Poolside AI
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 4,056 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 4,056 words
Add missing citations, update stale details, or suggest a clearer explanation.
Poolside AI is an artificial intelligence company that builds foundation models and developer tools for software engineering, with a focus on enterprise and regulated-industry customers that deploy the system on premises or inside their own clouds. The company was founded in May 2023 by Jason Warner, the former chief technology officer of GitHub and a former Redpoint Ventures managing director, and Eiso Kant, a serial software-tooling founder previously known for source{d} and Athenian.[1][2] Headquartered in Paris with offices in the United States and the United Kingdom, Poolside raised a $500 million Series B led by Bain Capital Ventures in October 2024 at a $3 billion valuation, and in October 2025 entered a follow-on round valuing the company at roughly $12 billion, with Nvidia committing up to an additional $1 billion.[3][4][5] Poolside trains its models with a technique it calls reinforcement learning from code execution feedback (RLCEF) and operates a training cluster of roughly 10,000 NVIDIA H100-class GPUs, with plans for a multi-gigawatt data-center campus in West Texas.[6][7] The company's flagship offerings are the Malibu and Point foundation models and, in 2026, the Laguna model family, packaged as Poolside Assistant for developers and as an enterprise platform for fine-tuning and on-premises deployment.[8][9]
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 two formally incorporated the company in San Francisco in early 2023 under the working name Snowball, then renamed it Poolside after discovering a trademark conflict with Amazon's Snowball edge-computing line.[2] The final name was drawn from informal conversations with a hyperscaler executive who used "poolside" as shorthand for relaxed strategic discussions during fundraising; Warner has said the founders kept it because the word evoked the unhurried, generational view of artificial general intelligence they wanted to convey.[2] An initial seed round of $26 million was announced in May 2023, led by Redpoint Ventures, with Warner stepping down from his managing-director role to take the chief-executive position full time.[15]
In August 2023, Poolside disclosed a $100 million seed extension, bringing total seed funding to $126 million.[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] At the same time the company announced that it had established a French subsidiary and would move its headquarters to Paris, drawn by access to European AI talent, lower compute costs through partnerships with Paris-based cloud provider Scaleway, and the personal involvement of Niel as an investor.[15][16] The company had fewer than ten employees at that point and had not yet released a product.[15]
| Round | Date | Amount | Lead investor(s) | Post-money valuation |
|---|---|---|---|---|
| Seed | May 2023 | $26 million | Redpoint Ventures | Not disclosed |
| Seed extension | August 2023 | $100 million | Xavier Niel, Felicis Ventures | Roughly $526 million[17] |
| Series B | October 2024 | $500 million | Bain Capital Ventures | $3 billion[4][18] |
| Series C (announced) | October 2025 | Up to $2 billion (round in progress) | Nvidia (lead) | $12 billion[5][19] |
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] The round closed and was announced on 2 October 2024 at $500 million and a $3 billion post-money valuation, more than five times the valuation at the seed extension a year earlier.[4][18] 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 Partners and Fin Capital.[4][18] Cumulative venture funding after the Series B was reported at approximately $626 million.[18]
In October 2025, Bloomberg reported that Poolside was raising a follow-on round of up to $2 billion at a $12 billion valuation, with Nvidia committing at least $500 million and as much as $1 billion contingent on the rest of the round being filled.[5] TechCrunch and other outlets confirmed the figures and described the deal as a four-fold valuation increase in twelve months.[19] The In-Q-Tel investment arm of the United States Central Intelligence Agency was disclosed during the 2025 cycle as a Poolside backer alongside Bain, Redpoint and Nvidia.[20]
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 and became chief technology officer in 2019, overseeing engineering during the period in which the company shipped GitHub Actions and the original GitHub Copilot pair-programming product.[1][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]
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] In 2014 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]
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.[8]
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] Poolside has described Malibu as supporting context windows that exceed one million tokens, allowing it to ingest substantial portions of a customer codebase at a time.[21] The model targets enterprise developer workflows that depend on cross-file reasoning rather than only line completion.[9][21]
Point is a smaller, low-latency model optimised for in-IDE code completion. Poolside markets it as producing real-time suggestions with sub-200-millisecond first-token latency in customer environments and reports that it returns completions up to ten times faster than comparable competitor models.[9][21] Point supports a context window of more than one million tokens at inference time and is intended for use cases that require continuous suggestions as a developer types.[9][21]
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, a 128,000-token context window, and native tool-calling and reasoning support; it was trained from scratch on 30 trillion tokens using 6,144 NVIDIA Hopper GPUs.[8] Laguna XS.2 is a 33-billion-parameter MoE with 3 billion activated parameters, 256 experts plus one shared expert, a 131,072-token context window, 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] XS.2 uses FP8 KV-cache quantisation and was reported to run on a Mac with 36 GB of RAM through Ollama, with the base version released under the Apache 2.0 licence for fine-tuning.[8]
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; an internal tool called AutoMixer trained about 60 proxy models to choose the final data mix, of which roughly 13 percent of XS.2's 4.4-trillion-token mixture was synthetic.[8]
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 deterministic results of running generated code: whether it compiles, whether it passes unit tests, how efficient it is and whether it satisfies static security checks.[23][24]
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] The same post said Poolside indexes more than 350 million public repositories at the data-ingestion layer and that the execution environment can run roughly 10,000 code executions per minute against approximately one million container images hosted on Amazon Web Services.[23][24][25] 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]
Poolside describes the RLCEF reward loop as iterating across millions of tasks drawn from real codebases. Models are asked to make a change, the change is built and executed, and the resulting signals (build success, test pass-rate, lint results, performance measurements) are used to update the model.[23][24] The company markets RLCEF as a way to generate synthetic training data at scale without exposing customer code, since the rewards are computed against public or synthetic targets rather than private corpora.[6][23]
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] Poolside has said that it filters this corpus to remove low-quality code, secrets and personally identifiable information, and excludes content carrying restrictive licences from the training pipeline; Laguna M.1's training mix was reported at 30 trillion tokens.[8][23] 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]
When Poolside announced the Series B in October 2024 it said the round would allow it to scale a dedicated training cluster to 10,000 GPUs, which contemporaneous reporting identified as Nvidia H100-class accelerators.[6][18][26] Bain Capital Ventures, the lead investor, partnered with the company on compute procurement and was credited in coverage for helping arrange financing for hardware purchases.[18] An earlier multi-year GPU access agreement with Iris Energy, an Australian operator of data centres running Nvidia accelerators, was reported by Bloomberg in September 2024.[4][17] Poolside also built training capacity through Scaleway, the French cloud arm of Iliad, after relocating its headquarters to Paris in 2023.[15]
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] Poolside hired Lance Smith as vice president of data centres to lead an infrastructure subsidiary handling the build.[7]
The original announcement named CoreWeave as the anchor compute partner. CoreWeave was to deploy a cluster of more than 40,000 NVIDIA GB300 NVL72 systems beginning in December 2025, with CoreWeave serving as the lead tenant for the first 250-megawatt phase under a 15-year lease and 500 megawatts reserved for future expansion.[7][27] In early 2026 the two companies disclosed that they had mutually ended the Project Horizon partnership, citing differing strategic timing after CoreWeave secured an $8.5 billion debt facility for its own buildout; Poolside said it would seek new partners to advance the campus.[28]
Poolside Assistant is the company's developer-facing application. It includes IDE plugins for VS Code, Visual Studio, the JetBrains IDE family and the Coder remote-development platform, plus a command-line interface, and provides chat, inline edits, test generation, repository search and multi-file refactoring features over a customer's own codebase.[21][22] In April 2024 Poolside described the Assistant product as covering "write/edit code snippets, answer developer questions, provide workspace-based suggestions," powered by Malibu for deep reasoning and Point for low-latency completion.[21]
Poolside's enterprise platform exposes a multi-agent orchestration layer that lets agents plan changes, execute code in sandboxed containers, run tests and open pull requests.[21][29] The platform is shipped with role-based access controls for both human and agent identities, an encrypted credential vault with pattern-based redaction in logs, configurable filesystem and network policies, and step limits enforced per session.[29] Action-by-action session trajectories are exportable and integrate with security-information-and-event-management systems with reported sub-five-minute event propagation; retention is configurable for 30, 90 or 365 days.[29]
Poolside's enterprise stack is designed for deployment inside customer infrastructure rather than as a hosted service. As of 2026 the product page lists four deployment patterns: a cloud VPC option installed in the customer's own cloud account, an on-premises enterprise rack of pre-provisioned hardware for large teams, a compact on-premises tower aimed at small or classified-environment teams, and a bring-your-own-hardware option installed on customer-supplied Nvidia GPU clusters using Helm charts.[29] All four patterns can be operated air-gapped, and the company markets a Security Technical Implementation Guide-hardened operating system, an Authority to Operate certification and Impact Level 5 deployability for United States Department of Defense workloads.[29]
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]
Poolside has been deliberately reticent about naming customers, but the company has confirmed that it sells to large enterprises in defence, financial services, industrial manufacturing, software and the public sector.[30] It targets organisations with five thousand or more developers.[21][30] Reported and announced relationships include:
In 2025 the company expanded into Israel, hiring Tzachi Cohen, a former engineer at AMD and Habana Labs, to lead an inference team in the country and beginning outreach to Israeli defence firms.[20] Poolside's pricing model is custom and consumption-based, structured around variables such as the volume of code processed, GPU type, performance and security testing, and benchmarking commitments rather than a flat per-seat fee.[31]
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] In a March 2025 interview, Warner said that most companies should build AI applications on top of someone else's foundation models rather than train their own, and that Poolside's choice to train its own models was justified only because advancing intelligence was its core capability.[30] He compared the value of frontier intelligence to the value of electricity and said the company saw software development as the fastest economic route toward general-purpose AI.[30]
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 acquired by OpenAI in 2025; and Augment.[21][30] Poolside's differentiation rests on its custom models, RLCEF training pipeline and on-premises distribution, none of which the leading hosted competitors have prioritised in the same way.[21][30] 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.[29]
As of late 2025 Poolside had grown to roughly 150 employees worldwide, with offices in San Francisco, Paris and London and a research-and-development cadence that included monthly concentrated work weeks held in Paris.[1][2] The chief revenue officer is Paul St. John, a former chief revenue officer of GitHub who joined in April 2024.[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 late 2025 Poolside named Lance Smith vice president of data centres to run Project Horizon's infrastructure subsidiary.[7]
The company operates as a remote-first organisation across three time zones, a structure that Warner has cited publicly in talks and podcasts as drawing on his decade of remote-engineering leadership at Heroku and GitHub.[11][12] 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 of Poolside has been split between coverage that emphasises the scale of its fundraising relative to a still-unreleased public product and coverage that treats the company as a credible vertical-AI play. Fortune in September 2024 described the founders as having "said little about their product or business" before the Series B announcement, and noted that Poolside had demonstrated a code-generation product to investors but had not opened a public sign-up.[4] TechCrunch's reporting in June and October 2024 framed the round as a bet on a domain-specific approach to coding AI in a market dominated by general-purpose foundation models.[17][18] Bloomberg's October 2025 report on the Nvidia investment characterised Poolside as one of a small group of AI startups that had quadrupled in valuation within a year.[5]
Crunchbase News observed in October 2024 that the Series B was unusually large for a coding-focused startup and that the round depended on a thesis that on-premises foundation models would be a durable enterprise category rather than a transitional one.[18] The Information and other outlets reporting on the 2025 Nvidia commitment described it as a follow-on bet that tied Poolside's compute roadmap closely to Nvidia's GB300 generation.[5][19] Independent research notes from Sacra and Contrary Research in 2025 valued Poolside on the strength of its custom-model strategy and its access to defence and financial-services budgets, while flagging compute costs and the difficulty of building proprietary data centres as the main downside risks.[2][21]
Public scrutiny of Poolside has focused on three areas. First, the company had not released a generally available product to individual developers as of mid-2025 and remained limited to enterprise sales channels, which made independent benchmark comparisons against GitHub Copilot, Cursor and Claude Code difficult outside controlled customer trials.[4][21] Second, the Project Horizon dissolution with CoreWeave in early 2026 raised questions about whether Poolside could deliver multi-gigawatt compute on its planned timeline; coverage in trade press characterised the unwinding as a setback for the campus's first phase.[28] Third, the company's exposure to United States defence and intelligence customers, including In-Q-Tel as an investor and the United States Department of Defense as a contracted customer, has prompted commentary about export-control and dual-use considerations as Poolside expanded into the Israeli defence market in 2025.[20]