CtrlK
Poolside (stylized as poolside) is an American artificial intelligence company that builds foundation models specifically designed for software development. Founded in May 2023 by Jason Warner, the former CTO of GitHub, and Eiso Kant, founder of source{d}, Poolside develops large language models trained using a proprietary technique called Reinforcement Learning from Code Execution Feedback (RLCEF). This method allows models to learn by actually writing and executing code rather than simply studying existing code text.
Poolside has raised over $626 million in venture funding as of its October 2024 Series B round, achieving a $3 billion valuation. In October 2025, Nvidia announced plans to invest up to $1 billion in the company as part of a $2 billion funding round that would value it at $12 billion [1]. The company targets enterprise customers with large engineering teams, including major banks and defense contractors, and its models are available through Amazon Web Services (AWS) via Amazon Bedrock [2].
Jason Warner and Eiso Kant first met in 2017 when Warner, then CTO of GitHub, tried to acquire Kant's company, source{d}, to serve as the AI engine at the core of GitHub. Although that acquisition did not go through, the two became close collaborators and spent the following six years discussing how AI would transform software development [3].
Warner had overseen the creation of GitHub Copilot during his time at GitHub and had also held senior engineering roles at Heroku and Canonical. After leaving GitHub, he served as Managing Director at Redpoint Ventures before deciding to build an AI company of his own. Kant, meanwhile, had founded source{d} (widely recognized as the first company dedicated to applying AI to source code) and later built Athenian, an engineering analytics platform [4].
The two co-founded Poolside in May 2023 in the United States, with Warner taking the role of CEO and Kant serving as CTO. Their core thesis was that software engineering represents the most promising path toward artificial general intelligence (AGI), and that purpose-built coding models trained through code execution would outperform general-purpose large language models applied to coding tasks [5].
Poolside initially raised a $26 million seed round in May 2023 led by Redpoint Ventures [6]. In August 2023, the company extended that seed round to $126 million in a deal co-led by Felicis Ventures and French billionaire Xavier Niel, with participation from Bain Capital Ventures, Air Street Capital, Bpifrance, Rodolphe Saade, Motier Ventures, and Scribble Ventures [7].
As part of this round, Poolside relocated its headquarters to Paris, France, establishing offices at 9 Rue des Colonnes in the 2nd Arrondissement. The company has maintained a dual presence across the US and Europe, with additional offices in San Francisco, Miami, Chicago, and Madrid. The move to Paris was partly motivated by access to strong AI research talent in Europe and by the investment from Xavier Niel, one of France's most prominent technology investors [7].
After raising its seed round, Poolside entered a prolonged stealth period lasting roughly one year. During this time, the company focused on building its technical infrastructure, training its foundation models, and assembling a research team with talent from organizations including DeepMind, Yandex, Amazon, and Uber. The company released very little public information during this period, which generated significant attention from technology media and investors [8].
In October 2024, Poolside emerged from stealth with a $500 million Series B round led by Bain Capital Ventures. This round valued the company at approximately $3 billion, making it a unicorn. The investor roster included Nvidia, DST Global, StepStone Group, Schroders Capital, Premji Invest, Dorsal Capital, BAM Elevate, Adams Street, Fin Capital, eBay Ventures, Citi Ventures, Capital One Ventures, HSBC Ventures, LG Technology Ventures, and SentinelOne's S Ventures, along with existing investors Felicis Ventures and Redpoint Ventures [9][10].
Following the Series B, the company began testing its product with large enterprise customers in both Europe and North America, focusing on organizations with 5,000 or more software developers [8].
In October 2025, Nvidia announced plans to invest between $500 million and $1 billion in Poolside as part of a $2 billion funding round. This investment would quadruple the company's valuation to approximately $12 billion, with more than $700 million already committed from existing investors [1]. Nvidia had previously participated in Poolside's Series B round.
By early 2025, Poolside had reached approximately $50 million in annualized revenue with a team of roughly 256 employees. By February 2026, the headcount had grown to approximately 322 [11].
On November 18, 2025, Poolside acquired Fern Labs, a London-based company specializing in multi-agent orchestration for production environments. Fern Labs was founded by Ash Edwards, Alex Goddijn, and Taylor Young, all former Palantir engineers. The acquisition brought Fern Labs' Bridge platform, a multi-agent orchestration layer designed for high-stakes deployments, into Poolside's product stack. This allowed Poolside to accelerate deployment of agentic AI systems inside private customer environments while maintaining strict security boundaries [12].
Poolside's primary technical differentiator is RLCEF, a training methodology that teaches models to write software by actually running code and learning from execution outcomes. Unlike conventional approaches that train models primarily on static text corpora of source code, RLCEF places models in environments containing hundreds of thousands of real codebases where they attempt to solve coding tasks, run the resulting code, and receive feedback based on whether the code compiles, passes unit tests, and meets efficiency and security standards [13].
As CTO Eiso Kant has described the approach: "Our models aren't just learning by consuming code; they're learning by coding themselves" [13].
The RLCEF system consists of two main components:
A key advantage of RLCEF is its ability to generate synthetic training data at scale. By continuously generating and evaluating coding tasks, Poolside can produce billions of training examples without relying on customer interactions or being constrained by the finite supply of publicly available source code. The company has containerized roughly one million public repositories with their executable test suites to serve as the training environment [13][14].
Poolside operates two proprietary foundation models:
| Model | Purpose | Context window | Key capabilities |
|---|---|---|---|
| Malibu | Complex software engineering tasks | 1,000,000+ tokens | Code generation, test writing, refactoring, documentation, deep reasoning |
| Point | Real-time code completion | 100,000 tokens | Sub-200 millisecond inline suggestions, context-aware predictions |
Malibu is the company's flagship model, optimized for multi-file code generation, test creation, refactoring, and documentation tasks. It is designed for interactive chat-based software engineering workflows and supports over one million tokens of context, allowing it to process large codebases in a single session [15].
Point is a smaller, quantized model engineered for speed. It provides real-time code completion suggestions directly within integrated development environments (IDEs), with response times under 200 milliseconds. Point uses advanced context awareness to predict what a developer will write next based on the surrounding code [15].
Both models are trained on over 500,000 open-source codebases and can be fine-tuned on individual customer environments, including proprietary code, internal documentation, and organizational coding standards [13].
Poolside's internal infrastructure for training and iterating on its models is called the Model Factory. Rather than treating model training as a one-off process, the Model Factory automates the entire pipeline from data preparation through model deployment, enabling continuous experimentation and improvement [16].
The Model Factory has evolved through six major iterations:
| Iteration | Focus | Description |
|---|---|---|
| 1 | Basic training | Standard forward and backward passes on predetermined architectures and datasets |
| 2 | Automated evaluations | Continuous benchmarking during pre-training rather than only after completion |
| 3 | Reinforcement learning | Modular RL system with code execution capabilities using containerized repositories |
| 4 | Architecture ablations | Systematic experimentation with model architectures using mu-P for scaling predictions |
| 5 | Data refinement | Automated data cleaning and quality assessment with lineage tracking |
| 6 | Data mixing | Dynamic data streaming with real-time blend adjustments during training |
The infrastructure runs on a Kubernetes-based orchestration system managing a 10,000-GPU H200 cluster. Key components include [16]:
Poolside's post-training process combines supervised fine-tuning (SFT) and reinforcement learning:
The company also employs a practice called "vibe checking," in which teams across the organization provide qualitative feedback on production models beyond automated metrics [16].
Poolside's primary product is a generative AI assistant for software engineering, built on a three-layer architecture [15]:
The assistant integrates with popular development environments including Visual Studio Code, IntelliJ, and web browsers. It supports tasks across the software development lifecycle, including code generation, testing, refactoring, documentation, and code review [15].
Poolside deploys its models within customer security boundaries, meaning no customer code leaves the customer's environment. This approach is a deliberate contrast to cloud-hosted AI coding tools where code is sent to external servers for processing. Models can be custom fine-tuned on each customer's proprietary codebase, internal documentation, and coding standards [15].
The company targets enterprise customers with large software development teams (typically 5,000+ developers), including major financial institutions, defense contractors, and retail organizations [8].
In December 2024, Poolside and Amazon Web Services announced a multi-year strategic partnership. Under this agreement, Poolside's Malibu and Point models became available through Amazon Bedrock, AWS's managed service for deploying foundation models. AWS was the first cloud provider to offer fully managed access to Poolside's models [2].
As part of the partnership, Poolside adopted AWS Trainium chips for model inference, which the company reported delivered approximately 40% cost savings compared to its previous infrastructure while reducing inference costs for customers [13].
In October 2025, Poolside announced Project Horizon, a 2-gigawatt AI campus planned for 568 acres of land on the Longfellow Ranch in Pecos County, Texas, part of the Permian Basin region. The campus is being developed in partnership with CoreWeave, which serves as the anchor tenant for the first 250-megawatt phase under a 15-year lease, with 500 megawatts reserved for future expansion [17].
Key details of Project Horizon include:
| Aspect | Details |
|---|---|
| Total capacity | 2 gigawatts across eight 250 MW phases |
| Land area | 568 acres |
| GPU deployment | 40,000+ Nvidia GB300 NVL72 GPUs (initial phase) |
| Power source | Natural gas from adjacent Permian Basin hub |
| First GPUs online | December 2025 |
| Full construction completion | First quarter of 2027 (estimated) |
| Connectivity | Dual long-haul fiber routes for high-bandwidth, low-latency connections |
The facility uses aero-derivative turbines equipped with selective catalytic reduction (SCR) systems for emissions control, along with grid-interconnect redundancy and battery storage. Rather than traditional sequential construction, Poolside employs a hybrid modular approach where major components such as electrical systems, cooling systems, and compute hardware are manufactured off-site and deployed continuously, with GPUs coming online monthly [17].
Lance Smith, VP of Data Centers, leads the Poolside Infrastructure Company division overseeing the project [17].
Poolside scaled to approximately 10,000 GPUs within its first twelve months of operation. The company currently operates a 10,000 H200 GPU cluster for its Model Factory training infrastructure. With the Project Horizon campus, the company plans to significantly expand its compute capacity with next-generation Nvidia GB300 NVL72 systems [13][16].
| Name | Role | Background |
|---|---|---|
| Jason Warner | CEO and co-founder | Former CTO of GitHub; former VP of Engineering at Heroku; former engineering lead at Canonical; Managing Director at Redpoint Ventures. BS in Computer Science from Penn State University; MS from Rensselaer Polytechnic Institute [4] |
| Eiso Kant | CTO and co-founder | Founder of source{d} (first company to apply AI to source code); founder of Athenian (engineering analytics); based in Portugal [4] |
| Paul St. John | Chief Revenue Officer | Former VP of Global Sales at GitHub; joined Poolside in April 2024 [18] |
| Margarida Garcia | Chief Operating Officer | Founding team member; previously VP of Operations [8] |
| Lance Smith | VP of Data Centers | Leads Poolside Infrastructure Company and Project Horizon [17] |
Poolside's research team includes engineers and scientists recruited from DeepMind, Yandex, Amazon, and Uber [15].
Poolside's funding history reflects rapid investor confidence in its approach to AI-powered software development.
| Round | Date | Amount | Lead investors | Valuation | Key participants |
|---|---|---|---|---|---|
| Seed | May 2023 | $26M | Redpoint Ventures | Undisclosed | Early backers |
| Seed extension | August 2023 | $100M ($126M total seed) | Felicis Ventures, Xavier Niel | Undisclosed | Bain Capital Ventures, Air Street Capital, Bpifrance, Rodolphe Saade, Motier Ventures, Scribble Ventures |
| Series B | October 2024 | $500M | Bain Capital Ventures | $3B | Nvidia, DST Global, StepStone, Schroders, Premji Invest, eBay Ventures, Citi Ventures, Capital One Ventures, HSBC Ventures, LG Technology Ventures, S Ventures (SentinelOne), Felicis, Redpoint |
| Reported round | October 2025 | Up to $2B (in progress) | Nvidia | $12B | Nvidia contributing $500M to $1B; $700M+ from existing investors |
Total disclosed funding through the Series B stands at $626 million [9].
Poolside operates in the rapidly growing AI-assisted software development market, which is projected to expand from $5 billion in 2024 to $27 billion by 2032 at a compound annual growth rate of approximately 24% [15]. The company faces competition from both established technology companies and well-funded startups.
| Competitor | Approach | Key differentiator |
|---|---|---|
| GitHub Copilot | Uses general-purpose LLMs (GPT-4, Claude, o1) | Largest user base (20M+ users); deep integration with GitHub ecosystem |
| Cursor | AI-native IDE built on VS Code fork | Rapid revenue growth ($300M+ ARR by mid-2025); project-wide context |
| Windsurf (formerly Codeium) | AI-powered IDE with free tier | Acquired by OpenAI for $3B in May 2025; 700K+ users |
| Magic | Long-context architecture (100M tokens) | Handles complex projects with minimal human intervention |
| Cognition (Devin) | Autonomous AI developer agent | Fully autonomous coding capabilities; strong SWE-bench performance |
| Amazon Q Developer | AWS-native AI assistant | Tight integration with AWS services and enterprise infrastructure |
Poolside differentiates itself from competitors in several ways [5][13][15]:
The demand for AI coding tools is driven by several factors. Industry analysts project a global developer shortfall of 85 million by 2030. Research indicates that software developers spend only about 52 minutes per day actively writing code, with the remainder devoted to meetings, code review, documentation, and other tasks. Gartner has estimated that 75% of enterprise software engineers will use AI code assistants by 2028, up from 10% in 2023 [15].
Poolside's CEO Jason Warner has stated that most companies should not attempt to build their own foundation models unless advancing intelligence is a core capability. The company positions itself not as a standalone coding tool but as a component within broader enterprise development ecosystems, differentiating through its specialized training methodology and on-premise deployment model [19].