By PoolsideLearning Objectives
- Understand what Poolside Code is and how RLCEF (Reinforcement Learning from Code Execution Feedback) works
- Evaluate Poolside's approach of building code-specific foundation models versus fine-tuning general LLMs
- Assess the company's enterprise-only strategy and massive valuation
What Is Poolside Code?
Poolside is building AI foundation models designed specifically for software engineering — not fine-tuned general-purpose LLMs, but models trained from scratch on code with a novel technique called RLCEF (Reinforcement Learning from Code Execution Feedback). Founded by Jason Warner (former CTO of GitHub who oversaw the launch of Copilot) and Eiso Kant (who founded the first company to apply AI to source code), Poolside is one of the most heavily funded AI coding startups.
The company is currently enterprise-only — targeting organizations with 5,000+ developers (banks, defense contractors, large tech companies). Models are available through Amazon Bedrock and EC2 but not through any public consumer-facing product.
💡Key Concept
RLCEF (Reinforcement Learning from Code Execution Feedback): Most coding AI is trained by predicting the next token in code. Poolside goes further: models explore solutions across 130,000+ real-world codebases, then receive execution feedback — did the code compile? Did unit tests pass? Was it efficient? Was it secure? This is like RLHF (the technique that made ChatGPT helpful) but for code — the model learns from whether its code actually works, not just whether it looks correct.
Models
| Model | Purpose | Details |
|---|---|---|
| Malibu | Complex software engineering tasks | Code generation, test writing, refactoring, documentation; 1 million+ token context window |
| Point | Real-time code completion | Context-aware next-step prediction; algorithm and design pattern optimization |
Both models can be fine-tuned per customer's software environment — meaning a bank's Poolside deployment understands that bank's specific codebase, frameworks, and coding standards.
How It Works
- Code-specific training: Models built from scratch on code (not adapted from text-focused LLMs)
- 130,000+ real codebases: Training data spans production-quality open-source repositories
- Execution feedback loop: Every code generation attempt is compiled, tested, and evaluated for efficiency and security
- Per-customer fine-tuning: Models adapted to each enterprise's specific software environment
- Multi-agent orchestration: Acquired Fern Labs (London) for "Bridge" — an orchestration layer for high-stakes production environments
Infrastructure
Poolside is building Project Horizon — a 2 gigawatt AI campus in West Texas with 40,000+ NVIDIA GB300 NVL72 GPUs. Initial capacity began coming online in December 2025. This gives Poolside dedicated training infrastructure independent of cloud providers.
Poolside vs. AI Coding Competitors
| Company | Approach | Key Difference |
|---|---|---|
| Poolside | Code foundation models from scratch with RLCEF | Enterprise-only; per-customer fine-tuning; code execution feedback in training |
| OpenAI Codex | General-purpose LLM adapted for code | Broader accessibility; not code-specific from the ground up |
| Claude Code | Anthropic's coding agent | More accessible (individual developers); strong reasoning but general-purpose base |
| Devin (Cognition) | Autonomous coding agent | Agent product layer; Poolside focuses on the foundation model layer underneath |
| Magic AI | Code-specific foundation model with massive context | Similar ambition; Magic focuses on context window; Poolside on execution feedback |
Company Details
| Detail | Info |
|---|---|
| Company | Poolside AI |
| Founded | 2023 |
| CEO | Jason Warner (former CTO of GitHub; oversaw Copilot launch) |
| CTO | Eiso Kant (founded source{d}; first to apply AI to source code) |
| Headquarters | Paris, France (incorporated) and San Francisco (operations) |
| Employees | ~256-322 |
| Latest Funding | $500 million Series B (October 2024) at $3 billion valuation |
| Raising | ~$2 billion Series C at reported $12 billion valuation (October 2025; not confirmed as closed) |
| Total Raised | $626 million+ (with up to $2 billion more in progress) |
| NVIDIA Investment | Up to $1 billion (in the raising round) |
| Revenue | ~$50 million (third-party estimate) |
| Infrastructure | Project Horizon: 2 GW AI campus; 40,000+ GB300 NVL72 GPUs |
| Website | poolside.ai |
Strengths
- Code-first from the ground up — models built specifically for software engineering, not adapted from general-purpose text LLMs
- RLCEF training — models learn from code execution feedback across 130,000+ real codebases; code that compiles and passes tests is rewarded
- Per-customer fine-tuning — models adapted to each enterprise's specific codebase, frameworks, and standards
- GitHub CTO pedigree — Jason Warner's experience overseeing Copilot gives deep understanding of what developers need
- Massive funding — $626 million+ raised with NVIDIA contributing up to $1 billion; dedicated 2 GW training facility
Limitations and Considerations
- No public access — enterprise-only; targeting organizations with 5,000+ developers. Individual developers cannot use Poolside
- No public benchmarks — no SWE-bench, HumanEval, or other standard scores published; capabilities cannot be independently verified
- Pre-product for most — unless you are a large enterprise customer, Poolside is a company to watch, not a tool to use today
- $12 billion valuation unconfirmed — the Series C reportedly in progress; final close and valuation not publicly confirmed
- Revenue modest for valuation — ~$50 million revenue against a potential $12 billion valuation implies significant future expectations
Key Takeaways
- Poolside is building code foundation models from scratch using RLCEF — where models learn from whether their code compiles, passes tests, and is efficient across 130,000+ real codebases
- Enterprise-only (5,000+ developer organizations); available on Amazon Bedrock; per-customer fine-tuning for each organization's specific codebase
- Founded by GitHub's former CTO (who launched Copilot); backed by up to $1 billion from NVIDIA; building a 2 GW dedicated training facility
- Currently a company to watch rather than a tool to try — no public access, no published benchmarks, but the approach and team are among the most ambitious in AI coding
