterMAXTiered rating and ranking system for GPU clouds [11]
InferenceXOpen-source continuous inference benchmark, originally launched as InferenceMAX [12]

The Information reported in April 2026 that SemiAnalysis expected to exceed $100 million in revenue for the year, drawn from subscriptions and AI supply chain research sold to startups, investment firms, and corporate teams [4]. That figure is the company's own projection as relayed by the publication, not an audited result.

Notable reports

Several SemiAnalysis pieces carrying Patel's byline became reference points well beyond the semiconductor trade press.

DateReportCo-authorsWhy it mattered
4 May 2023Google "We Have No Moat, And Neither Does OpenAI"Afzal AhmadPublished a leaked internal Google memo arguing open-source AI would outcompete both firms [13]
10 July 2023GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoEGerald WongSource of the widely repeated but unconfirmed GPT-4 parameter figures [3]
28 Aug 2023Google Gemini Eats The World: Gemini Smashes GPT-4 By 5X, The GPU-PoorsDaniel NishballPrompted a public rebuttal from Sam Altman [14]
22 Dec 2024MI300X vs H100 vs H200 Benchmark Part 1: TrainingSemiAnalysis teamLed to a meeting with AMD chief executive Lisa Su [15][16]
31 Jan 2025DeepSeek DebatesAJ Kourabi, Doug O'Laughlin, Reyk KnuhtsenRecast the DeepSeek V3 cost story during the market selloff [17]

The Google memo

On 4 May 2023 Patel and Afzal Ahmad published a leaked internal Google document under the headline "We Have No Moat, And Neither Does OpenAI." The post explained that the text "was shared by an anonymous individual on a public Discord server" and that SemiAnalysis had verified it originated from a Google researcher [13]. The authors distanced themselves from its argument: "The document is only the opinion of a Google employee, not the entire firm. We do not agree with what is written below ... We simply are a vessel to share this document which raises some very interesting points" [13].

GPT-4 architecture

On 10 July 2023, Patel and Gerald Wong published "GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE" [3]. The piece argued that inference cost, not training cost, was the binding constraint on frontier models, and that dense transformers could not scale economically much further. Its paywalled section described GPT-4 as a sparse mixture-of-experts model. Contemporary coverage reported the specifics as roughly 1.8 trillion parameters across 120 layers, 16 experts of about 111 billion parameters each with two routed per forward pass, training on about 13 trillion tokens, and a training cost near $63 million [18]. OpenAI has never confirmed any of these figures, and they should be treated as an unverified third-party estimate rather than a disclosed specification.

Gemini and the "GPU-Poors"

On 28 August 2023, Patel and Daniel Nishball published "Google Gemini Eats The World: Gemini Smashes GPT-4 By 5X, The GPU-Poors," arguing that Google's compute position gave Gemini an advantage over GPT-4 [14]. Altman replied on X that it was "incredible Google got that SemiAnalysis guy to publish their internal marketing/recruiting chart," and Patel responded that the data came from a Google supplier and that SemiAnalysis had built the chart itself [14].

AMD software

In December 2024, after months of testing, SemiAnalysis published "MI300X vs H100 vs H200 Benchmark Part 1: Training," which argued that Nvidia's CUDA software advantage remained intact and that bugs and usability gaps in AMD's ROCm stack made the AMD Instinct MI300X far harder to train on than its paper specifications suggested [15]. AMD chief executive Lisa Su responded publicly the next day, writing to Patel on X: "Thanks @dylan522p for the constructive conversation today. Feedback is a gift even when it's critical. We have put a ton of work into customer and workload optimizations but there is lots more we can do to enable the broad ecosystem" [16].

The DeepSeek cost debate

When DeepSeek's V3 and R1 releases triggered a selloff in AI-linked equities in late January 2025, the widely quoted claim was that the model had been trained for about $6 million. On 31 January 2025, Patel and colleagues published "DeepSeek Debates," which argued that the number "is attributed to just the GPU cost of the pre-training run, which is only a portion of the total cost" [17]. The report estimated that DeepSeek had access to roughly 50,000 Hopper-generation GPUs including about 10,000 H800s and 10,000 H100s alongside orders for H20s, that total server capital expenditure was around $1.6 billion, and that operating those clusters accounted for about $944 million [17]. Patel's framing, that the training run was genuinely efficient but sat on top of a far larger hardware base, became the standard counterpoint during the DeepSeek market crash.

Export controls, Huawei, and China

Patel is a frequent and generally critical commentator on United States AI chip export controls. In an April 2025 ChinaTalk interview with Jordan Schneider, Patel and Doug O'Laughlin argued that policy had been sequenced badly, with Patel saying that regulators "got the order of operations completely wrong, and many loopholes were left by the Biden administration" [19]. He pointed to continued equipment sales to Chinese fabs including SMIC, to Chinese dominance of optical transceiver manufacturing, and to stockpiles that insulated Huawei Ascend production from restrictions [19].

SemiAnalysis research under his byline has tracked those stockpiles in detail, reporting that Huawei obtained millions of Ascend dies from TSMC through intermediaries and that Samsung had been a leading supplier of HBM into China, giving Huawei a die and memory bank that supported Ascend output through 2024 and 2025 [20]. The firm's conclusion has generally been that controls constrain Chinese accelerator volumes over time, not that they have failed outright.

Benchmarks and rating systems

Two SemiAnalysis products under Patel's direction have become de facto reference points for buyers of AI compute.

ClusterMAX, launched on 26 March 2025 with Patel as lead author, rates GPU clouds on a tiered scale of Platinum, Gold, Silver, Bronze, and Underperforming (spelled "UnderPerform" in the launch edition), based on customer-facing criteria such as software quality, networking, storage, security, and support [11]. The first edition claimed roughly 90 percent coverage of the GPU rental market by volume [11]. The November 2025 version 2.0 rated 84 providers out of 209 tracked; the April 2026 version 2.1 added further entrants without a published count, and CoreWeave remained the only neocloud at the Platinum tier [21].

InferenceMAX, launched on 9 October 2025 and later renamed InferenceX, is an open-source benchmark that runs inference workloads nightly across accelerators including the GB200 NVL72, B200, H200, H100, and AMD's MI355X, MI325X, and MI300X, publishing results to a free public dashboard [12]. Nvidia's developer blog described it days later as an "independent third-party evaluation from SemiAnalysis" [22]. At the GTC 2026 keynote in San Jose on 16 March 2026, Jensen Huang presented InferenceX results on stage; SemiAnalysis's own conference write-up called "a Jensen callout for InferenceX during the keynote" a highlight of the event [5][23].

Public commentary

Patel is a regular podcast guest and conference speaker. He appeared with Nathan Lambert on episode 459 of the Lex Fridman Podcast, published 2 February 2025, in a long discussion of DeepSeek, export controls, and AI megaclusters [24]. In November 2025 he joined Dwarkesh Patel to interview Microsoft chief executive Satya Nadella inside Microsoft's Fairwater 2 data center, an episode published on 12 November 2025 [25]. He has spoken at industry events including the Design Automation Conference and the AI Infra Summit, where he was billed for a main-stage session on 16 September 2026 [26][27].

Investing, ETFs, and disputes

Patel's activities have expanded beyond research in ways that have raised questions about the firm's independence.

In February 2026, The Information reported that SemiAnalysis was in early talks to raise hundreds of millions of dollars for a venture fund, and that Patel had raised a $50 million special purpose vehicle as part of a $700 million fundraising by the GPU cloud provider FluidStack [28]. The fund subsequently took formal shape: on 29 July 2026, SemiAnalysis Capital Fund I, LP, a Delaware limited partnership, filed a Form D with the Securities and Exchange Commission for a $400,000,000 venture capital offering, with none sold as of the filing date. Patel signed the filing as manager of the general partner [32]. On 29 June 2026, Tema ETFs announced an exclusive partnership with SemiAnalysis to build a suite of semiconductor and AI infrastructure exchange-traded funds informed by the firm's research, with Patel quoted describing SemiAnalysis as "deeply specialized in AI, semiconductor, datacenter and cloud infrastructure intelligence and research globally" [29].

In March 2026 the firm became involved in litigation with a former employee. Reporting by Ian Cutress of More Than Moore, based on the filings in San Francisco County Superior Court, describes SemiAnalysis suing former employee Wei Zhou on 27 March 2026, and Zhou filing his own action on 31 March 2026 [6]. SemiAnalysis filed first, pre-empting a suit Zhou had told it he intended to bring; its complaint (SemiAnalysis LLC v. Wei Zhou, case CGC-26-635328, Superior Court of California, County of San Francisco) pleads nine causes of action, including trade secret misappropriation, conversion, breach of contract, breach of fiduciary duty, defamation, and unfair competition [31]. Zhou's filing alleges he was terminated in retaliation for refusing to incorporate material non-public information, connected to the $50 million investment vehicle Patel managed personally, into models sold to clients [6]. Patel's response, quoted by Cutress, was that "In January 2026, SemiAnalysis terminated an employee due to severe misconduct," citing workplace disruption, inappropriate comments, and alcohol-related issues [6]. As of the most recent reporting consulted, neither set of allegations has been tested in court, and none has been proven.

Separately, Jon Stevens, chief executive of the GPU cloud provider Hot Aisle, published a critical essay on 2 December 2025 arguing that SemiAnalysis blurs the roles of publisher, paid consultant, and investor, and that its disclosure practices fall short of the norms applied to sell-side analysts [30]. Stevens states in the piece that his evidence is observational, drawn from public exchanges and filings rather than named internal sources [30]. Hot Aisle is itself rated by ClusterMAX, placed in the Bronze tier in the April 2026 edition, which gives Stevens a commercial interest in the rating system he criticises [21].

References

  1. ^Dylan Patel: Founder, CEO, and Chief Analyst, SemiAnalysis - SemiAnalysis.
  2. ^Emma Cosgrove, How analyst Dylan Patel makes connections from Nvidia to the semiconductor industry and the tech it enables - Business Insider, 14 August 2024. Syndicated mirror: aol.com/analyst-dylan-patel-gets-inside-090002369
  3. ^Dylan Patel and Gerald Wong, GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE - SemiAnalysis, 10 July 2023.
  4. ^Abram Brown, A look at Dylan Patel's SemiAnalysis, an AI newsletter and research firm that expects $100M+ in 2026 revenue - The Information, 17 April 2026 (item summary via Techmeme, 19 April 2026).
  5. ^SemiAnalysis, Nvidia - The Inference Kingdom Expands - 24 March 2026.
  6. ^Ian Cutress, Dylan Patel's SemiAnalysis Is Being Sued - More Than Moore, 4 April 2026.
  7. ^Dylan Patel of SemiAnalysis on the $200B AI CapEx, Chip Wars, and Why Google Might Have No Profits in 2027 - Latent Space, 28 February 2026.
  8. ^SemiAnalysis - Substack publication page, accessed August 2026.
  9. ^About - SemiAnalysis.
  10. ^Models and Research - SemiAnalysis.
  11. ^Dylan Patel, Kimbo Chen, Daniel Nishball and others, The GPU Cloud ClusterMAX Rating System: How to Rent GPUs - SemiAnalysis, 26 March 2025.
  12. ^Kimbo Chen, Dylan Patel, Daniel Nishball and others, InferenceMAX: Open Source Inference Benchmarking - SemiAnalysis, 9 October 2025.
  13. ^Dylan Patel and Afzal Ahmad, Google "We Have No Moat, And Neither Does OpenAI" - SemiAnalysis, 4 May 2023.
  14. ^Maggie Harrison Dupre, OpenAI Rages at Report That Google's New AI Crushes GPT-4 - Futurism, 31 August 2023.
  15. ^SemiAnalysis, MI300X vs H100 vs H200 Benchmark Part 1: Training, CUDA Moat Still Alive - 22 December 2024.
  16. ^Lisa Su, post on X - 23 December 2024.
  17. ^Dylan Patel, AJ Kourabi, Doug O'Laughlin and Reyk Knuhtsen, DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts - SemiAnalysis, 31 January 2025.
  18. ^GPT-4 architecture, datasets, costs and more leaked - The Decoder, July 2023.
  19. ^Jordan Schneider, Dylan Breaks Huawei and Tariffs Right - ChinaTalk, 21 April 2025.
  20. ^SemiAnalysis, Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck - 8 September 2025.
  21. ^Best GPU Clouds 2026: ClusterMAX 2.0 + 2.1 - ClusterMAX by SemiAnalysis, April 2026.
  22. ^NVIDIA Blackwell Leads on SemiAnalysis InferenceMAX v1 Benchmarks - NVIDIA Technical Blog, 13 October 2025.
  23. ^Nvidia GTC 2026: CEO Jensen Huang keynote - CNBC, 16 March 2026.
  24. ^DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters, Lex Fridman Podcast #459 - 2 February 2025.
  25. ^Dwarkesh Patel, Satya Nadella: How Microsoft is preparing for AGI - Dwarkesh Podcast, 12 November 2025.
  26. ^Dylan Patel - Design Automation Conference 2026 speaker page.
  27. ^Dylan Patel - AI Infra Summit 2026 speaker page.
  28. ^Sources: Dylan Patel's SemiAnalysis is in early talks to raise hundreds of millions for a VC fund; Patel raised a $50M SPV toward Fluidstack's $700M fundraising - Techmeme summary of The Information, 18 February 2026.
  29. ^Tema ETFs and SemiAnalysis Launch Exclusive Semiconductor ETF Partnership - Tema ETFs press release, 29 June 2026.
  30. ^Jon Stevens, Influence as a Service: SemiAnalysis Under the Microscope - 2 December 2025.
  31. ^Superior Court of California, County of San Francisco, SEMIANALYSIS LLC VS. WEI ZHOU, case no. CGC-26-635328 - filed 27 March 2026.
  32. ^U.S. Securities and Exchange Commission, Form D, SemiAnalysis Capital Fund I, LP (CIK 0002145695) - filed 29 July 2026.

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

v1 · 2,925 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Reviewer note: Biographical claims, report bylines, quotations, litigation docket and the July 2026 SEC Form D confirmed against primary sources on 2026-08-01.

Suggest edit