AI code generation is removing the software engineering bottleneck that previously limited full-stack hardware startups, allowing companies like Anduril to iterate far faster with…

SpaceX's success inspired a wave of space startups; Exosat targets sovereign satellite internet demand, reflecting a broader space stack buildout including orbital solar and space…

Icarus's solar-powered U2-style aircraft provides persistent overwatch and drone communications — a critical enabler for modern drone warfare — winning early seven-figure contract…

Icarus's high-altitude solar UAV provides persistent overwatch and comms relay for drone operations, addressing a critical DoD gap; early seven-figure contracts signal new defense…

Nine Mothers' computer-vision-guided shotgun turret addresses an existential DoD need: protecting scarce elite special forces from commodity drone attacks deep behind enemy lines,…

Nine Mothers' CV-guided shotgun turret is nearly the sole solution for protecting high-value special operators from cheap drone attacks deep behind enemy lines; special forces pro…

Knox Metals is reshoring metal manufacturing in Detroit to serve the new defense tech startup ecosystem; legacy suppliers cannot match the speed new startups require, creating a S…

Nox Metal is revitalizing US metal manufacturing in Detroit to serve new defense tech startups that need faster supply chains than legacy providers can deliver, achieving software…

Nox Metal is revitalizing hollowed-out US metal manufacturing in Detroit to serve new defense tech startups that need speed and responsiveness legacy suppliers cannot provide, ach…

A100 GPU hourly costs are appreciating rather than depreciating due to insatiable AI compute demand exceeding supply, signaling sustained pricing power and a multi-year data cente…

Nvidia A100/H100 GPU hourly costs are rising due to extreme demand, while each architecture generation reduces floating-point precision (FP32→FP8) because LLM workloads don't need…

A100/H100 GPU hourly costs are rising due to structural supply-demand imbalance, while Nvidia's architecture roadmap (FP32→FP16→FP8→FP2) reveals precision headroom that startups l…

As Nvidia architectures push to lower precision (FP32→FP8), Lamb Labs is building custom processors exploiting the insight that LLMs don't need full floating-point precision, crea…

Lamb Labs is building new compute processors targeting the industry-wide shift to lower precision (FP8/FP2) where Nvidia's general-purpose architecture carries overhead; custom si…

Dipole Labs is replacing electronic data center switches with fully optical switches to eliminate the interconnect bottleneck that limits GPU cluster performance as GPU speeds out…

Electronic switches are becoming the bottleneck as GPU speeds scale; Dipole Labs' all-photonic switch removes optical-electrical-optical conversion, enabling direct GPU-to-GPU pho…

Electronic switches are now the bottleneck as GPU speeds scale; Dipole Labs' all-photonic switch removes this constraint by routing GPU-to-GPU traffic entirely with light, unlocki…

Salesforce's system-of-record moat holds because agents will use software more than humans; Slack's collaboration data and new AI harness position CRM to capture the agent-as-cust…

Salesforce is evolving from system of record into an AI harness where agents do work inside the platform, leveraging Slack's collaboration data as a moat; this could sustain SaaS…

Salesforce's system-of-record moat is intact and strengthening as AI agents become primary software users; Slack's collaboration data gives Salesforce a harness advantage where ag…

Juicebox evolved from LLM-powered search to full-cycle recruiting agents that contact candidates and schedule interviews, doubling or tripling revenue per customer as enterprises…

Juicebox evolved from LLM-powered candidate search to an autonomous agent that contacts, schedules, and screens candidates; customers pay 2-3x more because the agent completes the…

Juicebox evolved from LLM-powered people search to an autonomous recruiting agent that contacts candidates and schedules interviews, driving 2-3x revenue expansion per customer as…

The data/RL environment category has exploded from Scale AI's 2016 niche to over a dozen YC companies generating $10M–$100M+ annually, as frontier labs spend billions on proprieta…

Scale AI created the data-labeling category in 2016; today the market has exploded with labs spending ~$1B annually on proprietary data and RL environments, spawning a dozen YC co…

Mercor expanded from data labeling into reinforcement learning environments, capturing part of the ~$1B lab spend on post-training data; RL environments for finance, coding, and r…

Robotics foundation models (PI) require vertical-specific fine-tuning on proprietary real-world data to achieve the real-time reliability needed for tasks like data center cabling…

Boost Robotic builds robots for data center cabling that require fine-tuning foundation models on custom environment data to achieve real-time reliability in specific physical set…

Robotics foundation models must be fine-tuned on proprietary vertical data (thousands of hours of cabling footage) to achieve the real-time reliability required for physical tasks…

PI's robotics foundation models are the new starting point for all YC robotics companies, but none work out-of-the-box — every vertical (data center cabling, box packing) requires…

PI's generalist robotics models (π0) serve as the foundation that every YC robotics startup fine-tunes for vertical tasks; out-of-the-box performance is insufficient for real-time…

Physical Intelligence's robotics foundation models serve as a starting point, but every YC robotics company must fine-tune them on proprietary data for their specific vertical to…

All YC robotics companies fine-tune Physical Intelligence base models; Ultra's thousands of hours of box-packing video creates a defensible data moat for warehouse automation.