AI Wiki

We Have No Moat, And Neither Does OpenAI

AI CompaniesLarge Language ModelsOpen Source AI
19 min read
Updated
Suggest editHistoryTalk32 citationsFact-checked
RawGraph

"We Have No Moat, And Neither Does OpenAI" is an internal Google document from April 2023 arguing that neither Google nor OpenAI held a durable competitive advantage in large language models, because an open-source community working on leaked model weights was closing the quality gap faster than either company could defend it. It was published on 4 May 2023 by the semiconductor research newsletter SemiAnalysis, which said the text had been shared by an anonymous person on a public Discord server and originated with a researcher inside Google [1]. Bloomberg subsequently identified the author as Luke Sernau, a senior engineer at Google [2][3]. The document circulated very widely, and its central phrase became shorthand for the argument that open-weight models would commoditize frontier AI.

What the document is

SemiAnalysis prefaced the text with a short note describing its provenance. The publication said the text was "a very recent leaked document, which was shared by an anonymous individual on a public Discord server who has granted permission for its republication," adding: "It originates from a researcher within Google. We have verified its authenticity." SemiAnalysis said its only changes were "formatting and removing links to internal web pages," and stated plainly that "The document is only the opinion of a Google employee, not the entire firm" [1]. The newsletter also distanced itself from the contents: "We do not agree with what is written below, nor do other researchers we asked," promising a separate subscriber-only response piece [1]. Dylan Patel repeated the disclaimer when he posted the link, writing that it was the opinion of one Googler and that SemiAnalysis was "simply sharing" [4].

The byline on the SemiAnalysis post is Dylan Patel and Afzal Ahmad, who were the publishers rather than the authors [1]. The distinction matters, because early coverage and social media frequently credited the analysis to SemiAnalysis rather than to the anonymous Googler.

Authorship and Google's position

Bloomberg reported on 5 May 2023 that the document was written by Luke Sernau and posted to an internal Google system in early April 2023 [2][3][5]. Fortune, reporting the same day, described Sernau as a senior engineer at Google and said the document was published internally in early April before SemiAnalysis released it [5]. SiliconANGLE also credited Bloomberg with identifying Sernau [3].

Google DeepMind chief executive Demis Hassabis was asked about the document by The Verge in July 2023. He said "I think that memo was real," but added that engineers at Google write various documents that sometimes leak, that he "wouldn't take it too seriously" because "these are just opinions," and that he disagreed with its conclusions [30]. Fortune reported in May 2023 that neither Sernau nor Alphabet had responded to its request for comment [5]. SemiAnalysis characterized the text as one employee's opinion rather than company policy [1].

The argument

The document opens by describing the Google versus OpenAI rivalry as the wrong frame. "We've done a lot of looking over our shoulders at OpenAI," it begins, before asserting that "the uncomfortable truth is, we aren't positioned to win this arms race and neither is OpenAI. While we've been squabbling, a third faction has been quietly eating our lunch." That faction is open source: "Plainly put, they are lapping us" [1].

It then lists problems that Google treated as open research questions but which, in the author's view, had already been solved outside: foundation models running on a Pixel 6 at 5 tokens per second, personalized fine-tuning on a laptop in an evening, unrestricted model releases, and a multimodal ScienceQA state of the art trained in an hour [1].

The cost comparison is the line most often quoted. "While our models still hold a slight edge in terms of quality, the gap is closing astonishingly quickly. Open-source models are faster, more customizable, more private, and pound-for-pound more capable. They are doing things with $100 and 13B params that we struggle with at $10M and 540B. And they are doing so in weeks, not months" [1]. The 540B figure matches the parameter count of Google's PaLM. Three consequences follow, listed as bullets: "We have no secret sauce"; "People will not pay for a restricted model when free, unrestricted alternatives are comparable in quality"; and "Giant models are slowing us down," with a recommendation to treat sub-20B variants as a priority rather than an afterthought [1].

The mechanism the document credits is the March 2023 leak of Meta's LLaMA weights. "At the beginning of March the open source community got their hands on their first really capable foundation model, as Meta's LLaMA was leaked to the public," it says, noting the model had no instruction tuning, no conversation tuning, and no RLHF, and that the community added all of them within weeks. "The barrier to entry for training and experimentation has dropped from the total output of a major research organization to one person, an evening, and a beefy laptop" [1].

The author compares this to what had already happened in image generation, where "low-cost public involvement was enabled by a vastly cheaper mechanism for fine tuning called low rank adaptation, or LoRA, combined with a significant breakthrough in scale," namely latent diffusion for images and Chinchilla for text. Having an open model, the document argues, put Stable Diffusion on a different path from DALL-E, producing "product integrations, marketplaces, user interfaces, and innovations that didn't happen for Dall-E" [1].

LoRA and stackable fine-tuning

The technical core of the argument is LoRA. The document describes it as working "by representing model updates as low-rank factorizations, which reduces the size of the update matrices by a factor of up to several thousand," and says the resulting ability to "personalize a language model in a few hours on consumer hardware is a big deal." It calls the technique "underexploited inside Google" [1].

The key property, in the author's telling, is composability: "Part of what makes LoRA so effective is that - like other forms of fine-tuning - it's stackable. Improvements like instruction tuning can be applied and then leveraged as other contributors add on dialogue, or reasoning, or tool use. While the individual fine tunings are low rank, their sum need not be, allowing full-rank updates to the model to accumulate over time" [1]. Training giant models from scratch, by contrast, "not only throws away the pretraining, but also any iterative improvements that have been made on top" [1]. From this the document draws its most contrarian conclusion: "Focusing on maintaining some of the largest models on the planet actually puts us at a disadvantage" [1].

A separate section is headed "Data quality scales better than data size." It observes that open projects were training on small, curated datasets built with synthetic methods and scavenged from other projects, and that this "suggests there is some flexibility in data scaling laws," meaning that the data side of the scaling laws the field was working from was softer than assumed [1].

The timeline

The document appends a chronology of roughly seven weeks of open-source activity, which is the evidentiary base for the whole argument [1].

Date (2023)Event as described in the document
Feb 24Meta launches LLaMA, open sourcing the code but not the weights, at 7B, 13B, 33B and 65B parameters
Mar 3LLaMA is leaked to the public within a week
Mar 12Artem Andreenko gets the model running on a Raspberry Pi, too slowly to be practical
Mar 13Stanford releases Alpaca, adding instruction tuning; Eric Wang's alpaca-lora repo trains it "within hours on a single RTX 4090"
Mar 18Georgi Gerganov uses 4-bit quantization to run LLaMA on a MacBook CPU. "It is the first 'no GPU' solution that is fast enough to be practical."
Mar 19A cross-university collaboration releases Vicuna, evaluated with GPT-4 as judge; training cost $300
Mar 25Nomic creates GPT4All, a model and an ecosystem; training cost $100
Mar 28Cerebras trains the GPT-3 architecture on the Chinchilla-optimal compute schedule, so the community is no longer dependent on LLaMA
Mar 28LLaMA-Adapter adds instruction tuning and multimodality in one hour with 1.2M learnable parameters, reaching a new state of the art on multimodal ScienceQA
Apr 3Berkeley launches Koala; more than 50% of the time users either prefer Koala to ChatGPT or have no preference; training cost $100
Apr 15Open Assistant releases an RLHF model and dataset scoring 48.3% against ChatGPT's 51.7% on human preference

The document notes that the Vicuna evaluation method was "suspect," and that Vicuna's training data was obtained by sampling ChatGPT dialogues posted on sites such as ShareGPT, circumventing API restrictions [1].

The recommendation

The strategic conclusion is that Google should stop trying to out-secret an ecosystem it cannot outrun. "Who would pay for a Google product with usage restrictions if there is a free, high quality alternative without them?" the document asks, and answers that Google should not expect to catch up: "The modern internet runs on open source for a reason" [1]. It argues that secrecy is already leaking through researcher attrition, that individuals face far weaker license constraints than corporations, and that the beneficiary of the situation is Meta: "Paradoxically, the one clear winner in all of this is Meta. Because the leaked model was theirs, they have effectively garnered an entire planet's worth of free labor" [1].

Google, the document says, should do to LLMs what it did with Chrome and Android and own the ecosystem instead: "The more tightly we control our models, the more attractive we make open alternatives... But this control is a fiction." It recommends that "Google should establish itself a leader in the open source community, taking the lead by cooperating with, rather than ignoring, the broader conversation," including "publishing the model weights for small ULM variants," and concludes that "We cannot hope to both drive innovation and control it" [1].

The epilogue dismisses the competitor the rest of Google was watching: "And in the end, OpenAI doesn't matter. They are making the same mistakes we are in their posture relative to open source... Open source alternatives can and will eventually eclipse them unless they change their stance" [1].

Reception in 2023

The document spread quickly. The Hacker News submission of the SemiAnalysis post, filed the morning of 4 May 2023, accumulated 2,455 points and 1,039 comments [6]. Simon Willison published a detailed excerpt-and-commentary post the same day, writing that the document looked real to him and that "even if it isn't, I think the analysis within stands alone." He called it "the most interesting piece of writing I've seen about LLMs in a while" [7]. LWN covered it that day for a Linux and free-software audience [8]. Bloomberg, Fortune and SiliconANGLE all ran stories on 5 May [2][3][5].

The Latent Space podcast convened an emergency Twitter Space on 4 May with Willison and Travis Fischer; the hosts reported that 2,800 people listened in, and published a cleaned-up recording the next day [9]. That episode is also the clearest contemporaneous record of the private pushback. The hosts reported that a CEO had told them Sernau did not seem to work in AI-related parts of Google and that "software engineers don't understand moats." They cited Emad Mostaque of Stability AI, himself an open-model advocate, as repeatedly stating that "Closed models will always outperform open models" because closed models can wrap open ones, and as listing the moats he did see: "Unique usage data, Unique content, Unique talent, Unique product, Unique business model." The hosts added their own objection: it was not clear what race the document thought Google and OpenAI were in, since neither was especially concerned with running models locally on phones [9].

Substantive counterarguments arrived over the following weeks.

  • On the evidence base, a Berkeley group published "The False Promise of Imitating Proprietary LLMs" on 25 May 2023. Training imitation models on ChatGPT outputs, the authors found they "close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data," and that crowd raters missed this because such models mimic ChatGPT's style but not its factuality. Their conclusion was that "there exists a substantial capabilities gap between open and closed LMs" [10]. Vicuna and Koala, two of the document's headline examples, were exactly this kind of model.
  • On the supply chain, MIT Technology Review argued on 12 May 2023 that the open-source boom was "built on Big Tech's handouts" and could be withdrawn. It quoted Joelle Pineau, then managing director of Meta AI, declining to commit: "Is this the same strategy that we'll adopt for the next five years? I don't know, because AI is moving so quickly" [11].

How the predictions have held up

Assessments of the document have diverged, and named commentators disagree about which parts were right.

Writing in August 2024, Matthew Lynley of Supervised judged the memo partially correct: on-device inference through quantized small models had become common, low-rank fine-tuning had become the standard personalization method, and price competition had intensified, but the prediction that people would not pay for restricted models behind an API had not materialized, and Google had held its position through a dual strategy of frontier models plus open releases [12].

Linux Foundation chief executive Jim Zemlin, writing on 11 February 2025 in the weeks after DeepSeek-R1, took the opposite view. He recalled that comparable open models had lagged closed ones by months in 2023 but said that "that slight time-to-market advantage has shrunk to 24 hours," citing the speed with which Hugging Face and others replicated OpenAI's Deep Research agent [13]. Colin Eberhardt of Scott Logic wrote on 27 July 2026 that the memo "generated interest at the time of writing, but soon faded from view," and that "Three years later, that memo reads less like an opinion and more like a prediction," pointing to GLM-5.2 and Kimi K3 as the moment the gap closed entirely [14].

Measured rather than narrative assessments are less emphatic. Epoch AI reported on 29 May 2026 that between 1 January and 28 May 2026 the most capable open-weight models lagged frontier closed models by an average of four months, an 8-point gap on its Epoch Capabilities Index, and that this was slightly wider than the roughly three-month lag it had measured in October 2025. Epoch cautioned that its estimate may understate the gap, since open models can perform worse on private benchmarks and leading labs do not always release their strongest systems [15]. Stephen O'Grady of RedMonk, writing on 15 May 2026 without reference to the memo, described the same pattern as a cycle rather than a convergence: "Closed leaps forward, open is hot on its heels. And the cycle appears to be getting faster" [16].

Against the specific claims, the record is mixed:

Claim in the documentSubsequent development
Open models will close the quality gapZ.ai released GLM-5.2 weights under an MIT license on 16 June 2026, a 753B-parameter mixture-of-experts model reported to beat GPT-5.5 on several long-horizon coding benchmarks at about one-sixth the cost [17][18]. Epoch AI still measured a four-month average lag as of May 2026 [15]
Focus on the sub-20B regime; giant models are a disadvantageEpoch AI measured frontier training compute growing 4-5x per year through May 2024, and estimated in January 2025 that the frontier open-weight trend scales at about 4.7x per year [19][20]. Moonshot AI's Kimi K3, announced in July 2026 at 2.8 trillion parameters with weights published on 27 July, was reported as the largest open-weight model released to that point [21][31]
Data quality scales better than data sizeMicrosoft's phi-1, published 20 June 2023, trained a 1.3B-parameter model on 6B tokens of textbook-quality web data plus 1B tokens of synthetic textbooks and reached 50.6% pass@1 on HumanEval [22], an early instance of the curated-data approach the document described
Google should lead in open source and publish small-model weightsGoogle released Gemma on 21 February 2024 at 2B and 7B parameters [23]. By the Gemma 4 release in April 2026, Google had moved the family to an Apache 2.0 license and reported over 400 million cumulative downloads and more than 100,000 community variants [24]
OpenAI does not matter unless it changes its stanceOpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0 on 5 August 2025, its first open-weight language models since GPT-2 [25]
Meta is the one clear winnerTechCrunch reported in July 2025, relaying the New York Times, that senior figures at Meta Superintelligence Labs had discussed moving away from the open Behemoth model toward a closed one, though Meta said its position on open source was unchanged; Mark Zuckerberg then signaled that not all of the company's superintelligence models would be open sourced [26][27]. O'Grady's May 2026 sample found that every one of the 17 Chinese frontier models he counted was open weight or fully open, against 64% closed among 42 US models [16]

Two framing points are worth separating. The document was about model weights and research velocity, not about revenue, and it made only one commercial prediction, that people would not pay for restricted models. Lynley singled out that prediction as the one that had, in his words, "not end[ed] up entirely correct, though that could still change in the future" [12]. Its chronology also needs care: the LLaMA weights that triggered the events it describes were leaked in March 2023 without Meta's authorization, and Meta's own launch that February released the code but not the weights [1]. Meta first published Llama weights itself, free for research and commercial use, with Llama 2 on 18 July 2023, more than three months after the document was written [29].

Relationship to SemiAnalysis

SemiAnalysis published the document, wrote its introductory note, and said it had verified authenticity, but did not write the analysis and explicitly disagreed with it [1][4]. The post nonetheless became one of the things the newsletter is best known for. Speaking on the Latent Space podcast in November 2023, Dylan Patel said it was "sometimes bad to like have the infamousness of, you know, only people caring about this or the GPT-4 leak or the Google has no moat leak," contrasting that visibility with the supply-chain research that makes up most of the firm's work [28]. The post was resubmitted to Hacker News on 27 January 2025, drawing a smaller second thread of 95 points and 53 comments [32].

References

  1. ^Dylan Patel and Afzal Ahmad, "Google 'We Have No Moat, And Neither Does OpenAI'", SemiAnalysis, 4 May 2023. newsletter.semianalysis.com/...no-moat-and-neither
  2. ^"Google Is Falling Behind in AI Arms Race, Senior Engineer Warns", Bloomberg, 5 May 2023. bloomberg.com/...i-warning-we-have-no-secret-sauce
  3. ^Kyt Dotson, "Google and OpenAI struggling to keep up with open-source AI, senior engineer warns", SiliconANGLE, 5 May 2023. siliconangle.com/...ource-ai-senior-engineer-warns
  4. ^Dylan Patel (@dylan522p), post on X, 4 May 2023. x.com/...1654066074382651394
  5. ^Will Daniel, "A senior Google engineer just referenced Warren Buffett's decades-old economic moat theory", Fortune, 5 May 2023. fortune.com/...ificial-intelligence-warren-buffett
  6. ^Hacker News discussion, item 35813322, 4 May 2023 (2,455 points, 1,039 comments). news.ycombinator.com/item
  7. ^Simon Willison, "Leaked Google document: 'We Have No Moat, And Neither Does OpenAI'", 4 May 2023. simonwillison.net/...no-moat
  8. ^"Google: 'We Have No Moat, And Neither Does OpenAI'", LWN.net, 4 May 2023. lwn.net/...930939
  9. ^"No Moat: Closed AI gets its Open Source wakeup call, ft. Simon Willison", Latent Space, 5 May 2023. latent.space/...no-moat
  10. ^Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine and Dawn Song, "The False Promise of Imitating Proprietary LLMs", arXiv:2305.15717, 25 May 2023. arxiv.org/...2305.15717
  11. ^Will Douglas Heaven, "The open-source AI boom is built on Big Tech's handouts. How long will it last?", MIT Technology Review, 12 May 2023. technologyreview.com/...oogle-openai-eleuther-meta
  12. ^Matthew Lynley, "Revisiting that old Google AI memo", Supervised, 9 August 2024. supervised.news/...isiting-that-old-google-ai-memo
  13. ^Jim Zemlin, "'We Have No Moat': Open Source AI's Breakneck Innovation", Linux Foundation, 11 February 2025. linuxfoundation.org/...ce-ais-breakneck-innovation
  14. ^Colin Eberhardt, "The Rise of Open Weights; And the fall of commercial AI?", Scott Logic, 27 July 2026. blog.scottlogic.com/...rise-of-open-weights
  15. ^"Open models lag state-of-the-art closed models by 4 months", Epoch AI, 29 May 2026. epoch.ai/...open-closed-eci-gap
  16. ^Stephen O'Grady, "Open and Closed: The Pursuit of Frontier Models", RedMonk, 15 May 2026. redmonk.com/...open-ai-models
  17. ^Simon Willison, "GLM-5.2 is probably the most powerful text-only open weights LLM", 17 June 2026. simonwillison.net/...glm-52
  18. ^"Z.ai's open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost", VentureBeat, June 2026. venturebeat.com/...g-benchmarks-for-1-6th-the-cost
  19. ^"Training compute of frontier AI models grows by 4-5x per year", Epoch AI, analysis covering 2010 to May 2024. epoch.ai/...ntier-ai-models-grows-by-4-5x-per-year
  20. ^"Frontier open models may surpass 1e26 FLOP of training compute before 2026", Epoch AI, 15 January 2025. epoch.ai/...open-models-threshold
  21. ^"China's 2.8-trillion-parameter Kimi K3 beats Claude Fable 5 in Frontend Code Arena benchmark", Tom's Hardware, July 2026. tomshardware.com/...2-8-trillion-parameter-kimi-k3
  22. ^Suriya Gunasekar et al., "Textbooks Are All You Need", arXiv:2306.11644, 20 June 2023. arxiv.org/...2306.11644
  23. ^"Gemma: Google introduces new state-of-the-art open models", Google, 21 February 2024. blog.google/...gemma-open-models
  24. ^Nia Castelly, Amanda Casari and Olivier Lacombe, "Gemma 4: Expanding the Gemmaverse with Apache 2.0", Google Open Source Blog, 2 April 2026. opensource.googleblog.com/...averse-with-apache-20
  25. ^"Introducing gpt-oss", OpenAI, 5 August 2025. openai.com/...introducing-gpt-oss
  26. ^"Meta built its AI reputation on openness, that may be changing", TechCrunch, 14 July 2025. techcrunch.com/...on-openness-that-may-be-changing
  27. ^"Zuckerberg signals Meta won't open source all of its 'superintelligence' AI models", TechCrunch, 30 July 2025. techcrunch.com/...-its-superintelligence-ai-models
  28. ^"The State of Silicon and the GPU Poors, with Dylan Patel of SemiAnalysis", Latent Space, 17 November 2023. latent.space/...semianalysis
  29. ^"Meta and Microsoft Introduce the Next Generation of Llama", Meta, 18 July 2023. about.fb.com/...llama-2
  30. ^Nilay Patel, "Google DeepMind CEO Demis Hassabis on the path from chatbots to AGI", Decoder, The Verge, 10 July 2023. theverge.com/...s-google-deepmind-ai-alphafold-agi
  31. ^Moonshot AI, Kimi-K3 model repository, Hugging Face (weights published 27 July 2026). huggingface.co/...Kimi-K3
  32. ^"Google 'We Have No Moat, And Neither Does OpenAI'", Hacker News item 42838112, 27 January 2025. news.ycombinator.com/item

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 · 3,711 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: All 28 quotations from the leaked document matched word for word against the SemiAnalysis publication, and authorship, circulation figures and retrospective assessments checked against named sources on 2026-08-01.

Suggest edit