Y Combinator Hard Tech Share Hits 20% as AI Slashes Engineering Costs
Y Combinator
Autonomous coding tools have eliminated software engineering as the primary startup bottleneck, shifting capital and revenue velocity toward physical infrastructure and experienced solo founders.
AI-assisted summary of The State of Startups in 2026 — original source. Check the original for full context.
Source published:
Key ideas
- Hard tech has expanded from 8% to 20% of Y Combinator batches, coinciding with a jump in median Demo Day revenue from $8,000/month historically to $20,000/month. (0:00)
- Solo founders now make up nearly 19% of YC cohorts—an all-time high—as AI coding agents allow domain experts to launch and reach initial traction before recruiting co-founders later in the lifecycle. (30:00)
- Enterprise software has shifted from human-operated SaaS tools to autonomous full-job agents, increasing the share of YC companies executing end-to-end workflows from 10% to more than 25%. (15:00)
- Selling proprietary training data and specialized Reinforcement Learning (RL) simulation environments to frontier AI labs has become a secretive megabusiness, with over a dozen YC startups generating tens to hundreds of millions in revenue. (22:38)
Detailed notes
## Background Published by Y Combinator on September 18, 2026, on their podcast The Lightcone, this episode features partners Garry Tan, Jared Friedman, Diana Hu, and Harj Taggar analyzing structural shifts across recent startup cohorts. Drawing from YC's vantage point of working with thousands of founders, the episode explores how startups are pivoting from bits to physical atoms, why nearly one in five companies now has a solo founder, and how AI-native teams are reaching meaningful revenue faster than ever.
## TL;DR Analyzing recent batch metrics, the YC partners reveal that hard tech startups—spanning robotics, defense, photonics, and domestic metallurgy—have surged from 8% to 20% of the cohort as foundational AI and code generation eliminate traditional software engineering bottlenecks. Concurrently, enterprise SaaS is evolving from passive systems of record into full-stack autonomous agents that execute entire jobs, driving median batch revenue from $8,000 to $20,000 per month. The hosts also highlight a secretive multi-million-dollar boom in startups supplying reinforcement learning environments and physical world datasets to frontier AI research labs. Finally, they unpack demographic shifts showing that solo founders now comprise roughly one-fifth of the batch, while experienced engineering veterans in their 30s and 40s are outperforming by orchestrating swarms of coding agents like human software teams.
## Atmosphere & style The discussion carries the fast-paced, collegial energy of an insider partner meeting, combining proprietary batch data with candid storytelling and technical depth. The hosts exhibit an infectious enthusiasm, regularly marveling at the routine ritual of wiring up new frontier models to effortlessly solve previously stubborn bugs overnight. It feels both intensely practical and visionary, offering listeners a behind-the-scenes view of an ecosystem undergoing a generational, AI-accelerated transformation.
## Bulletin - `00:00` Hard tech has expanded from 8% to 20% of Y Combinator batches, coinciding with a jump in median Demo Day revenue from $8,000/month historically to $20,000/month. - `30:00` Solo founders now make up nearly 19% of YC cohorts—an all-time high—as AI coding agents allow domain experts to launch and reach initial traction before recruiting co-founders later in the lifecycle. - `15:00` Enterprise software has shifted from human-operated SaaS tools to autonomous full-job agents, increasing the share of YC companies executing end-to-end workflows from 10% to more than 25%. - `22:38` Selling proprietary training data and specialized Reinforcement Learning (RL) simulation environments to frontier AI labs has become a secretive megabusiness, with over a dozen YC startups generating tens to hundreds of millions in revenue.
## Key takeaways - `00:00` The startup ecosystem is pivoting heavily from bits to atoms, with robotics, domestic manufacturing, defense tech, and power infrastructure driving hard tech to 20% of the YC batch. - `03:00` Advanced AI code generation has transformed hardware economics by eliminating the need for massive firmware and software engineering teams, allowing PhD-led teams to build full-stack physical systems rapidly. - `05:00` Geopolitical conflict, space commercialization, and the reshoring of American manufacturing have opened major opportunities for agile defense tech contractors and dual-use component fabricators. - `10:00` AI compute shortages are driving up rental costs for aging GPUs like A100s and accelerating startups developing optical switches, ternary silicon architectures, and data center energy systems. - `15:00` Incumbent software systems of record must build native agent execution harnesses rather than merely exposing open APIs or MCP endpoints that risk making their core data layers easily disintermediated. - `20:00` Enterprises are demonstrating immediate willingness to pay premium prices for full-stack autonomous agents that complete entire operational jobs rather than traditional tools that merely track work. - `20:00` Founders leveraging up to 20 parallel AI coding agent sessions are scaling companies from zero to seven figures in annualized revenue within the three-month duration of the YC batch. - `22:38` Frontier AI research labs are allocating upwards of $1 billion into purchasing specialized vertical domain data and custom RL simulation environments to surpass web pre-training scaling limits. - `25:00` Physical robotics foundation models require low-latency, real-time closed-loop control, driving high demand for vertical-specific demonstration datasets rather than one-size-fits-all general models. - `30:00` The founder playbook is evolving: solo founders increasingly prove product traction using AI tooling before recruiting late-stage co-founders under adjusted equity structures. - `33:00` Experienced second-career founders and engineering managers are outperforming younger peers because directing swarms of coding agents maps directly onto managing professional software engineering teams. - `35:00` The primary competitive advantage in tech has shifted from raw software development ability to domain taste and knowing precisely what to build.
## Detailed walkthrough ### `00:00` The Shift to Atoms and Accelerated Startup Growth
The episode opens with the partners reflecting on unexpected shifts in the startup ecosystem, from an influx of experienced older founders in their late 30s to 50s and solo founders flourishing at Y Combinator, to the sheer speed at which new AI models turn previously intractable problems into working products overnight. YC works with thousands of founders every year, giving the partners early visibility into emerging trends before they become obvious to the broader public. Diana shares the results of an internal analysis of all companies accepted over the past 12 to 18 months, highlighting two dramatic shifts: a massive swing toward hard tech companies that touch physical atoms rather than pure software bits, and a significant acceleration in the revenue velocity of YC companies during the batch itself.
Historically, the median YC company enters the batch at zero revenue—completely pre-product and pre-revenue—and would reach roughly $8,000 in monthly revenue by Demo Day. In recent batches, that figure has soared to a median of $20,000 in monthly revenue. At the same time, hard tech has expanded from occupying just 8% of the batch to an unprecedented 20%, driven by widespread breakthroughs in physical manufacturing, robotics, compute infrastructure, and geopolitical shifts.
- **Hard Tech Batch Share:** Surged from 8% to 20% of the entire batch over the past 12 to 18 months. - **Batch Revenue Velocity:** The median revenue achieved by Demo Day rose from $8,000/month historically to $20,000/month today. - **Robotics:** Rose from 1% of the batch to roughly 6%–7%. - **Industrial Manufacturing (Reshoring to the US):** Increased from 4% to approximately 10% of the batch. - **Defense Tech:** Expanded from 1.5% to about 5% of the batch. - **Semiconductor Stack & Photonics:** Grew from 1% a year prior to nearly 4% of the batch. - **Power Infrastructure:** Jumped from 1% to nearly 3% of the batch to support massive data center energy demands.
`03:00` Technical Depth, PhD Founders, and AI Compounding Hardware
As startups shift toward physical engineering, the founder profile inside YC has correspondingly changed. In the Summer 2026 batch, one in six founders holds a PhD. Jared and Garry explain that developing complex technologies like silicon photonics or specialized hardware demands deep academic research backgrounds, and these highly technical teams are outperforming historical benchmarks.
Crucially, the rise of advanced AI and automated code generation ("CodeGen") has completely transformed the economics and development cycles of hard tech. In prior decades, physical engineering was throttled by the need to hire massive software engineering organizations to build firmware, controls, and operating stacks while competing against tech giants like Meta or Google for scarce engineering talent. Garry points out that Palmer Luckey frequently highlighted this dynamic at Anduril: with modern code generation, small, focused teams can produce full-stack hardware and software at extreme velocity. As Garry summarizes:
> “Even three or four years ago, you would talk about software engineering and the top-tier software engineers as one of the limiting reagents to being able to do really, really top-tier, full-stack hardware.”
Smart foundation models are now accelerating underlying scientific workflows, allowing early-stage teams to make breakthroughs much earlier and cheaper than previously thought possible.
```mermaid flowchart LR accTitle: AI Acceleration of Hard Tech Startups accDescr: Diagram illustrating how foundational AI and code generation reduce engineering bottlenecks for hardware companies. subgraph Traditional["Historical Bottleneck"] direction TB A["Hard Tech Idea"] --> B["Massive Software Engineering Teams Needed"] B --> C["High Burn & Competition with Big Tech"] C --> D["Slow Development Cycles"] end subgraph Modern["Modern AI-Enabled Paradigm"] direction TB E["Hard Tech Idea + Deep Expertise (1 in 6 PhDs)"] --> F["AI Code Generation & Model-Driven Research"] F --> G["Small, Lean Technical Core"] G --> H["Rapid Full-Stack Hardware Prototyping"] end ```
`05:00` Macro Drivers: Space, Defense, and the Reshoring of American Manufacturing
The partners identify three massive macro trends propelling growth in the world of atoms: the precedent set by generational companies like SpaceX, the resurgence of state-level geopolitical conflict and drone warfare, and the urgent necessity of reshoring industrial supply chains. The success of SpaceX’s public offering has ignited a new wave of space founders tackling the entire orbital stack, including Summer 2026 batch company Exosat, which is building sovereign satellite communications networks (akin to a sovereign Starlink) tailored for orbital data centers like StarCloud that need dedicated power and communications.
Geopolitical conflict has pushed defense tech to the forefront of founders' minds. Garry notes two recent investments: Icarus (from the Fall batch), which develops solar-powered high-altitude reconnaissance aircraft functioning like a modern U-2 spy plane for persistent battlefield overwatch and communications, and Nine Mothers (from the Spring batch), which manufactures computer-vision-guided shotgun turrets to protect elite Special Forces units behind enemy lines from cheap, commodity drone swarms. Garry reflects on the failure of traditional defense primes who treat procurement like cost-plus consulting engagements, highlighting how modern startups can bypass that capture:
> “To be able to see new startups that can actually take advantage of all of the AI, all of the tech, all of the new ways of building things to build things that, frankly, the defense primes can't build. That's a really powerful megatrend right now.”
Beyond direct military contractors, dual-use supply chain businesses are expanding at software-like speeds. A prime example is Knox Metal, a company revitalizing domestic metallurgy out of vacant factory spaces in Detroit. Decades of outsourcing hollowed out American manufacturing, leaving defense tech startups starved for high-velocity domestic component suppliers. Traditional legacy metal fabricators operate too slowly to meet startup demand; Knox Metal has stepped in to serve as the agile manufacturing backbone for the entire defense hardware ecosystem, drawing direct parallels to how Stripe became the default infrastructure layer for the early Web 2.0 wave.
`10:00` Compute Bottlenecks, New Silicon, and the Physical Demands of AI
The explosion in AI demand is creating severe physical bottlenecks in the data center ecosystem, leading to bizarre economic conditions where aging hardware is actually gaining value over time. An A100 GPU per hour is appreciating in cost simply because compute supply cannot keep up with runaway market demand. In response, a wave of startups is taking on the physical infrastructure layer: building data centers, solving immense power constraints, and integrating novel battery and energy systems. Other founders are going even deeper into the stack to challenge NVIDIA at the silicon level. Tyler worked with a company called MatX (Lam Labs) that is designing new processors for compute, while another startup, Bota, is building custom hardware architectures based on ternary representation. This is driven by an underlying architectural reality: newer generations of NVIDIA hardware have systematically reduced floating-point precision—stepping down from FP32 to FP16, FP8, and beyond—because large language models simply do not require full floating-point precision, making even FP2 workable.
Beyond silicon, network interconnects have become the critical ceiling on performance. In standard data centers, when GPU A needs to communicate with GPU B, it routes traffic through an electronic switch. As GPU speeds accelerate rapidly, these electronic switches have turned into the primary bottleneck. A current batch startup called Dipole Labs is addressing this by building the first fully optical switch, routing photons directly from GPU to GPU to bypass electronic switching limits entirely. At the same time, the broader hardware push is being propelled by the impending "ChatGPT moment" for robotics. Startups are building across the entire stack—vertical industrial robotics, deployment infrastructure, training data marketplaces, and foundational robotics labs. As Kwan from Physical Intelligence (Pi) pointed out in a recent episode, robotics is split evenly between AI and hardware. Recent leaps like Astra are already moving task completion benchmarks from roughly 10% (as seen with models like Fable) up into the 60% to 70% range, signaling that scaling laws for physical embodiment are rapidly solidifying.
```mermaid flowchart TD accTitle: Physical Bottlenecks in AI Hardware accDescr: Diagram showing how AI demand exposes data center bottlenecks across power, interconnects, and silicon, leading to hardware innovations.
subgraph Demands["AI Market Pressures"] Demand["Skyrocketing Compute Demand"] --> AgingHW["Appreciating GPU Costs\n(e.g., NVIDIA A100s)"] end
subgraph Bottlenecks["Infrastructure Chokepoints"] AgingHW --> Energy["Power & Energy Limits"] AgingHW --> SwitchLimit["Electronic Switch Bottlenecks"] AgingHW --> SiliconLimit["Floating Point Overkill\n(FP32/FP16)"] end
subgraph Innovations["Startup Solutions"] Energy --> DCStartups["New Data Centers & Batteries"] SwitchLimit --> Optical["Dipole Labs\n(Fully Optical Switches)"] SiliconLimit --> NewArch["MatX & Bota\n(Ternary / Custom Silicon)"] end ```
For Y Combinator, this shift marks a dramatic and welcome return to Silicon Valley's original roots. While YC has funded hard tech companies since 2014, securing venture capital for them was historically an uphill battle against the industry's addiction to high-margin software. Garry reflects on that frustrating era: “Like I remember awesome stuff that we were super excited, like rockets and planes and chips and data centers and stuff like that. And then VCs would just be like, oh, we only do B2B SaaS.” Because hardware-intensive companies cannot be easily bootstrapped, they rely on downstream investors willing to fund massive capital expenditures. That investor mindset pivoted with shocking speed earlier in the year—coinciding with the drop in SaaS stocks and the surge of coding agents like Claude Code around the Winter batch Demo Day—prompting venture capital to eagerly embrace atoms over pure bits once again.
- NVIDIA A100 GPUs per hour are appreciating in rental cost despite being older hardware, driven entirely by compute shortages. - Bota is designing custom hardware architectures around ternary representation to exploit the finding that LLMs can function on minimal precision (even FP2). - Dipole Labs is building all-optical data center switches to eliminate the electronic bottleneck between high-speed GPUs. - Robotics benchmarks are jumping rapidly, with Astra reportedly completing 60% to 70% of physical evaluation tasks compared to previous marks around 10%.
---
`15:00` The SaaS Recovery, Agent Harnesses, and Full-Stack End-to-End Workflows
While hard tech has enjoyed a surge of investor attention, enterprise SaaS has simultaneously shown surprising resilience. Public software stocks staged a sharp recovery, led by blowout earnings from Snowflake and strong growth from Salesforce. The survival of traditional software giants comes down to the enduring power of the system of record. When autonomous software agents proliferate, they do not eliminate software; they consume it far faster and more intensely than human users ever could. This dynamics reframes enterprise architecture: if founders build software that autonomous agents want to use, system-of-record incumbents can preserve immense moats—provided they capture the workflow layer.
```mermaid flowchart LR accTitle: The System of Record Harness Dynamic accDescr: Flowchart illustrating how systems of record must build agent harnesses to avoid disintermediation.
subgraph Core["Incumbent Core"] SOR["System of Record\n(e.g., Salesforce, Slack)"] end
subgraph Risk["Vulnerability"] SOR -->|"Raw MCP / Open API"| Disinter["Data Extracted Elsewhere\n(Zero Moat / Easy Switching)"] end
subgraph Defense["Winning Strategy"] SOR -->|"Native Agent Interface"| Harness["Proprietary Harness\n(Slack AI Harness, Codex)"] Harness --> Agents["Agents Execute Work\nInside the Data Moat"] end ```
The battlefield has consequently shifted toward an aggressive war over "AI harnesses." Salesforce's release of an AI harness natively inside Slack illustrates the strategy: incumbents must provide the execution environment where agents operate rather than just exposing a raw Model Context Protocol (MCP) endpoint that risks making their data layer trivial to bypass. The speakers note that system-of-record companies must actively become harnesses where work gets completed, echoing benchmark lessons from ARC-AGI-3 (where pre-Astra models lagged purely because they lacked the proper agent harness, but later achieved scores north of 90% when equipped with custom scaffolding). As one of the partners frames the existential stakes for enterprise incumbents: “It’s like basically, if you are a system of record, you either will be preyed upon, like you'll release an MCP, and then maybe like the data around the data, the data goes elsewhere, like becomes very trivial to switch. Or you kind of have to be a harness.”
This evolution is reflected directly in YC's batch composition and revenue performance. Rather than SaaS being dead, it has transitioned from single-point tooling to full-stack, end-to-end agents that perform entire professions. Five to eight years ago, SaaS meant selling an interface that required a human operator; today, software autonomously executes the workflow. The share of accepted YC companies tackling full end-to-end task completion has leaped from 10% to more than 25% of the batch. These startups take over complete operational loops—such as acting as the insurance broker, handling clinical intake, or running end-to-end medical billing. This functional shift is driving unprecedented financial velocity for early-stage teams:
| Metric | Historical Performance | Current Batch Benchmark | |:--- |:--- |:--- | | **Acceptance MRR (Median)** | $0 | $0 | | **Demo Day MRR (Median)** | ~$8,000 | ~$20,000 | | **Full-Stack / End-to-End Batch Share** | ~10% | >25% |
Nearly a year after agentic coding became truly viable with the launch of models like Claude 3.5 Opus, buyers are no longer looking for tools to help them work faster—they are paying for software that simply takes the work off their plates and finishes it completely. While skeptics might dismiss rapid revenue acceleration as mere AI hype, the sustained contract values indicate genuine demand for end-to-end execution.
`20:00` The Shift from Systems of Record to Full-Job Agents
The discussion opens on why enterprises are eager to pay large sums for AI software. While critics might dismiss enterprise spending as hype—paying for AI merely because "it's like the cool thing to do"—the partners argue that the bull case they predicted two years ago is playing out: autonomous agents are inherently more valuable than traditional SaaS. A traditional system of record merely tracks work, whereas an agent actually executes the job from start to finish. When software performs the entire workflow, companies happily write massive checks because the return on investment is immediate and obvious.
The partners point to Juicebox, an AI recruiting tool that has experienced rapid growth over the past two years, as a prime example of this transition. Originally, Juicebox operated as an LLM-powered people search engine: recruiters could enter a specification for a candidate profile, and the tool returned relevant candidates, leaving the recruiter to handle cold outreach manually. Recently, Juicebox released an agent that takes over the outreach directly, contacting prospective hires and handling interview scheduling. This change is poised to double or triple revenue per account because clients demand more agents. Far from feeling threatened, human recruiters welcome this automation: > “I think they're finding that the recruiters themselves are actually really excited to use the agents because it frees them up to do the work that they feel is like unique and interesting.”
The discussion highlights an unprecedented acceleration in founder velocity during the YC batch: - Startups are now scaling from zero to seven figures ($1,000,000+) in annualized revenue within the three-month duration of the batch, a milestone that historically took 18 months or longer. - This velocity is driven by agentic coding: founders who are "super AI pilled" regularly run as many as 20 parallel coding agent sessions at once, enabling them to build mature, enterprise-grade software in weeks.
```mermaid flowchart LR accTitle: Evolution of Enterprise AI Tools accDescr: Diagram contrasting traditional SaaS tracking with search tools and modern autonomous agent workflows. subgraph Traditional["Legacy SaaS"] A["System of Record<br>(Tracks Human Work)"] end subgraph V1_AI["Early AI (Search)"] B["LLM Candidate Search<br>(Finds Profiles)"] --> C["Manual Reach-out & Admin"] end subgraph V2_AI["Autonomous Agents"] D["Recruiting Agent<br>(Juicebox)"] --> E["Autonomous Outreach"] E --> F["Automated Scheduling"] F --> G["Human Focus on Culture Fit"] end ```
`22:38` The Hidden Megabusiness: Data and RL Environments for AI Labs
The conversation turns to one of the most lucrative yet secretive categories in modern tech: startups selling specialized proprietary data and Reinforcement Learning (RL) training environments to frontier AI research labs. Unlike typical consumer or B2B SaaS startups that aggressively market their milestones, these companies operate quietly because transparency invites immediate competition. What began years ago with Scale AI and later Mercor has exploded into an enormous industry.
Garry and the team share that within just the past two years, YC has backed more than a dozen companies generating over $10 million in annual revenue selling data or RL environments to frontier labs—with several generating hundreds of millions of dollars. Garry specifically highlights AfterQuery and DataCurve as standout performers, noting: > “I mean, there are probably too many to name that are honestly like maybe don't even want to be mentioned because once you have something that's working, you almost don't want people to know.”
The partners note that frontier foundation model creators are reaching limits on standard web data and are spending aggressively on RL environments to unlock higher-order reasoning. These environments require deep domain expertise, advanced computer science, and complex systems engineering: - Startups build specialized RL environments for deep verticals like financial modeling, allowing models to explore vast state spaces and maximize domain-specific benchmarks. - Frontier AI research labs now collectively allocate upwards of $1 billion into procuring external training data and custom RL simulation environments.
`25:00` Physical World Data and Robotics Foundation Models
The demand for specialized training data is moving rapidly from pure software into the physical world. Frontier labs are heavily motivated to ground AI in physical reality, creating an intense appetite for egocentric video datasets, teleoperation logs, and multi-step robotic demonstrations. This has triggered eight- and nine-figure contracts between major research labs and physical-data startups.
Within recent YC batches, multiple startups have captured this surge: - **Praxis Robotics** (Summer 2026 batch) focuses on robotic data acquisition. - A stealth company in the batch operates a worldwide distributed data collection network inside industrial manufacturing facilities. - **DeepReach** collects unique real-world operational data through a network of local entrepreneurs around the globe. - **Human Archive** (Winter 2026 batch) gathers rich physical interaction datasets.
At the same time, tools like River AI, Tinker, Claude Code, and open-source fine-tuning setups are lowering the barrier for teams to train or adapt models on their proprietary domain data. The group debates whether general robotics foundation models will mirror LLMs: > “Perhaps in order to get robots to work in a specific vertical, like let's say robots that do operations in data centers... It's better to get a model that's fine-tuned and trained on custom data that just works in that environment, because the thing that's also challenging for robotics, they need to be in real time and respond very quickly to the stimuli and have an action plan, which is different than LLMs.”
The structural differences between text LLMs and robotic systems dictate how startups deploy them: - Text LLMs can afford latency and asynchronous execution (letting an agent plan and come back later), but robotics requires instant, low-latency sensory processing and continuous closed-loop control to avoid physical hazards. - Startups like Boost Robotics (which automates complex cable routing in data centers) cannot rely solely on off-the-shelf general models. - YC companies leveraging Physical Intelligence's π (Pi) foundation models consistently fine-tune them on vertical-specific data; company Ultra, for instance, trains base Pi models on thousands of hours of targeted physical manipulation footage. - This dynamic mirrors consumer media loops: just as ByteDance leverages billions of TikTok user interaction signals to train cutting-edge generative video models like SeedDance, proprietary task data creates an unassailable moat.
`29:15` The Resurgence of the Experienced Solo Founder
Garry and Diana shift the discussion to founder demographics, noting that the profile of top-performing entrepreneurs has fundamentally changed over the past year. While many new incubators and venture funds claim to specialize in solo entrepreneurs, Garry remarks that YC remains the definitive destination for solo founders.
Diana reveals striking admissions data: historically, solo founders comprised only about 5% of accepted YC batch companies. Today, that proportion has surged past roughly one in five (approaching 20%), driven by seasoned industry veterans and experienced technical leaders who can leverage AI agent suites to build, ship, and scale enterprise products without needing to assemble large initial teams.
`30:00` The Surge of Solo Founders and the New Co-Founder Playbook
The YC batch has seen a historic shift, with solo founders making up 18 to 19 percent of companies—nearly one-fifth of the entire cohort, marking the highest spike Y Combinator has ever recorded. Historically, launching an elite startup required pairing complementary extremes: a charismatic hustler who could persuade investors and customers, teamed up with a world-class technologist who could actually construct the software. Having co-founders dramatically improved a founding team's odds because individuals possessing both top-tier selling and deep technical execution skills were vanishingly rare. Today, AI has altered that balance. The hosts note that knowing what to build and how to prompt is becoming far more valuable and difficult than simply having a CTO who writes code from scratch.
While solo founders appear to be an emerging archetype, the partners point out that YC has always backed legendary solo founders who proved to be massive outliers. Apoorva Mehta started Instacart solo, Brian Armstrong entered YC as a single founder before being joined by Fred Ehrsam at Coinbase, and Parker Conrad was admitted solo before bringing on Laks Srini as his CTO. These founders were historically the exception because they could simultaneously invent, build, and sell. Today, AI eliminates the hard technical barrier, allowing solo operators to get an initial product off the ground without needing to be an extraordinary software engineer.
Even with the rise of AI-enabled solo builders, the hosts emphasize that having elite co-founders remains an immense advantage. However, the operational playbook has shifted. Rather than the traditional model where two founders sit in a room from day one with an immediate 50/50 equity split, many founders now launch alone, prove product traction with AI tooling, and recruit co-founders later in the company’s lifecycle under different equity arrangements. As Harj Taggar observes:
> “I think that will be the trend i think we'll see a little like more single founders in the batch which you're already seeing like starting the batch but at least of the things that succeed, I still expect that they're going to be adding co-founders as the company progresses.”
```mermaid flowchart LR accTitle: Evolution of the Solo Founder Journey accDescr: Diagram illustrating how solo founders use AI to gain traction before recruiting late-stage co-founders. subgraph Traditional["Traditional Path"] T1["Two Co-Founders<br>(Hustler + Hacker)"] --> T2["50/50 Equity Split<br>at Day Zero"] T2 --> T3["Build MVP & Launch"] end subgraph Modern["Modern 2026 Path"] M1["Solo Domain Expert"] --> M2["Prompt & Deploy MVP<br>via Coding Agents"] M2 --> M3["Demonstrate Traction"] M3 --> M4["Add Elite Co-Founders<br>Later in Lifecycle"] end ```
- **Solo founder representation**: Solo founders now account for 18% to 19% of YC batches, an all-time high. - **Historical precedents**: Canonical solo-starter examples in YC history include Apoorva Mehta (Instacart), Brian Armstrong (Coinbase), and Parker Conrad (Zenefits). - **Shifting bottleneck**: Raw coding ability is no longer the primary gating factor; knowing *what* to build and how to steer tools has become the scarce skill.
---
`33:00` Why Experienced Founders and Engineering Managers Excel with Coding Agents
Alongside solo founders, experienced and second-career founders in their thirties and forties are experiencing a major resurgence. The panel points to figures like Peter Steinberger—who entered his 40s with deep software experience, became thoroughly "AI-pilled," and began rapidly trying ideas—as prime examples of this trend. For decades, tech culture fetishized the 19-year-old college dropout, but the hosts argue that seasoned veterans who have "been around the block" hold an enormous competitive edge because they possess domain taste and know where the pitfalls and hidden complexities lie. As Garry Tan puts it:
> “It's like, basically, if you've been around the block, you know where the dragons are, you sort of have taste, and then those people in particular are like unusually powerful right now.”
With the elimination of classic gatekeeping—such as requiring elite venture connections or an immediate technical co-founder—the primary requirement is knowing what to build. Garry challenges opinionated online critics to stop making excuses, boot up OpenCode, and prove their ideas in software.
Furthermore, the hosts highlight an unexpected operational parallel: directing swarms of autonomous coding agents requires almost the exact same skillset as managing human engineering organizations. Seasoned engineering leaders from places like Shopify, who spent years managing large engineering departments, adapt seamlessly to agentic workflows. They know how to assign modular tasks, define architectural boundaries, provide clear feedback, and unblock stalled processes—often managing fleets of agents far more productively than a brilliant teenager without management experience. The only difference, they joke, is that dealing with coding agents requires managing 99.9% fewer human emotions.
- **The Peter Steinberger prototype**: Veteran technologists in their 40s who combine decades of software taste with early adoption of coding agents are shipping at unprecedented speed. - **Management parallels**: Guiding AI coding agents maps directly onto managing professional software engineers (setting specifications, scoping tickets, handling blockers). - **Lower friction**: Managing agents provides the throughput of an engineering team without interpersonal emotional overhead.
---
`35:00` Coding in the Time of AGI
When asked for concrete advice for an aspiring founder looking to start a company in the current landscape, the panel’s answer is straightforward: just start prompting. The state of AI models is advancing so fast that the experience of building software changes month to month. Garry shares the visceral feeling of testing GPT-6 on release day, describing how models suddenly become palpably smarter and resolve stubborn, long-standing bugs without human intervention. Founders can now hand an agent its own past chat history and unresolved tickets, tell it to revisit what it previously couldn't solve, and watch it systematically debug the codebase.
The hosts marvel at the routine morning ritual of wiring up a newly released frontier model and watching problems that seemed impossible just thirty days prior immediately start working. Garry Tan reflects on the sheer pace of this paradigm shift:
> “And like to think that that might be this thing that we get to do for the next 18, 24 months, 36 months, I don't know when it ends, but that's coding in the time of AGI, I guess.”
The hosts wrap up the episode by reinforcing that the present moment is one of the most exciting windows in the history of technology to build a startup, encouraging viewers to take advantage of the tools and show what they can create.
## Notable quotes > `33:00` It's like, basically, if you've been around the block, you know where the dragons are, you sort of have taste, and then those people in particular are like unusually powerful right now. — Garry Tan
## Mentioned Y Combinator, Garry Tan, Jared Friedman, Diana Hu, Harj Taggar, Palmer Luckey, Anduril, SpaceX, Exosat, StarCloud, Icarus, Nine Mothers, Knox Metal, Stripe, NVIDIA, MatX, Lam Labs, Bota, Dipole Labs, Physical Intelligence, Kwan, Claude Code, Snowflake, Salesforce, Slack, Claude 3.5 Opus, Juicebox, Scale AI, Mercor, AfterQuery, DataCurve, Praxis Robotics, DeepReach, Human Archive, River AI, Tinker, Boost Robotics, Ultra, ByteDance, TikTok, SeedDance, Apoorva Mehta, Instacart, Brian Armstrong, Fred Ehrsam, Coinbase, Parker Conrad, Laks Srini, Zenefits, Peter Steinberger, Shopify, OpenCode, GPT-6
## Video description <p>YC works with thousands of founders every year, which gives us an early look at how startups are changing. Right now, the shift is striking: startups are moving from bits to atoms, nearly one in five YC companies has a solo founder, and companies are reaching meaningful revenue faster than ever.In this episode of The Lightcone, Garry, Jared, Diana, and Harj dig into what’s driving these changes and what they mean for founders. They discuss how AI is making it possible for smaller teams to take on more ambitious problems, why experienced founders are having a resurgence, and why knowing what to build is becoming more important than simply knowing how to build it.</p>
[Source](https://anchor.fm/s/8c1524bc/podcast/play/125938592/https%3A%2F%2Fd3ctxlq1ktw2nl.cloudfront.net%2Fstaging%2F2026-8-18%2F432171154-44100-2-7e57e7b587c8f.mp3)