Startups Are Moving From Bits to Atoms
In a Nutshell
**Startups are shifting from software to hard tech—robotics, defense, semiconductors, and infrastructure—now comprising 20% of recent YC batches versus 8% before, with median revenue growth jumping from $8K to $20K MRR.** AI is accelerating this by enabling faster iteration, while macro trends like SpaceX's success and visible conflicts are driving founders toward atoms. Experienced solo founders in their 30s-50s are surging to 18-19% of batches, leveraging agentic tools to build full-stack products that automate entire jobs at software-like speeds.
These notes were generated by AI and may contain inaccuracies.
Some of the most powerful founders lately are in their late 30s, 40s, or 50s, marking a resurgence of the experienced founder.
Founders can now wire up a new model and see things working that were broken just a month prior.
YC analyzed companies accepted in the last 18 months and identified two major trends. The number of hard tech companies has risen from 8% to 20% of the batch. The median revenue growth has accelerated: previously companies reached $8K MRR by batch end, but now the median is $20K MRR.
Hard tech companies are those that touch atoms rather than just bits. Robotics has grown from 1% to 6-7% of the batch. Industrial manufacturing, bringing production back to the US, has increased from 4% to 10%. Defense has risen from 1.5% to 5%. Semiconductor stack and photonics companies have grown from 1% to nearly 4%. Power infrastructure companies have increased from 1% to nearly 3%.
One in six founders in the current summer batch has a PhD, a higher rate than historically. Technical expertise is required for fields like silicon photonics.
AI models are accelerating scientific research, enabling startups to achieve breakthroughs earlier. Previously, hard tech was limited by supply chains and software engineering talent. Now, with codegen tools, companies like Anduril can move faster without needing thousands of top-tier engineers.
Three macro trends are driving the shift toward atoms. SpaceX's successful IPO has inspired a generation of founders to build in space. Companies like Exosat are developing sovereign Starlink solutions, while Beyond Reach Labs is creating solar panels for satellites. StarCloud and similar ventures will need power infrastructure in space.
Founders have grown up with visible conflicts on social media and want to contribute meaningfully. Icarus, funded last fall, builds solar-powered U-2 spy planes providing overwatch and communications for drone operations, securing seven-figure contracts with the Department of War. Nine Mothers, funded last spring, produces anti-drone defense systems using computer vision for special forces protection.
Many defense startups are dual-use, selling to both government and private sectors. Knox Metals is rebuilding America's metal supply chain in Detroit, addressing the hollowed-out metal industry. Their customers include new defense tech startups that need rapid metal production, which legacy suppliers cannot provide at the required pace.
Knox Metals is growing at software-like rates. New defense tech startups prefer working with agile suppliers that match their speed, similar to how early YC companies chose Stripe over legacy vendors.
AI demand is driving data center construction. Nvidia A100 GPUs are appreciating in value due to supply constraints. Startups are addressing the full stack from site construction to power solutions including energy and batteries. Lamb Labs is building new processors as alternatives to Nvidia. Bot is developing custom hardware using ternary representation, aligning with Nvidia's trend of reducing floating-point precision across generations from FP32 to FP16 to FP8.
Dipole Labs is building the first fully optical switch for data centers, replacing electronic switches that have become bottlenecks as GPU speeds increase. The optical approach routes photons directly from GPU to GPU.
Robotics companies are building infrastructure for deployment and data collection, anticipating a ChatGPT moment. Astra has achieved 60-70% on benchmarks where previous systems reached only 10%. The industry is discovering new scaling laws for robotics.
YC has funded hard tech since 2014, but VCs previously avoided these companies in favor of B2B SaaS. Silicon Valley venture capital is returning to its roots of funding hard tech after a decade focused on software.
Investor interest in hard tech increased rapidly when SaaS stocks declined earlier this year and Claude Code gained traction. Interest was visible by demo day. Some SaaS stocks like Salesforce and Snowflake have since recovered with strong earnings.
System of record companies remain valuable when they function as harnesses that agents use. Salesforce benefits from its position as a collaboration system of record and its Slack integration.
System of record companies must either become harnesses where work happens or risk losing their data moat through MCP releases. The model plus harness combination produces results, as seen with Astra reaching over 90% on RAG benchmarks after custom harness integration.
Companies building full-stack, end-to-end solutions have grown from 10% to over 25% of the batch. These products complete entire jobs rather than serving as point solutions. Agentic coding since Opus 4.5 has enabled this shift. Companies solving insurance brokerage, clinical intake, and medical billing are seeing significant revenue growth.
Products that automate entire jobs command higher prices than systems of record. Large enterprises write substantial checks early when the value proposition is compelling.
Juicebox began as an LLM-powered people search tool. Their new agent product contacts candidates, schedules interviews, and handles additional workflow steps. Per-account revenue is expected to double or triple as customers adopt more agents. Recruiters appreciate the automation of rote outreach, freeing them for culture-fit assessment and human elements.
Companies are reaching seven figures in revenue during the three-month batch, compared to 18 months historically. Agentic coding enables rapid product maturity through parallel coding agent sessions.
Over the last two years, YC has funded more than a dozen companies each generating over $10 million annually selling data or RL environments to AI labs, with some reaching hundreds of millions. This category has grown from a niche to a major segment.
Once something is working, there's a natural tendency to keep it private. Data remains one of the critical legs of scaling laws, and without it, models cannot improve substantially beyond current capabilities. RL environments are seeing significant customization for specific use cases, such as finance applications where practitioners combine domain expertise, computer science, and systems work to go infinitely deep. RL has emerged as a major engine for model development, even as benchmark maxing becomes more expensive, and appears to be a permanent approach for large model companies.
Major labs are reportedly spending around a billion dollars on RL environments. A real business has emerged around creating these environments, with long-term horizon tasks being developed. This trend is extending into robotics as labs seek to solve physical world AI deployment. Real-world environments, including egocentric data and teoptic tasks, are becoming a significant data category, with labs making eight and nine-figure deals with companies in this space.
YC companies in the summer 2026 batch working on this include Practis Robotics and a company collecting data from a global network of industrial production sites. Another company, Deep Reach, works with local entrepreneurs worldwide to gather human archive data. Winter 2026 also saw multiple companies in this space.
All systems of record may need to become AI harnesses and potentially train their own models. Tools like River AI and Tinker are becoming interesting for this purpose. Open-source tools like OpenClaw allow users to train custom models with automatic data cleaning. While not yet a major factor, this capability is expected to grow significantly.
Having proprietary data enables special-purpose training of open-weight models that can outperform frontier models on specific tasks. This trend is expected to be even more significant in robotics, where foundation models have different characteristics than LLMs.
Robotics foundation models operate in physical 3D space with far more degrees of freedom than language models that model reality as language. For robots to work in specific verticals like data center operations, fine-tuned models trained on custom data are more effective than general models. Real-time response requirements and action planning present unique challenges for robotics that differ from LLMs, which can process tasks asynchronously.
Boost Robotic builds robots for data center cabling, where fine-tuned models are necessary because real-world conditions can change unpredictably. All YC companies using Physical Intelligence models for robotics deployment are fine-tuning these models rather than using them out of the box, even for specific applications like data center cable installation.
Ultra starts with Physical Intelligence models but enhances them with thousands of hours of footage of specific tasks like putting items in boxes, making the system highly specialized for that function. The pattern of using proprietary data to improve models mirrors what's happening with coding agents and content platforms, where companies leverage their unique data to create better specialized models.
The most capable founders recently have been in their late 30s, 40s, and 50s, representing a resurgence of experienced founders. Solo founders now comprise 18-19% of accepted YC companies, up from 5% a year ago, marking the highest spike recorded. Nearly one-fifth of the current batch consists of solo founders.
The traditional requirement for complementary co-founder skills is shifting because knowing what to prompt and what to build has become more valuable than pure technical implementation skills. Historically successful solo founders include Max Mullen with Instacart, Brian Armstrong with Coinbase, and Parker Conrad, who later added a CTO after initial acceptance.
While solo founders are increasing in YC batches, successful companies still tend to add co-founders as they progress. The trend shows founders bringing on co-founders later in the company lifecycle after achieving initial traction, which may result in different equity arrangements compared to traditional 50/50 splits at founding.
Founders who have been around multiple times are performing exceptionally well. Peter Steinberger exemplifies this pattern, having worked as a dev manager and on previous startups before becoming deeply engaged with AI tools. Experienced founders bring knowledge of potential pitfalls and refined taste for what to build.
The traditional gatekept requirements for startup success, such as needing specific co-founders or investor connections, are becoming less relevant. The critical factor now is knowing what to build, and experienced founders with strong opinions are particularly well-positioned to succeed.
Managing coding agents shares similarities with managing human teams. Founders with experience managing engineering teams, such as Peter Steinberger, Boris, and Toby from Shopify, excel at orchestrating large teams of coding agents. Their experience allows them to manage agents more effectively than less experienced founders, partly because they approach agent management with greater patience and understanding.
The immediate advice is to start prompting with current AI tools. Opening GPT-6 revealed agents that are palpably smarter, capable of resolving issues that previously required manual deep dives. The capability to wake up and wire up a new model that resolves previously intractable problems represents a unique historical moment expected to continue for 18-36 months.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
