Instructure, the company behind the widely used Canvas learning platform, has disclosed that it recently suffered a cybersecurity incident and is now investigating its impact.

The U.S.-based education technology company is best known for developing Canvas, a widely used learning management system that helps schools, universities, and organizations manage coursework, assignments, and online learning.

“Instructure recently experienced a cybersecurity incident perpetrated by a criminal threat actor. We are actively investigating this incident with the help of outside forensics experts,” reads a statement from Steve Proud, Chief Security Officer.

“We are working quickly to understand the extent of the incident and actively taking steps to minimize its impact. Maintaining your trust is our highest priority, and we are committed to transparency throughout this process.”

Instructure says that it will provide new information regarding its investigation as it becomes available.

Since May 1, some services, including Canvas Data 2 and Canvas Beta, have been under maintenance, with customers warned they may experience issues with tools that rely on API keys.

The company has not stated whether this maintenance is related to the security incident.

BleepingComputer contacted Instructure earlier today with questions about the incident, but has not received a response.

BleepingComputer previously published and retracted an earlier report about this incident after determining it was based on incorrect information from a prior disclosure.

Targeting education technology firms

Threat actors have increasingly targeted education technology firms due to the large amounts of personal information they hold on students and teachers.

In January 2025, educational software provider PowerSchool disclosed a breach in which a threat actor claimed to have stolen data belonging to 62 million students.

In September 2025, Instructure disclosed a separate breach resulting from a social engineering attack that allowed attackers to access data in its Salesforce instance. At the time, a threat actor known as ShinyHunters claimed responsibility for the incident and listed the company on a data leak site.

Threat actors have also targeted Infinite Campus in similar campaigns, with claims of data theft from the company’s Salesforce environment.

AI chained four zero-days into one exploit that bypassed both renderer and OS sandboxes. A wave of new exploits is coming.

At the Autonomous Validation Summit (May 12 & 14), see how autonomous, context-rich validation finds what’s exploitable, proves controls hold, and closes the remediation loop.

Claim Your Spot

Microsoft has confirmed that Windows 11 is getting a new modern Run dialog with dark mode support and faster performance in a new preview build 26300.8346.

The Run dialog has been around since the Windows 95 era, and it is one of those small Windows features that many power users still rely on every day.

You just need to press Win + R, type a command, open a file path, launch a tool, or quickly jump to a location without opening File Explorer first.

With the new version, Microsoft is trying to modernize Run without changing what makes it useful.

Unlike the legacy Run, modern Run matches Fluent Design, supports dark mode out of the box, and is actually faster than the legacy Run.

That is interesting because modern counterparts usually have a reputation for slower performance.

Microsoft noted that designing a modern Run dialog for Windows 11 was challenging, as the company had to maintain the same performance and retain the minimal user interface, similar to the original Run that shipped with Windows 95.

“When we set out on creating the new experience, we knew the existing dialog was fast. We also knew we needed to be sure we deeply understood how you all used it. Modernize, be opinionated, and evolve it,” Microsoft explained in a blog post.

“To help evolve, we added a measure briefly to the dialog to see what was being used and to measure time-to-show. This confirmed a few key things that helped the design process.”

Microsoft says performance was one of the most important factors when designing the modern Run dialog. That’s because quite a lot of people use the existing Run dialog to paste text from the clipboard, then copy it again to remove text formatting.

This experience mostly works because of how fast the existing Run is. The legacy Run dialog takes approximately 103ms to appear after you press the Win + R keyboard shortcut.

Interestingly, the modern Run is actually faster. It has a median time-to-show of just 94ms.

“This was a huge team effort – we’ve collaborated tightly with partners across the platform to get these UI surfaces loading snappy. Improvements we’ve made to the platform don’t just make Run fast, but they help make the whole OS more efficient,” the company said.

Microsoft says it expects these numbers to improve as well as there is still room for improvement,

Microsoft drops ‘Browse’ feature in new Run

Microsoft looked at how people use the existing Run dialog before deciding what should stay and what could be removed. One example is the Browse button, which lets you browse a specific directory to open a program.

According to Microsoft, the Browse button usage is less than 0.0038%. This number is based on a sample of 35 million users who open Windows Run.

As a result, Microsoft has dropped the Browse button from the modern Run. The company argues that it researched how Run was being used and how fast it was, which helped form a baseline to build the modern Run.

Modern Run also supports ~\, which allows you to quickly access your home directory. It also shows icons in the list, which should make entries easier to identify without making the dialog feel too heavy.

How to enable or disable modern Run

While modern Run looks great and works well in our test, some of you may not like the idea, which is why the feature is entirely optional and tied to Advanced Settings in Windows.

According to Microsoft, modern Run does not get turned on automatically. Instead, you need to open Settings > Advanced Settings and manually enable modern Run, which replaces the legacy Run.

There are also plans to add more features to modern Run, and Microsoft says it is collecting feedback before rolling it out more broadly.

Other changes rolling out with today’s preview update

In addition to the Run dialog upgrade, Microsoft is improving Windows Share UI for AAD users.

Until now, if you wanted to add an app to the share dialog, you had to open the MS Store, install the app first, and then find it in the Share list. Now, you can install apps directly from the Share UI.

Last but not least, Magnifier now gives you more control over how you zoom, including preset zoom levels of 5%, 10%, 25%, 50%, 100%, 150%, 200%, and 400%.

These changes will roll out to everyone in the coming months, but for now, you’ll need to download Windows 11 Build 26300.8346 from the new Experimental Channel.

AI chained four zero-days into one exploit that bypassed both renderer and OS sandboxes. A wave of new exploits is coming.

At the Autonomous Validation Summit (May 12 & 14), see how autonomous, context-rich validation finds what’s exploitable, proves controls hold, and closes the remediation loop.

Advertisement

Claim Your Spot

The company that built TikTok’s recommendation algorithm, the system that predicts with unsettling accuracy what a person wants to watch next, is now using a related class of AI to predict how molecules will behave inside a human body. ByteDance’s drug discovery unit, Anew Labs, presented its first AI-designed therapy at the American Association of Immunologists’ annual meeting in Boston in mid-April, showing data on a small molecule designed by generative AI to inhibit IL-17, a cytokine involved in autoimmune diseases including psoriasis, rheumatoid arthritis, and ankylosing spondylitis.

The molecule targets a protein-protein interaction, a category of drug target that the pharmaceutical industry has spent decades calling undruggable because the binding surfaces are too large and too flat for conventional small molecules to disrupt. Anew Labs says its AI found a way in.

The presentation in Boston was the first time ByteDance showed the world what its drug unit has built. It will not be the last. The company is registered to exhibit at the BIO International Convention in San Diego in June, and its head of computational chemistry will present at the Free Energy Workshop in Barcelona next week.

The unit

The 💜 of EU tech

Advertisement

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Anew Labs operates from Shanghai, Singapore, and San Jose, California, with 36 core team members listed on its website and a scientific advisory board that includes Liu Yongjun, former president of Innovent Biologics, Ji Ma, a former principal scientist at Amgen, and Hua Zou, scientific director of protein chemistry at Takeda California.

The advisory board reads like a recruitment list from the companies that dominate biologics and immunology, disciplines where the targets Anew Labs is pursuing have historically required injectable antibody therapies costing tens of thousands of dollars per year. The unit’s ambition is to replace those injections with oral pills, using generative AI to design small molecules that can do what antibodies do but in a form that patients can swallow.

Chris Li, head of biology, presented one of Anew Labs’ four pipeline drug candidates in Boston. The molecule is a pan-spectrum IL-17 inhibitor, meaning it is designed to block multiple forms of the IL-17 cytokine rather than a single variant. Existing IL-17 therapies, including Novartis’s secukinumab and Eli Lilly’s ixekizumab, are injectable antibodies that generated billions in annual revenue by treating psoriasis and other inflammatory conditions. An oral small molecule that achieves comparable efficacy would be commercially transformative, both cheaper to manufacture and easier for patients to take.

The challenge is that IL-17’s binding surface with its receptor is a protein-protein interaction, a broad, shallow interface that gives small molecules very little to grip. The gap between what AI can do in a laboratory and what it delivers to patients remains the defining tension of health technology, and IL-17 is precisely the kind of target where that gap is widest.

The model

In March, Anew Labs published a preprint on bioRxiv describing AnewOmni, a generative AI framework trained on more than five million biomolecular complexes. The model is designed to work across molecular scales, from small chemical compounds to peptides to nanobodies, assembling chemically meaningful building blocks at atomic resolution.

In the preprint, the researchers demonstrated that AnewOmni could design functional molecules targeting KRAS G12D, one of the most studied oncology targets in the world, and PCSK9, a cholesterol-related protein, achieving success rates between 23 and 75 per cent with only low-throughput laboratory validation. The model uses programmable graph prompts that allow researchers to steer the generation process by specifying chemical, geometric, and topological constraints.

The technical approach is significant because it attempts to solve a problem that has limited AI drug discovery across the industry: most generative models work well at one molecular scale but fail when asked to design across scales. A model that designs small molecules cannot typically also design peptides or protein-based therapeutics.

AnewOmni claims to be the first framework to succeed at functional molecular design across all scales, which, if validated in clinical settings, would give Anew Labs a platform capability rather than a single-programme capability. Isomorphic Labs, the DeepMind spinoff backed by Eli Lilly and Novartis, released its own drug design tool in February that doubles the accuracy of AlphaFold 3, and has partnership agreements with combined milestone values of up to $3 billion. The race to build the definitive AI drug design platform is global, and ByteDance has entered it with a model that, on paper, addresses a limitation that its competitors have not yet publicly solved.

The context

ByteDance is not the first technology company to move into drug discovery. Anthropic acquired Coefficient Bio for $400 million in an acqui-hire that brought fewer than ten people into the AI company’s biological research efforts. Google’s DeepMind has been working on protein structure prediction since AlphaFold’s breakthrough, which won the 2024 Nobel Prize in Chemistry. Microsoft has invested in biology-focused AI through its partnership with Paige, a computational pathology company.

Nvidia has built BioNeMo, a platform for training and deploying biomolecular AI models. The pattern is consistent: the companies with the most advanced AI infrastructure are redirecting a portion of that capability toward biology, because drug discovery is a problem shaped like the problems AI is good at, searching vast combinatorial spaces for rare solutions that satisfy multiple constraints simultaneously.

What distinguishes ByteDance’s entry is the source of its AI expertise. TikTok’s recommendation engine is, at its core, a system that models human behaviour by processing enormous quantities of data and predicting which combinations of content will produce the desired response.

Anew Labs’ generative models do something structurally similar: they process enormous quantities of molecular data and predict which combinations of atoms will produce the desired biological response. The mathematical architectures are not identical, but the organisational capability, the ability to train large models on massive datasets, iterate rapidly, and deploy at scale, is transferable. ByteDance’s AI infrastructure, built to serve 1.5 billion TikTok users, is now being applied to a problem where the users are molecules and the engagement metric is binding affinity.

The test

More than 173 AI-discovered drug programmes are now in clinical development globally, with 15 to 20 entering large-scale trials this year. Whether AI will revolutionise drug development depends on how it is used, and the industry’s 90 per cent clinical failure rate has not yet demonstrably improved.

Insilico Medicine’s rentosertib, a first-in-class drug for idiopathic pulmonary fibrosis where both the target and the molecule were discovered using AI, showed positive Phase IIa results published in Nature Medicine. The Recursion-Exscientia merger created the most comprehensive AI drug discovery platform in the industry, but then discontinued its lead AI-discovered candidate after long-term data did not confirm earlier efficacy trends. The pattern across the field is promising early data followed by the same biological reality that has always made drug development difficult: molecules that work in a dish do not always work in a body.

Anew Labs has four pipeline candidates and a generative platform that, if its preprint results hold, can design functional molecules across scales. It has the backing of a parent company valued at roughly $300 billion with AI infrastructure that dwarfs most pharmaceutical companies’ computational resources. It has advisors from Innovent, Amgen, and Takeda.

What it does not yet have is clinical data. The IL-17 molecule presented in Boston was preclinical. The distance from a poster at an immunology conference to an approved oral therapy that replaces injectable antibodies is measured in years and billions of dollars, and most molecules that start that journey do not finish it. The most ambitious AI-biology startups are the ones whose founders understand that the algorithm is the beginning, not the end.ByteDance built an algorithm that changed how a billion people consume content. Whether the same company can build an algorithm that changes how a disease is treated is a question that no conference presentation can answer. Only a clinical trial can.

The flight from John F. Kennedy International Airport to the East 34th Street Heliport in Midtown Manhattan took seven minutes. By car, depending on the time of day, the same journey takes between 60 and 120 minutes. On Friday, Joby Aviation landed its all-electric air taxi at the heliport as part of a demonstration hosted by VertiPorts by Atlantic, the infrastructure company that operates the site. The aircraft, registration N545JX, had been flying across New York’s existing heliport network all week, touching down at the Downtown Skyport and the West 30th Street and East 34th Street heliports in Midtown after departing JFK. The flights were demonstrations, not commercial service. No passengers paid for a ticket. But the distinction between demonstration and operation is narrower than it has ever been. Joby cleared Stage 4 of the FAA’s five-stage type certification process in late March. Stage 5 is the final conformity inspection and operational demonstration. If it passes, and the company says it expects to by late 2026, the type certificate it receives will be the first ever issued for an electric vertical takeoff and landing aircraft in the United States.

The aircraft

The Joby S4 is a tiltrotor with six electric motors, four on the wings and two on the V-tail, that give it the vertical lift of a helicopter and the forward flight efficiency of a fixed-wing aircraft. It carries one pilot and up to four passengers, cruises at approximately 200 miles per hour, and has a range of roughly 150 miles on a single charge. The propellers tilt from vertical to horizontal after takeoff, allowing the aircraft to transition from hovering to cruising flight. A full recharge takes under 20 minutes. The aircraft weighs 4,800 pounds at maximum takeoff weight, roughly the same as a large SUV, and its wingspan of 39 feet means it fits on a standard helipad. The noise profile, according to Joby, is 100 times quieter than a conventional helicopter at the same distance, a claim that New York’s noise-sensitive residential neighbourhoods will have the opportunity to test if commercial operations begin.

Backed by $500 million from Toyota, which is also providing manufacturing expertise, Joby has built the aircraft through a vertically integrated model that includes in-house development of the electric motors, flight control software, and battery management systems. Joby has partnered with Air Space Intelligence to build the air traffic management systems that will coordinate eVTOL flights across urban airspace, a problem that becomes critical the moment more than a handful of aircraft are operating simultaneously above a city of eight million people.

The route

The New York demonstrations are the second stop on Joby’s 2026 Electric Skies Tour, a national campaign timed to the United States’ 250th anniversary. The tour launched in March with a flight over the Golden Gate Bridge in San Francisco, piloted by Andrea Pingitore, and moved to New York in late April. Joby has not disclosed which cities are next. The routes demonstrated in New York, JFK to Downtown Skyport, JFK to West 30th Street, JFK to East 34th Street, trace the commercial network the company plans to operate. Through partnerships with Delta Air Lines and Uber, Joby intends to offer an integrated travel experience: a passenger books a trip on the Uber app, takes a ground vehicle to the nearest vertiport, flies to the airport in minutes, and boards a Delta flight. The reverse works on arrival.

New York was selected in March as one of eight projects under the federal eVTOL Integration Pilot Programme, established by executive order under President Trump’s Unleashing American Drone Dominance directive. The programme, which spans 26 states, allows selected projects to begin supervised operations during a three-year pilot period, bypassing the traditional certification timeline for operational approvals. The Port Authority of New York and New Jersey and the New York City Economic Development Corporation are both partners. The infrastructure already exists: New York’s heliport network, built for helicopter traffic, can accommodate eVTOL aircraft with minimal modification. Dedicated vertiport infrastructure is being developed in other cities, but Joby’s New York strategy relies on the heliports that are already there, which means it does not need to wait for new construction to begin commercial service.

The economics

Joby has not announced official pricing. The company has said its target is pricing comparable to Uber Black, approximately $3 to $6 per mile. A trip from JFK to Midtown, roughly 15 miles by air, would cost approximately $200 per seat at the Uber Black rate, which is comparable to what Blade charges for its existing helicopter service on the same route and competitive with ride-share services that can charge $150 to $250 depending on traffic and surge pricing. The difference is time. Seven minutes versus 90 minutes changes the value proposition even if the price is similar. As fleet size increases and operations scale, Joby expects pricing to move toward Uber X rates of $2 to $3 per passenger mile, though that projection depends on manufacturing volume, battery costs, and utilisation rates that have not been achieved at commercial scale.

The financial picture for the broader eVTOL industry is mixed. European eVTOL startups have faced repeated delays amid certification hurdles, and at least six manufacturers have entered insolvency since 2023, including Lilium and Volocopter. Volocopter’s planned air taxi flights at the Paris Olympics were scrapped over certification failures. Joby’s principal American competitor, Archer Aviation, is also progressing through FAA certification and has said it expects to see first commercial revenue in 2026. But Joby is the most advanced Western eVTOL company by certification stage, and its partnerships with Toyota, Delta, and Uber provide manufacturing capacity, distribution channels, and a booking platform that no competitor currently matches.

The question

For a decade, the eVTOL industry operated primarily in PowerPoint. Concept renders of sleek aircraft gliding between rooftop vertiports, projected market sizes in the hundreds of billions, and timelines that slipped by years with each quarterly update. Joby’s own certification target has moved from 2023 to 2024 to 2025 to late 2026. The difference now is that the aircraft is flying, in public, over one of the most complex airspace environments in the world, on the routes it intends to serve commercially, using the infrastructure it intends to use. The FAA has confirmed that the propulsion system reliability and fly-by-wire redundancy meet Stage 4 requirements. What remains is Stage 5: the final conformity inspection that leads to the type certificate.

The question that the New York demonstrations answer is not whether the technology works. It does. An electric aircraft took off vertically from JFK, flew across the East River at 200 miles per hour, and landed at a helipad next to the FDR Drive without incident. The industry has been promising this for years. Joby has now shown it. The question the demonstrations do not answer is whether the economics work at scale, whether noise levels are acceptable to communities beneath the flight paths, whether the air traffic management systems can handle hundreds of daily flights over Manhattan, whether battery degradation will affect range and recharge times in commercial service, and whether passengers will pay $200 for a seven-minute flight when the same journey by car costs less and only takes longer. The car takes longer, but it does not require a booking on an app, a ride to a heliport, a security process, and a ride from a helipad to the final destination. The total door-to-door time advantage narrows when you count the ground segments on both ends. Joby’s bet is that the time advantage is large enough, and the experience is compelling enough, that a meaningful number of New Yorkers and travellers will choose the air. Late 2026 will determine whether they are right.

Perplexity’s Personal Computer arrived on Mac first because it was the best for locally deploying the agentic AI platform, the company has declared.

During Apple’s quarterly financial results. CEO Kevan Parekh referenced to the growing use of the Mac as a base of operations for AI platforms. Parekh namechecked Perplexity as one such company, taking advantage of Apple Silicon and its unified memory structure.

Ahead of Apple’s results release and analyst call, Perplexity CEO Aravind Srinivas talked about the Mac during its Ask NYC enterprise and finance event. Specifically, how Personal Computer, its agentic AI platform that works with local files, was made for use on Mac.

Personal Computer is a version of Perplexity’s Computer that uses multiple agents to perform tasks for the user. A key difference is that it handles tasks from a local computer, such as a Mac mini, instead of the cloud.

The local nature of the tool also means it has access to the user’s files on the Mac mini, with it able to create and edit them depending on the task at hand.

Deployment capacity and interconnectedness

Deemed one of the best and most accessible ways to deploy Personal Computer, the Mac mini has been seized upon by users for the purpose of bringing Perplexity’s AI system into their homes and offices. It’s a combination of the cloud-based service and the local files owned by the user working in a single hybrid setup.

To Perplexity, the Mac mini is one of the “best ways to deploy” Personal Computer “at full capacity.” With the nature of work spanning the iPhone and the Mac, Personal Computer “builds on the continuity Apple users already expect” to get their tasks done.

This continuity between the iPhone and Mac is important, especially since users can command Personal Computer to do things on their Mac from their iPhone.

Since its launch in March, Computer has managed to complete more than $2.8 billion in labor-equivalent work. In the case of Perplexity itself, the team dogfoods with Personal Computer, claiming to increase revenue by five times while increasing headcount by 34%.

The event also included a number of other announcements affecting Personal Computer. The experience is being brought to Microsoft Teams, with Personal Computer able to be messaged directly or pulled into a channel without leaving a conversation.

App connections and workflows

There’s also a beta of Computer in Excel with a native side-panel, so the model and data can be viewed side-by-side in the spreadsheet tool. Perplexity is also working with 1Password to allow Computer to perform actions in password-protected tools, while preventing the model from seeing the user’s credentials.

Workflows in Computer will help bundle prompts, context, and the output format for specific enterprise tasks into a single starting point. The idea is to reduce the amount of technical work a user has to get done before the AI platform gets going.

A library of workflows is being built up for commonly-run tasks, with there being more than 70 at this time. They can also be shared between team members, scheduled separately, customized, and also run asynchronously.

Personal Computer functionality is available to all Perplexity Pro, Max, and Enterprise subscribers using a Mac. Pricing starts from $17 per month for Pro, $167 per month for Max.

As for anyone wanting to pick up a Mac mini or a Mac Studio to enjoy this for themselves, they may have a little trouble doing so. During the call, it was revealed that Apple is seeing huge demand and supply constraints that will stick around for several months.

Y Combinator published its Summer 2026 Request for Startups in late April, just days before the application deadline. The document lists 15 categories of companies that YC’s partners want to fund. Eight of them require capital, hardware, or both. The list includes AI for low-pesticide agriculture, counter-swarm drone defence, inference chips for space, lunar manufacturing from molten regolith, and semiconductor supply chain software for a process that crosses a dozen countries and takes five months to complete.

Each category is written by a named partner, and each reads less like a startup prompt than a thesis on why the economics of a particular industry have just shifted. The most influential startup accelerator in the world, the institution that funded Airbnb, Stripe, and Dropbox, is telling founders that the next decade of billion-dollar outcomes will come not from building software but from using AI to enter the physical, regulated, and capital-intensive industries that software alone never touched.

The thesis

The RFS opens with Garry Tan, YC’s chief executive, writing about agriculture. AI can now identify individual weeds and pests in real time, he argues, and when combined with robotic precision treatments, the result is farming that uses dramatically less pesticide while improving yield. The category is not agtech in the way Silicon Valley has historically understood it, which meant software dashboards for farm management. It is agtech that involves building physical robots, training vision models on biological data, and deploying hardware in fields.

Tyler Bosmeny’s entry on counter-swarm defence compares the companies he wants to fund to Cloudflare rather than Raytheon, software-defined defence systems that neutralise drone swarms at a fraction of the cost of traditional missile systems. The United States Department of Defense proposed more than $70 billion for drone and counter-drone systems in its latest spending plan, and defence tech is experiencing its strongest investment cycle in decades. Adi Oltean asks for founders who will 3D-print structures from molten lunar regolith and extract raw materials including silicon, aluminium, iron, and titanium through electrolysis on the moon.

The hard-tech categories are not aspirational filler. They reflect a structural change in what venture capital is willing to fund. Defence tech startups raised a record $49.1 billion in 2025, nearly double the prior year. Anduril, the autonomous weapons company, raised $4 billion at a $60 billion valuation in March. SpaceX has demonstrated that hardware-intensive businesses can produce venture-scale returns. The old assumption that hardware could not generate the margins or the speed that venture capital requires has collapsed, and YC’s RFS is the clearest institutional acknowledgement that the collapse is permanent.

The software that remains

Seven of the 15 categories are software, but none of them resemble the SaaS playbook that defined the previous decade. The category YC calls Software for Agents asks founders to rebuild every major software category for a world where the next trillion users are not people but AI agents. That means APIs, machine-readable documentation, command-line interfaces, identity systems, permissions layers, and payment infrastructure designed for autonomous programmes rather than human beings. Google rebranded its entire AI platform around agents at Cloud Next 2026, consolidating Vertex AI into the Gemini Enterprise Agent Platform and launching a $750 million fund to finance agentic deployments. Gartner predicts that 40 per cent of enterprise applications will include task-specific AI agents by the end of this year, up from less than 5 per cent in 2025.

The Company Brain category asks for a system that pulls knowledge out of every fragmented source inside a company, structures it, keeps it current, and turns it into what YC calls an executable skills file for AI. This is not enterprise search. It is a living map of how a company works: how refunds are processed, how pricing exceptions are decided, how engineers respond to incidents. The Dynamic Software Interfaces category is its mirror image, asking founders to rebuild software so that agents can operate it natively rather than scraping interfaces built for humans.

The SaaS Challengers category names the targets explicitly: ERP, chip design software, industrial control systems, and supply chain management. These are the categories where incumbent vendors charge the most and innovate the least, and where AI-native replacements could capture enormous markets if they can clear the switching costs.

The physical turn

The RFS entry on semiconductor supply chains may be the most revealing. A single advanced AI chip goes through approximately 1,400 process steps, crosses a dozen countries, and takes five months to manufacture. That supply chain is managed, as the RFS puts it, with spreadsheets, SAP, and phone calls. Diana Hu, the YC partner who wrote the entry, is asking for founders who will replace that infrastructure with software that can track, optimise, and predict across the most complex manufacturing process on earth.

The category sits at the intersection of every force currently reshaping the technology industry: the US-China chip export controls, the reshoring of semiconductor fabrication, the explosion in AI chip demand, and the geopolitical fragility of supply lines that route critical components through Taiwan, South Korea, the Netherlands, and Japan.

The space categories are similarly grounded in economics rather than aspiration. Reusable rockets from SpaceX and Stoke Space are about to produce a massive increase in the capacity to put objects in orbit, which means an equally massive increase in demand for the electronics that operate there.

YC wants inference chips optimised for mass, thermal performance, and radiation hardness. SpaceX and Blue Origin are already racing to put data centres in orbit, and the AI hardware that runs inference workloads in terrestrial data centres does not survive the thermal and radiation environment of space. The market for space-rated inference silicon does not exist yet. YC is betting that it will.

What changed

Y Combinator’s Spring 2026 RFS, published just three months earlier, listed eight categories. The Summer edition nearly doubled that to 15. The Spring list included AI for product management, government AI, AI-native hedge funds, and stablecoins. Those are recognisably software businesses with AI bolted on. The Summer list includes lunar regolith manufacturing and counter-drone defence systems. The shift between the two documents is the most dramatic reorientation in YC’s public investment thesis since the accelerator began publishing requests for startups.

The change reflects what has happened to venture capital more broadly. In the first quarter of 2026, $297 billion flowed into startups globally, 2.5 times the prior quarter and the most venture funding ever recorded in a three-month period. Accel raised a $5 billion fund on the back of returns from Anthropic and Cursor.

Andreessen Horowitz raised $15 billion. Thrive Capital closed more than $10 billion. The money is not looking for the next enterprise SaaS dashboard. It is looking for the companies that will apply AI to the industries where the margins are highest, the incumbents are slowest, and the barriers to entry have historically been physical rather than digital. YC’s RFS is the most explicit version of that thesis because it names the industries by name: agriculture, defence, space, semiconductors, medicine, manufacturing.

The stablecoin category, one of the few holdovers from the Spring list, reveals a different kind of ambition. YC describes stablecoins as sitting between the regulated and unregulated worlds, creating room for services that combine the strengths of both: yield-bearing accounts, tokenised real-world assets, and infrastructure that moves money faster and cheaper across borders. The AI Personalised Medicine category asks for agents that analyse genomic data, electronic health records, and wearable output to generate patient-specific treatment protocols rather than population-level guidelines. Neither category requires building physical hardware. Both require operating in industries where regulation, liability, and institutional trust are the barriers, not code.

The signal

YC’s Request for Startups is not a prediction. It is a commitment. The partners who write the entries are the partners who will evaluate the applications, and the categories they describe are the companies they will fund. When Garry Tan writes about agriculture robots and Tyler Bosmeny writes about counter-drone systems and Adi Oltean writes about 3D-printing on the moon, they are telling founders what the next YC batch will look like. The document is the closest thing the startup ecosystem produces to a forward-looking investment mandate from its most influential institution.

The mandate says that software is now the substrate, not the moat. The models are commoditising. The infrastructure is scaling. The interfaces are being rebuilt for agents. What remains scarce is the ability to apply that substrate to the physical world: to build the robot that replaces the pesticide, the chip that survives the radiation, the defence system that costs less than the drone it destroys, the supply chain software that tracks 1,400 process steps across 12 countries, the molecular model that designs a drug for a target the industry called undruggable.

Y Combinator built its reputation by funding two founders in a garage writing code. Its Summer 2026 RFS is a document that says the garage is no longer enough. The founders it wants now are the ones who can write the code and then build the thing.