The State of Startups in 2026

1. Key Themes

The Hard Tech Renaissance Is Structural, Not Cyclical

YC's own portfolio data shows a dramatic shift in what founders are building: hard tech companies grew from 8% to 20% of the batch. The breakdown by category is stark — robotics went from 1% to 6-7% of the batch, industrial manufacturing from 4% to 10%, defense from 1.5% to 5%, semiconductors/photonics from 1% to 4%, and power infrastructure from 1% to 3%. "So all these numbers across the physical atom stacks have somewhere triple or quintupled" 00:03:07, said Diana. Garry Tan frames this as generational: "This is the age of the machine, I think" 00:03:14.

AI Coding Agents Are Collapsing the Cost of Building Hardware Companies

The bull case for hard tech isn't just capital rotation away from software — it's that AI itself makes hard tech companies viable in ways they weren't before. "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... you don't need to hire a thousand great engineers versus Google or Meta" 00:04:31, Garry Tan noted, citing Anduril's use of code generation. Jared reinforces this: "the super smart models that we have now are actually accelerating scientific research and making it possible for startups to have bigger research breakthroughs earlier" 00:05:01.

PhDs and Deep Technical Founders Are Surging — and Winning

Technical depth is being rewarded at unprecedented rates. "In the current summer batch, one in six of the founders actually has a PhD... And those founders, I think, have disproportionately been doing like especially well" 00:03:44, said Jared.

Software Isn't Dead — It's Being Redefined Around Full Job Completion

The narrative that AI is killing SaaS is wrong according to Diana; instead, SaaS is transforming into full-stack agentic execution. "The percentage of companies we accepted that do sort of full stack end-to-end work or a task has gone from just 10% to over 25% of the batch" 00:17:58. This is driving revenue: median startup revenue by end of batch jumped from $8K to $20K MRR, and some companies are going "from zero to seven figures in revenue during the batch... in a span of three months" 00:21:59, versus roughly 18 months previously.

The "System of Record" Must Become an "AI Harness" or Die

Garry Tan's thesis: companies holding valuable enterprise data face a binary choice. "Basically if you are a system of record, you either will be preyed upon... you'll release an MCP and then maybe... the data goes elsewhere, like becomes very trivial to switch. Or you kind of have to be a harness. You have to be the way people not just read and write, but actually do their work inside your system of record" 00:17:01. Salesforce/Slack is the live case study — Benioff's advantage is that "a lot of the most AI appealed people and companies in the world still use Slack... if the harness is in there... then you have like this mega data moat" 00:16:29.

Data and RL Environments for Labs Is a Massive, Stealthy Category

A category few talk about publicly is quietly enormous. "In the last two years, YC has funded more than a dozen companies that are each making more than $10 million a year selling data or RL environments to the labs... in many cases, hundreds of millions of dollars... these are companies that were just a couple of years old" 00:23:20, said Jared. Diana adds: "Reportedly, the big labs are spending about a billion dollars on this" 00:24:58.

Solo Founders Are Thriving Like Never Before

"One of the shocking stats from analyzing the [current] companies from a year ago, we used to only have about 5% of the companies... be solo founders. And now we're over 18, 19%, which is a huge... the highest spike that we've seen" 00:29:48, said Diana. Garry Tan's framing: "A lot of people seem to say that they want to be YC for solo founders. But it turns out YC is the YC for solo founders" 00:00:00.

The Return of the Experienced Founder (30s, 40s, 50s)

Garry Tan highlights a resurgence of older, seasoned builders: "Frankly, some of the most powerful and badass founders that we've been seeing lately, there might be in their late 30s, 40s, even 50s" 00:00:00. Peter Steinberger is cited as the archetype: "he's, I believe, in his early 40s and he'd been a dev manager. He'd worked on startups before... he just tried a lot of stuff and then he knows what to build" 00:32:53.

GPUs Are Appreciating in Cost — an Inversion of Normal Tech Depreciation

Diana surfaces a counterintuitive compute economics fact: "GPUs from NVIDIA, let's say like an A100 GPU per hour is actually appreciating in cost, which is unusual... The price is going up because it's just too much demand and not enough supply" 00:09:48.

Robotics Foundation Models May Need to Be Vertically Fine-Tuned, Unlike LLMs

Diana's hypothesis: robotics won't follow the LLM path of one dominant general foundation model. "LLM is the whole thing is you model reality as language. And for robotics, you model reality in the physical 3D space, which has way more degrees of freedom... it's better to get a model that's fine-tuned and trained on custom data that just works in that environment" 00:27:06. Jared confirms this empirically: "all the YC companies that are using physical intelligence as models to deploy robotics, they're all fine-tuning the Pi models... you have to actually fine-tune it for your specific case... in order for it to work" 00:28:17.

2. Contrarian Perspectives

Defense Primes Are Structurally Incapable of Competing With AI-Native Startups

Garry Tan takes a direct shot at legacy defense contractors: "Classically, there was just a lot of, frankly, capture from the big defense primes that are just doing sort of cost plus. They think of themselves as consultants... startups 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" 00:07:18. He backs this with concrete traction: Icarus (solar-powered spy plane) hit "seven-figure contracts with the new Department of War" 00:06:23 and Nine Zero (anti-drone defense) sold directly to Special Forces.

Co-founder Dogma Was Never as True as the Industry Claimed

The conventional wisdom that startups need complementary co-founders (a "hustler" and a "hacker") is being challenged by both historical fact and current data. Harj Taggar points out this pattern isn't even new: "Ravi Gupta with Instacart, Brian Armstrong with Coinbase, at least when the batch started, were single founders" 00:30:58, and "Parker Conrad got into YC as a single founder" 00:31:15 (later hiring Laks Srini as CTO). The real shift, per Harj: "The bar for being able to like have the idea, be able to sell it and be able to build it all by yourself was just really, really high" 00:31:22 — and AI is lowering that bar dramatically, not eliminating the pattern.

Old Metal Suppliers Are Being Displaced Not by Technology But by Speed and Willingness to Partner

Jared draws a surprising analogy between the Web 2.0-era fintech disruption and today's defense supply chain: "The existing suppliers that are these sort of like sleepy old businesses, mostly run by old people, just like can't keep up with the pace that the new defense tech startups want to build at... it reminds me a bit of like when the Web 2.0 boom happened... like Stripe, for example, you could use a legacy credit card vendor, but like it's just like way better to work with Stripe" 00:09:04. The contrarian point: hardware manufacturing businesses (Knox Metal) can now grow "at software growth rates" 00:08:50 — a category previously assumed to be capital-intensive and slow.

RL/Data-for-Labs Companies Deliberately Hide Their Success

Unlike most startups that promote traction, this category actively avoids visibility. Garry Tan: "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" 00:23:51 — implying some of the most lucrative YC companies right now are intentionally invisible to the market and even to competitors.

SaaS "Death" Is Overstated — Moats Can Hold If Companies Adapt Fast Enough

Against the AI-kills-SaaS narrative, Harj Taggar notes real counter-evidence: "since that, a bunch of the SaaS stocks have actually recovered and are doing better than ever. Like Salesforce is like the prime example of that. I think Snowflake recently had like... blowout earnings" 00:15:03. The nuance: it's not survival by default, but survival by becoming the agent's harness rather than a passive system of record.

3. Companies Identified

Anduril — Defense technology company. Cited by Garry Tan as proof that AI code generation accelerates full-stack hardware development: "having CodeGen means that suddenly even all the things that they do at Anduril can happen much, much faster" 00:04:05.

Icarus — Solar-powered U-2-style spy plane startup providing overwatch and communications for drone warfare, YC-funded. "They've been able to get to seven-figure contracts with the new Department of War" 00:06:23.

Nine Zero (Nine Mothers) — Anti-drone defense company selling a "shotgun turret with CV" to Special Forces. "It's actually almost the only way that you could protect special forces deep behind enemy lines" 00:06:49.

Knox Metal — Metal manufacturing company rebuilding America's metal supply chain in Detroit. Growing at "software growth rates" 00:08:50 by serving defense tech startups faster than legacy suppliers.

Exosat — Building a "sovereign Starlink solution" 00:05:20, summer '26 batch.

Beyond Reach Labs — Building solar panels for satellites in space, winter '26 batch, anticipating data centers in space needing power.

Lam Labs — Building new processors/silicon as an alternative to NVIDIA, worked with by Tyler at YC.

Bot — Building custom hardware architecture using ternary representation for AI models, exploiting the trend toward lower floating-point precision in LLM architectures.

Dipole Labs — Building the first fully optical switch for data centers (replacing electronic GPU-to-GPU switches), current batch. "It will actually be much faster than the electronic switches that we use now" 00:12:14.

Boost Robotic — Building robots for data center operations (e.g., cabling), fine-tuning foundation models on custom vertical data.

Praxis Robotics — Summer '26 batch robotics/data company.

Deep Reach — Company (worked with by Brad at YC) collecting data via a network of local entrepreneurs globally.

Human Archive — Winter '26 batch, part of the robotics/embodied data wave.

Ultra — Robotics company starting from the Pi foundation model, fine-tuned with "thousands of hours of footage of putting things in boxes" 00:28:36 for warehouse/logistics tasks.

Juicebox — AI recruiting tool. Started as "LLM powered people search" and evolved into a full agent that contacts candidates and schedules interviews. "That's going to just on a like per account basis... double or triple like the revenue they make from a single customer" 00:21:06, per Harj Taggar.

Scale (Scale AI) — YC-funded in 2016, pioneered the data-labeling-for-labs category "when this was like a tiny little niche thing" 00:22:50.

Mercor — Followed Scale into the data-for-labs category, per Jared, helping turn it into a major sector.

Turing / DataCurve ("data curve") — Referenced by Garry Tan among the standout data/RL environment companies: "the big ones, I mean, I think after query and data curve both really, really great" 00:23:51 (Surge and DataCurve).

River AI and Tinker — Cited by Garry Tan as emerging platforms enabling companies to train their own proprietary models rather than relying purely on frontier labs: "that's where things like River AI or Tinker start becoming really interesting" 00:26:16.

Salesforce — Cited as proof that system-of-record SaaS moats can hold if the company builds its own AI harness (Slack AI harness). "The moats are intact for now" 00:15:32, per Garry Tan.

Snowflake — Cited by Harj Taggar for recent "blowout earnings" 00:15:03 as counter-evidence to the SaaS-is-dead narrative.

SpaceX — Referenced as a catalyst inspiring a new generation of space-focused founders following its successful IPO.

Stripe — Used as historical analogy for how new, faster-moving vendors displace legacy incumbents even in infrastructure categories.

4. People Identified

Peter Steinberger — Cited by Garry Tan as the canonical example of the resurgent experienced founder: "he's, I believe, in his early 40s and he'd been a dev manager. He'd worked on startups before. And then... he sort of uniquely got extremely AI pilled with the clankers early. But then he just tried a lot of stuff and then he knows what to build" 00:32:53.

Palmer Luckey — Referenced by Garry Tan re: Anduril's use of AI code generation to accelerate hardware development.

Ravi Gupta (referenced as "Poova") — Cofounder of Instacart, cited as historical example of a successful single founder entering YC.

Brian Armstrong — Cofounder/CEO of Coinbase, cited as another historical single-founder YC success story.

Parker Conrad — Founder (Rippling, previously Zenefits), cited as having entered YC as a solo founder before bringing on Laks Srinivasan as CTO.

Laks Srinivasan — CTO brought on by Parker Conrad after YC acceptance; referenced by Garry Tan as an example of the "add co-founder later" pattern.

Boris Cherny — Referenced by Jared as someone whose engineering-management background translates well into effectively directing large fleets of coding agents.

Toby Lütke — CEO of Shopify, referenced by Jared alongside Boris Cherny as an experienced engineering leader well-suited to managing AI agent teams: "people like Peter or you or Boris Cherny and like Toby from Shopify who have had whole careers like managing teams of engineers actually like take to this super well" 00:34:15.

Quan (from Pi / Physical Intelligence) — Referenced as a prior podcast guest whose robotics thesis (half AI, half hardware) the hosts say they believe in.

5. Operating Insights

Mine Your Own Chat Logs for a Free Roadmap of Unsolved Problems

Garry Tan shares a specific, immediately actionable tactic for founders using coding agents: "You just be like, actually, could you just go back to the list of things that you couldn't figure out? Like look at your... all of our last chats and... anything that looks like you didn't figure out, like try to figure it out now. And it'll do it, like every single time" 00:35:06. This treats each new model release as an opportunity to systematically re-run a backlog of previously "impossible" bugs — a repeatable operating cadence, not a one-time trick.

Recruiting Agents Should Augment the "Rote" Part of the Job, Not Replace the Judgment Part

Harj Taggar's detailed breakdown of Juicebox reveals a template for where agentic automation creates the most value without cannibalizing the human role: automate outreach/scheduling (the part professionals dislike) while preserving judgment-heavy tasks like culture-fit assessment. "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" 00:21:51 — a reusable framework for any vertical AI product: identify the rote 80% versus the judgment-based 20%, automate the former, and sell the tool as liberation rather than replacement to drive adoption.

Treat Coding Agent Management as a People-Management Skill, Not a Technical One

Jared's observation reframes what makes someone effective at scale with AI agents: it's leadership experience, not necessarily raw coding skill. "Managing coding agents is actually like in some ways not that different from managing people... [experienced managers] can like spin up huge teams of coding agents and manage them maybe more effectively than even like a really smart 19 year old who hasn't had those years of experience" 00:34:15. Garry Tan adds the practical edge: "we can be a little bit less abusive to our agents... You have to catch their emotions like 99.9% less" 00:34:42 — meaning management overhead per agent is lower than per human, so the multiplier effect for experienced managers is even larger.

6. Overlooked Insights

The FP32→FP2 Precision Collapse Is Quietly Creating a New Silicon Category

Buried in the compute discussion is a detail with outsized implications: NVIDIA's own architecture generations have been steadily reducing floating-point precision (FP32 → FP16 → FP8...) because "it turns out that the LLM architecture doesn't need the full precision floating point" 00:11:33, with Garry Tan adding "FP2 is even somewhat usable" 00:11:44. This single fact is the entire thesis behind why a company like Bot can build ternary-representation custom silicon as a viable NVIDIA alternative — it implies the entire chip industry's multi-decade obsession with precision was calibrated for a workload (general computing) that AI doesn't actually need, opening a wedge for specialized, cheaper silicon that the market hasn't fully priced in yet.

Co-founder Equity Norms Are Quietly Breaking Down Without Anyone Noticing

In a single aside, Harj Taggar flags a structural shift with major cap-table implications that gets no further discussion: solo founders who later add co-founders "as the company progresses" 00:32:23 means "maybe like the equity ownership is different or maybe the dynamic is just slightly different to the traditional way" 00:32:01 versus the classic 50-50 split from day one. This is a quietly forming new normal for startup formation — later-added co-founders will increasingly hold minority stakes rather than equal partnerships, which changes incentive design, retention risk, and how investors should evaluate founding-team stability at the seed stage. Neither Jared nor Garry push further on this, but it implies a coming wave of governance and vesting structure innovation that cap table tools and lawyers aren't yet built for.