PearX Demo Day: Stable 3D Worlds, On-Device AI Chips and Private Assistants

Pear VC put 16 startups on stage at its latest PearX demo day in San Francisco, held in the week before Oct. 5, 2026. TechCrunch attended in person and reported its picks for the companies drawing the most investor conversation on Oct. 5 TechCrunch.

PearX is a 12-week program run by Pear VC, a firm focused on pre-seed and seed startups. Cohorts are capped at 20 startups. Investments through the program can run as high as $2 million and do not use standard terms TechCrunch.

Speridlabs was among the companies named as drawing attention. It builds spatial foundation models, AI models that understand 3D space, for robotics, gaming and special effects. Its product Mundus was described as a 3D Midjourney that keeps geometry persistent when a part is changed.

2D image generators can swap texture or style frame by frame. Production 3D assets cannot tolerate that drift. Mesh consistency means the shape holds together, scene layout means objects stay in place, and physical plausibility means the result could exist in the real world. Persistent geometry means an edit updates without breaking those properties. For robotics, that makes synthetic training data usable. For gaming and effects, it speeds work on assets that must still go through rigging, lighting and physics.

Saia was also named. The company is building a chip that runs AI inference, the step where a trained model makes predictions, directly on device out of flash storage to bypass traditional memory. Saia claims eight times the capacity while using four times less power than Nvidia's Jetson. Founder Ayaan Govil is 20 years old. The company plans to start fabricating test chips next year and aims for mass production by 2028 TechCrunch.

At the edge, meaning on the device rather than in the cloud, performance is often limited by memory bandwidth, how fast data can be moved, and power limits, rather than raw calculation speed. Running directly from flash requires solving data path and endurance issues, how data moves and how long storage lasts under heavy use. The timeline follows a normal pattern for silicon. Test chips next year and 2028 volume leave time for bring-up, yield learning and software work. The 8x capacity and 4x power figures are company claims to be tested on silicon, not independent benchmarks.

Ren was named as well. It is developing a secure AI personal assistant that keeps data on-device where possible or uses a secure private cloud. Veros, the fourth company detailed in the report, is an AI-native trust and estate planner.

Assistants need access to calendar, mail, files and system APIs, the connections that let software use other software. Trust and estate work involves long-lived, high-liability documents across jurisdictions and family structures. Both need retrieval, permissions handling and auditability, clear records of what was accessed, more than raw model quality. On-device processing with private cloud for overflow is now a common setup for sensitive workloads. Legal planning tools also need version control, attorney review workflows and integration with custodians and courts.

The broader context here sits apart from the pitches themselves. In my view, the batch gives a small read on where pre-seed risk is currently accepted: world models that produce stable 3D structure, inference hardware that removes the memory bottleneck, and agents trusted with private data or legal outcomes. Each is hard in a different way. Spatial consistency is a research problem. Edge silicon is a capital and execution problem. Private assistants and estate planning are distribution and liability problems.

Stepping back to look at PearX itself, the program stays small by design. A 12-week, sub-20-company cohort with checks up to $2 million on non-standard terms gives Pear VC early access without forcing founders into one standard accelerator contract. At pre-seed, company formation, technical validation and first design partners count more than demo-day presentation.

Looking past demo day itself, what these approaches could enable is concrete. Cheaper, lower-power on-device inference expands where models can run without connectivity. Stable generative 3D shortens the loop from prompt to simulatable or playable asset. Private-by-default assistants and domain-specific planners make it easier to delegate work professionals currently avoid handing to general chatbots. None of that is guaranteed at demo day. The work after demo day counts.