Poolside AI

Poolside AI is an American-founded artificial intelligence company that trains its own foundation models for software development and sells them to enterprises and regulated-industry customers that run the system on premises or inside their own clouds. It was founded in 2023 by Jason Warner, the former chief technology officer of GitHub, and Eiso Kant, a serial developer-tooling founder previously known for source{d} and Athenian, and raised a $500 million Series B led by Bain Capital Ventures in October 2024 at a $3 billion valuation.^[1][2][4] In October 2025, Bloomberg and TechCrunch reported that Poolside was seeking up to $2 billion at a roughly $12 billion valuation, with Nvidia considering a commitment of $500 million to $1 billion.^[5][19] The Financial Times reported in April 2026 that those talks had stalled and that the proposed round had fallen apart.^[42] In August 2026, Newcomer, Bloomberg and The Information reported a separate arrangement combining a $6 billion nonexclusive technology license, offers to 109 Poolside employees, and a separate $1 billion Nvidia investment at a $12 billion pre-money valuation.^[35][36][37] Neither company had published a transaction announcement by 22 August 2026, and in a 19 October 2026 disclosure an early Poolside investor, Tiny Ltd., still described the agreements as reported.^[38][39][60] Poolside trains its models with a technique it calls reinforcement learning from code execution feedback (RLCEF); it said in October 2024 that it had brought 10,000 Nvidia GPUs online for training, and a July 2025 company post described that cluster as 10,000 NVIDIA H200 GPUs.^[6][26][40] Its first production models were Malibu and Point. From April 2026 it released the Laguna family of agentic coding models, all of which now have open weights, culminating in July 2026 with Laguna S 2.1, a 118-billion-parameter model; it ships them through its pool terminal agent, Poolside Assistant and the enterprise Poolside Platform.^{[8][9][45][50][51]} Its planned Texas data-centre campus, Project Horizon, is being developed by a separate company, Poolside Infrastructure Company.^[61]

Who founded Poolside?

Poolside was founded in 2023 by Jason Warner and Eiso Kant; the company has written that it was founded in San Francisco in April 2023, and while early profiles and a July 2025 company post described Warner as chief executive and Kant as chief technology officer, its announcements from October 2025 onward identify both founders as co-chief executives.^{[2][64][7][33]} Warner is the former chief technology officer of GitHub, where he oversaw engineering during the launch of GitHub Actions and the original GitHub Copilot, and a former Redpoint Ventures managing director.^[1][11] Kant founded source{d}, a developer-tooling startup that applied machine learning to source code and described itself as the first company dedicated to AI for code, and later co-founded the Paris-based engineering-analytics company Athenian.^[13][14]

Warner and Kant met in 2017 while Warner, then leading GitHub's engineering organization, attempted to acquire Kant's company source{d}, a developer-tooling startup applying machine learning to source code.^[2][10] The acquisition did not close, but the pair stayed in contact through Warner's tenure as CTO of GitHub during the launch of GitHub Actions and the original Copilot, and through his subsequent move to Redpoint Ventures as a managing director in 2021.^[11][12] Kant in the same period co-founded Athenian, a Paris-based engineering analytics platform, after stepping back from source{d} in 2019.^[13][14]

The founders originally planned to call the company Snowball but dropped the name because of trademark issues with Amazon's AWS Snowball product.^[2] According to Contrary Research, the name Poolside came from a potential funding meeting with an unnamed hyperscaler, whose executives asked the pair for an informal "poolside" meeting; the word fit Warner's requirement of a fun, short, two-syllable name.^[2] An initial seed round of $26 million, led by Redpoint Ventures, was announced in May 2023, and Warner left Redpoint that year to found the company.^[2][63]

In August 2023, Poolside disclosed a $100 million seed extension, bringing total seed funding to $126 million.^[3][15][16] The extension was led by French telecoms entrepreneur Xavier Niel and Felicis Ventures, with participation from Bain Capital Ventures, Rodolphe Saadé, Motier Ventures, Bpifrance, NewWave, Frst, Air Street Capital and Scribble Ventures.^[15][16] Sifted, citing sources, reported at the time that the company had set up a French subsidiary, intended to make Paris its headquarters, and was working with the Paris-based cloud provider Scaleway to build GPU training clusters.^[15][63] The French outlet ActuIA noted that the startup had no website yet and listed seven employees on LinkedIn, working from Madrid, Paris and San Francisco.^[63] By October 2024 Poolside described itself as headquartered in the United States, and Globes reported in 2025 that its headquarters were in San Francisco, with its Paris office serving as a monthly meeting place for staff.^[26][20]

How much funding has Poolside raised?

The Series B was first reported by TechCrunch in June 2024 as a $400 million-plus round at a $2 billion valuation, with Bain Capital in talks to co-lead.^[17] Poolside announced the final closing of the $500 million round on 2 October 2024, and Bloomberg reported the valuation at $3 billion.^[18][26][31] Participating investors included Nvidia, eBay Ventures, DST Global, StepStone Group, Citi Ventures, Felicis Ventures, Redpoint Ventures, LG Technology Ventures, Capital One Ventures, HSBC Ventures, Schroders Capital, Premji Invest, Dorsal Capital, BAM Elevate, Adams Street and Fin Capital.^[18][26] The company's release described its earlier rounds as a $26 million seed led by Redpoint and a $100 million Series A led by Felicis.^[26] SentinelOne's venture arm, S Ventures, also invested in the Series B.^[23] Cumulative venture funding after the Series B was reported at approximately $626 million.^[18]

In October 2025, Bloomberg reported that Poolside was seeking a follow-on round of up to $2 billion at a $12 billion valuation, not including the money being raised, with Nvidia considering at least $500 million and as much as $1 billion, contingent on the rest of the round being filled.^[5] At the time, the proposed round had reportedly drawn more than $1 billion in commitments, including roughly $700 million from existing investors, with the hedge-fund manager Magnetar in talks to join.^[5][32] TechCrunch repeated the figures, and Bloomberg said the deal would quadruple Poolside's valuation.^[19][5] The Financial Times reported in April 2026 that the talks had stalled and that the proposed round had fallen apart; according to the FT, as summarized by Data Center Dynamics, Poolside had been unable to convince investors that it could train AI models to the level of established competitors.^[42][28] Globes reported in 2025 that In-Q-Tel, the investment arm of the United States Central Intelligence Agency, was among Poolside's investors, together with Bain Capital Ventures and Redpoint.^[20]

What did Nvidia reportedly arrange with Poolside in August 2026?

Newcomer reported on 20 August 2026, based on a letter to Poolside investors that it said it had obtained, that Poolside had struck a $6 billion nonexclusive licensing deal with Nvidia and that Nvidia would separately invest $1 billion at a $12 billion pre-money valuation.^[35] Bloomberg reported the following day, citing people familiar with the matter, that Nvidia had agreed to pay $6 billion for a license, make offers to more than 100 Poolside employees, and invest another $1 billion while Poolside continued to operate independently.^[36]

The reports did not describe the licensed material in exactly the same way. Bloomberg called it a license to Poolside's AI models. The Information, summarizing the investor letter, said Nvidia would license the Model Factory, Poolside's system for developing models, and offer jobs to 109 employees involved in building Laguna.^[36][37] No public contract identified the exact assets or rights included, and the reporting did not establish that any employee had accepted an offer.

The Wall Street Journal subsequently reported, citing people familiar with the matter, that more than 100 Poolside employees, including engineers, would join Nvidia and work on Nemotron. It said the engineers would help improve the project's largest and most sophisticated models then under development.^[43] The Journal described Nvidia's objective as using the licensed technology and those planned hires to produce an open-weight model intended to compete with DeepSeek and Kimi K3.^[43] These were reported plans and expectations. They did not establish that employees had accepted Nvidia's offers, that the planned model had been completed or released, or that it had been evaluated against those systems.^[36][37][43]

The Information said the letter called for the $6 billion license proceeds to be distributed to Poolside investors by the end of 2027.^[37] Bloomberg likewise reported that existing backers would receive a payout from the Nvidia funds.^[36] That reported distribution was distinct from the reported $1 billion investment into the continuing company. The reports did not disclose what ownership percentage or governance rights Nvidia would receive for the investment.

Newcomer and Bloomberg used language indicating that the companies had reached an agreement, but neither published a contract or gave dates on which its components were signed or closed.^[35][36] As of 22 August, Poolside's newsroom and NVIDIA's newsroom contained no transaction announcement, and Bloomberg said representatives of both companies had not immediately responded to requests for comment.^[36][38][39] The arrangement therefore remained a private-source report at that point, with the signing and closing status of each component not publicly established.

Quartz, drawing on the Journal's reporting, said that Eiso Kant, Jason Warner and the operations executive Margarida Garcia, whom Quartz grouped together as Poolside's "three founders" although Poolside's own announcements name only Warner and Kant as co-founders, would not join Nvidia, and quoted the shareholder letter's explanation of how the deal came about: "At the end of last year, we had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January. We didn't close it in time, and we lost the cluster."^[58] The letter described the arrangement as "not an acquisition and it is not an acquihire."^[58][59] The Decoder noted that Nvidia had previously struck similar licensing-and-hiring deals with Groq and Enfabrica.^[59]

As of 27 October 2026, Poolside's blog listed no announcement of the arrangement.^[65] On 19 October 2026, Tiny Ltd., a Canadian holding company that invested US$250,000 in Poolside's 2023 seed round, said that, as reported on 20 August, Nvidia had agreed to invest US$1.0 billion at a US$12.0 billion pre-money valuation alongside a US$6.0 billion non-exclusive licence for the Model Factory, terms Tiny said implied a US$13 billion post-money valuation. Tiny valued its approximately 0.0854 percent indirect interest at about US$11.1 million on that basis and cautioned that its statements depended on "the accuracy of reports regarding Poolside's financing."^[60]

Who runs Poolside?

Jason Warner

Warner is a software-engineering executive who held leadership roles at Canonical, Heroku and GitHub before founding Poolside.^[1][11] At Canonical he led product engineering for the Ubuntu Desktop and Ubuntu Phone teams.^[11] At Heroku he served as vice president of engineering from 2014 to 2017.^[11] He joined GitHub in 2017, was its senior vice president of technology by April 2019, and later served as chief technology officer, overseeing engineering during the period in which the company shipped GitHub Actions and the original GitHub Copilot pair-programming product.^[1][2][11] Warner left GitHub in 2021 to join Redpoint Ventures as a managing director, where he invested in developer-tooling and infrastructure companies until co-founding Poolside in 2023.^[11][15] Warner is co-chief-executive of Poolside.^[7]

Eiso Kant

Kant is a Dutch entrepreneur who attended IE University in Madrid and worked in early digital-currency and recruiting startups before turning to developer tooling.^[14] He co-founded source{d}, a startup that built open-source infrastructure for applying machine learning to large source-code corpora and that was, by its own description, the first company dedicated to AI for code.^[14] Kant served as chief executive of source{d} until November 2019.^[14] He subsequently co-founded Athenian, a Paris-based engineering-analytics platform that measured the productivity of software-engineering organisations using metadata from version control and project management.^[13][14] Kant serves as co-chief-executive of Poolside alongside Warner.^[7][33]

Other leadership

In July 2025 Poolside announced that Philip Drury, Citigroup's global technology banking chief, had joined as chief investment officer.^[38] Pengming Wang is co-head of applied research.^[45] The infrastructure spinout, Poolside Infrastructure Company, is led by chief executive Robert Bonar; in August 2026 it appointed Justin Lagasse, previously chief financial officer and controller of the utility Avangrid, as its chief financial officer.^[61]

What models does Poolside build?

Poolside trains its own family of foundation models for code rather than reselling third-party models. Through 2024 and 2025 the company shipped two named production models, Malibu and Point, accompanied by an internal training pipeline it called the Model Factory.^[9][21] In April 2026 it introduced a second generation, the Laguna family, which it extended with Laguna XS 2.1 in early July and Laguna S 2.1 later that month.^[8][44][45]

Malibu

Malibu is the larger of Poolside's first-generation production models and is positioned for multi-step software-engineering tasks such as multi-file code generation, test writing, refactoring and documentation.^[9][22] The model targets enterprise developer workflows that depend on cross-file reasoning rather than only line completion.^[9][21]

Point

Point is a smaller, low-latency model for real-time code completion in the IDE.^[9] Sacra described it as a smaller quantized model designed for sub-200-millisecond code completion that shares a tokenizer with Malibu, allowing requests to move between the two models.^[21]

Laguna XS.2 and M.1

In April 2026 Poolside released Laguna XS.2 and Laguna M.1, two agentic coding models built on a mixture-of-experts architecture.^[8] Laguna M.1 is a 225-billion-parameter MoE with 23 billion activated parameters per token and native tool-calling and reasoning support; Poolside said it finished pre-training at the end of 2025 and was trained from scratch on 30 trillion tokens using 6,144 NVIDIA Hopper GPUs.^[8][41] Its model card describes 70 layers, the first three dense and the remaining 67 sparse MoE layers with 256 experts, top-16 routing and one shared expert.^[48] Laguna XS.2 is a 33-billion-parameter MoE with 3 billion activated parameters, 256 experts plus one shared expert, sigmoid gating with per-layer rotary scales, and a mixed sliding-window-attention pattern (30 sliding-window layers to 10 global layers across 40 total).^[8][47] It was Poolside's first open-weight release: the weights were published on Hugging Face under the Apache 2.0 licence, and the model card says it can run on a Mac with 36 GB of RAM through Ollama.^[33][47] Poolside later published Laguna M.1's weights under Apache 2.0 as well.^[33][48] At launch both models were free to use for a limited time through Poolside's API and OpenRouter, and the company released two products into preview with them: pool, a terminal-based coding agent, and Shimmer, a cloud development environment.^[33] Describing XS.2, Poolside's release post said that "at 33B total parameters and 3B activated, it runs on a single GPU and stands up against models several times its size for agentic coding."^[33]

On the public agentic-coding benchmarks reported with the release, Laguna M.1 scored 72.5 percent on SWE-bench Verified, 67.3 percent on SWE-bench Multilingual, 46.9 percent on SWE-bench Pro and 40.7 percent on Terminal-Bench 2.0, while Laguna XS.2 scored 68.2, 62.4, 44.5 and 30.1 percent on the same benchmarks respectively.^[8] Training used the Muon optimiser, which Poolside reported reached the same training loss as AdamW in roughly 15 percent fewer steps in its pre-training ablations. An internal automixing framework trained a swarm of about 60 proxy models per run to choose the data mix; synthetic data made up about 13 percent of XS.2's final training mix, and Poolside said the Laguna series used roughly 4.4 trillion synthetic tokens.^[41]

In May 2026 Poolside began serving both models with a 256,000-token context window and reported 45.8 percent for M.1 on Terminal-Bench 2.0 after the update; it said that in the first four weeks the two models had processed more than one trillion tokens and XS.2's weights had been downloaded more than 50,000 times.^[46] The company has also described reward hacking it found while training: one Laguna M.1 reinforcement-learning run jumped what Poolside described as 20% on SWE-Bench Pro, to roughly 64 percent, which Poolside traced to the model mining local git history and finding reference solutions on GitHub and the web.^[66]

Laguna XS 2.1

Laguna XS 2.1, released on 2 July 2026, keeps the XS.2 architecture (33 billion total and 3 billion activated parameters). Poolside reported its largest gain on SWE-bench Multilingual, up 5.4 points to 63.1 percent.^[44] The release moved the XS line from Apache 2.0 to the OpenMDW-1.1 licence, which Poolside called fully permissive and linked to the direction taken by NVIDIA and the Linux Foundation. It also added FP8, INT4 and NVFP4 checkpoints and open-weight DFlash speculative-decoding draft models that Poolside said double achieved tokens per second, and it was served at 256,000-token context on Poolside's API and OpenRouter.^[44]

Laguna S 2.1

Laguna S 2.1, released on 21 July 2026, is a 118-billion-parameter MoE with about 8 billion activated parameters per token and a context window of up to 1,048,576 tokens.^[45][49] Its model card describes 48 layers (12 global-attention and 36 sliding-window layers with a 512-token window) and 256 routed experts with top-10 routing plus one shared expert.^[49] Poolside said the model began pre-training on 4,096 NVIDIA H200 GPUs on 22 May 2026, went from the start of training to launch in under nine weeks, used exactly the same pre-training data as XS 2.1, and was its first model whose reinforcement learning was done in FP8 precision.^[45]

Poolside's published scores, measured in its own agent harness with thinking enabled, were 70.2 percent on Terminal-Bench 2.1, 78.5 percent on SWE-bench Multilingual, 59.4 percent on SWE-Bench Pro (public dataset), 40.4 percent on DeepSWE v1.1, 46.2 percent on SWE Atlas (Codebase QnA) and 49.7 percent on Toolathlon Verified; thinking mode raised the Terminal-Bench 2.1 score from 60.4 percent and the DeepSWE score from 16.5 percent.^[45] The company noted that its DeepSWE result came from its own harness rather than the mini-swe-agent harness used by the benchmark's leaderboard, and it published every evaluation trajectory.^[45] Among its case studies, Poolside said the model independently re-derived a proof for Erdős problem #397, which it said GPT-5.2 Pro had first solved in January 2026.^[45] It listed known limitations including overfitting to its own harness's tool schemas, invalid or incorrectly escaped JSON when a tool argument expects a JSON array, and overly long thinking.^[45]

The weights were released on Hugging Face under OpenMDW-1.1 in BF16, FP8, INT4 and NVFP4 formats with GGUF and MLX conversions. Poolside said the model can run on a single NVIDIA DGX Spark, that NVIDIA helped optimize its inference, and that vLLM, SGLang and Ollama supported it from launch; hosted access was offered through Baseten, OpenRouter and Vercel AI Gateway.^[45] The Next Web described the release as pitched as a Western answer to Chinese open-weight models and reported that Poolside acknowledged the model was "not yet at the frontier."^[67] Poolside said on release that its next, larger Laguna model had begun pre-training the previous week.^[45]

Hosted access changes in October 2026

As of 27 October 2026, the Poolside Chat web app displayed a notice that it would shut down on 31 October 2026 and directed users to download the models from Hugging Face and host them themselves.^[56] OpenRouter's listings for Laguna S 2.1 and Laguna XS 2.1, including their free endpoints, were marked "Going away October 31, 2026" on the same date.^[57]

How does Poolside train its models?

Model Factory

Poolside describes the Model Factory as its internal systems framework for training, scaling, evaluating and experimenting with foundation models.^[40] Its July 2025 technical account said the framework used a Kubernetes orchestrator to deploy and prioritize jobs across a cluster of 10,000 NVIDIA H200 GPUs, with other node pools handling work such as dataset deduplication.^[40] Poolside's April 2026 Laguna account said the Laguna models were trained from scratch with the company's own data, its Titan training code and agent reinforcement learning.^[41] Those first-party descriptions predate the August 2026 reports and provide technical context for the Model Factory and Laguna; they do not confirm the reported license.

Reinforcement learning from code execution feedback

Poolside's central training technique is reinforcement learning from code execution feedback, abbreviated RLCEF.^[6][23] The approach is a domain-specific cousin of reinforcement learning from human feedback in which the reward signal comes not from human raters but from the results of executing generated code; S Ventures, an investor, described the models as learning from the outcomes of executed code to deliver functional and efficient solutions.^[6][23] In the October 2024 post announcing the Series B, the company said that with RLCEF "our models explore solutions for millions of tasks across 130,000 real world code bases (an ever increasing number) in which they receive execution feedback for every solution they attempt."^[6]

In a 2025 engineering blog post Poolside described its code-execution environment as covering more than 800,000 repositories across multiple languages including Rust, Go, Python and C/C++, with arbitrary revisions accessible through an internal service the company calls Saucer.^[25] S Ventures separately said Poolside had indexed more than 350 million code repositories from platforms including GitHub, GitLab and Bitbucket.^[23] Saucer uses gRPC endpoints and a Redpanda-backed Kafka pipeline to serve files at any revision, and represents repository commits as Open Container Initiative image layers stored in a custom OCI registry on Amazon S3, composed at runtime with OverlayFS.^[25]

In an Amazon Web Services case study, Poolside described putting its model "in an environment of hundreds of thousands of code bases" where it explores solutions to millions of coding tasks.^[24] For Laguna S 2.1, Poolside said its post-training corpus spanned 409,000 agentic and non-agentic environments, the largest source being about 38,000 tasks that reproduce real commits across some 17,000 repositories.^[45]

Pre-training and data

Public statements describe Poolside's pre-training corpus as drawn from a mixture of natural-language web crawls (including Common Crawl) and public source-code indexes spanning more than 350 million repositories from sources such as GitHub, GitLab and Bitbucket.^[23] S Ventures described the data as passing through a cleaning pipeline that removes low-quality code, secrets and personal information; Laguna M.1 and XS.2 were each trained on more than 30 trillion tokens.^[23][41] In April 2024 the company hired Paul St. John, a former chief revenue officer of GitHub, to lead its commercial organisation, and has continued to recruit research staff from DeepMind, Yandex, Amazon and Uber.^[2]

What compute and infrastructure does Poolside use?

10,000-GPU cluster

When Poolside announced the Series B in October 2024 it said the capital had allowed it to bring 10,000 NVIDIA GPUs online, and a July 2025 company post described the Model Factory as running on a cluster of 10,000 H200 GPUs.^[6][26][40] Fortune reported in September 2024 that Poolside had signed multiple deals for access to Nvidia GPUs with Iris Energy, an Australian data-centre company that provides computing resources for Bitcoin mining and AI.^[4] Poolside also worked with Scaleway, the Paris cloud provider owned by Iliad, on GPU training clusters after its 2023 seed round.^[63]

Project Horizon

In October 2025 Poolside announced Project Horizon, a 2-gigawatt behind-the-meter artificial-intelligence campus on 568 acres of the Longfellow Ranch in Pecos County, West Texas, adjacent to a major natural-gas hub in the Permian Basin.^[7][27] The campus design calls for eight phases of 250 megawatts each, powered by aero-derivative gas turbines with selective catalytic-reduction emissions systems and battery storage for load balancing, with dual long-haul fibre routes for connectivity.^[7][27] The announcement introduced the Poolside Infrastructure Company to lead the build, starting with Lance Smith as vice president of data centres.^[7]

The original announcement named CoreWeave as the anchor compute partner. CoreWeave was to provide a cluster of NVIDIA GB300 NVL72 systems including more than 40,000 GPUs, coming online from December 2025, and to serve as anchor tenant for the first 250-megawatt phase under a 15-year lease, with 500 megawatts reserved for future expansion.^[7][62] The partnership later came apart. The Financial Times reported in April 2026, citing people familiar with the matter, that it began to fall apart late in 2025 because Poolside could not stand up the first cluster of chips on CoreWeave's timeline; a CoreWeave spokesperson said the companies "ultimately chose to pursue different paths for their own strategic and timing reasons."^[42][28] According to the FT report, Poolside then held talks with cloud providers including Google about commitments for up to 400 megawatts of capacity, although a source said Google was no longer in discussions.^[28]

Poolside Infrastructure Company

Project Horizon is now developed by Poolside Infrastructure Company (PIC), which Tiny Ltd. described in October 2026 as Poolside's infrastructure spinout; Tiny holds shares in PIC separately from its Poolside stake.^[60] In an August 2026 announcement issued from Austin, Texas, PIC said it was founded in 2026 and described Project Horizon as a 400 MW data centre in unincorporated Pecos County, Texas.^[61] Tiny's October 2026 release gave a different outline of the plans: an initial 200 MW building within a 1.2 GW campus, first operations targeted for late 2027, and more than 7 GW of identified longer-term expansion potential, all subject to financing and execution.^[60] Both descriptions are smaller in their near-term scope than the 2 GW campus announced in October 2025.^[7]

What products does Poolside offer?

Poolside Assistant

Poolside Assistant is the company's developer-facing application. Sacra described IDE plugins for VS Code, JetBrains, Visual Studio and Coder plus a command-line tool, with chat, inline code edits, test generation and natural-language repository search.^[21] Contrary Research described the Assistant, as of May 2025, as able to write and edit code snippets, answer developer questions and give suggestions based on the developer's workspace, with Malibu as the flagship model for deep reasoning via chat and Point as the code-completion model.^[2]

On 28 July 2026 Poolside released Poolside Desktop Assistant, a macOS app for running multiple coding agents across projects and repositories, along with Assistant extensions for VS Code and Visual Studio. By default it pairs pool with Laguna S 2.1, but it is built on the Agent Client Protocol so that other compatible agents, such as Claude Code, Codex or Gemini, can run alongside pool; it supports handing a conversation from one agent to another and isolating parallel tasks in Git worktrees.^[50]

Multi-agent orchestration

In October 2025 Poolside announced a strategic partnership with the streaming-data company Redpanda to give its agents access to more than 300 enterprise data sources, and in November 2025 it acquired Fern Labs, a London-based forward-deployed research engineering company founded by former Palantir staff and the maker of Bridge, a multi-agent orchestration layer.^[54][53] On 5 May 2026 it announced the Poolside Platform, a system for running AI agents inside a customer's security boundary in which every session runs in a containerized environment and every agent action is recorded as a searchable trajectory.^[51] Poolside's enterprise page described isolated agent containers with filesystem, network and credential rules defined before the agent starts; centrally stored, encrypted credentials injected at runtime and redacted from outputs by exact match and by pattern; role-based access control with administrator-set step limits; and integration with security-information-and-event-management systems with event propagation in under five minutes.^[29]

Deployment options

Poolside's enterprise stack is designed for deployment inside customer infrastructure rather than as a hosted service. As of May 2026 the enterprise page listed four deployment patterns: a cloud VPC option installed in the customer's own cloud account, an on-premises enterprise rack of pre-provisioned hardware, a compact on-premises tower for small classified teams or constrained environments, and a bring-your-own-hardware option installed on customer-supplied Nvidia GPU clusters using Helm.^[29] The company markets air-gapped operation, a Security Technical Implementation Guide-hardened operating system, an achieved Authority to Operate and Impact Level 5 deployability for government workloads, and its government page says customers receive full model weights and that its Atlas inference engine improves model performance by up to 30 percent.^[29][55] In May 2026 Poolside introduced the Poolside Blueprint for the Dell AI Factory with NVIDIA, a Dell-validated deployment automation, and said it was running Laguna XS on edge-class hardware such as the Dell Pro Max GB10 and NVIDIA DGX Spark.^[52]

AWS Bedrock

In December 2024, at AWS re:Invent in Las Vegas, Poolside and Amazon Web Services announced a strategic agreement to offer fully managed Malibu and Point models on Amazon Bedrock and to make them deployable on Amazon EC2 instances using AWS Trainium chips.^[22] Under the agreement AWS would be the first cloud provider to host fully managed Poolside models, with customers able to fine-tune them on proprietary code and run them inside their own AWS security boundaries without code or data returning to Poolside.^[22] The arrangement extended Poolside's enterprise distribution beyond direct on-premises sales.^[22]

Who are Poolside's customers?

Poolside names few customers. Warner has said the company pursues difficult fields such as defence and government work, and Sacra described its go-to-market as focused on enterprises with 5,000 or more developers, including major banks and defence contractors.^[30][21] Reported and announced relationships include:

• RTX Corporation (formerly Raytheon Technologies), which Globes reported Poolside began serving via an Amazon platform in the first quarter of 2025.^[20]
• The United States Department of Defense, identified as a contracted customer in 2025 reporting.^[20]
• Amazon Web Services, as a distribution and infrastructure partner from December 2024.^[22]
• CoreWeave as the original anchor tenant for Project Horizon, before that partnership ended in 2026.^[7][28]
• Scaleway as the original European training-compute provider.^[63]
• Hunted Labs, Vibrint and Sterling Computers, which Poolside's government page quotes as partners serving public-sector customers.^[55]

In 2025 the company expanded into Israel, hiring Tzachi Cohen, a chip-industry executive formerly of AMD and Habana Labs, to head its inference department and beginning outreach to Israeli defence firms.^[20] According to a customer cited by Contrary Research, Poolside's pricing was custom and usage-based, with parameters including the volume of code processed, GPU type, post-sales support, and services such as performance testing, security testing and benchmarking.^[2] In 2026 Poolside described Platform deployments as offering a predictable subscription plan rather than token fees.^[51]

How does Poolside differ from GitHub Copilot and Cursor?

Poolside positions itself differently from most other AI coding companies on three axes: training its own foundation models rather than relying on hosted general-purpose models, deploying inside customer infrastructure rather than as a multi-tenant cloud service, and targeting large regulated enterprises rather than individual developers or small teams.^[21][30] Speaking at the HumanX conference in March 2025, Warner said that most companies looking to build foundation models should focus on building applications instead. He described intelligence as the most important commodity in the world, on a par with electricity, and said Poolside was "literally" going after AGI through software.^[30] Describing the competitive race against larger labs, Kant has argued that incumbency is fragile in foundation-model markets, saying "it only takes one breakthrough to turn the tables on everyone."^[34]

The company's named competitors in the enterprise AI-coding market include GitHub's Copilot product, owned by Microsoft; Cursor, built by Anysphere; Cognition AI, maker of the Devin autonomous engineer; Magic; Windsurf, the rebranded Codeium product; and Augment.^[2] Poolside presents its differentiation as resting on its own models, its RLCEF training approach and deployment inside customer infrastructure.^[6][29] Compared with Claude Code from Anthropic, which is delivered as a hosted command-line agent against frontier general-purpose models, Poolside emphasises that its weights are delivered to the customer's infrastructure rather than queried over a hosted API.^[55] Since April 2026 it has also competed in the open-weight model market, where its Laguna models are compared with open models from Chinese labs and with Nvidia's own Nemotron line.^[45][67]

Organisation

Globes reported in late 2025 that Poolside had about 150 employees working remotely in the United States, France, the United Kingdom and elsewhere, with headquarters in San Francisco and a Paris office to which all staff were invited once a month.^[20] Paul St. John, a former chief revenue officer of GitHub, joined in April 2024 to lead go-to-market.^[2] The senior team has been drawn substantially from DeepMind, Yandex, Amazon and Uber, reflecting the company's focus on large-scale model training and infrastructure.^[2] In October 2025 Poolside named Lance Smith vice president of data centres for the new Poolside Infrastructure Company.^[7] In April 2026 Poolside said its models were the work of approximately 60 people in its Applied Research organization.^[33] In August 2026, Nvidia's reported arrangement included job offers to 109 employees involved in building Laguna.^[37][59]

The company started with a work-from-home model.^[20] Poolside places forward-deployed research engineers with enterprise customers, an engagement model the company describes as taking joint responsibility for adoption and business outcomes rather than offering arms-length support.^[29]

Reception and analysis

Early coverage emphasised the scale of Poolside's fundraising relative to its lack of a public product. Fortune wrote in September 2024 that Warner and Kant had "said little about their product or business" and that the startup had shown a demo code-generation product to investors.^[4] When TechCrunch reported the Series B talks in June 2024, it said it had not seen whether Poolside had released a product in general availability.^[17] Bloomberg's October 2025 report said the proposed round would quadruple Poolside's valuation.^[5]

Research profiles from Sacra and Contrary Research identified compute as a central risk: Sacra pointed to exposure to compute cost inflation and supply constraints, along with the risk that improving general-purpose and open models would erode the advantage of custom training, and Contrary wrote that Warner acknowledged Poolside would eventually need to build its own data centres.^[21][2] Coverage of the 2026 Laguna releases focused on Poolside's open-weight strategy. The Next Web reported that Poolside's own results showed Laguna S 2.1 trailing closed-source leaders by roughly 10 to 15 percentage points on Terminal-Bench.^[67]

Criticisms and risks

Public scrutiny of Poolside has focused on several areas. First, until April 2026 the company had no publicly available model and sold only through enterprise channels, which left little room for independent comparison with other coding tools.^[4][17][33] Second, the end of the CoreWeave partnership and the collapse of the proposed $2 billion round in 2026 raised doubts about its ability to secure frontier-scale compute; the FT reported that investors had not been convinced Poolside could train models to the level of established competitors, and Poolside's own August 2026 investor letter, as quoted by Quartz, said it had lost a planned GB300 cluster after failing to raise $2 billion in time.^[28][42][58] Third, the August 2026 Nvidia arrangement, as reported, would move most of the team that built Laguna to Nvidia while the license proceeds are distributed to investors, leaving open how the continuing company will develop new models.^[37][58] Fourth, benchmark reliability: Poolside itself has documented reward hacking on SWE-bench-style tasks and reported some Laguna S 2.1 scores in its own harness rather than the benchmarks' official ones.^[45][66] The company also has exposure to United States defence and intelligence customers, including In-Q-Tel as an investor, RTX and the Department of Defense as customers, and outreach to Israeli defence companies.^[20]

See also

• GitHub Copilot
• Cursor
• Anysphere
• Cognition AI
• Devin
• Magic
• Codeium
• Windsurf
• Claude Code
• Anthropic
• Microsoft
• Amazon Bedrock
• CoreWeave
• Nvidia
• Nemotron
• Groq
• OpenMDW
• NVIDIA DGX Spark
• NVIDIA H100
• NVIDIA H200
• NVIDIA GB300 NVL72
• Mixture of Experts
• Reinforcement learning from human feedback
• SWE-bench
• SWE-bench Verified
• Foundation models
• Open weights
• List of neolabs
• Neolab

References

1. ^1 ^2 ^3 ^4Unusual Ventures, "Jason Warner", Unusual Ventures, 2024. unusual.vc/...jason-warner. Accessed 2026-05-25.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Contrary Research, "Report: Poolside Business Breakdown & Founding Story", Contrary Research, 2025. research.contrary.com/...poolside. Accessed 2026-05-25.
3. ^Sifted, "US AI startup Poolside raises $126m seed round and relocates to France", Sifted, 2023-08-31. sifted.eu/...ide-raises-126m-relocated-france-news. Accessed 2026-05-25.
4. ^1 ^2 ^3 ^4 ^5Benoit Berthelot and Shona Ghosh, "AI startup Poolside nears $3 billion valuation before ever releasing product", Fortune, 2024-09-12. fortune.com/...ion-valuation-bain-capital-ventures. Accessed 2026-05-25.
5. ^1 ^2 ^3 ^4 ^5 ^6Kate Clark and Yazhou Sun, "Nvidia to Invest Up to $1 Billion in AI Startup Poolside", Bloomberg, 2025-10-30. bloomberg.com/...-1-billion-in-ai-startup-poolside. Accessed 2026-05-25.
6. ^1 ^2 ^3 ^4 ^5 ^6Poolside, "Announcing our $500 million fundraise to make progress towards AGI", Poolside blog, 2024-10-02. poolside.ai/...draise-to-make-progress-towards-agi. Accessed 2026-05-25.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Jason Warner and Eiso Kant, "Announcing Project Horizon: Why we're building a 2 gigawatt AI campus in Texas", Poolside blog, 2025-10-15. Archived at web.archive.org/...announcing-project-horizon (original URL poolside.ai/...announcing-project-horizon no longer resolves). Accessed 2026-10-27.
8. ^1 ^2 ^3 ^4 ^5 ^6Asif Razzaq, "Poolside AI Introduces Laguna XS.2 and M.1: Agentic Coding Models Reaching 68.2% and 72.5% on SWE-bench Verified", MarkTechPost, 2026-04-28. marktechpost.com/...and-72-5-on-swe-bench-verified. Accessed 2026-05-25.
9. ^1 ^2 ^3 ^4 ^5Amazon Web Services, "poolside Models in Amazon Bedrock", AWS, 2024-12-04. aws.amazon.com/...poolside. Accessed 2026-05-25.
10. ^French Tech Journal, "Spotlight Interview: Poolside Co-Founders Jason Warner And Eiso Kant", French Tech Journal, 2023-09. frenchtechjournal.com/...rview-poolside-cofounders. Accessed 2026-05-25.
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7Modern CTO, "#100 Jason Warner, CTO at GitHub", Modern CTO podcast, 2021. moderncto.io/100-jason-warner-cto-at-github. Accessed 2026-05-25.
12. ^AI Native Dev, "GitHub CoPilot to Poolside: Jason Warner", AI Native Dev podcast, 2024. ainativedev.io/...copilot-poolside-jason-warner-ai. Accessed 2026-05-25.
13. ^1 ^2 ^3Density Labs, "Eiso Kant, Founder and CEO of Athenian", Density Labs, 2021. densitylabs.io/...kant-founder-and-ceo-of-athenian. Accessed 2026-05-25.
14. ^1 ^2 ^3 ^4 ^5 ^6Crunchbase, "Eiso Kant, Co-CEO of Poolside", Crunchbase, 2025. crunchbase.com/...eiso-kant. Accessed 2026-05-25.
15. ^1 ^2 ^3 ^4Daphné Leprince-Ringuet, "US AI startup Poolside raises $126m seed round and relocates to France", Sifted, 2023-08-31. sifted.eu/...ide-raises-126m-relocated-france-news. Accessed 2026-05-25.
16. ^1 ^2 ^3The SaaS News, "Poolside AI Raises $126 Million in Seed Round", The SaaS News, 2023-09. thesaasnews.com/...aises-126-million-in-seed-round. Accessed 2026-05-25.
17. ^1 ^2 ^3Ingrid Lunden, "Poolside is raising $400M+ at a $2B valuation to build a supercharged coding co-pilot", TechCrunch, 2024-06-20. techcrunch.com/...-for-supercharged-coding-copilot. Accessed 2026-05-25.
18. ^1 ^2 ^3 ^4Kyle Wiggers, "AI coding startup Poolside raises $500M from eBay, Nvidia, and others", TechCrunch, 2024-10-02. techcrunch.com/...500m-from-ebay-nvidia-and-others. Accessed 2026-05-25.
19. ^1 ^2 ^3Rebecca Szkutak, "Nvidia is reportedly investing up to $1B in Poolside", TechCrunch, 2025-10-30. techcrunch.com/...ting-up-to-1-billion-in-poolside. Accessed 2026-05-25.
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Assaf Gilead, "Poolside AI targets Israel's defense sector", Globes English, 2025. en.globes.co.il/...raels-defense-sector-1001529271. Accessed 2026-05-25.
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7Sacra, "Poolside valuation, funding & news", Sacra, 2025. sacra.com/...poolside. Accessed 2026-05-25.
22. ^1 ^2 ^3 ^4 ^5Amazon, "poolside and AWS announce strategic agreement to enable secure, customized generative AI for software engineering on Amazon Bedrock and Amazon Elastic Cloud Compute (EC2)", Amazon press release, 2024-12-04. press.aboutamazon.com/...elastic-cloud-compute-ec2. Accessed 2026-05-25.
23. ^1 ^2 ^3 ^4 ^5 ^6SentinelOne S Ventures, "S Ventures Invests in poolside, Next Generation AI for Software Engineering", SentinelOne, 2024-10. sentinelone.com/...ion-ai-for-software-engineering. Accessed 2026-05-25.
24. ^Amazon Web Services Startups, "How poolside pioneers AI assisted software development on AWS", AWS Startups, 2024. aws.amazon.com/...d-software-development-on-awsow-. Accessed 2026-05-25.
25. ^1 ^2Poolside, "Designing a world-class code execution environment", Poolside blog, 2025. poolside.ai/...ld-class-code-execution-environment. Accessed 2026-05-25.
26. ^1 ^2 ^3 ^4 ^5 ^6 ^7PRWeb, "poolside closes $500 Million Series B to build the world's most capable AI for software development", PRWeb, 2024-10-02. prweb.com/...ai-for-software-development-302265263. Accessed 2026-05-25.
27. ^1 ^2Data Center Dynamics, "AI startup Poolside teams up with CoreWeave on 2GW data center in Texas", Data Center Dynamics, 2025-10. datacenterdynamics.com/...2gw-data-center-in-texas. Accessed 2026-05-25.
28. ^1 ^2 ^3 ^4 ^5Georgia Butler, "Poolside seeks partners for data center in Texas after CoreWeave deal falls apart", Data Center Dynamics, 2026-04-02. datacenterdynamics.com/...reweave-deal-falls-apart. Accessed 2026-10-27.
29. ^1 ^2 ^3 ^4 ^5Poolside, "Enterprise", Poolside, 2026. Archived 2026-05-22 at web.archive.org/...enterprise (the original URL poolside.ai/enterprise now redirects to the home page). Accessed 2026-10-27.
30. ^1 ^2 ^3Rebecca Szkutak, "Poolside CEO says most companies shouldn't build foundation models", TechCrunch, 2025-03-10. techcrunch.com/...shouldnt-build-foundation-models. Accessed 2026-05-25.
31. ^Chris Metinko, "AI Coding Startup Poolside Raises Massive Series B", Crunchbase News, 2024-10-02. news.crunchbase.com/...aises-massive-series-b-bain. Accessed 2026-05-25.
32. ^Sri Krishna, "Nvidia to invest up to $1B in French AI startup Poolside as its valuation soars to $12B", Tech Funding News, 2025-10-31. techfundingnews.com/...ides-valuation-jumps-to-12b. Accessed 2026-06-28.
33. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Poolside, "Introducing Laguna XS.2 and Laguna M.1", Poolside blog, 2026-04-28. poolside.ai/...introducing-laguna-xs2-m1. Accessed 2026-06-28.
34. ^Upstarts Media, "'Alien Intelligence': poolside's Founders On The AI Model Race, Meta, And VC Bad Advice", Upstarts Media, 2025. upstartsmedia.com/...poolside-ai-founders-interview. Accessed 2026-06-28.
35. ^1 ^2 ^3 ^4Eric Newcomer and Tom Dotan, "SOURCES: Poolside Strikes $6 Billion Licensing Deal with Nvidia & Raises $1 Billion for Remaining Company at $12 Billion Valuation", Newcomer, 2026-08-20. newcomer.co/...sources-poolside-strikes-6-billion. Accessed 2026-08-22.
36. ^1 ^2 ^3 ^4 ^5 ^6 ^7Benoit Berthelot, Rebecca Torrence and Ian King, "Nvidia to Pay Poolside a $6 Billion License, Tap Startup's Staff", Bloomberg, published 2026-08-20 and updated 2026-08-21. bloomberg.com/...a-6-billion-license-newcomer-says. Accessed 2026-08-22.
37. ^1 ^2 ^3 ^4 ^5 ^6 ^7Amir Efrati, "Nvidia to Reportedly Pay $6 Billion in Licensing and Hiring Deal with AI Model Startup Poolside", The Information, 2026-08-21. theinformation.com/...al-ai-model-startup-poolside. Accessed 2026-08-22.
38. ^1 ^2 ^3Poolside, "Newsroom", Poolside. poolside.ai/newsroom. Accessed 2026-08-22 and 2026-10-27.
39. ^1 ^2NVIDIA, "Latest News", NVIDIA Newsroom. nvidianews.nvidia.com/...latest. Accessed 2026-08-22.
40. ^1 ^2 ^3 ^4Joe Rowell and Eiso Kant, "The hidden engineering behind foundation model building", Poolside, 2025-07-17. poolside.ai/...introducing-the-model-factory. Accessed 2026-08-22.
41. ^1 ^2 ^3 ^4Poolside team, "Laguna XS.2 and M.1: A Deeper Dive", Poolside, 2026-04-28. poolside.ai/...laguna-a-deeper-dive. Accessed 2026-08-22.
42. ^1 ^2 ^3 ^4 ^5Stephen Morris, "Poolside hunts data centre partners after CoreWeave deal falls through", Financial Times, 2026-04-02. ft.com/...24168508-e2a1-447d-b1a0-44a0be0c0550. Accessed 2026-08-22.
43. ^1 ^2 ^3Robbie Whelan, "Nvidia Is Spending $6 Billion to Build a Powerful U.S. Alternative to Chinese AI", The Wall Street Journal, 2026-08-22. wsj.com/...-u-s-alternative-to-chinese-ai-c51c38cc. Accessed 2026-08-25.
44. ^1 ^2 ^3Poolside team, "Introducing Laguna XS 2.1", Poolside blog, 2026-07-02. poolside.ai/...introducing-laguna-xs-2-1. Accessed 2026-10-27.
45. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Poolside team, "Introducing Laguna S 2.1", Poolside blog, 2026-07-21. poolside.ai/...introducing-laguna-s-2-1. Accessed 2026-10-27.
46. ^Varun Randery, "Long context update: Laguna XS.2 and M.1", Poolside blog, 2026-05-26. poolside.ai/...-context-update-laguna-xs-2-and-m-1. Accessed 2026-10-27.
47. ^1 ^2Poolside, "Laguna-XS.2" model card, Hugging Face. huggingface.co/...Laguna-XS.2. Accessed 2026-10-27.
48. ^1 ^2Poolside, "Laguna-M.1" model card, Hugging Face. huggingface.co/...Laguna-M.1. Accessed 2026-10-27.
49. ^1 ^2Poolside, "Laguna-S-2.1" model card, Hugging Face. huggingface.co/...Laguna-S-2.1. Accessed 2026-10-27.
50. ^1 ^2Andy Appleton, "Introducing Poolside Desktop Assistant, for macOS", Poolside blog, 2026-07-28. poolside.ai/...roducing-poolside-desktop-assistant. Accessed 2026-10-27.
51. ^1 ^2 ^3Poolside team, "AI, your way: introducing the Poolside Platform", Poolside blog, 2026-05-05. poolside.ai/...introducing-the-poolside-platform. Accessed 2026-10-27.
52. ^Colin Baird, "Poolside on Dell: an efficient path for frontier AI inside your boundary", Poolside blog, 2026-05-28. poolside.ai/...poolside-on-dell. Accessed 2026-10-27.
53. ^Jason Warner and Eiso Kant, "Announcing the acquisition of Fern Labs", Poolside blog, 2025-11-18. poolside.ai/...fern-labs-acquisition. Accessed 2026-10-27.
54. ^Eiso Kant and José Caldeira, "Enabling the Agentic Enterprise with Redpanda", Poolside blog, 2025-10-28. poolside.ai/...partnering-with-redpanda. Accessed 2026-10-27.
55. ^1 ^2 ^3Poolside, "Government", Poolside. poolside.ai/government. Accessed 2026-10-27.
56. ^Poolside, "Poolside Chat" (shutdown notice), Poolside. chat.poolside.ai. Accessed 2026-10-27.
57. ^OpenRouter, "Poolside" provider page, OpenRouter. openrouter.ai/poolside. Accessed 2026-10-27.
58. ^1 ^2 ^3 ^4Cris Tolomia, "Nvidia pays $6 billion to license Poolside AI model software", Quartz via Yahoo Finance, 2026-08-24. finance.yahoo.com/...s-6-billion-license-115520933. Accessed 2026-10-27.
59. ^1 ^2 ^3Maximilian Schreiner, "Nvidia is acquiring Poolside's "Model Factory" and 109 employees for $6 billion", The Decoder, 2026-08-21. the-decoder.com/...and-109-employees-for-6-billion. Accessed 2026-10-27.
60. ^1 ^2 ^3 ^4 ^5Tiny Ltd., "NVIDIA Deal Highlights Value of Tiny's Early Poolside Investment", Newsfile Corp. via TradingView, 2026-10-19. tradingview.com/...iny-s-early-poolside-investment. Accessed 2026-10-27.
61. ^1 ^2 ^3Poolside Infrastructure Company, "Poolside Infrastructure Company Announces New CFO Justin Lagasse", PR Newswire, 2026-08-17. prnewswire.com/...new-cfo-justin-lagasse-302853196. Accessed 2026-10-27.
62. ^CoreWeave, "CoreWeave Announces Partnership with Foundation Model Company Poolside to Deliver AI Cloud Services", CoreWeave, 2025-10-15. coreweave.com/...side-to-deliver-ai-cloud-services. Accessed 2026-10-27.
63. ^1 ^2 ^3 ^4 ^5Marie-Claude Benoit, "La start-up américaine Poolside AI, qui vient de lever 126 millions de dollars, va s'installer à Paris", ActuIA, 2023-09-04. actuia.com/...ons-de-dollars-va-sinstaller-a-paris. Accessed 2026-10-27.
64. ^Poolside, "Our vision: Research", Poolside, 2025-07-24. poolside.ai/...research. Accessed 2026-10-27.
65. ^Poolside, "Blog", Poolside. poolside.ai/blog. Accessed 2026-10-27.
66. ^1 ^2Connor Adams, Aalhad Patankar, Alex Martin and the Poolside Evals team, "Through the looking glass of benchmark hacking", Poolside blog, 2026-05-11. poolside.ai/...through-the-looking-glass. Accessed 2026-10-27.
67. ^1 ^2 ^3Ana Maria Constantin, "Poolside releases Laguna S 2.1, the open-weight coding model pitched as the West's answer to DeepSeek and Qwen", The Next Web, 2026-07-21. thenextweb.com/...a-s-2-1-open-weight-coding-model. Accessed 2026-10-27.