Speridlabs exits stealth to build a spatial AI foundation model

Why it matters

If Spatial AI models can reliably parse and generate 3D scenes, they could unlock better robotics, AR interfaces, simulation, and digital twin tooling. Operators and investors watch these launches for signals on where foundation models are headed beyond text and 2D vision.

Speridlabs came out of stealth on May 21 and said it is building Spatial AI, a foundation model that understands and generates the 3D world, in a thread on X. The lab added that it is backed by Base10 Partners (@Base10Partners) and Pear VC (@pearvc), and linked to a launch essay, "The shape of intelligence", along with a short video teaser.

https://www.youtube.com/watch?v=cLqCAS2Rn3A

In the post, the team argues that most of what is called computer vision is a 3D problem in disguise, and that the field has optimized for narrow 2D benchmarks like detection, segmentation, tracking, and reconstruction rather than a single world model. As one symptom, they point to autonomous driving stacks that run more than 40 separate AI models stitched together by heuristics. The goal, they say, is one shared intelligence layer: a model that holds a coherent 3D representation and stays consistent across viewpoint, occlusion, and time, with reconstruction, understanding, editing, navigation, planning, and generation as different queries to the same system. "In the physical world, plausible is not enough. You also need true."

Speridlabs outlines why a spatial foundation model has not emerged yet: data, representation, evaluation, and systems. On data, they cite a gap between text/image/video corpora and 3D, estimating that public 3D scene datasets total around 100,000 scenes with roughly 5,000 added per day, far short of the scale used to train language and image models. On representation, they note 3D lacks a universally accepted primitive that is compact, learnable, editable, and good for reasoning. On evaluation, they argue that spatial consistency across viewpoints, long-horizon reasoning, partial observation, and occlusion are not yet measured by standard benchmarks. On systems, they say 3D workloads are heavy across training, serving, tooling, viewers, formats, and pipelines.

What has changed now, per the lab: there is a foundation-model training culture, generative models are strong enough to serve as priors, compute is expensive but accessible outside the largest labs, and demand for consistency across viewpoints and time is broad across industries.

The roadmap is staged. First, build stable spatial priors from real-world capture, with reconstruction as the first foundation layer. Next, train a controllable 3D generative model that also serves as a data engine to address the 3D data gap. On top of those, build the world model itself, first static, then dynamic. Each stage is intended to be useful on its own and a step toward the end state.

Speridlabs also says it will be open by default: publishing, benchmarking, and releasing weights to help establish Spatial AI as a public scientific field. The X post and essay did not include model specifications, APIs, pricing, or release timing.