Every Y Combinator batch offers a snapshot of what a new generation of founders believes will matter next.

Y Combinator’s Summer 2026 batch offers an early look at the next wave of AI startups, with founders targeting everything from energy and semiconductors to robotics, healthcare and space infrastructure.

Some are tackling the constraints emerging as AI scales—power, compute, security and the infrastructure needed to bring increasingly capable systems into the physical world. Others are applying AI to industries that traditional software never fully transformed, from government relations and private-equity diligence to hotels and healthcare.

The result is a batch pursuing ideas that would have seemed unusually ambitious for seed-stage startups only a few years ago.

Taken together, these companies point to a broader shift in where founders see opportunity in AI. As models become more capable, the opportunity is expanding beyond applications built on top of them to the infrastructure and industries being reshaped around them.

The next wave of AI may be defined not just by what can be built with the technology, but by what needs to be rebuilt around it.

AI Is Becoming An Infrastructure Problem

As AI systems become more capable, some founders are looking beyond the models themselves to the physical constraints underneath them—from power and data centers to chips and entirely new forms of computing.

Atomarine is betting that if data-center capacity can’t expand fast enough on land, some of it could move offshore. The company is developing floating data-center campuses on barges, using seawater for cooling, with the longer-term ambition of powering offshore compute with nuclear ships.

The founders see the ocean as a way around one of AI’s emerging physical constraints: “AI demand is growing faster than grids can expand, making the ocean one of the few places left where infrastructure can scale at the required speed.”

During YC, Atomarine says it signed letters of intent for compute demand, secured marine and software partners, and began building a floating GPU demonstration on a chartered barge at San Francisco’s Pier 50.

Torus is tackling another infrastructure constraint: the engineering capacity required to build power stations, data centers and industrial facilities. The company uses AI to automate engineering work that its founders say can take traditional firms weeks to complete.

For cofounder Marcus Lima, this is his second time through YC, after participating in the Summer 2021 batch with Heimdal. He returned in part for the pressure YC puts on founders to prioritize execution.

“YC is a forcing function for a company,” Marcus says. “It locks in a culture of execution from day one. Nothing accelerates and sharpens a company to focus only on what’s important like YC.”

Torus ultimately envisions what its founders call a “click-to-build world,” where much of the engineering work behind critical infrastructure can be automated.

Lamb Labs is developing Model Processing Units, or MPUs, an architecture that hard-codes an entire AI model into silicon. Rather than moving model weights repeatedly between memory and compute, its approach is designed to keep them on-chip.

The company didn’t originally plan to build hardware. Founders Niki, an AI researcher, and Thomas, a theoretical physicist, started with software optimization before repeatedly running into the same constraint.

“We became a chip company,” they say.

Lamb Labs says its architecture has demonstrated generation speeds above 20,000 tokens per second in some configurations, as the company bets that increasingly specialized hardware will be needed to support AI inference at scale.

Kara is questioning something even more fundamental: whether silicon should remain the dominant material at all.

The company is developing diamond semiconductor technology, betting on diamond’s ability to handle the heat and power density associated with increasingly demanding computing systems.

Its founders summarize the thesis in one line: “Today, we put diamond beside the chip to keep it cool. Tomorrow, diamond will be the chip.”

At the beginning of YC, Kara described itself through several possible products. Paul Graham pushed the founders to articulate the larger idea behind them. Their answer became much simpler: Kara is building the diamond semiconductor.

Two other companies in the batch are questioning an assumption deeper still: does the computer need to be made from conventional materials at all?

Parasma is training living human brain cells to perform computational tasks. Its founder came to biological computing through gaming and neuromorphic computing, eventually contributing to research involving teaching brain cells to play Doom. During YC, Parasma’s ambition expanded from building applications on biological computing toward developing more of the underlying stack itself.

Frontier Computing is also exploring biological neural tissue as a computing platform. Its founder began coding at eight and growing brain tissue at 17 after receiving an Emergent Ventures grant. The company is now developing biological neural systems as accelerators for machine learning, based on the idea that neurons naturally place memory and computation close together.

YC’s emphasis on speed has even affected Frontier’s wet-lab roadmap. The company says it has plotted a path toward 10 times more compute capacity by mid-2027 than it originally planned.

AI Is Moving Into The Physical World

As AI systems move from generating answers to taking actions, another challenge emerges: how to train, deploy and secure them in environments where mistakes can have real consequences.

Olam Labs is building multi-agent simulations where AI systems can run companies, simulate towns, play games and interact over long periods. The goal is to test capabilities conventional benchmarks struggle to capture, such as whether an autonomous agent can operate inside a complex economy or cooperate with other agents over time.

Cofounder Om Buddhdev has been building simulated worlds since childhood, from Roblox games at 11 and Minecraft economy servers to competitive gaming. He later worked across product, infrastructure and agents at AI Dungeon, where exposure to model evaluations helped inspire Olam.

The company started as something its 21-year-old founders were building for fun. They spotted the YC deadline on the final day, recorded their application video with minutes remaining and submitted it with roughly 30 seconds left. They got in.

Maingen is applying a similar thesis to the physical economy. Its founders describe the product as a “flight simulator for industrial AI”—interactive simulations where agents can practice operating factories, power plants and production lines without the consequences of making mistakes in the real world.

Founder Phillip Yan initially built preventative-maintenance software for solar farms, but the team found that the agents they wanted to deploy struggled in messy physical environments. During YC, Maingen pivoted toward reinforcement-learning environments for AI labs.

“Everyone talks about the 10x software engineer, but where’s the 10x process engineer or controls engineer?”

OS3 is building affordable robots for businesses around a simple idea: “The model is not the product, deployment is.”

That means asking decidedly unglamorous questions: How long does the robot run? How much does it cost? Can someone repair it quickly?

The founders learned the lesson firsthand after buying an imported humanoid robot for roughly $60,000. During testing, it lost power and collapsed onto founder Riso’s leg, damaging his apartment floor.

They decided to build their own.

OpenVector , meanwhile, transforms any existing camera into an automated 24/7 workforce without requiring new hardware.

During YC, the founders encountered use cases they hadn’t anticipated. The Hong Kong metro system, they say, explored using the technology to visualize passenger density; a gym wanted workout summaries generated from existing cameras; and a factory wanted inventory automatically updated as deliveries arrived.

The company’s bet is that physical AI will scale faster if businesses can use hardware they already have.

Fabraix is building automated adversarial testing for AI agents, based on the idea that increasingly autonomous systems need security testing that can adapt as quickly as attackers do.

One founder’s relationship with hacking goes back further: he says he was nearly expelled from university after hacking its registration system and gaining access to every account on campus.

His YC journey required persistence of another kind.

Fabraix was his tenth application.

During the batch, the company also discovered an unconventional sales strategy: break a prospective customer’s agent first, then send the company a report showing exactly how.

Apparently, breaking someone’s AI can be a surprisingly effective way to get a meeting.

AI Is Reshaping Industries Software Never Fully Transformed

Another group of S26 founders is betting that AI can finally automate industries where traditional software stopped short of actually doing the work.

Locke is trying to rethink how companies interact with government. Its founder previously worked across the White House, political campaigns, Deloitte’s government-relations team and startups.

At one point, he found himself on a call until 1 a.m. helping the CEO of a defense-tech company figure out how to register his business on a government website.

The experience crystallized the problem: highly technical founders often have no idea how to navigate government.

The founder says Locke went from zero to nearly $800,000 in annualized run rate in eight weeks and was on track for $1 million by Demo Day.

LATO is attacking another relationship-driven industry: private-equity diligence. Founder Tymek Staniszewski spent six years in private equity and remembers one $50 million transaction where his firm paid $400,000 for commercial diligence, received a 200-page report, used two slides—and still lost the deal because the research arrived too late.

LATO uses AI voice and research agents to automate parts of that process. Longer term, the founders want to simulate markets so investors can test decisions before making them.

LATO got into YC on its third attempt.

Rex is applying the same shift to enterprise finance. After five years building finance and revenue infrastructure and working with more than 100 finance-operations teams, its founders concluded that AI wouldn’t simply make finance employees more productive—it could eventually execute entire workflows. Rex entered YC focused on order-to-cash, but customer demand quickly pulled it into procure-to-pay and other processes.

Axelrod is going after an industry where outdated software may actually be the opportunity: hotels.

For cofounder Saman, the problem is personal. He spent four years working at the front desk of his father’s hotel and estimates that roughly half of every shift went to manually copying information between systems that couldn’t communicate.

Many of those systems were built long before the API era. Instead of waiting for them to modernize, Axelrod’s agents interact with them the way an employee does: looking at the screen and clicking the buttons.

The team’s counterintuitive insight is that the absence of APIs may not be an integration problem. It may be the moat.

“Every week we think bigger,” the founders say.

What once seemed futuristic—AI-native hotels—now feels achievable.

Arbital is betting that the lines between traditional and crypto markets will continue to blur—and that the next generation won’t think of themselves as stock or crypto traders, but simply as traders. The platform brings discovery and execution across stocks, crypto and other assets into one place. Its founders previously worked at crypto derivatives exchange Aevo, where they say they helped grow cumulative trading volume from roughly $1 billion to more than $100 billion. During YC, Arbital expanded its vision from bringing institutional-style market making to retail traders toward becoming the first place traders go to discover ideas, see what others are trading and execute across asset classes.

New Infrastructure Needs New Institutions

Building the physical infrastructure for AI creates another set of questions: who finances it, who insures it and who ultimately controls it.

PRINCEPS is building a specialist insurance company for the compute economy. The founders initially focused on software for hyperscaler grid connections, but their work across power systems and infrastructure financing exposed a different gap: as more capital flows into GPUs, data centers and power infrastructure, the risks surrounding those assets are becoming increasingly complex.

Rather than build another tool for existing insurers, PRINCEPS decided to become the insurer itself—what the founders describe as “coverage at the speed of compute.”

Exosat is building satellite connectivity with a focus on giving governments and enterprises more control over their communications infrastructure. For its founder, the problem is personal: growing up in rural areas, he remembers struggling with painfully slow satellite internet.

Exosat initially planned to build very-low-Earth-orbit direct-to-cell satellites capable of delivering inference directly to devices. But conversations with customers pushed the company toward a different vision: a politically neutral satellite network designed to give customers more choice over the infrastructure they depend on.

His reasoning for joining YC was unusually mathematical. If YC could improve his odds of building a Starlink-scale company by more than 7%, he calculated, the equity was worth it.

“There’s not much difference between having $5B or $6B, ” he says.

Personalization Moves From Promise To Product

Another group of S26 founders is betting that better AI and better measurement can move healthcare and consumer health away from population averages and toward the individual.

Omanta wants to give each patient something resembling a personal scientific R&D team. Starting with oncology, the company integrates medical and molecular data, searches for therapeutic options and works alongside physicians to pursue individualized strategies.

Its team previously built multimillion-dollar therapeutic programs for individual patients spanning cancer genomics, personalized vaccines, immunotherapies and drug repurposing, work they say helped an N-of-1 cancer patient who had exhausted standard care reach complete remission.

During YC, Omanta’s vision expanded from serving a handful of patients at extraordinary depth to using AI and automation to make what the founders call a “personal research lab” accessible to many more patients.

GutGutGoose started with founder Leon spending nearly $600 a month importing probiotics from the U.S. to Australia—only to find they sometimes did nothing or made him feel worse.

His gastroenterologist offered an explanation that eventually became the company’s premise: every gut is different.

Leon, who went to university at 14 to study financial modeling and later built a pharmacy business, began wondering whether the same mathematical thinking used to model markets could model the ecosystem inside the human gut.

GutGutGoose now sequences a person’s microbiome and uses metabolic modeling to select probiotic strains for that individual. The founders report an 83% colonization rate in their alpha cohort.

Lumeria is building an at-home imaging device designed to track how a person’s skin changes over time.

The problem is personal for both founders. One founder’s sister dealt with severe eczema for more than eight years before discovering that dust in her carpet was the trigger. The other founder developed eczema herself and waited months for dermatological care while receiving multiple misdiagnoses.

The founders entered YC expecting to launch around January 2027. Then a YC partner asked a simple question: “What is stopping you from launching next week?”

They accelerated the timeline, found beta testers and launched. Lumeria says it generated more than $100,000 in revenue in under 48 hours, along with more than 2.5 million views.

“Speed stopped being the risk and became the strategy,” the founders say.

Perhaps the most revealing thing about Y Combinator’s Summer 2026 batch isn’t how much AI is in it. By now, that is almost expected.

What stands out is where founders believe the next opportunities are emerging.

As models become more capable and intelligence becomes cheaper to deploy, the bottlenecks are shifting outward. Power matters more. Compute infrastructure matters more. Security, specialized hardware, real-world deployment and access to proprietary data matter more. And industries that resisted traditional software suddenly look more open to automation.

That shift is also changing what an AI startup can look like. It can be a semiconductor company, a robotics company, an insurer, a healthcare system—or a floating data center.

The defining question is becoming less “What can we build with AI?” and more “What becomes possible because AI now exists?”

Y Combinator’s Summer 2026 batch offers one early answer: AI may have started as a software revolution, but its next chapter is increasingly being built in the world around it.