Key takeaways
- Capital is concentrating in fewer deals, a record low of new investments in Israel in H1 2026
- Explicitly named categories get funded quickly, and become crowded
- 10 requests for startups in h2 2026, as well as the latest YC Summer 2026 RFS
Every edition of this series has asked the same question: what do investors want founders to build next? This one asks a different question instead: of everything YC, a16z and others asked for this year, what actually got funded?
There’s a practical reason for that shift. As of mid-July, no new “Fall 2026” request-for-startups list has been published. YC’s site still shows Summer 2026 as its most recent set of asks, and a16z hasn’t updated its “Big Ideas” memo since December. But there is still a signal: a16z’s Speedrun SR007 cohort is now underway, built around themes like agent-native infrastructure, AI-native services, voice agents, and AI selling to AI, all extensions of ideas YC has been pushing since early 2026.
When multiple accelerators converge on the same themes without publishing new lists, it usually means the market has already moved.
So instead of summarizing a list that doesn’t exist yet, this edition does something different: it checks the last two sets of requests, Winter 2026 and Summer 2026, against real funding activity from the first half of the year. You can track previous editions published in 2024 and 2025 part 1 and part 2 as well as the most recent one from February 2026. All the requests below are new.
Why Israel is a useful stress test
To ground this analysis, I looked at Israeli venture funding in H1 2026. Israel is a small, dense and well-documented ecosystem, which makes it possible to map actual deals against specific theses rather than relying on anecdotes.
The headline numbers look strong. According to IVC–LeumiTech Q2 2026 and my weekly tracking of funding deals in Israel via the #FIRGUN Newsletter data, startups raised roughly $7.6 billion in H1 2026. Q2 alone reached over $4.2 billion, making 2026 one of the strongest years on record by quarterly average.
But the structure of that funding tells a more nuanced story:
- Capital is concentrated. The top nine rounds in Q2 captured over 60% of funding.
- Late-stage dominates. D+ rounds took nearly half of all capital, while Seed–A accounted for just ~17% and pre-seed barely registered.
- Fewer new bets. Only 35.5% of VC/CVC activity went into first-time investments — a record low.
- Deal count is down. Q2 saw just 87 deals, far below historical averages.
There is plenty of capital, but much less permission to be merely interesting.
The sector mix has also shifted dramatically. Cybersecurity now accounts for roughly a third of all funding, while defence, space, quantum and semiconductors have grown rapidly. Life sciences, automotive and other previously large sectors have shrunk as a share of capital.
That shift mirrors, almost line for line, what YC has been asking for: more defence, more hardware, more deep tech, more infrastructure, and less generic SaaS. Mapping H1 Israeli funding rounds against the Summer 2026 requests reveals a clear pattern: many of the most prominent categories were not just validated, they were flooded with capital.
Looking across written requests and actual funding activity, a few patterns emerge:
- Explicitly named categories get funded quickly, and become crowded. Agent security and enterprise AI are now highly competitive.
- Adjacent categories often offer better opportunities. Quantum and AI-native discovery systems sit next to major themes but face less competition.
- Gaps matter. Categories with strong theses but little funding, like agriculture, may represent open ground.
The key is to treat requests for startups as forecasts to be tested, not instructions to be followed.
Update: YC Just Published Its Fall 2026 Requests for Startups
August 2026 update: When I originally published this post in July, Y Combinator had not yet released a Fall 2026 Requests for Startups list, so I put together ten areas I believed were particularly interesting for founders based on where technology, customer demand and venture funding were heading. YC has now published its official Fall 2026 RFS, and the overlap is striking. Physical-world AI, defence, human identity and trust, AI-native compliance, software maintenance and consumer AI all feature prominently in both lists. More broadly, YC’s requests reinforce a shift I’ve been writing about for some time: AI is moving from something that generates content and assists humans to infrastructure that increasingly acts in the digital and physical world.
Here are YC’s 13 requests, paraphrased with some additional context on why I think each matters.
1. AI tutors that grow with the student
YC wants personalised AI tutors capable of delivering something approaching private-tutor-quality education at consumer scale, particularly around fundamental skills such as reading, writing and arithmetic. The bigger opportunity isn’t simply another AI homework assistant, but a system that develops a persistent understanding of how a child learns and adapts its difficulty, explanations and teaching style over years. Importantly, YC frames this as augmenting teachers rather than replacing them, a distinction that will matter enormously for adoption by parents, schools and educators.
2. The new defence industrial base
YC’s defence request is particularly explicit: the U.S. military needs low-cost interceptors, next-generation sensors, drones, payloads, resilient logistics, advanced manufacturing and modular hardware built around open architectures. The underlying shift is economic as much as technological. When inexpensive autonomous systems can create expensive threats, defence systems designed around million-dollar responses become increasingly unsustainable, making cost per interception, manufacturing speed and rapid iteration strategic advantages. For Israeli founders, this should sound familiar. The combination of military experience with autonomy, sensing, cyber and hardware expertise makes defence one of the areas where Israel should continue producing important companies.
3. A cloud for “small software”
AI coding agents have made it incredibly cheap to create software for tiny audiences, sometimes an audience of one, but the infrastructure hasn’t caught up. AWS, Azure and the traditional cloud stack were designed for companies deploying products to thousands or millions of users, not organisations potentially running thousands of small applications created by individual employees and AI agents. YC wants infrastructure that makes these applications as easy to deploy, secure, permission and share as a Google Doc. If the cost of creating software approaches zero, the number of applications could explode, creating demand for an entirely different deployment and management layer.
4. Multiplayer AI
Most AI today is strangely solitary: you prompt an agent, work with it privately and perhaps send somebody else the output. YC believes agents should become collaborative environments where multiple people can enter the same session, observe what an agent is doing, redirect it and hand responsibility between humans and AI. Think Google Docs or Figma, but with agents as active participants. This becomes particularly interesting as agents move from tasks lasting seconds to workflows lasting hours or days, whether that’s a sales team collectively working on a deal, engineers sharing coding agents or lawyers collaborating with an agent on a transaction. AI may still be waiting for its multiplayer moment.
5. Compute at sea
Perhaps the most eye-catching request is moving AI infrastructure offshore. Compute increasingly faces constraints around electricity, land, cooling, grid connections and planning permission, and YC proposes modular data-centre vessels that can operate together as floating compute infrastructure. It sounds futuristic, but it fits a broader trend I’ve written about on VC Cafe: data centres are beginning to escape the traditional data-centre building, with experiments involving floating, underwater, modular, nuclear-powered and potentially space-based infrastructure. The larger point is that the bottleneck in AI is increasingly physical, and that is creating startup opportunities far beyond models and software.
6. Mass-market AI consumer products
Despite the explosion of generative AI products, relatively few AI-native applications have earned a permanent place on hundreds of millions of consumers’ home screens. YC believes the economics are changing as inference costs fall and models become capable enough to deliver persistent consumer value across learning, money, health, travel, entertainment and communication. But putting AI inside an app isn’t enough. The winners will still need what great consumer companies have always needed: retention, habit, distribution, identity and, where appropriate, network effects. Falling token costs create the opportunity, but they don’t create the product.
7. AI for an ageing population
Ageing populations are creating enormous demand for care while many countries face chronic shortages of caregivers. YC sees opportunities across conversational voice interfaces, safety monitoring, caregiving robotics and software helping families coordinate appointments, medication, emergencies and care. AI could be particularly powerful here because conversational interfaces remove many of the menus, forms and workflows that make traditional software difficult for older users. It’s a good example of AI potentially expanding the addressable market for software rather than simply making existing software more efficient.
8. Operating systems for physical-world work
Most software was built for people sitting behind computers, but most workers globally don’t. Construction, logistics, maintenance, manufacturing and field services increasingly involve three forms of labour: humans, robots and AI agents, creating an opportunity for a new software layer that coordinates all three. The next generation of vertical software may therefore look less like systems of record and more like operating systems deciding whether a task should be assigned to a person, robot or agent and then monitoring the outcome. Companies coordinating physical work can also accumulate proprietary datasets showing how tasks actually happen in the real world, potentially creating a valuable data moat.
9. Crypto infrastructure for the agent economy
YC remains bullish on crypto, but the thesis is increasingly about infrastructure rather than speculative tokens. Stablecoins, tokenised assets and programmable payment rails provide useful infrastructure for moving money globally, and AI agents make this more interesting because machines need ways to transact, manage permissions and potentially pay other machines. Existing financial infrastructure was overwhelmingly designed around humans and businesses, not autonomous software. Agentic commerce could therefore become one of crypto’s more compelling practical use cases.
10. Data for the physical world
Modern AI became powerful partly because the internet provided enormous quantities of text, images and code, but the physical world doesn’t have an equivalent dataset. Falling sensor costs, robotics and better models now make it possible to collect much richer information from agriculture, energy, logistics, construction and industrial environments. The opportunity isn’t simply gathering more data: better sensing allows companies to model physical systems, predict what happens next and eventually control them. Some of the next great proprietary AI datasets may therefore come from sensors rather than websites.
11. Proving you’re human
As synthetic voices, images and video become increasingly difficult to distinguish from reality, many of the internet’s existing trust mechanisms start to break. Seeing someone’s face on a video call is no longer sufficient proof that the person is real, creating demand for a new human-verification layer that doesn’t force users to sacrifice privacy. The applications extend far beyond deepfake fraud to financial transactions, messaging, dating, social networks, marketplaces and reviews. We’ve spent years asking humans to prove to websites that they aren’t bots. We may now need infrastructure allowing humans to prove to each other that they aren’t machines.
12. AI-native compliance infrastructure
Compliance remains surprisingly manual, particularly in regulated industries where companies still rely heavily on spreadsheets, specialist teams and fragmented software to monitor regulatory changes, maintain licences, prepare audits and generate reports. YC sees an opportunity to rebuild this stack around AI, but the more interesting companies won’t simply add copilots to existing compliance workflows. They’ll rethink what compliance looks like when regulatory monitoring, anomaly detection, reporting and audit trails can happen continuously, potentially turning compliance from a periodic human process into always-on infrastructure.
13. Self-maintaining APIs
AI has dramatically reduced the cost of creating software, but maintaining all that software could become the next bottleneck. When an API changes today, providers publish documentation or changelogs and customers are responsible for updating their applications. YC imagines agents reversing that relationship: when a provider introduces a breaking change, its agent could identify affected customer code, make the necessary modifications, test them and open a pull request. Think Dependabot for the API economy. As AI generates an ever-growing volume of software, an entire new category could emerge around keeping that software alive.
Ten requests for startups for Fall 2026
1. The reliability layer for enterprise agents
Security is only part of the problem. Enterprises need to know whether agents completed tasks correctly, used the right data and behaved within policy.
There is room for evaluation, simulation and observability platforms built around real business outcomes. One of the challenges lies in the indeterministic nature of LLMs. Run the same prompt 100 times and you might get very different responses. It might be ok for creative writing, but it’s not ok for business decisions like approving a loan or refunding a product. Startups are trying to solve this with harnesses or extensive stress testing of agents, or by analysing the logs (after the fact) and being notified of anomalies. Agent reliability is another one of those categories that may already feel a bit crowded, but the problem is big and still remains.
2. Maintenance infrastructure for AI-generated software
AI makes it easy to create software, but not to maintain it.
The next bottleneck is managing complexity: dependencies, architecture drift and security. Founders can build systems that continuously map, test and explain codebases. I’ve written about ‘Cheap prototype, expensive maintenance’. All those vibe coded projects built with Lovable, Base44 and Claude Code artefacts were incredibly exciting to launch with prompts. But as they are used with real users and stress tested at scale, the bugs start to creep up, making maintenance precarious. Code review or automatic bug patching won’t do the trick here.
3. The economics control plane for AI
Companies still struggle to connect compute usage to business outcomes.
There is an opportunity to build financial infrastructure for inference: routing, cost measurement, budgeting and decision-making about when intelligence is worth buying. The rise of open-source LLMs, many of them Chinese, is currently being scrutinised by the US government, so there’s a need to better manage tokens, usage and the economic efficiency of AI. I wrote about this in my latest post on Tokenmaxxing being the wrong metric to track. I’ve seen ‘model-routing’ startups take a gatekeeping approach where every prompt will be analysed and routed to the best/cheapest model for the task. I don’t think it will work from a privacy perspective, so the door is open for better solutions.
4. Identity, payments and reputation for an agent economy
Agents are beginning to transact, but commerce infrastructure assumes humans.
We need identity, delegated authority, payments, escrow and reputation systems designed for machines. According to Cloudflare, between June 2025 and April 2026, human traffic to sites of businesses in many industries fell ~40%. Users are still buying products online, but they’re relying more and more on AI search and LLMs to do their research. The next step will be to let the agents transact, but there’s a whole tooling layer required to do that safely.
5. AI-native service companies in regulated markets
Instead of selling tools, become the service provider.
Tax, insurance, healthcare administration and compliance are all ripe for AI-native companies that deliver outcomes directly. I’ve written about AI native services in VC Cafe based on data from Sequoia’s Julian Bek, and others, and I foresee this trend will continue to grow.
6. Systems of action for the physical economy
Despite massive inefficiencies, sectors like construction and manufacturing remain underfunded.
The opportunity is to build systems that not only analyse data but take action, scheduling, sourcing, pricing and execution. Agentic AI for non-sophisticated ICPs is attractive because their margins are low (so they are desperately looking for more efficiency). In addition, the manufacturing, construction or even retail ICP would most likely choose to ‘buy’ (vs. build) due to lack of tech talent in their organisation. The caveat is that it’s not ‘one size fits all’ and go to market may feel more like a services company than a software company selling ‘saas’ licences/ seats.
7. Cheap defence against cheap autonomy
The economics of defence are shifting. Cheap drones require cheap countermeasures. Israel’s battle again Hezbo**a and the Ukraine-Russia war have demonstrated the scale of the drone threat.
Opportunities exist in sensing, electronic warfare and autonomous defence systems optimised for cost per threat neutralised. This is a category that is getting a lot of funding already and may very well become crowded, but it seems like the demand is there not just from homeland security, but also for protecting critical infrastructure.
8. Evidence infrastructure for personalised medicine
Personalised medicine requires trust, validation and longitudinal data.
Startups can build the infrastructure that makes recommendations explainable, auditable and clinically useful. We’ve seen a lot of funding going into drug discovery, but advancements in models and the availability of genomic data open interesting possibilities in the space of personalised medicine.
9. The task economy for training and supervising AI
As AI systems become more capable, they require better data and human oversight.
The next generation of data companies will organise expert knowledge into structured, high-quality tasks. This is a category that got popularised by the likes of Mercor and Micro1, but it continues to be relevant, especially as organisations aim to refine their models based on proprietary data.
10. Consumer AI that earns repeat behaviour
Consumer AI is not dead, but the bar is higher. Since the introduction of vibe coding, the number of apps on the app stores has exploded, but usage hasn’t increased massively.
The opportunity is to build products where AI enables something new, but long-term value comes from habit, identity and relationships. Founders that manage to find a niche, create a habit and get the engagement and retention needed to build a big business, may strike gold here. Gaming is part of this story.
Here you have it in one image:
What can you take away from this as a founder
The most important insight from H1 2026 is not which categories are hot. It is how quickly capital moves once a thesis is validated.
By the time a category appears on multiple investor wish lists, dozens of teams are already building in it. That also puts the VCs in a tricky situation, where even though the category seems ‘hot’, competition will make it tough to scale.
The better strategy is to identify the underlying shift, from copilots to agents, from tools to outcomes, from software to systems, and then find the implication that has not yet been fully funded.
The gap between the request and the deal flow is where the next opportunities will come from. No matter what you choose, keep in mind that the timeline for startups to exit didn’t shorten significantly, and while many startups eventually pivot, you should be willing to commit the next 7-10 years of your life working on the problem you’re focusing on, so choose wisely!
Shameless plug, if you’re an Israeli founder building in these areas, I’d love to hear from you. At Remagine Ventures, we’re often the first believers in the startup, providing pre-seed funding and helping the teams traverse the most risky part of their startup journey.
For more Requests for Startups (RFS) check out the Y Combinator Summer 2026 list that came out a few hours after the publishing of this post.
More requests for startups from Pear VC and Engage.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- Weekly Firgun Newsletter – September 4 2026 - September 4, 2026
- Wonderful’s $5B Bet to Own the Enterprise AI Layer - September 3, 2026
- Why Israeli Unicorns Are Built in America - September 2, 2026
