Context

Y Combinator partners Garry Tan, Jared Friedman, Diana Hu, and Harj Taggar pull data from the thousands of founders YC works with each year to describe three shifts they're seeing before they become obvious elsewhere: a surge back into "hard tech" (companies touching physical atoms, not just software), a near-quadrupling of solo-founder acceptance, and startups reaching meaningful revenue dramatically faster than in past years. The episode matters to PMs and founders because it's grounded in YC's own batch-acceptance and revenue data, not just anecdote, and it argues the underlying cause is the same in all three trends: AI-assisted building has changed what one person, or one small team, can actually take on.

The Big Idea

AI hasn't just made software easier to build, it has changed what kind of company is fundable and who can found it: hard tech companies are surging because coding agents remove the huge-engineering-team requirement that used to make them slow and expensive, and solo founders are surging because knowing what to build now matters more than having the traditional hustler-plus-technologist founding pair.

YC's own numbers back this up directly: hard tech acceptance rose from 8% to 20% of the batch, and the median company's revenue by the end of a batch jumped from about $8,000 to about $20,000 in monthly recurring revenue, over the same period solo-founder acceptance rose from about 5% to nearly 19%.

Key Insights

Hard tech acceptance jumped from 8% to 20% of the batch

Diana Hu's data shows specific categories driving this: robotics went from 1% to 6-7% of the batch, industrial manufacturing from 4% to 10%, defense from 1.5% to 5%, and compute infrastructure (semiconductors, photonics) from roughly 1% to 4%, plus a new power-infrastructure category near 3%. The partners attribute this partly to AI-assisted engineering removing what used to be the binding constraint on hard tech: needing hundreds of elite engineers to build the software layer around a physical product, which Garry Tan compares to Anduril's experience, where code generation collapses a task that once needed a large team down to one or two strong engineers.

Three macro forces are pulling founders toward atoms

Jared Friedman names three separate drivers behind the hard-tech surge, not one. First, high-profile hard tech IPOs (SpaceX is the example given) created a visible proof point that inspired founders to build in adjacent spaces like satellite infrastructure. Second, a generation of founders who came up during a period when war and defense were prominent in public discourse are actively choosing defense startups, and getting real traction: Friedman cites Icarus (a solar-powered surveillance plane with seven-figure Department of War contracts) and Nine Mothers (anti-drone defense for special forces) as examples of small teams outcompeting traditional defense primes stuck on cost-plus contracting. Third, AI compute demand itself is pulling founders into the physical infrastructure that supports it: data center construction, power, and even alternative chip architectures competing with Nvidia.

Fast-growing "dual use" hardware suppliers show hardware can now grow like software

Knox Metals, a company rebuilding domestic metal manufacturing in Detroit, is cited as growing at what the partners call "software growth rates," a phrase one of them attributes to a Paul Graham tweet. The explanation given is structural: Knox Metals' customers are new defense-tech startups that need metal fast, and legacy metal suppliers (described as "sleepy old businesses mostly run by old people") can't move at the pace these new customers demand, creating an opening the same way Stripe outcompeted legacy payment processors for web 2.0-era startups that needed a faster-moving vendor.

Full-stack, "does the job" products are replacing point-solution SaaS

The partners report that the share of YC companies building full end-to-end workflow products, ones where an AI agent actually performs the job (medical billing, insurance underwriting, clinical intake) rather than just providing software a human operates, rose from about 10% to over 25% of the batch. This is presented as the direct cause of the revenue acceleration: a product that fully automates a job is worth more to a buyer than a system of record that merely tracks that job, so enterprises are willing to write larger, earlier checks for products that deliver complete outcomes. Juicebox, an AI recruiting tool, is used as a concrete example: it started as an LLM-powered candidate search tool, then added an agent that also contacts and schedules candidates, which the partners say is expected to double or triple per-account revenue as customers buy more of the agent's capability.

Selling data and RL environments to AI labs has quietly become a major, low-visibility category

The partners describe a category of company, selling training data or reinforcement-learning environments to frontier AI labs, that started with a single company (Scale, funded by YC in 2016) and has since grown into more than a dozen YC-funded companies each earning over $10 million a year, some reaching hundreds of millions, within just a couple of years of founding. They note this category is unusually invisible because these companies have a disincentive to publicize their growth (once a data or RL-environment approach is working well, a company doesn't want competitors to notice), and the underlying demand is real: the partners cite an unverified but widely repeated figure that major AI labs are collectively spending roughly a billion dollars a year on this category, including a newer sub-segment supplying real-world robotics data (teleoperation footage, egocentric video) that labs need to get AI working in physical environments.

Companies are reaching seven figures in revenue within a single three-month batch

The partners describe seeing YC companies go from zero to seven-figure annual revenue within the roughly three-month span of a single batch, something that historically took at least 18 months. They attribute this to founders being deeply fluent with coding agents (running many parallel agent sessions to reach product maturity fast) combined with the shift toward full-stack, job-replacing products that command higher prices per customer from day one.

Solo-founder acceptance nearly quadrupled, from about 5% to nearly 19% of the batch

Diana Hu's data shows this is the largest swing YC has tracked in this metric. The partners' explanation is that the classic founding-team requirement, pairing a "hustler" who can sell and fundraise with a technologist who can build, existed because both skills were independently hard to acquire and rarely found in one person. Now that building has gotten dramatically easier with AI-assisted coding, the harder, scarcer skill has shifted to knowing what to build, meaning a single founder with strong judgment and enough technical fluency to direct coding agents can now do what used to require a full co-founding pair. The partners are careful to note this doesn't make co-founders worthless: strong co-founding teams still correlate with success, and many successful solo-founder companies (they cite Instacart, Coinbase, and Rippling as historical examples that started with a single founder) still added co-founders later, once traction was established, rather than staying solo permanently.

Experienced founders in their 30s, 40s, and 50s are having a resurgence

The partners specifically call out a wave of "badass" founders well past the stereotypical young-founder age, citing Peter Steinberger (an experienced developer and startup veteran in his early 40s) as a canonical example: someone who had already worked as a dev manager and on prior startups, got early and deep into AI coding tools, and now has enough taste and pattern recognition from experience to know precisely what to build. Their argument is that classic startup gatekeeping (needing a co-founder, needing name-brand investors) has weakened at the same time that judgment and taste, disproportionately held by people with more career experience, have become the scarcer, more valuable skill.

Mental Models & Frameworks

The harness war: system-of-record companies must become the interface agents work through

The partners describe an emerging pattern where software companies that used to be passive systems of record (Salesforce, Slack) are racing to become the active "harness" through which AI agents actually do work, not just where data about work gets stored. Their framing: a system of record either gets preyed upon (its data becomes accessible via something like an open MCP integration, making it trivial for customers to switch away) or it becomes the harness itself, the place where agents read, write, and complete real work, which preserves its competitive moat. They point to Salesforce's own Slack AI harness as an early instance of this shift and note the same logic drives their expectation that data-heavy platforms will increasingly need to train their own custom models rather than rely purely on frontier models plugged in from outside.

Proprietary interaction data becomes a training-data moat

The partners argue that any company sitting on a large stream of real usage data, coding transcripts revealing who the strongest coders are, video engagement data revealing what content is compelling, is positioned to use that data to train specialized models that outperform general frontier models on that specific task. They frame this as a durable advantage distinct from simply having a good product: the data itself becomes a flywheel for building a better, narrower model than any competitor without equivalent proprietary data could match.

Practical Application

Evaluate a hard-tech idea by asking what coding agents removed as a blocker

Before dismissing a hardware or physical-infrastructure idea as too capital- or team-intensive, check specifically whether the previously binding constraint was software engineering headcount. If a coding agent can now do the work that used to require a large engineering team (as it has for companies like Anduril), the economics of building that hard-tech company may have changed even if the physical/manufacturing side of the challenge hasn't.

Build for agents as customers, not just humans

Given the partners' observation that Salesforce's growth is increasingly driven by agents using its software rather than only humans, when scoping a new product or integration, explicitly ask whether an AI agent, not just a human user, would want to use it, and design the interface and API surface around that usage pattern from the start rather than retrofitting it later.

If you're a solo founder, don't treat that as a permanent structural choice

Following the pattern the partners describe in founders like the Instacart, Coinbase, and Rippling examples, treat starting solo as a way to move fast on an unproven idea, not as evidence you shouldn't eventually bring on a co-founder. Plan to actively evaluate adding a co-founder once you have real traction, since the data suggests most enduring solo-founder successes eventually add strong partners rather than staying solo indefinitely.

Weigh whether your data stream could become a model-training asset

If your product generates a large volume of real usage data relevant to a task (code, video engagement, physical task footage), evaluate whether that data could eventually justify training a specialized model rather than only ever consuming a general-purpose one. The partners' argument is this is becoming a realistic moat-building move for data-rich companies, not just something reserved for frontier labs.

Bottom Line

YC's partners argue the same underlying shift, AI-assisted building collapsing the team size and skill combination needed to execute an idea, explains three separate trends in their own batch data at once: hard tech is surging because the engineering-headcount barrier fell, solo founders are surging because knowing what to build now matters more than knowing how to code it, and revenue is arriving faster because full-stack, job-replacing products are simply worth more to buyers than point-solution software ever was.