YC S26’s top decile: The 24 companies RuntimeWire would call first after Demo Day

Why it matters

YC's strongest S26 pitches show AI companies taking responsibility for complete operations, while revenue and live deployments separate them from pipeline-heavy peers.

YC S26's top decile: The 24 companies RuntimeWire would call first after Demo Day

We reviewed the pitches, investor materials, public traction, founder histories and what companies have actually put into production. These are the 24 meetings we would take first.

Y Combinator held its Summer 2026 Demo Day on September 10th in San Francisco, presenting the batch to an invite-only audience of roughly 1,500 investors and media. Public S26 trackers were sitting around 240 companies as Demo Day wrapped, making 24 companies a useful approximation of the top decile.

RuntimeWire reviewed YC company pages and launch posts, company websites, founder material, press coverage, investor outreach arriving around Demo Day, and the YC investor materials available to us. We used those private materials to identify companies worth investigating, then looked for public evidence before relying on a claim in this ranking.

Our hierarchy was fairly simple: revenue already collected carried the most weight. Signed contracts and production deployments came next. Usage and measurable customer outcomes mattered considerably. Pilots carried less weight. We discounted pipelines, RFQs, letters of intent and projected annualized revenue heavily.

For hard tech and biotech, where revenue can arrive years after the underlying breakthrough, we gave more weight to functioning hardware, experimental validation and unusually relevant founder history.

Every operating figure below should be understood as company-reported unless otherwise specified. RuntimeWire has not audited these companies' financial statements.

Company logos below are pulled from YC investor materials reviewed for this story.

1. Datoric

The secure training-data engine for physical AI

Datoric has one of the cleanest traction stories in the batch.

The two-person company says it has built a network of more than 300,000 active contributors collecting multilingual speech, egocentric video and action-conditioned data for AI labs. More importantly, Datoric says it generated nearly seven figures of revenue in the last 30 days.

There is plenty of competition in AI training data. Datoric's interesting bet is that provenance and security become more valuable as labs move from text datasets toward human behavior captured in the physical world. Every recording is tied back to its contributor, device, consent and collection history.

The question we would ask first is customer concentration. If the revenue is spread across multiple model builders and gross margins remain attractive, this starts looking like infrastructure rather than a project shop.

2. SpaceFlow

AI procurement that is already responsible for real money

SpaceFlow says its system manages $400 million a year of supplier spend, with multiple six-figure contracts signed.

That makes the company unusually easy to understand economically. SpaceFlow isn't asking procurement teams to buy another dashboard. Its agents execute work against existing enterprise systems, with the ERP remaining the system of record.

The company reports 3x faster RFQ cycles, a 42% reduction in off-contract spending at one customer, and $120,000 in annual savings at another. One customer cancelled three planned procurement hires.

If SpaceFlow can continue increasing spend under management much faster than headcount, it begins to resemble an AI-native procurement services company with software economics hiding inside it.

3. Cosmic Robotics

The giant construction robot that actually goes to work

Cosmic has something many robotics startups still lack: robots working on real construction sites.

Its Cosmic-1 heavy-lift robot has been deployed on major US solar farms across tens of thousands of panel installations, where the company says it has more than doubled labor productivity.

The company is now pushing into data-center construction while working under contract with NASA on lunar construction robotics.

There is a nice compounding loop here. Customers pay Cosmic to perform economically useful work on Earth. Those deployments generate the operational data needed to improve the machines. The same systems eventually move toward increasingly difficult construction environments.

Fleet reliability, utilization and hardware margins are the questions now. The robot itself has already crossed the threshold from impressive demo to working equipment.

4. Traceforce

CrowdStrike for the AI apps employees installed themselves

Traceforce is protecting 3,000+ employee devices across six enterprises and says it has more than 40 active pilots, including Fortune 500 companies.

Its wedge is straightforward: security gateways can monitor traffic they know about. They have a harder time seeing an employee running Claude Code, Cursor, an MCP server or another agent directly on a laptop.

Traceforce operates at the endpoint and inventories which AI applications, MCP connections and skills are active, then gives security teams controls over what those systems can reach.

The founders previously ran engineering at Clumio, the cyber-resilience company acquired by Commvault.

Traceforce says its active pilot pool represents more than $3 million of potential ARR. We would focus heavily on conversion. If a meaningful share of those pilots become deployments, this could move up the ranking quickly.

5. Radley

Instead of selling AI to radiologists, Radley wants to own the radiology practice

Radley is pursuing one of the more interesting structures in the batch: use AI to rebuild the service business itself.

The founders say they are acquiring a radiology practice that would bring Radley to $4.6 million in annualized revenue, alongside $2.2 million in contracts.

They have some relevant scar tissue. The team met while helping Carbon Health scale to more than 100 clinics.

The important metric after the acquisition closes will be productivity: how many scans can Radley process per radiologist and per operations employee while maintaining clinical quality?

If the answer improves significantly, Radley captures the value inside the provider instead of negotiating software pricing with one.

6. Rex

An AI-native BPO for getting enterprises paid

Rex says its agents currently manage more than $500 million in receivables, including money owed by Fortune 100 companies.

The product automates order-to-cash work that still lives across invoices, customer portals, email chains and exceptions that traditional workflow automation has trouble handling. Synthesia is a named production customer.

The founders spent four years building Sequence, an a16z-backed finance-operations company, before attacking the same problem from an operating-services angle.

This is the kind of AI business model worth watching: price against outsourced labor and finance operations rather than another software seat.

7. Allia Health

An EHR distribution wedge turning into a national mental-health group

Allia says 8,000+ clinicians across more than 1,000 practices already use its system, supporting roughly 300,000 patients.

It is now using that installed base to assemble an AI-native national clinical group. More than 600 clinicians have joined, and Allia reports $50 million-plus in annualized care capacity under signed national insurance contracts.

Care under the new structure begins in September, which is why we treat the $50 million number as capacity rather than realized revenue.

Still, the distribution advantage is difficult to ignore. Most healthcare startups have to acquire clinicians after building the business model. Allia already has thousands operating inside its software.

8. Dialogus

This is what production voice AI looks like when the volume gets serious

Dialogus jumped into the top ten after we found its current production numbers.

YC says the company already powers voice operations for Fortune 500 enterprises. Dialogus's own customer materials say its systems handle millions of calls a year.

One carrier ran 42,536 calls in a single day through the platform. Dialogus reports 94% were resolved without a human.

That matters because voice-agent demos have become easy. Telephony, reliability, compliance, integrations, escalation and tens of thousands of simultaneous real customer interactions remain difficult.

Dialogus is trying to own that production layer, including the infrastructure that observes outcomes and automatically tests improvements to the agents themselves.

Those numbers put it firmly inside our top decile.

9. DeepReach

A distributed capture network for the physical-AI data shortage

DeepReach has placed 475 proprietary capture devices with more than 100 data partners across seven countries.

Those partners collect first-person data inside warehouses, farms, workshops, kitchens, repair shops and other environments that robotics labs would struggle to reproduce at scale.

The company says the network generated nearly 150,000 clips in three months, with volume roughly doubling month over month, and that its data is already being used in production by frontier-model and robotics companies.

The structural bet is compelling: physical AI needs breadth. Ten thousand hours from one lab environment can teach something different from ten thousand hours spread across thousands of real workplaces.

Disclosure: RuntimeWire founder Ryan Merket has an investment in DeepReach through a syndicate SPV on AngelList.

10. OpenRelay

An inference network that tries to make the chip disappear

OpenRelay says it is already processing 100 billion inference tokens a week across 22 physical locations, using eight accelerator SKUs.

Its goal is to route AI workloads across Nvidia, AMD, TPU, Trainium and other available capacity without requiring customers to choose the underlying hardware.

That could become increasingly valuable if inference infrastructure fragments across chips and clouds.

The founders have unusually relevant backgrounds: one helped build managed inference infrastructure at Voltage Park and was part of the Amazon Bedrock team at AWS.

At this level of usage, the next questions are financial ones: revenue, take rate, gross margin and how much of the demand is organic customer traffic.

11. Nori

A $1,688 humanoid with $440,000 of sales in six weeks

Nori is the company we most clearly underweighted in our first pass.

The company is building a bimanual mobile robot for $1,688, dramatically below the price of most general-purpose humanoids.

YC says Nori has generated more than $440,000 in sales six weeks after launch and has already deployed its first robot.

The hardware business alone is interesting. The data strategy may be more important.

Nori wants thousands of affordable robots operating in homes and businesses, with users teaching them new tasks. Each deployed machine then becomes a potential source of training data for more general robotic policies.

There is an enormous amount still to prove around manufacturing, reliability and support at that price. Demand is no longer hypothetical.

12. Stoa

Trying to build the price-discovery layer for AI hardware

Stoa recorded more than $300 million in RFQs during its first month.

We emphasize the unit: RFQs are requests for quotes. They are not completed transaction volume or revenue.

The opportunity remains fascinating.

GPU inventory still changes hands through broker networks, private spreadsheets and opaque negotiations. Stoa wants to build the market infrastructure around that activity: verified dealers, firm bids, contracts, shipping and settlement.

If enough transactions actually clear on the platform, Stoa begins accumulating something potentially valuable beyond fees: a proprietary history of what AI hardware is worth.

That could eventually matter to buyers, sellers, lenders and anyone financing GPU fleets.

13. Markov

Expert computer-use data for models learning to do real work

Markov says it has sold more than 33,000 hours of expert computer-use data to frontier AI labs.

Its open datasets have also crossed 200,000 Hugging Face downloads.

The specialization matters. Computer-use models need examples of people operating difficult software over long sequences: CAD, architecture, engineering tools and other applications where good training trajectories are expensive to produce.

The risk is commoditization. If general crowd-work networks can cheaply generate equivalent data, margins compress.

If expert workflows remain scarce and model labs repeatedly come back for difficult domains, Markov has a much better business.

14. Atlas Discovery

An AI-native pharma company already producing unusual evidence

Atlas says it has generated six figures in contracted revenue while signing an eight-figure LOI with an AI lab.

That LOI gets a discount in our ranking. The rest is harder to dismiss.

Three rare-disease foundations are testing drugs surfaced by Atlas's research agents, and the company has published results around clinical-trial prediction and drug repurposing.

Atlas reports nearly a 5x improvement over frontier LLMs on its drug-repurposing benchmark, along with research presented at ICLR, CSHL and ICML.

Biotech will eventually be judged by biology rather than benchmark scores. For a company this young, Atlas has already created several independent ways to find out whether the system works.

15. WonderTx

One of the batch's highest-credibility biotech moonshots

WonderTx is working on small molecules that could replace injectable medicines with pills.

Founder Abraham Heifets previously co-founded Atomwise, one of the earliest companies to apply deep learning to structure-based drug discovery.

The technical claim worth paying attention to is experimental rather than financial. WonderTx says that across four structurally and biologically different targets where its models had no target-specific training data, it successfully selected binders that were subsequently validated in the lab. Some were confirmed through crystallographic structures.

Drug development remains slow, capital intensive and unforgiving.

A founder who has already built a major AI-drug-discovery company, combined with wet-lab evidence on zero-shot targets, gets the meeting.

16. Avoca Systems

AI patient access that is already live across 250 radiology clinics

Avoca says its radiology-specific patient-access system is running across more than 250 clinics, including networks operating 150-plus locations.

The company reports some unusually tangible customer outcomes: call abandonment falling from roughly 30% to 2%, eligible-call booking conversion above 80% in major deployments, and patient satisfaction averaging 4.6 out of 5.

Avoca integrates into radiology systems rather than asking providers to replace them, automating phone calls, scheduling, referrals and follow-up.

Most of the company's current deployment history comes from Australia. Its US expansion is the next test.

If the results survive that transition, Avoca has a strong vertical-software wedge into a large and operationally messy market.

17. Manifold

Warehouse robots sold like temp labor

Manifold's business model may be as important as its robot.

The company deploys autonomous systems for warehouse picking, loading, unloading and palletizing without asking customers for large upfront equipment purchases. Customers pay per pick, at a price intended to sit below human labor.

YC says Manifold is already deploying robots across multiple warehouses under programs covering more than five million picks.

That structure attacks one of the biggest barriers to warehouse automation: convincing an operator to redesign a facility and approve a large capital project before knowing whether the robot works.

We would want to see actual completed-pick volume, robot uptime and contribution margin. The deployment commitment is enough to move Manifold into the 24.

18. Marble

A restaurant operating system with measurable restaurant economics

Marble is already live across multi-unit restaurant operations in the US and Canada, including A&W Restaurants and several independent groups.

Its system combines computer vision for inventory, demand forecasting and agents that handle purchasing, prep planning, invoice reconciliation, scheduling and other back-of-house workflows.

The company reports a 40% reduction in food waste, elimination of stockouts in deployments and more than 30 manager hours saved per week.

Restaurant software is a punishing business because customers churn when ROI is vague.

Marble's advantage is that its value proposition can be measured in food, labor and cash every week.

19. Torus

AI for the engineers building data centers, power plants and industrial infrastructure

Torus is working with multiple Fortune 500 companies on engineering for critical infrastructure, including data centers and energy installations.

The company describes itself as "Legora for physical engineering firms": an AI environment for generating engineering packages, diagrams and other design artifacts across disciplines.

That is a hard market to enter because mistakes have physical and financial consequences. The same problem can create defensibility once customers trust the system.

Founder Marcus Lima previously built Heimdal, a carbon-capture company. CTO Rahul Thayil has spent years working on engineering and automation systems.

We want to see project-level economics and proof that the claimed productivity gains persist on real builds. The customer set gets Torus into the ranking.

20. Locke

An AI-native government affairs firm with real early revenue

Locke says it passed a $500,000 annual revenue run-rate within six weeks of launching, working with customers across defense, healthcare and GovTech.

The company combines AI agents with human policy staff to monitor regulation, engage policymakers and pursue government contracts.

The business is interesting because government affairs is traditionally expensive, relationship-driven and labor intensive.

Locke becomes much more valuable if its policy agents let revenue grow substantially faster than the staff required to deliver the work.

Founder Parth Badhwar has worked on both sides of the problem, including policy roles at the White House and Deloitte before joining startup GTM.

The early sales velocity is difficult to ignore.

21. COACH

Gong for the sales conversations that happen in people's living rooms

Thousands of field-sales reps already use COACH, according to YC.

The company says it is live with the five largest in-person sales companies in France and has produced an average 32% increase in conversion rates.

Reps record face-to-face meetings from the mobile app. COACH identifies what the best sellers do differently and turns that into individualized coaching for the rest of the team.

The founders have direct experience with the problem: they previously built a 44-person field-sales organization that generated EUR 3.5 million in 14 months.

There are meaningful privacy and adoption questions around recording in-person sales conversations.

There is also a measurable reason for customers to tolerate the friction if the conversion lift holds.

22. Isengard Industries

A founder-market-fit bet on mass-produced defense systems

Isengard has less public current-company traction than most businesses above it.

The reason it remains in the 24 is Francisco Serra-Martins.

Before Isengard, Serra-Martins co-founded and ran Terminal Autonomy, which YC says scaled to $60 million in revenue serving customers that included the US Army, NATO and Ukrainian forces.

Isengard now has a 30-person team building locally manufacturable strike and counter-UAS systems.

That history matters in defense, where production, procurement and deployment knowledge often matter as much as a prototype.

We would want current orders, manufacturing economics and delivery capacity before moving Isengard much higher.

The founder has already demonstrated that he can sell and manufacture defense systems at meaningful scale.

23. Fabraix

An autonomous red team for AI agents

Fabraix says its Nyx security agent has found vulnerabilities in customer-facing AI systems at dozens of Fortune 500 companies.

On AgentHarm, an offensive AI-security benchmark, the company reports a 78% attack success rate, compared with 67% for GPT-5.6 Sol.

The product continuously attacks replicas of customer environments, placing malicious inputs into pages, documents, tools and other surfaces to see where agents break.

The market case is straightforward: AI agents can change behavior when the underlying model, prompt, permission or connected tool changes. Security testing once a year will not keep up.

The diligence question is commercial rather than technical now: how many companies turn a successful security assessment into recurring spend?

24. TryNearby

A local creator network where the distribution engine may be the moat

More than 100 restaurants subscribe to TryNearby, which automatically pairs local creators with local businesses and handles the logistics around the visit.

The company reports 37% month-over-month growth, greater than 90% monthly restaurant retention, more than 2,500 published creator videos and roughly 35 million views.

Its agents already handle more than 10,000 creator text messages a month.

The founder mix is unusually relevant. Yousef Abdelfattah co-founded FaZe Clan and spent years building creator audiences. His co-founders bring experience from creator infrastructure, Replit and consumer growth.

The question is whether restaurant unit economics remain attractive as the network expands city by city.

The retention number is the reason TryNearby holds the final slot.

What S26 is telling investors

There is a coherent pattern running through much of this list.

AI startups are increasingly taking responsibility for the operation itself.

SpaceFlow runs procurement work. Rex handles order-to-cash. Radley owns the radiology practice. Allia is assembling the clinical group. Locke employs policy staff. Marble reaches from inventory counts into purchasing and scheduling.

That expands the economic surface area available to the startup. A company replacing outsourced labor, administrative headcount or an operating department can eventually capture considerably more value than a narrow software tool.

The second major theme is physical AI.

Datoric, DeepReach and Markov are all selling varieties of scarce human-behavior data. Cosmic and Manifold are putting robots into economically useful environments. Nori approaches the same problem from the other direction: make the robot cheap enough that deployment itself creates the training network.

And then there is infrastructure beneath the models. OpenRelay is trying to arbitrage accelerator availability. Stoa wants to make GPU hardware tradable and priceable.

It makes S26 feel less like another batch of wrappers around foundation models and much more like a collection of companies figuring out what happens when AI touches the rest of the economy.

Why some spectacular pitches didn't make it

We were deliberately harsh on LOIs.

A $100 million letter of intent can matter. It can also represent almost no revenue.

The same applies to pipelines, marketplace RFQs, unsigned purchasing interest and annualized capacity that has yet to turn into delivered service.

Stoa's $300 million of RFQs remains impressive enough to make the ranking because an active marketplace can compound into price discovery. Allia's $50 million of contracted care capacity stays because it sits alongside 8,000 clinicians already using the underlying system. Atlas's eight-figure LOI gets attention because there is contracted revenue and biological work happening around it.

In each case, we tried to separate the exciting number from the thing underneath it.

The meetings we'd take tomorrow

If we had room for only five meetings after Demo Day, we would start with Datoric, SpaceFlow, Cosmic Robotics, Traceforce and Radley.

They represent five very different bets, which is part of what makes the batch interesting.

One has nearly seven figures of monthly data revenue. One is operating against $400 million of supplier spend. One has a ten-thousand-pound robot installing solar panels. One is already sitting on thousands of employee endpoints. One is trying to rebuild an entire medical specialty from the inside.

Six months from now, the ranking should look different.

That's the point.

Demo Day measures how well a company can compress its story into a few minutes. The more interesting race starts the morning after: who can turn the story into evidence fastest?

Those are the companies we'll keep watching.

RuntimeWire's ranking is an editorial assessment based on information available as of September 10th, 2026. It is not investment advice. Company-reported operating and financial figures have not been independently audited unless explicitly stated otherwise.