Startups Are Moving From Bits to Atoms

YC Batch Trends & Hard Tech Resurgence

• Hard Tech Share Expansion: The proportion of hard tech (physical "atoms" over software "bits") in accepted YC batches rose from 8% to 20% over the last 12–18 months.
  • Robotics: Increased from 1% to 6–7% of the batch.
  • Industrial Manufacturing (US-based): Rose from 4% to 10%.
  • Defense: Grew from 1.5% to 5%.
  • Compute Infrastructure (Semiconductors/Photonics): Increased from 1% to nearly 4%.
  • Power Infrastructure: Grew from 1% to nearly 3%.
• Founder Expertise Shift: The current summer batch features a 1-in-6 founder rate holding a PhD, marking a historic high for technical depth in hard tech.
  • AI Acceleration: Advanced models (e.g., CodeGen) have reduced the need for massive engineering teams, allowing startups to execute full-stack hardware faster than previously possible.
• Macro Drivers for Hard Tech:
  • Space Economy: Following SpaceX's IPO, a new wave of space-focused startups has emerged (e.g., Exosat for sovereign Starlink, Beyond Reach Labs for orbital solar).
  • Defense Modernization: New administration approaches favor agile startups over legacy defense primes, driving demand for dual-use tech (e.g., Icarus solar spy planes, Nine Mothers anti-drone systems).
  • Supply Chain Reshoring: Companies like Knox Metal are rebuilding US metal manufacturing in Detroit to serve high-growth defense tech, achieving "software-like" growth rates.
  • Compute Scarcity: GPU demand has reversed traditional depreciation trends (e.g., NVIDIA A100 prices appreciating), spurring startups in data center build-out and alternative silicon (e.g., LAM Labs, Bot).

Software Evolution & Agentic Workflows

• Revenue Acceleration: Median YC company revenue at the end of the batch jumped from $8k to $20k MRR (from pre-product/pre-revenue).
  • Cause: A shift toward "full-stack" agents that execute entire workflows (e.g., insurance brokering, medical billing) rather than point solutions.
  • Validation: Startups like Juicebox saw revenue per account double or triple after launching agents that handle end-to-end recruiting tasks (sourcing, contacting, scheduling).
• System of Record Transformation:
  • AI Harnesses: Traditional SaaS systems of record must evolve into "AI harnesses" where models can read, write, and execute tasks internally (e.g., Slack AI, Salesforce) to maintain moats against switching costs.
  • Agent-Centric Sales: Software value is shifting toward products that act as agents; YC batch acceptance for end-to-end task automation rose from 10% to over 25%.
• Data & RL Environment Markets:
  • Scale: YC has funded 12+ companies in the past two years selling data/RL environments to labs, with many generating >$10M/year (some hundreds of millions).
  • Labs Spending: Major AI labs are reportedly spending ~$1B annually on RL environments and custom data.
  • Use Cases: High-value data categories include finance-specific RL environments and egocentric/tele-op robotic data (e.g., Praxis Robotics, DeepReach).

Founder Demographics & The Solo Surge

• Solo Founder Boom: The percentage of accepted solo founders spiked from 5% to ~19% (highest on record).
  • Catalyst: AI coding agents lower the barrier to building, allowing single founders to execute technical and operational tasks that previously required co-founders.
  • Co-founder Trend: Successful solo founders increasingly add co-founders later in the lifecycle rather than starting with them.
• Experienced Founder Resurgence: Successful founders are increasingly in their late 30s, 40s, or 50s.
  • Advantage: Experienced founders possess "taste" (knowing what to build) and management skills transferable to agent teams (managing coding agents akin to managing engineers).
  • Examples: Peter Steinberger (early 40s, ex-dev manager) cited as a canonical example of an experienced, AI-pilled founder.

Forward-Looking Technical Trends

• Robotics Foundation Models (PI):
  • Fine-Tuning Requirement: Out-of-the-box physical intelligence (PI) models are insufficient; successful robotics startups (e.g., Ultra, Boost Robotic) fine-tune models on specific vertical data (e.g., data center cabling, boxing).
  • Real-Time Constraints: Unlike LLMs, robotics requires immediate reaction to physical stimuli, necessitating specialized training data.
• Precision Evolution: New hardware architectures (e.g., Bot) are moving toward ternary representations and lower floating-point precision (FP2), leveraging the fact that LLMs do not require full precision.
• Optical Interconnects: Startups like Dipole Labs are replacing electronic switches with all-photon switching to resolve bottlenecks where switch speeds lag behind GPU speeds.
• Model Training: Systems of record are expected to increasingly train proprietary models on internal data to create specialized, frontier-performance models.