Microsoft's Foundry IQ Knowledge Graphs Gave Enterprise Agents 93.5% Accuracy vs. 26.1% for Standard RAG. The Context Strategy Shift Is Bigger Than Any One Product.
On November 17, Bessemer Venture Partners and Innovation Endeavors co-led a $57.5 million Series B for LanceDB. The company builds the data format underpinning training pipelines for video, robotics, and world models — and its customer list reveals why multimodal training data infrastructure is the next bottleneck in AI.
On November 17, 2026, LanceDB announced a $57.5 million Series B co-led by Bessemer Venture Partners and Innovation Endeavors, with Theory Ventures, Dimension, Swift Ventures, CRV, and Y Combinator also participating. Total funding now stands at $98.5 million.
The customer list in the announcement is the more interesting data point: Netflix, Microsoft AI, Runway, Luma AI, and Midjourney. These are not early adopters who took a meeting — they are production users running LanceDB in the pipelines that train their most important AI models. Runway trains video generation models on Lance. Midjourney stores its training datasets in Lance. Microsoft AI uses it for research infrastructure. That customer list is not built in a demo — it is built by solving a real, painful, expensive problem in production.
The problem LanceDB solves is not the one most people think about when they think about AI infrastructure. It is not about serving models faster. It is about training better models in the first place — and training data infrastructure is about to become the most important bottleneck nobody is talking about.
What Training Data Infrastructure Actually Is
The public narrative around AI infrastructure centers on inference: how fast can a model generate a response, how cheaply can you run it at scale, how do you retrieve relevant context in real time. That is the vector database problem, and the market has funded several hundred million dollars of solutions for it.
Training data infrastructure is a different problem, and it is upstream of inference entirely. Before a model can answer a question, it must be trained. Before it can be trained, its training data must be collected, cleaned, versioned, and loaded into GPU memory thousands of times over the course of a training run. The data loading pipeline is not the glamorous part of AI development, but it is frequently the bottleneck.
Here is a concrete example of why it matters: a video generation model training run might require iterating over 100 million video frames, repeatedly, in random order, over the course of weeks. The training job on 1,000 GPUs will stall on the data loader if the data format cannot support the access pattern — random access to individual samples at the speed the GPUs can consume them. This is not a retrieval problem; the data is not changing between training runs. It is a format and access problem: how do you store 100 million video frames so that a training job can read them at GPU memory bandwidth, in arbitrary order, without bottlenecking the compute?
LanceDB's answer is the Lance format — an open-source columnar data format built specifically for ML workloads. Lance is not the first attempt at this problem: TensorFlow's TFRecord format and PyTorch's WebDataset format both tried to solve it, and both have significant production deployments. What Lance adds is a combination of features that neither predecessor provides: fast random access, versioning, native multimodal support (video, images, text, embeddings, and structured data can all live in a single Lance dataset), and clean integrations with both PyTorch and TensorFlow data loaders.
Theory Ventures' Series B write-up makes the case plainly: "Multimodal AI runs on Lance." The claim is that as AI moves from text-only to video, robotics, and world models — the next generation of AI training workloads — the data format underneath those models will increasingly be Lance.
The Customer List Is the Thesis
LanceDB's Series B announcement named five production customers: Netflix, Microsoft AI, Runway, Luma AI, and Midjourney. Each of them is operating at the frontier of a specific multimodal AI domain.
Netflix runs recommendation and content understanding models on video data at a scale that few organizations can match. Netflix's recommendation model consumes metadata and viewing behavior across hundreds of millions of users and thousands of hours of video content. Its content understanding pipeline extracts semantic features from video frames to enable search, curation, and automatic tagging. Both workloads require training on large video datasets and need a format that handles the data loading overhead efficiently.
Runway and Luma AI are training video generation models — arguably the most data-intensive AI workload currently in production. A single video generation training run requires iterating over millions of video clips, each consisting of hundreds or thousands of frames, with associated text descriptions for conditional generation. The data loading overhead for these runs is substantial: without a format optimized for random access, GPU utilization drops sharply as the training job waits for data.
Midjourney trains image generation models on proprietary datasets built from curated images. Its model training pipeline is one of the most valuable in generative AI — the stylistic quality of Midjourney's output is a direct function of the quality and curation of its training data. The data layer underneath that training process is infrastructure worth protecting.
Microsoft AI is the customer that most validates LanceDB's enterprise deployment model. Microsoft runs ML research and product development across dozens of teams, with strict data governance requirements that make a fully-managed third-party service impractical. LanceDB Enterprise's design — deployed inside the customer's own cloud account, not on LanceDB's infrastructure — is specifically architected for customers like Microsoft who cannot put proprietary training data in a vendor's data center.
The pattern across these customers is consistent: they are all building at the frontier of video or multimodal AI, they all have training pipelines that process massive volumes of unstructured data, and they all have data governance requirements that preclude fully-managed third-party solutions. That is the customer segment LanceDB designed for, and the Series B round is funding the go-to-market expansion into more organizations with the same profile.
Why Lance Is Not Competing With Parquet
The comparison to Apache Parquet is inevitable — both are columnar formats — but it is misleading about the actual competitive landscape.
Parquet was designed for analytical queries: scan all rows, aggregate by column, return a result. It excels at this workload. Data warehouses (Databricks, Snowflake, BigQuery) are built on Parquet for this reason. Delta Lake and Apache Iceberg add ACID transactions and schema evolution on top of Parquet to make it suitable for data lake architectures.
None of this applies to ML training workloads. Training jobs do not scan all rows; they randomly sample batches. Training jobs do not aggregate by column; they read entire rows (or multi-modal records) for individual training examples. The access patterns are different, and Parquet's performance characteristics reflect its design for analytics, not random-access batch sampling.
LanceDB's claim of 100x faster random access than Parquet is based on its own benchmarks, which the company acknowledges are preliminary. The claim should be treated cautiously — ML training benchmarks are notoriously sensitive to hardware configuration, dataset characteristics, and batch size. What is not in dispute is that Parquet was not designed for the random-access batch sampling pattern that characterizes ML training, and that this creates measurable data loading overhead in production training pipelines.
The practical comparison table:
| Format | Optimized For | Random Access | Versioning | Multimodal | ML Framework Integration |
|---|---|---|---|---|---|
| Apache Parquet | Analytical queries | Poor | No (native) | No | Via pandas/Arrow |
| Delta Lake | Data lakes / ACID | Poor | Yes | No | Via pandas/Arrow |
| TFRecord | TF training | Moderate | No | Partial | TensorFlow native |
| WebDataset | PyTorch training | Moderate | No | Partial | PyTorch native |
| Lance | ML training | Excellent | Yes | Yes | PyTorch + TF native |
Lance is not trying to replace Parquet in data warehouses. It is trying to become the standard format for ML training data storage, versioning, and loading — a category that did not have a standard format before Lance.
The Architecture: Why Open Source and In-Cloud Deployment Both Matter
LanceDB made two architectural decisions that are not obvious from the outside but are central to its enterprise sales motion.
The first decision: make Lance fully open source under Apache 2.0. This means the format is not a lock-in mechanism. Any company can store data in Lance files, any framework can read them, and any cloud provider can offer storage optimized for Lance access patterns. This is a deliberate choice — it makes Lance the neutral format, like Parquet, rather than a proprietary format with vendor dependencies. The enterprise trust this generates is worth more than the alternative of keeping Lance closed: closed formats cannot become ecosystem standards.
The second decision: deploy LanceDB Enterprise inside the customer's cloud account rather than as a fully managed service. This is the opposite of the Pinecone or Weaviate approach, which offer fully managed services where customer data lives on vendor infrastructure. For the specific customers LanceDB is targeting — frontier AI labs, large enterprises with proprietary training data — this is not a nice-to-have. It is a requirement. Training data for frontier models is among the most valuable intellectual property in those organizations; storing it on a third-party's servers is not a risk they will accept.
The combination of open format and in-cloud deployment creates a sales motion that is different from most infrastructure companies: LanceDB is not selling access to a service, it is selling the managed operational infrastructure to run an open format reliably and efficiently inside the customer's own environment. The customer retains full control of their data; LanceDB provides the tooling, integrations, and support to make it work at scale.
The Competitive Landscape: Where the Threat Comes From
LanceDB's most plausible competitive threat is not from vector databases moving upstream — it is from the hyperscalers building native training data infrastructure.
AWS has SageMaker Feature Store, GCP has Vertex AI Feature Store, and Azure ML has its own dataset registry. Each of these offers some training data versioning and management within the broader managed ML platform. None of them is specifically optimized for the Lance access pattern or the multimodal data types that LanceDB targets — but all of them have the cloud distribution and enterprise relationships that LanceDB is building.
The second threat is from Databricks, which has been systematically acquiring data infrastructure companies and building toward a unified data and AI platform. Databricks' acquisition of Neon for approximately $1 billion and its integration of Delta Lake demonstrate the playbook: build or acquire the data format, run the managed service, and capture the enterprise data gravity. Databricks could add Lance support natively, or acquire LanceDB directly. At $98.5M in total funding and a customer base of frontier AI labs, LanceDB is a credible acquisition target within the next 24-36 months.
The third threat is from open-source communities standardizing on a different format. TensorFlow's TFRecord and PyTorch's WebDataset each have large existing codebases and community inertia. Switching costs for established training pipelines are real, and format migrations in ML engineering are notoriously painful. LanceDB's adoption by Runway, Midjourney, and Luma AI — all of whom were presumably building new pipelines rather than migrating existing ones — is more straightforward than convincing organizations with legacy TFRecord pipelines to migrate.
The similarity to what Mercor is building with expert data is worth noting: both companies are positioning proprietary data assets as long-term moats. LanceDB's open-source approach inverts this — the data format itself is not proprietary, but the operational knowledge, integrations, and customer relationships accumulated around it are. That is a different kind of moat, and arguably a more durable one.
The Use-of-Funds Signal
CEO Chang She's public statements describe the Series B as funding three areas: large-scale data curation tools, feature engineering infrastructure, and data loading optimization for video, world models, and robotics.
This use-of-funds breakdown reveals the product roadmap. Data curation tools suggests LanceDB is moving beyond storage into the pipeline work that happens before data gets into a Lance dataset — cleaning, deduplicating, labeling, and filtering raw video and image data. Feature engineering infrastructure suggests LanceDB is building compute-adjacent services that transform raw data into training-ready features. These are significant expansions beyond a data format; together, they describe an end-to-end ML data platform.
The specific mention of world models and robotics is notable. World models — neural networks that learn a predictive model of the physical environment for robotics and autonomous vehicle applications — represent the next generation of data-hungry AI training workloads. Training a world model on sensor data from a robot fleet requires storing, versioning, and efficiently loading petabytes of sensor recordings: video, lidar, IMU, and proprietary sensor outputs, all timestamped and synchronized. LanceDB's multimodal format is a natural fit for this use case, and the mention suggests active conversations with robotics companies alongside its current media and software customer base.
What Enterprise AI Teams Should Take From This Round
The LanceDB Series B is not primarily relevant to teams running RAG applications or deploying AI agents for enterprise workflows. Those teams are building in the inference layer, and their infrastructure choices (vector databases, orchestration frameworks, model APIs) are orthogonal to what LanceDB provides.
It is relevant to two groups:
Teams building or fine-tuning models on proprietary data. If your organization is using proprietary data to fine-tune foundation models, your training data storage and versioning infrastructure matters for reproducibility, compliance, and cost efficiency. LanceDB's open-source format, versioning, and in-cloud deployment model is worth evaluating alongside Parquet + Delta Lake and native cloud ML platforms.
Teams tracking where the AI infrastructure investment cycle is heading. The sequence of infrastructure rounds — Fractile's $720M for inference memory bandwidth, Modal's $750M for serverless GPU compute, LanceDB's $57.5M for training data format — maps the layers of the AI stack that are still being built out. The pattern indicates that inference-time infrastructure (vector databases, orchestration) has largely been funded, and capital is now moving to training-time infrastructure: data format, feature engineering, compute orchestration for training runs. That shift has implications for where technical talent and tooling investment will concentrate over the next 24 months.
The AI middleware tax analysis showed that the per-call cost structure of inference pipelines creates compounding costs that most teams underestimate. Training data infrastructure creates a parallel dynamic on the training side: poor data loading efficiency wastes GPU time at $2-5 per GPU-hour, and that waste compounds across training runs that may each span weeks. The economic case for investing in training data infrastructure is the same as the economic case for optimizing inference pipelines — but the training-time optimization is still largely unmade in most organizations outside the top tier of AI labs.
What the Bessemer Lead Signals About Category Trajectory
Bessemer's decision to co-lead LanceDB's Series B is worth reading as a category signal. Bessemer's infrastructure portfolio includes Twilio (communications API), Shopify (e-commerce infrastructure), and Cloudflare (CDN and network security). Each of these companies became the standard underlying infrastructure for a large software category, with network effects and switching costs that made them dominant once embedded in production.
Bessemer's pattern-match on LanceDB appears to be: open data format → standard adoption → managed enterprise service → platform economics. The Lance format, like S3 or Parquet, is most valuable when it becomes the assumed substrate — when every ML framework has a Lance reader, every cloud provider offers Lance-optimized storage, and every MLOps tool integrates Lance versioning. At that point, LanceDB Enterprise becomes the natural choice for teams that want managed operations on the standard format, the same way Databricks became the natural choice for managed Delta Lake.
Whether that platform trajectory materializes depends on how quickly the frontier AI labs and research institutions converge on Lance as the standard training data format. The November 17 customer list — Netflix, Microsoft AI, Runway, Luma AI, Midjourney — is a credible signal that the convergence has begun.
Takeaway: LanceDB's $57.5M Series B is not primarily an AI infrastructure funding story — it is an indication that the market is finally building the training-time data layer that has been the missing piece beneath frontier AI development. The Lance format's combination of fast random access, multimodal support, and open-source licensing positions it as a potential Parquet-equivalent for ML training data. The customer list (Netflix, Microsoft AI, Runway, Midjourney) validates that the format solves a real problem in production. For enterprise AI teams, the relevant question is not whether to use LanceDB today — most are not training frontier models — but whether the training data infrastructure category is about to repeat the investment cycle that created a $500M+ vector database market three years ago.
Frequently Asked Questions
What is LanceDB and what does it do differently from vector databases like Pinecone?
LanceDB is a training data infrastructure platform, not a vector database. The distinction matters: vector databases like Pinecone, Weaviate, and Qdrant are built for inference-time retrieval — taking an embedding, finding the nearest neighbors, and returning results at query time. LanceDB is built for training time — storing, versioning, and loading the raw multimodal data (video frames, images, text, sensor data) that gets fed into model training pipelines. Its underlying format, Lance, is a columnar format designed for fast random access on large multimodal datasets, with native support for PyTorch and TensorFlow data loaders. The key operational difference: vector databases solve the problem of making a trained model's outputs searchable; LanceDB solves the problem of getting high-quality data into the model training pipeline efficiently and reproducibly. LanceDB also supports vector search and can serve as a hybrid retrieval layer, but that is a secondary use case, not the reason Netflix and Midjourney are customers.
Who invested in LanceDB's $57.5 million Series B and why?
LanceDB's Series B was co-led by Bessemer Venture Partners and Innovation Endeavors, with Theory Ventures, Dimension, Swift Ventures, CRV, and Y Combinator also participating. The round brings LanceDB's total funding to $98.5 million across a seed, Series A, and Series B. Bessemer has historically backed infrastructure companies that become essential plumbing for large software categories — Twilio, Shopify, LinkedIn, and Cloudflare are in its portfolio. Its Series B lead signals that Bessemer views training data infrastructure as a category with comparable platform dynamics: a format or protocol layer that becomes entrenched once it is embedded in production training pipelines. Theory Ventures led the Series A, and its blog post from November 2026 describes the investment thesis as 'multimodal AI runs on Lance' — positioning LanceDB as the data layer that enables video, robotics, and world model training at scale.
Why do Netflix, Microsoft AI, Runway, Midjourney, and Luma AI use LanceDB?
Each of these customers is building at the frontier of video or multimodal AI. Netflix is training recommendation and content understanding models on video data at massive scale. Microsoft AI runs research and product teams building multimodal models. Runway and Luma AI train video generation models. Midjourney trains image generation models on large datasets. All of them share the same infrastructure problem: they need to store, version, and efficiently load large volumes of multimodal data for training runs that can span weeks and cost millions of dollars. The Lance format's key advantages for these workloads are fast random access (the company claims 100x faster than Parquet on random-access patterns), native multimodal support (video, images, text, embeddings all in a single dataset), and versioning that lets teams roll back to any snapshot of a training dataset. These are table-stakes requirements for a production multimodal training pipeline, and LanceDB is one of the few systems purpose-built for them.
What is the Lance format and how does it compare to Parquet or Delta Lake?
Lance is an open-source columnar data format designed specifically for machine learning workloads, developed by LanceDB's founders. The comparison to Parquet and Delta Lake is useful but imprecise. Apache Parquet is a columnar format optimized for analytical queries — it excels at large scans over structured data and is the standard format in data warehouses. Delta Lake adds ACID transactions, versioning, and schema enforcement on top of Parquet files in a data lake. Neither was designed for ML training workloads, which have a different access pattern: rather than scanning all rows, training jobs do random access to individual samples (a batch of 32 video frames from different parts of a 100-million-frame dataset). Lance's design optimizes for this access pattern. The company claims 100x faster random access than Parquet; this claim is based on its own benchmarks and should be validated independently. Delta Lake and Apache Iceberg provide better analytical query performance. The practical implication: Lance is not competing with Parquet for analytics; it is competing with custom ML data formats (TFRecord for TensorFlow, WebDataset for PyTorch) that ML engineering teams were building in-house before LanceDB existed.
Is LanceDB open source and what is LanceDB Enterprise?
LanceDB's core software is open source under the Apache 2.0 license, and the Lance format itself is a separate open-source project. The open-source LanceDB can be run self-hosted. LanceDB Enterprise is the commercial tier, which adds managed deployment (runs in the customer's own cloud account — AWS, GCP, or Azure — not on LanceDB's infrastructure), enterprise security controls, support SLAs, and performance features for large-scale training workloads. This architecture — open core, with an enterprise tier that runs in the customer's cloud — is a deliberate design choice. Training datasets for frontier AI models contain proprietary data that companies will not move to a third-party managed service. LanceDB Enterprise keeps data inside the customer's cloud boundary, which is a prerequisite for sales to Netflix, Microsoft AI, and similar customers with strict data governance requirements.
What does LanceDB's Series B mean for the broader AI infrastructure market?
LanceDB's $57.5M Series B is the latest signal in a broader shift of infrastructure investment from inference to training. The first wave of AI infrastructure investment went into vector databases for retrieval-augmented generation (Pinecone, Weaviate, Qdrant raised a combined $500M+ in 2022-2024). The second wave is now funding the upstream layer: the data infrastructure that enables training better models in the first place. LanceDB's round follows Applied Compute's $350M for custom AI deployment infrastructure, Fractile's $720M for inference memory bandwidth, and Modal's $750M for serverless GPU compute. Each of these rounds is addressing a different bottleneck in the AI production stack. LanceDB's specific bet is that the training data layer — currently solved by a patchwork of custom scripts, TFRecord files, and in-house formats — will consolidate around a standard open format the same way analytical data consolidated around Parquet. If that consolidation happens, LanceDB's position as the architect of the standard format gives it platform economics: every tool, framework, and cloud service that integrates Lance deepens the moat.