CHAPTERS
- 0:00 – 2:02
YC’s latest batch stats: hard tech surge and faster revenue ramps
Garry and Diana set the context: YC sees patterns early, and recent batches show two big shifts—more companies building in the physical world and startups reaching meaningful revenue much faster during the batch. They frame the episode around what builders should know about what’s changing in startups right now.
- •YC analyzes the last 12–18 months of accepted companies for trend shifts
- •Hard tech share rises sharply (from ~8% to ~20%)
- •Median company enters YC at ~zero revenue but now exits with higher MRR
- •Episode goal: what’s state-of-the-art and what founders should do next
- 2:02 – 3:29
What “bits to atoms” looks like: robotics, manufacturing, defense, compute, power
The hosts define hard tech as companies that touch atoms, then break down which physical categories are growing inside YC. They cite large percentage increases across robotics, industrial manufacturing, defense, semiconductors/photonics, and power infrastructure.
- •Robotics grows from ~1% to ~6–7% of the batch
- •Industrial manufacturing rises from ~4% to ~10%
- •Defense increases from ~1.5% to ~5%
- •Semiconductor/photonics and power infrastructure both expand quickly
- •Overall: ‘atom-stack’ companies have tripled/quintupled in share
- 3:29 – 4:05
The technical-founder boom: more PhDs and deeper domain expertise
YC is increasingly funding founders with heavy technical credentials because many of these new hard-tech areas require research-level depth. They note a striking stat about PhD prevalence and argue these founders are performing disproportionately well.
- •In the current summer batch, ~1 in 6 founders has a PhD
- •Deep research domains (e.g., silicon photonics) demand specialized backgrounds
- •YC is intentionally backing more highly technical teams
- •Observed correlation: expert founders performing strongly
- 4:05 – 5:52
Why AI makes hard tech easier: codegen removes a key bottleneck
Garry and Jared argue hard tech is becoming more feasible because AI accelerates software and research work that used to require huge teams. Code generation changes the economics of building full-stack hardware+software companies by reducing dependence on scarce elite engineers.
- •Hard tech used to be constrained by complex software plus supply chains
- •AI/codegen reduces the need for very large software teams
- •Startups can move faster without competing head-on with Big Tech hiring
- •Models accelerate research and enable earlier breakthroughs
- 5:52 – 8:01
Defense and manufacturing renaissance: startups beating the primes
They describe a renewed wave of defense and dual-use startups, driven by geopolitics and a willingness to modernize procurement. Garry shares examples of defense startups winning meaningful contracts and explains why agile startups can out-innovate legacy defense primes.
- •Founder motivation rises as war becomes more salient culturally
- •Examples: Icarus (solar-powered U-2-like platform) and Nine Mothers (anti-drone turret)
- •Startups winning seven-figure contracts; demand for modern drone-era capabilities
- •Critique of legacy defense primes (cost-plus, slower iteration)
- •Dual-use startups also build supply-chain and manufacturing capabilities
- 8:01 – 9:48
Rebuilding U.S. industrial capacity: Knox Metals and ‘software-like’ growth
The conversation shifts to industrial manufacturing and reshoring, highlighting Knox Metals as an example of new suppliers moving at startup speed. Garry connects this to historical patterns: new fast-moving ecosystems prefer buying from similarly fast-moving vendors.
- •Knox Metals aims to rebuild domestic metal manufacturing capacity
- •Defense-tech customer demand pulls new industrial suppliers forward
- •Legacy suppliers can’t match the pace expected by modern startups
- •Analogy to early YC: Stripe vs legacy payment providers
- •Hardware companies can now grow at ‘software growth rates’
- 9:48 – 11:44
AI compute becomes physical infrastructure: data centers, power, and new silicon
Diana explains how AI demand turns compute into an ‘atoms’ problem: construction, power, and supply constraints are now central. They discuss GPU scarcity, rapid data center buildouts, and startups building alternatives across hardware architecture and the semiconductor stack.
- •A100 GPU-hour pricing appreciating signals severe compute scarcity
- •Startups emerge across site construction, planning software, buildout, and energy
- •New chips and architectures pursue alternatives to Nvidia
- •Trend toward lower precision (FP32→FP16→FP8→lower) fits LLM needs
- •Examples: Lamb Labs (processors) and Bot (ternary/low-precision architecture)
- 11:44 – 12:32
Data-center bottlenecks and photonics: fully optical switching
Garry highlights interconnect limitations inside modern AI data centers—switching and routing can bottleneck GPU performance. A featured company aims to replace electronic switches with optical switching to keep up with GPU speeds.
- •Switches increasingly limit GPU-to-GPU communication throughput
- •Electronic switching struggles to keep pace with GPU progress
- •Dipole Labs building a fully optical (photon-based) switch
- •Goal: faster interconnect and less bottlenecking for AI workloads
- 12:32 – 15:24
Robotics nearing its ‘ChatGPT moment’: vertical robots and the deployment stack
They argue robotics is approaching a breakthrough era, though not fully there yet, and that a new scaling law may be forming. YC sees growth in startups across vertical robotics, deployment infrastructure, and data pipelines feeding robotics labs.
- •Industry sentiment: robotics is close to a major inflection point
- •Startups span vertical robots, deployment/infrastructure, and data providers
- •Benchmarks improving quickly (Astra vs earlier baselines)
- •Progress often depends on pairing models with the right ‘harness’
- 15:24 – 17:58
Software isn’t dead—systems of record must become AI ‘harnesses’
Jared and Garry push back on ‘SaaS is over’ narratives, arguing moats persist but software must evolve. Systems of record need to become the environment where agents do work, not just store data—otherwise they risk being unbundled via interoperability layers.
- •SaaS stocks recovering suggests classic moats still matter
- •Agents as customers: build software agents want to use
- •‘Harness wars’ emerging (tools that orchestrate models + workflows)
- •Systems of record must enable work execution, not just read/write data
- •Risk: releasing interfaces (e.g., MCP) could make switching easier without a harness
- 17:58 – 22:51
Why YC startups are growing faster: end-to-end agents that ‘do the job’
Diana shares batch-level evidence that agentic products are capturing more value by completing full workflows. This shift from point solutions to end-to-end task execution helps explain the jump in median revenue outcomes during YC and enterprises paying earlier.
- •Full-stack, end-to-end task companies rise from ~10% to >25% of the batch
- •Median end-of-batch MRR increases from ~$8K to ~$20K
- •Some companies go from $0 to seven figures in revenue within a 3-month batch
- •Agentic coding speeds product maturity and iteration cycles
- •Example: recruiting tools evolving from search to outreach/scheduling agents
- 22:51 – 27:07
The hidden boom: data and RL environments as massive businesses
They reveal a less visible but rapidly growing category: companies selling data and RL environments to frontier labs. Many stay quiet for competitive reasons, yet YC has funded numerous firms reaching $10M+/year quickly, with labs reportedly spending heavily.
- •In 2 years, YC funded a dozen+ companies each >$10M/year in data/RL env revenue
- •Some reach hundreds of millions in revenue within a few years
- •Data is a core ‘leg’ of scaling laws alongside compute
- •RL environments become a major spend area for labs; heavy customization by domain
- •Robotics increases demand for real-world/egocentric/teleop task datasets
- 27:07 – 29:14
Why robotics needs specialized models: fine-tuning, real-time constraints, vertical data
Diana and Jared explain why robotics differs from LLMs: higher-dimensional physical spaces and real-time safety needs. They argue vertical robotics will require specialized models fine-tuned on proprietary data, with examples of YC companies adapting foundation models to niche tasks.
- •Robotics models operate in 3D physical space with more degrees of freedom than language
- •Real-time reaction and safety constraints make ‘slow thinking’ less viable
- •Vertical deployments often require fine-tuned models rather than out-of-box use
- •Examples: data-center cabling robots; packing/boxing tasks fine-tuned on extensive footage
- •Proprietary data + near-frontier open-weight models enable strong specialization
- 29:14 – 34:58
The rise of the solo founder and the return of experienced builders
They discuss two people-trends: solo founders are becoming more common, and experienced founders (30s–50s) are resurging. AI tooling lowers the barrier to building, while management experience can translate into effectively running ‘teams’ of coding agents.
- •Solo founders rise from ~5% to ~18–19% of accepted companies
- •Historical precedents: founders often start solo, add co-founders later
- •AI reduces the need for exceptional ‘builder’ capability to get moving
- •Experienced founders have taste, pattern recognition, and ‘know what to build’
- •Managing coding agents resembles managing engineering teams
- 34:58 – 36:28
What founders should do right now: start building with the newest models
In closing, Garry gives direct advice: begin experimenting and prompting, because capability leaps can suddenly make previously stuck problems solvable. The group emphasizes this as a rare window where iteration speed is compounding and builders should take advantage.
- •Use new model releases to revisit previously unsolved bugs and blockers
- •Leverage agents to mine past chats/issues and attempt fixes again
- •Expect rapid capability jumps over the next 18–36 months (uncertain horizon)
- •Core call to action: ship faster and let tooling amplify your output
