Startups Are Moving From Bits to Atoms
Y Combinator
Experienced Founders And Solo Builders 0:00
The episode opens by noting a shift toward older, more seasoned founders, many in their late 30s through 50s, describing a resurgence of what they call the experienced founder. There is also a passing observation that although many people say they want to build a YC for solo founders, YC itself already fills that role. The hosts frame the current moment as unusual, since AI models that failed at a task a month ago can suddenly start working after only minor adjustments.
Hard Tech And Faster Growth 1:01
Diana shares data from an analysis of all companies accepted over the past 12 to 18 months. The share of hard tech companies, meaning startups that build physical things rather than just software, has risen from 8 percent to 20 percent of the batch. Growth speed has also jumped: the median YC company used to enter pre-revenue and reach about 8,000 dollars in monthly revenue by the end of the batch, but now the median company reaches 20,000 dollars in monthly revenue over the same three-month period.
Categories Driving The Hard Tech Rise 2:01
Robotics has grown from 1 percent to 6 or 7 percent of the batch. Industrial manufacturing, tied to rebuilding production in the United States, has grown from 4 percent to 10 percent. Defense startups have grown from about 1.5 percent to 5 percent. Companies building semiconductors or photonics for AI compute have grown from about 1 percent to nearly 4 percent, and power infrastructure companies, needed to support that compute, have grown from 1 percent to nearly 3 percent. Across these physical, atom-based categories, representation has tripled or quintupled.
Technical Founders And AI Leverage 3:30
The current summer batch shows one in six founders holding a PhD, far more than in the past, reflecting the deep technical expertise needed for fields like silicon photonics. AI coding tools are also removing a historic bottleneck in hard tech: previously, building something like a defense company required large teams of elite software engineers, but code generation now lets a small team do work that used to need hundreds of hires, changing the underlying economics of building physical products.
Three Macro Trends Behind Atoms 5:00
Three broader trends are named as drivers. SpaceX's successful IPO has inspired a new generation of founders building in space, including companies like Exosat, working on a sovereign Starlink-style solution, and Beyond Reach Labs, building solar panels for satellites. A second trend is a generation of founders shaped by recent war coverage, drawn to defense work. Examples include Icarus, building a solar-powered spy plane that also handles communications and has landed seven-figure Department of War contracts, and Nine Mothers, building a shotgun-style anti-drone turret with computer vision used by special forces.
Manufacturing Supply Chains And Compute Demand 8:00
Beyond direct defense sales, dual-use startups are rebuilding supply chains, such as Knox Metals, which is restoring American metal manufacturing in Detroit and growing at software-like speeds because new defense startups need metal faster than legacy suppliers can provide. The third macro trend is compute: demand for AI has made even older Nvidia GPUs appreciate in cost due to scarcity, driving startups into data center construction, power and battery solutions, and new chip designs, including Lambda Labs' new processors and a company called Bot building hardware using lower-precision ternary representations suited to how modern AI models actually run.
Optical Switches And Robotics Infrastructure 12:00
Dipole Labs is building a fully optical switch to replace the electronic switches that route data between GPUs in data centers, aiming to remove a growing speed bottleneck. Robotics is described as the next frontier, with companies building everything from vertical robotics applications to deployment infrastructure and data supply for robotics labs. A benchmark comparison is mentioned where an earlier model scored around 10 percent and a newer one, Astra, reached 60 to 70 percent, suggesting rapid progress toward a breakthrough moment in robotics similar to what happened with chat-based AI.
Venture Capital Returns To Hard Tech 13:59
YC has funded hard tech since 2014, but investors long preferred safer software bets, often saying they only did business software. That preference shifted quickly this year once software stocks dipped and Claude Code gained traction, though some software companies like Salesforce and Snowflake have since posted strong results, suggesting older software models may still hold value if they become essential infrastructure that AI agents themselves rely on.
Software Shifts From Records To Action 17:00
The share of accepted companies doing full end-to-end task automation, rather than simple point solutions, has grown from 10 percent to over 25 percent of the batch, and this shift is linked to the jump in median revenue from 8,000 to 20,000 dollars. Examples include a recruiting tool called Juicebox, which evolved from searching for job candidates into an agent that also contacts them and schedules interviews, increasing revenue per account without eliminating the recruiter's role, since human judgment on things like culture fit still matters.
Fast Revenue And The Data Business 22:00
Some companies are now jumping from zero to seven figures in revenue within a single three-month batch, something that used to take about 18 months, aided by founders running many parallel coding agent sessions to mature their products quickly. A separate fast-growing category sells data or RL environments, meaning simulated tasks used to train AI models, to major labs. YC has funded more than a dozen such companies each earning over 10 million dollars a year, some reaching hundreds of millions, with major labs reportedly spending around a billion dollars overall on this kind of data, including robotics-focused data drawn from real-world industrial and physical tasks.
Fine-Tuning Models On Proprietary Data 26:30
Using proprietary data to fine-tune models is becoming a bigger factor in building companies, since open-weight models are already close to frontier quality and can be specialized to outperform general models in narrow tasks. This effect looks even larger in robotics than in language models, because robots operate in physical 3D space with far more degrees of freedom than language, and need to respond in real time rather than pausing and resuming like an LLM can. A company called Boost Robotics, which builds robots for cabling work in data centers, illustrates this: rather than using a general robotics foundation model out of the box, it makes more sense to fine-tune on data specific to that environment. YC companies working with physical intelligence models reportedly all fine-tune those models rather than using them unmodified, and a company called Ultra starts from a base model but improves dramatically after training on thousands of hours of footage of tasks like packing boxes. The same pattern shows up elsewhere, such as using coding transcripts to identify top coders and train better coding models, or using platforms like TikTok's viewing data to train more compelling video generation tools like Seedance.
Experienced Founders Are Resurging 29:30
A growing number of the most capable founders now seen are in their late 30s, 40s, or even 50s, marking a resurgence of the experienced founder. Data from YC shows solo founders made up only about 5 percent of accepted companies a year ago, and now make up 18 to 19 percent, the sharpest spike YC has observed. Historically, strong companies needed a pairing of a persuasive, sales-minded person and a world-class technologist, but that combination is now less necessary since knowing what to build and what to prompt has become more valuable than raw coding ability. Some of the most famous single founders at YC's start, including Apoorva Mehta of Instacart and Brian Armstrong of Coinbase, succeeded alone because they were exceptional across selling, building, and vision, a rare combination. Now the bar for that kind of exceptional individual capability is lower, though most successful companies still eventually add co-founders after gaining early traction, rather than starting as a fully split partnership from day one.
Experience Now Beats Gatekeeping 33:01
Peter Steinberger, in his early 40s with a background as a dev manager, exemplifies how people who have already worked in startups and understand where the pitfalls are tend to do unusually well once they adopt AI tools seriously. Old gatekeeping norms, like needing a co-founder or elite investor backing, matter less now, since knowing what to build is the most important factor. Experience managing engineering teams also seems to translate into managing coding agents effectively, sometimes better than a very talented but inexperienced person could manage, since experienced managers know how to direct work without over-personalizing the process. The practical advice offered is simply to start prompting and building immediately, since new models keep solving problems overnight that were unsolvable weeks earlier, and there is little excuse left for waiting.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
