Worldset, Inc.

The bottleneck in science

People die from diseases we should be able to understand and eventually cure. Extending healthy human life is one of the highest-leverage goals humanity can pursue, yet biology remains constrained by the speed and throughput of scientific discovery.

We believe curing most diseases will require intelligence beyond what human researchers alone can provide. Transformers gave humanity a path to scale scientific reasoning far faster than we can scale the number of trained scientists. But intelligence is only useful when it can learn from physical reality.

AI has consumed most of the high-quality information available on the internet. Its largest remaining blind spots are domains where data is private, non-digital, expensive, poorly structured, or does not yet exist. Experimental biology has all five characteristics.

The bottleneck to unlocking immense scientific progress is training AI to run experiments in the physical world.

Science data centers

Frontier models cannot become capable experimental scientists by learning only from papers and textbooks. Experimental work produces largely private, non-digital, poorly structured data and is full of tacit knowledge that disappears in failed runs, undocumented adjustments, and unpublished results.

Worldset is building science data centers: modular, autonomous laboratories where AI agents can design, conduct, observe, and debug experiments. We capture the resulting agent trajectories, which include the goal and starting context, protocol drafts, tool calls, instrument states, raw observations, errors, interventions, revisions, and final outcomes. We then sell this proprietary training and evaluation data to AI labs.

A failed experiment with complete context can be more valuable for training than a successful result without the context of how it was achieved. By capturing every decision, action, observation, error, correction, and result, we produce datasets that do not yet exist anywhere else.

Today, scientific progress scales with scarce trained personnel. Every additional research program requires more PhDs. Human scientific reasoning is becoming less of a constraint on how much research we can run. The bottleneck is gradually shifting to experimental capacity: the number of automated instruments and facilities available for models to test hypotheses in the physical world.

AI labs already spend hundreds of billions on compute infrastructure. As public datasets are exhausted and model capabilities converge, a growing share of that capital moves toward acquiring proprietary, hard-to-generate data.

According to people from OpenAI, the big labs’ spending on data grew from less than $1B in 2023 to around $10B in 2026 and is on track to exceed $100B in 2029. The rapid growth of data companies like Mercor, AfterQuery, and Micro1 shows that AI labs will pay a lot for training data that can meaningfully improve model performance.

Biology is our entry point. After year one, we plan to expand into chemistry and materials science so frontier models can learn to operate across many fields of experimental science.

Our near-term goal is to build a network of data-center-scale research facilities. This will allow us to control a substantial share of the world’s experimental capacity and become the default physical infrastructure through which machine intelligence conducts science.

Standardized lab modules let us add capacity far faster than conventional laboratories. The first lab modules establish the interfaces and operating model. Once validated, modules built on the same architecture can be manufactured in a factory and then deployed, monitored, and improved as one fleet. We build the facilities end to end, covering site development, instrument procurement, software, operations, data capture, and evaluation.

Our long-term vision is to use this infrastructure to extend healthy human life and drastically accelerate discovery across a broad range of scientific disciplines.

Business model and economics

We will license curated datasets of experimental trajectories from our proprietary catalog and generate targeted datasets around specific model weaknesses. The initial buyers are frontier AI labs and neo-labs building AI scientists. Big pharma could become an additional customer segment if pharmaceutical companies start building specialized in-house models.

Research partnerships will keep model developers close to the production loop. We identify a model’s weakness through a benchmark, generate relevant trajectories to address it, and then feed the resulting data back into training. This lets us produce datasets that address demonstrated capability gaps rather than speculate about what customers need.

In this data market, there is no standard price per experiment. Some experiments take hours and require a few dollars’ worth of reagents, while others take weeks and require a thousand dollars’ worth. The price per dataset therefore depends on the complexity and cost of the experiments and whether the dataset is licensed exclusively or non-exclusively.

Pricing benchmarks from conversations with people at Handshake, DeepMind, Shift, and Scale AI:

Human-generated experimental data is commonly priced at approximately production cost plus 50% on the first sale.
The same non-exclusive dataset is often licensed to 3-10 buyers; after production costs are recovered, subsequent licenses are pure profit.
Exclusive rights typically cost 3-5 times as much as a non-exclusive license.
Data vendors achieve 80-90% gross margins on human-generated data.

Our data is cheaper to produce because we do not need to pay humans to generate it. We therefore target production costs of 10-15% of cumulative dataset revenue, implying an 85-90% gross margin.

The competitive landscape

When you ask your agent of choice about automated labs, a few names come up frequently. Most fall into one of four categories, with some of the well-known companies listed below:

Drug discovery companies (Isomorphic Labs, Insitro, Recursion) - they focus on generating their own IP and bringing drugs into the clinic. Selling their data would interfere with their core business.
Contract research organizations (Adaptyv, Ginkgo) - they generate narrow datasets from specific assays, such as antibody binding or transcriptomics. They repeat a small set of protocols rather than run flexible n=1 experiments. They sell biological datasets, not agent trajectories.
Companies building their own AI scientist (Lila, Periodic Labs, Medra) - their data is the moat for their models. Selling this data to competing AI labs would destroy their core business.
Data companies (Substrate, C5R) - Substrate focuses more on selling biological data, such as binding data, cell-perturbation profiles, and genomics datasets, than on agent trajectories. C5R uses agents to control experiments but relies on humans to run most of them rather than focusing on automation.

We are building the infrastructure and data layer that teaches frontier models how to perform experimental science. We record the problem, context, decisions, actions, instrument telemetry, failures, corrections, and outcome -- not merely the final biological measurement.

This makes us the neutral data and physical-feedback infrastructure layer for scientific AI. Frontier labs need an early, trusted partner that can help them turn a general model into a capable scientist. We can become that partner before each lab spends years building fragmented physical infrastructure internally.

Our durable advantages

Our durable advantages compound as the infrastructure fleet and dataset catalog grow:

Scientific Taste: Exclusive partnerships with the most relevant professors in all subfields of biology allow us to incorporate their research ideas and unique taste into the experiments we run.
Unique data: Agent decisions, instrument interactions, failed experiments, recovery behavior, and high-frequency telemetry down to the raw, unprocessed sensor data of scientific instruments are not available from others.
Replicability: Standardized execution, systematic evaluation, and precisely controlled environmental conditions make every experiment as replicable as biologically possible.
Scalability and throughput: We are designing our system to scale to $100B+ in CapEx deployment from the start. This would be around 10 times the combined experimental capacity of the top 20 pharma companies.
Customer relationships and model integration: Close research partnerships let us understand each model’s weaknesses, generate data in direct response, and become deeply integrated into customers’ training and evaluation loops.
Harness engineering: Better tool interfaces, context management, feedback, graders, and failure recovery make agents running experiments in our facilities more successful at reaching a goal than the underlying model would be on its own.
Compounding infrastructure: Every experiment improves the software, automation, evaluation systems, protocols, and future lab-module designs that produce the next dataset.

Long-term vision

The data business finances our near-term goal: a network of data-center-scale research facilities that accounts for a substantial share of the world’s experimental capacity. With this, we will become the default physical infrastructure through which machine intelligence conducts science. AI labs pay for the data that improves their models while bearing the much larger cost of the compute needed to train them.

We build facilities end to end, from site development and instrument procurement to software, operations, data capture, and evaluation. The first lab modules establish an architecture we can validate and use to build more modules in a factory. We can then deploy, monitor, and improve these modules as one fleet. This lets us add capacity far faster than conventional laboratories, creating a compounding loop of more modules, more experiments, and more data.

At scale, we can train our own scientific foundation models and pursue proprietary drug discovery. Running 1,000x more experiments than other drug discovery companies would let us explore more hypotheses and learn faster from failure. Owning this infrastructure creates a path to a trillion-dollar company spanning scientific data, AI models, and drug discovery. Our long-term vision is to use it to extend healthy human life and drastically accelerate discovery across a broad range of scientific disciplines.

Roadmap

3 months: Build the first lab module (done after 4 weeks at a cost of <$1M). Hire an excellent team of founding engineers and founding scientists. Run the first fully autonomous experiments in our lab modules. Start the first research partnerships with relevant professors. Start the first data partnerships with AI labs.
6 months: Build 5 more lab modules. Scale the partnerships with professors and AI labs.
9 months: Start designing and building a factory that can produce one lab module container per day (at an average of $3M per container).
12 months: Lay the groundwork and acquire the capital to build the first $1B data-center-scale research facility.
After year one: Expand into chemistry and materials science.

Founding team

Leon was chief of staff at Dash0, where he helped scale the company from pre-revenue to unicorn in 15 months, building out GTM, hiring, finance, and investor relations. Before that, he went through YC S24 and worked in strategy consulting at Bain & Company. He is a former professional esports player and ultramarathon runner.

Claudio built a lab-test platform as part of a two-person dev team, that processed over 40 million COVID tests across 5,000 sites and served one million weekly users. He later began studying molecular biotechnology at TUM. He has built automated workcells for AI-bio companies and actively contributes to PyLabRobot, the most widely used open-source lab-automation library.

We are running multiple work trials every week for founding engineers, AI researchers, and scientists to bring together an exceptional early team ready to build this with us.