Bulut Labs began as a side project in early 2021 — a notebook of ideas about what would happen if you trained a foundation model on the world rather than the web. We have spent five years sharpening that thesis, prototyping in our spare hours, and refusing to chase the language-model gold rush. We are entering public-build mode in 2026.
Architectures are public. Compute is rentable. Talent is poachable. The corpus we are building — slowly, with consent — is the only durable advantage. We treat it with the seriousness of a crown jewel and the patience of a librarian.
Every byte we train on must trace back to a consent ledger entry. Every model checkpoint we ship must trace forward to specific data sources. If we cannot defend a training decision in court, in front of a regulator, or to the person filmed, we should not have made it.
We do not announce shipping until we ship. We do not list customers we do not have. We do not put logos in our deck of investors who turned us down. The discipline starts on this website and works inward.
We have no headcount yet. When we do, we will hire small, expensive, opinionated people who care more about what they ship than what they list. Pedigree is a tiebreaker, not a filter.
The customers we want are people who run fleets, factories, ports and grids — people whose pager goes off when we are wrong. Everything we ship is engineered for that audience, not for benchmark leaderboards.
We are not selling a model. We are selling access to a research practice — a way of thinking about physical AI that customers can buy into early and shape from the inside. The lab itself is what we are building.
A side-project notebook with a single sentence on the cover: "Train on the world, not the web." Three friends start meeting weekly to argue about whether language models will run out of useful tokens. The notebook is still on a shelf in İstanbul.
A laptop-trained encoder on a few hundred GB of permissioned dashcam footage. Useless. Useful for proving the data engine architecture we want to build is actually buildable.
Two years of slow technical work — multimodal calibration, lidar tokenisation experiments, the first sketches of stratus. We watch the language-model boom from a respectful distance and stay out of it.
stratus training framework on paperThe rig becomes road-ready and starts producing time-synced multimodal data across İstanbul and three rural Anatolian provinces. Provenance ledger goes from prototype to production. The corpus crosses ten terabytes — small by frontier standards, but every byte permissioned.
Five years in, we are stepping out of stealth. Applying to AWS Activate and NVIDIA Inception. Recruiting design partners. Publishing the first version of this website. Looking, finally, for a few people who want to build the next era with us.
Target year for our first public release — a 14-B distillation under research licence. Contingent on funding, design-partner traction and the integrity of our roadmap. We will not announce until we ship.
İstanbul-based founder. Building Bulut Labs full-time since 2026; building toward it on nights and weekends since 2021. Background in software, fintech and data infrastructure. Writes most of the code, signs every contract, answers hello@bulutlabs.com personally.
Looking for a research-minded co-founder with deep multimodal-ML experience to lead the BULUT-1 model line. Equity meaningful, scope wide, conviction required. Email the founder directly with a link to the most interesting thing you have shipped.
CUDA-comfortable, distributed-training-curious engineer to own stratus end to end. Will work directly with the founder on architecture choices and have outsized influence on what the company looks like in 2027.
Builds, calibrates and operates the multimodal capture pipeline. Will spend a meaningful percentage of their first year in the field, on rooftops, in vehicles, with sensors. Hardware-friendly software engineers welcome.
| Entity | Bulut Labs · Türkiye-based |
| Founded | 2021 · part-time until 2026 |
| Stage | Pre-seed · bootstrapped |
| Investors to date | none |
| Board | founder-only · will expand with first round |
| Headcount | founder-led · first hire targeted Q4 2026 |
| Currently applying to | AWS Activate · NVIDIA Inception · Google for Startups |
| Will incorporate as Delaware C-Corp | when first institutional round closes |
Three conversations are open: design partners, cloud credits, and the first external hires. If any of them is yours, we should probably talk. Email the founder directly.