The YC W26 batch has 190 companies, 64% B2B, 74% AI-related. Fourteen reached $1M ARR by Demo Day, a record. Rebel Fund's ML algorithm, which has scored every YC company since 2012, rated 35% of W26 startups in the top 20% of all YC companies ever evaluated.

I attended Demo Day in person last week, spoke with many of the founders, and analyzed the batch company by company. Three patterns stood out:

• The hard-tech cohort is the largest in YC history
• Agent infrastructure is the most competitive category
• The emerging model to watch is service-as-software

While AI continues to be the predominant theme, what that AI is being applied to is getting broader: robots, satellites, batteries, radar, dive gear, professional services. The shift across the last three YC batches:

Hard Tech

Thirty-three companies tagged "Hard Tech" in YC's own directory. The cohort splits into two clusters.

Robotics and Physical AI

The larger cluster: around 21 robotics and physical AI companies.

Origami Robotics, Human Archive, Asimov, and One Robot are each building real-world physical data for training robot foundation models, each from a different angle. Origami is collecting manipulation data via a co-designed glove-robot system. Human Archive is deploying 1,000+ custom rigs to capture synchronized multimodal data. Asimov is collecting egocentric movement data from households and businesses across three continents. Congruent is doing the same for autonomous vehicles, building AI-native radar with its own training data and simulator. When four founders independently converge on the same bottleneck - real-world data - the bottleneck is probably real.

That data feeds a new deployment pattern. Capture the environment digitally, generate a simulation, fine-tune the robot policy in simulation, deploy. One Robot (whose team includes a NASA JPL veteran and a former Tesla Dojo engineer) is building the world models that make it possible.

The supply chain layer is forming alongside it. Hlabs is building the US-made robotics component supply chain: actuators, FOC motor controllers, Nvidia Jetson compute modules. The same "you can't fork a supply chain" observation from S25, now applied to robotics components specifically.

GrazeMate is deploying autonomous AI-powered drones to herd cattle, built by a founder who grew up on his family's 6,000-head cattle farm in Queensland. You can build the software from anywhere. You can't build the domain credibility.

Space, Defense, and the Industrial Base

The second cluster spans space and defense, with constraints ranging from manufacturing to physics.

Kyten Technologies, founded by ex-SpaceX and Starlink engineers who collectively deployed 5,000+ battery packs to orbit, is now building a domestic aerospace battery pack factory. The US drone and satellite industry needs a reliable American source for flight-grade batteries. Kyten's constraint is manufacturing expertise that took years on orbit to develop.

Beyond Reach Labs is building deployable solar arrays that expand from dining-table size to the scale of a football field in orbit, with a patented design that changes geometry in space. DAIVIN! is building tankless dive gear via real-time water electrolysis, targeting military and commercial diving applications. In both cases, the constraint is physics: the underlying science is genuinely hard to replicate.

On the defense side, Seeing Systems is building AI drone systems and has worked with multiple NATO forces. Milliray (formerly DroneTector) is building technology to detect and track hostile drones. AxionOrbital Space is taking an interesting angle on Earth observation: SAR-to-optical image translation. The bottleneck for using synthetic aperture radar data is that SAR images are unintelligible to most analysts. AxionOrbital is translating them into optical-quality images, enabling 24/7 Earth observation regardless of weather or lighting.

Across both clusters, the edge lives outside of software. In the S25 analysis, 11 companies were building for the physical world. In W26, 33 are.

As AI compresses the cost of building software, the constraints that can't be compressed become relatively more valuable.

→ All hard-tech companies

The Engine Room Fills Up

Around 39 developer tools and agent infrastructure companies are in the batch. This is the densest category.

In the S25 analysis, I mapped five AI infrastructure layers. In the F25 series, I noted 25 multi-agent orchestration companies in a single batch. W26 shows the next stage: a full agent stack taking shape, with each layer now populated by multiple teams.

• Identity and access: who the agent is and what it's allowed to touch (Agentic Fabriq, Didit)
• Payments: how agents spend money with oversight and audit trails (Sponge, Maven)
• Observability and ops: monitoring, debugging, and keeping agents running in production (Moda, Cascade)
• Safety and guardrails: runtime controls that prevent agents from acting beyond their scope (Salus, Sentrial)
• Developer tools: building, testing, and deploying agents faster (Canary, OpenSpec)
• Inference and compute: making agents run faster and cheaper at the hardware level (Piris Labs, RunAnywhere)

Three batches ago, each of these layers was a greenfield opportunity. Today, the stack is becoming predictable, and the layers look similar to what emerged in cloud infrastructure a decade earlier. The question has shifted from "what needs to be built?" to "who can build it in a way that's hard to replicate?"

Three patterns stand out among the companies that break through.

• Founder credibility that can't be hired away: Mendral is founded by the creators of Docker and Dagger, building an AI DevOps agent. The people who built the previous generation of developer infrastructure are now building the next one.
• Solving problems the stack itself created: Canary is building an AI QA engineer that reads source code to understand developer intent, addressing what Canary's founders describe as a 40%+ year-over-year increase in production incidents from AI-generated code. AI coding tools created a new category of bugs. Canary exists because of that.
• Attacking the stack from a different dimension: Piris Labs is building photonic interconnects to address the memory bandwidth bottleneck in AI inference at the hardware level, with early results showing 5x lower latency. When everyone else is optimizing software, a hardware approach is inherently differentiated.

The bar for what counts as differentiated in agent infrastructure is higher than it was a year ago. The density of the companies here is a sign the category matters, not that the companies in it don't.

→ All developer tools and agent infra companies

The Regulated Frontier

Healthcare (22 companies), Fintech (29), Legal (7), and Compliance (10) make up a significant portion of the batch. A pattern that started with the S25 "Rebuilding Problem" (what you can't retrofit in regulated industries) and continued with the F25 "Compliance Wedge" (regulatory complexity as a moat) has taken another step.

Several W26 companies have expanded from building tools for regulated industries to running the professional services themselves. The model: service-as-software. In practice, AI does 80% of the work. Humans provide the oversight and the licenses.

• Legal: General Legal is operating as an AI-native law firm for commercial contracts at flat fees ($500/contract). Stilta is building agentic AI for patent attorneys. LegalOS is running an AI-native immigration law firm with a 99%+ visa approval rate. These companies are capturing the margin of a professional services firm, not the margin of a SaaS tool.
• Healthcare: Mango Medical is building instant planning with AI agents for orthopedic surgery, with early enterprise validation from a top orthopedic company. Beacon Health is deploying AI employees for primary care workflows. Opalite Health is solving language barriers in healthcare, a federally mandated need.
• Fintech: End Close is automating reconciliation for high-volume payments companies, founded by the person who built the reconciliation division at Modern Treasury, which processed $1 trillion annually. Panta is an AI-powered commercial insurance brokerage targeting surplus and specialty lines, where licensed brokers handle strategy while AI handles the paperwork.

The constraint here is the license itself. The professional credential, the regulatory approval, the liability framework. These companies don't sell tools to professionals. They are the professional. The same pattern from the physical-world companies, applied to knowledge work.

Service-as-software represents about 4% of the batch. The risk profile is different from SaaS: these companies don't face customer churn in the traditional sense, but they do carry regulatory liability. If the model scales, it raises an interesting question: does software capture the value, or does the license?

→ All healthcare, fintech, legal, and compliance companies

Where W26 Points

YC's recently released Spring 2026 RFS is explicitly steering toward the physical world. Three of ten themes ask for companies that operate in atoms: "Modern Metal Mills," "Large Spatial Models," and "AI Guidance for Physical Work." A fourth, "AI-Native Agencies," validates the service-as-software model that emerged organically in W26.

When YC dedicates 30% of its new RFS to atoms, the investment thesis is no longer a subtext.

What to watch in X26:

• Does the hard-tech cluster further expand?
• Does agent infrastructure remain as dense?

The Takeaway

Across the batch, the same property surfaces. The companies with the strongest early positions own a constraint that can't be replicated with software: proprietary physical-world data, manufacturing expertise, regulatory trust, or a professional license.

The arc across three batches has been clear:

• S25 proved agents could work.
• F25 showed they could replace entire organizations.
• W26 asks: as the cost of producing software compresses, what new constraints emerge?

The companies emerging now suggest an answer: the constraint is owning something you can't download.

Connect with me on LinkedIn for commentary on startup patterns as they emerge.

Appendix: All 190 YC W26 Companies

Grouped by the three sectors covered in this analysis, plus all remaining companies. Each company appears once, in the group that best fits its primary business.

Hard Tech, Robotics, Space and Defense

33 companies tagged "Hard Tech" in YC's directory, plus 2 closely related robotics companies (Asimov, Human Archive). ↑ Back to section

Developer Tools and Agent Infrastructure

40 companies. ↑ Back to section

Healthcare, Fintech, Legal and Compliance

56 companies. ↑ Back to section

All Other Companies

59 companies.