The tool that trained GPT-4, now worth $1.7 billion.
Founded behind a karate studio in San Francisco. Adopted by OpenAI, Meta, and NVIDIA. Acquired for $1.7 billion. Weights & Biases built the invisible infrastructure that most of the world's AI runs on - and most people have never heard of it.
AI Engineers
Enterprise Orgs
Acquisition Price
OSS Integrations
In 2017, Lukas Biewald did a stint at OpenAI. What he found wasn't what the press releases suggested: world-class researchers, notebooks full of breakthroughs, and a lab at the frontier of AI. What he actually found were researchers tracking experiments in spreadsheets. Logging metrics by hand. Losing runs. Spending entire days trying to reproduce results that had worked last Tuesday.
Biewald had seen this before. He and his longtime collaborator Chris Van Pelt had spent a decade building CrowdFlower - a data labeling and crowdsourcing company that eventually sold to Appen for $300 million in 2019. They knew what good developer tooling could do. And they knew the machine learning world had almost none of it.
So they set up shop. Not in a gleaming SoMa office or a co-working space with a cold brew tap. Behind a karate studio in San Francisco's Mission District, chosen because the rent was cheap. The third co-founder, Shawn Lewis - former engineer at Google and the U.S. Naval Research Laboratory - joined through a distinctly non-corporate channel: he'd been in the same Y Combinator batch as Biewald's wife.
The company they built was named for the two parameters at the heart of every neural network: weights and biases. The parameters a model adjusts during training. The parameters that determine whether a model is useful or not. It's a name that functions as both a literal description and a subtle signal - this is a company that actually understands what it's building.
Their initial pitch was simple: add one line of import wandb to your Python script, and you'd get experiment tracking, visualization, and reproducibility for free. No setup. No fuss. It started spreading through AI research labs the way good dev tools always do - one engineer telling another that something actually works.
By the time OpenAI was training GPT-4, they were managing over 2,000 projects and millions of experiments through W&B. By 2024, more than a million engineers were using the platform. In May 2025, CoreWeave acquired the company for $1.7 billion - Biewald and Van Pelt's second major exit together.
"W&B's name refers to the two key parameters adjusted during neural network training - the weights that encode what a model knows, and the biases that shift how it applies that knowledge. A name that actually means something."- from the W&B founding philosophy
All three founders previously worked together at CrowdFlower (now Figure Eight), a data annotation company, before pivoting to build W&B. They are a rare team: technical enough to build the tools themselves, and experienced enough to have shipped a company before.
Stanford Mathematics & Computer Science graduate. Early career at Yahoo! Search Relevance. Co-founded CrowdFlower in 2007. Named to Inc.'s 30 Under 30. Host of the Gradient Dissent AI podcast. Internship at OpenAI in 2017 directly inspired W&B.
Former Google engineer. Co-founded CrowdFlower with Biewald in 2007 and ran it for over a decade. Also named to Inc.'s 30 Under 30. A repeat operator who has built and sold companies before and carries that experience into W&B's product strategy.
Former engineer at Google and the U.S. Naval Research Laboratory. Founder of Beep Networks. Connected to Biewald through a Y Combinator batch - via Biewald's wife, not through any professional introduction. Sometimes the best co-founder finds you sideways.
Machine learning is not clean. Models fail mysteriously. Hyperparameter runs multiply. Datasets get modified and no one tracks the change. W&B is the platform built to manage that mess - the system of record for everything that happens between "write the model" and "ship the model."
Experiment tracking for ML training runs. Logs metrics, hyperparameters, system stats, and media in real-time. Automated hyperparameter search (Sweeps), dataset and model versioning (Artifacts), collaborative dashboards (Reports), and centralized model registry. The flagship product that started it all.
Built for the GenAI era. Traces every LLM call, visualizes agent pipelines, runs evaluations across accuracy, latency, and cost, and monitors production deployments in real-time. Open source. Available on AWS Marketplace. The tool for engineers who build with language models rather than just train them.
API and playground access to leading open-source LLMs - DeepSeek R1, LLaMA 4, Phi 4, Qwen3 - powered by CoreWeave's GPU cloud. Announced in preview at Fully Connected 2025. Combines W&B's evaluation and observability tools with raw inference from CoreWeave's infrastructure.
Infrastructure observability that unifies CoreWeave compute monitoring with W&B model training visibility. The first major post-acquisition product that shows what the CoreWeave deal actually enables: a full-stack view from GPU utilization to model performance.
A central repository for models and datasets with versioning, aliases, lineage tracking, and governance controls. The equivalent of a git repository for trained models - except it also tracks who changed what, when, and why, across the entire organization.
The wandb Python library is free, open-source, and integrated into 20,000+ OSS repos including LangChain, LlamaIndex, and GPT4All. Often just one line of code. This is deliberate: W&B built its growth by giving individual engineers something genuinely useful for free, then letting them pull their organizations in.
Free tier for individuals. Teams plan at approximately $50/user/month. Academic plan free for researchers with expanded storage. Enterprise with custom pricing, SSO, compliance controls, and on-premises options. Available through AWS Marketplace and Google Cloud Marketplace.
Total raised: approximately $205M across four rounds. Final exit to CoreWeave valued the company at $1.7 billion - a tidy multiple for investors who got in at a $5M Series A.
W&B's growth was built on the individual engineer. One developer adds import wandb, likes what they see, tells a colleague, and three months later the entire ML team is on it. That's how a startup ends up inside OpenAI, Meta, NVIDIA, and Microsoft simultaneously.
As of 2025: 1 million+ developers and 1,400+ organizations globally. 30+ foundation model builders use W&B for training and evaluation.
Reached unicorn status ($1B+ valuation) in October 2021 with Series C funding led by Felicis and Insight Partners
Acquired by CoreWeave for ~$1.7B in May 2025 - one of the largest AI developer tooling acquisitions on record
Used by OpenAI during GPT-4 training to manage 2,000+ projects and millions of experiments
Integrated into 20,000+ open-source repositories including LangChain, LlamaIndex, and GPT4All
Surpassed 1 million developer users and 1,400+ enterprise customers worldwide
Certified as NVIDIA DGX-Ready Software and NVIDIA AI Enterprise partner since 2024
W&B Weave reached General Availability for enterprises in December 2024
GitHub repo surpassed 10,900+ stars; hosts Fully Connected global AI conference with 750+ engineers per event
W&B has integrations with over 50 ML frameworks and is embedded in every major cloud platform. Not as a checkbox integration - as a genuinely useful tool that developers choose to use.
Deep integration with Azure AI Foundry for automatic experiment tracking during GPT-4 and GPT-4o fine-tuning. Azure OpenAI Service support and one-click W&B deployment on Azure.
Certified NVIDIA DGX-Ready Software and NVIDIA AI Enterprise partner. Integrations with NVIDIA NIM microservices and DGX systems. Long-standing partnership since 2021.
W&B Weave listed on AWS Marketplace in the "AI Agents and Tools" category in 2025. Integrations with Amazon Bedrock and SageMaker.
Parent company since May 2025. W&B tools integrated into CoreWeave's AI cloud stack. Joint products include Mission Control, W&B Inference, and Weave Online Evaluations.
Official W&B organization on HuggingFace with joint model card integrations. W&B is a primary experiment tracking solution for the Hugging Face ecosystem.
Strategic cloud partner with GCP infrastructure integrations and presence on Google Cloud Marketplace.
The company was started behind a karate studio in the Mission District. Not a co-working space. Not a garage. A karate studio. The rent was cheap.
W&B's name is literally what it sounds like: weights and biases are the two core parameters adjusted during neural network training. Naming a company after the thing it actually cares about is rarer than it sounds.
Lukas Biewald and Chris Van Pelt previously sold CrowdFlower to Appen for $300 million in 2019. The $1.7B W&B exit is their second rodeo together.
Co-founder Shawn Lewis joined because he was in the same YC batch as Biewald's wife. The introduction was personal, not professional. Sometimes the best technical hires arrive sideways.
The wandb Python library is so ubiquitous it shows up in 20,000+ open-source repositories. Often it's a single line of code: wandb.init(). That's the entire integration.
W&B has been remote-first since 2017. Not because of COVID, not because of a policy change, but because that's how they started. The karate studio era ended early.
W&B Server v0.77 released with security enhancements to API keys, and general availability of media panel sync and looping features.
Fully Connected Conference London hosted 750+ AI engineers over two days (Nov 4-5, 2025). The annual event also runs in San Francisco and Tokyo.
CoreWeave and W&B announced three new AI cloud products in preview at Fully Connected 2025: Mission Control, W&B Inference, and Weave Online Evaluations.
CoreWeave (Nasdaq: CRWV) completed acquisition of Weights & Biases for approximately $1.7 billion. W&B continues to operate under its own brand.
CoreWeave announced the W&B acquisition. Separately, W&B Weave was listed in the new AWS Marketplace "AI Agents and Tools" category.
W&B Weave reached General Availability for enterprises, delivering production-grade LLM tracing, evaluations, monitoring, and guardrails.
Expanded Microsoft integration: W&B Models now integrates with Azure AI Foundry for automatic experiment tracking during GPT-4 and GPT-4o fine-tuning.
W&B certified as NVIDIA DGX-Ready Software and NVIDIA AI Enterprise partner. New integrations with NVIDIA NIM microservices and DGX systems announced.