It is the last day of summer, which feels like the right time to close the book on YC's Summer 2026 batch.
Team Ignite invested in 17 companies from S26. That matches the 17 we backed in Spring and the 16 we backed in Winter. The count was stable. What we bought was not.
This post does two things. First, our takeaways on the batch itself, and how they line up (and don't) with what we wrote after Winter and Spring. Second, the 17 companies: what each one does, where to find them, why we invested, and what has to go right. Every company fact below comes from public launch pages, company websites, and public profiles. Nothing here comes from a data room.
Part One: What We Saw in the Batch
1. The biggest batch since YC went quarterly, and it moved down the stack
By Demo Day eve, YC's public directory listed 234 Summer 2026 companies, and the cohort was roughly 20% larger than the batch before it. That makes it the largest batch since YC shifted to a quarterly cadence, though still well short of the mega-batches of the early 2020s. Size is not the story anyway. Composition is.
One comparison of Spring and Summer 2026 found AI infrastructure rising from 8% to 20% of the batch, AI products falling from 55% to 39%, industrials doubling from 12% to 24%, and autonomous agents falling from 45% to 33%. A separate analysis counted 64% of the batch as AI-native, with agent companies alone at 22%, and industrials jumping from 12.8% to 23.7%.
Put plainly: Spring was the batch of agents doing jobs. Summer was the batch of the plumbing and the physical infrastructure those agents run on. Model training, evaluation, routing, security, chips, data centers, robots, and the data that teaches robots.
2. More of the same? No. This is the second derivative of the same thesis
Read our last three batch posts back to back and the arc is clear.
- Winter 2026: we titled our preview "AI That Does the Job", and followed it with the 16 companies we backed.
- Spring 2026: we argued that software is becoming labor, and that the batch felt like a map of where founders think AI will first win budget authority. The companies we backed are here.
- Summer 2026: the answer to what comes next. Once software is labor, the scarce inputs are no longer the agents. They are the compute, the data, the permissions, the security, and the physical-world interfaces those agents need in order to work.
So this batch was not more of the same. It was the consequence of the same.
Our own portfolio shows the shift more sharply than the batch does. In Spring, most of our 17 checks went to applied AI owning a workflow: therapists, pathologists, specialty practices, construction estimating, customer success. In Summer, only four of our 17 are classic "AI owns the workflow" companies. The other thirteen sit in infrastructure, security, compute, physical AI, and defense.
We did not set out to rotate. The best founders rotated, and we followed them.
3. We warned about confusing a layer with a company. Then the batch was mostly layers
In June we wrote that the danger is confusing a layer with a company. Summer 2026 tested that warning hard, because the fastest-growing slice of the batch was layers.
The clearest example is "train your own model from your production traffic." At least five companies in the batch pitched some version of it. Orchestra trains specialized models from work traces, Riften routes company AI traffic and uses that work to develop private models, and Experiential Labs uses production traffic to improve company-owned models. Belvedir and Agnost both build toward the same destination from different starting points. We own two of the five. That is a crowded trade and it would be dishonest to pretend otherwise.
Robotics training data was the physical-world version of the same crowding. Capture networks, tactile gloves, data-quality APIs, in-store collection, and expert demonstrations all showed up in the same batch.
Our view: most of these become services businesses with good revenue and mediocre multiples. The few that become venture-scale will own something that compounds, such as a physical capture standard, a proprietary failure dataset, a distribution network, or a system of record. Everything else is a feature waiting to be bundled by the platform beneath it.
We still bought layers, deliberately. At our check size, being wrong on one layer costs us a small check. Missing the layer that becomes the control point costs us the fund.
4. Security is the tax on software-as-labor
Three of our 17 checks are security or governance companies. NebuSec defends code and cloud. Fabraix attacks customer-facing agents before adversaries do. Decawork governs what internal agents are allowed to do.
That is not a coincidence. If AI agents hold credentials, move money, touch production systems, and talk to customers, then every one of them is an insider with no background check. Meanwhile frontier models have gotten good enough at finding exploitable bugs that the offense and defense balance has shifted, and the most capable models are gated behind limited-access programs that most defenders will never get into.
In Spring we underwrote agents winning budget authority. In Summer we underwrote the security spend that budget authority forces. This is one of the few categories where we expect incumbents and startups to both win, because the attack surface is growing faster than anyone can cover it.
5. The AI compute supply chain is now a seed-stage category
For years, "AI infrastructure" at seed meant software running on someone else's GPUs. This batch went further down, into the physical and financial layers. On the physical side: floating data centers powered by nuclear energy delivery ships, diamond wafers and semiconductors, modular data centers for sites with stranded power, and biological neural networks as a computing substrate. On the financial side: a marketplace for buying and reselling future AI inference capacity, a derivatives exchange for compute, and insurance for data centers and compute providers.
We backed three companies here: hardware intelligence in chip verification, Marengo in data center design and permitting, and Stoa in price discovery and trading for GPUs. This connects straight back to Spring, when we backed Expanse for GPU utilization. We now have exposure to the compute chain from the design of the chip, to the design of the building, to the market where the hardware trades, to the scheduler that runs the job.
6. Atoms are back, and defense is a standing position, not a phase
Industrials doubling in a single batch is the most important number in this post. After Fall 2025 we wrote about drones on the hunt. In Spring we backed Arlo Industries for passive aerial sensing. In Summer we backed Vernius Systems for the radar seekers that let a cheap interceptor actually hit something.
That is deliberate. We are assembling a counter-drone stack across batches: detection in one company, terminal guidance in another. Cheap drones broke the cost curve of air defense, and the fix will be built by startups shipping components, not primes shipping programs.
On robotics we were more selective than the batch was. We backed touch data (6thSense), deployment infrastructure (Agency Tool Company), real-world capture networks (DeepReach), and camera-based automation (OpenVector). We passed on most full-stack robot builders. The picks-and-shovels layer of robotics is where a small early check still buys real ownership before the capital intensity arrives.
7. Prices went up again, and value capture got harder at the same time
In June we wrote that the new center of gravity seemed to be around a $30 million valuation cap, with $40 million or $50 million no longer reserved for the obvious monsters. In Summer, most of what we saw priced between $30M and $40M post-money, and a handful of the hottest physical-AI and marketplace rounds cleared $70M. That is a meaningful step up in one batch.
It is also the worst possible combination. Prices rose while the batch shifted toward layers, where value capture is hardest to prove. More founders than ever raised before a lead priced them, on uncapped MFN SAFEs, letting the market set the number later.
We leaned into that structure in a large share of our deals. An MFN means our SAFE picks up the best terms anyone else in the round gets (same as YC's $375k standard deal). When you are writing early, small, and fast into a rising market, structure is how you manage price risk without slowing down.
8. Founder quality was high, and so was credential inflation
This was the most technically credentialed batch we have seen. A world number one CTF hacking team. A CVPR spotlight author. A PhD whose dissertation is the core of the product. Two founders who already built and sold a robotics company.
It was also a batch where we had to verify more than usual. Headline credential lines on launch pages sometimes decomposed into internships under inspection. Round sizes and pipeline totals frequently did not reconcile with the underlying documents. At least one founder planted a prompt injection in their public profile, aimed at investor screening bots.
Two lessons. For founders: investors now run diligence with AI, and AI reads everything, including the things you assumed no human would check. For investors: verify the single claim the whole pitch leans on before you look at anything else.
9. What was missing
Consumer was effectively a rounding error. Consumer stayed a minority category at around 5% of the batch, consistent with the last several cohorts.
Traditional vertical SaaS for small businesses is close to gone, replaced by companies that sell the outcome or become the service business outright. This batch had an accounting firm where agents do the bookkeeping and a CPA signs off, an AI-operated collections agency, an insurance carrier using AI to underwrite and price risk, and a short-term rental property management company running its operations with agents.
That is the Spring thesis playing out: the winners stop sounding like software vendors as fast as possible.
We also think consumer is under-owned at seed right now. The next great consumer company will probably look like a toy when it shows up. It did not show up in this batch.
Part Two: The 17 Companies We Backed (Alphabetically)
6thSense
What they do: 6thSense builds wearable tactile sensing rigs that capture touch, pressure, vision, depth, and hand pose while people perform real tasks, then sells that data to robot-learning teams. The sensor layout is designed to mirror the tactile skin on robot hands. Website: 6thsense.dev YC Profile: ycombinator.com/companies/6thsense
Why we invested: Dexterous manipulation is the hardest open problem in robotics, and video alone cannot teach it. A robot that only sees cannot learn how hard to grip an egg versus a wrench. Touch is the missing modality, and the labs building humanoids and manipulators know it. The data-vendor shape has enormous precedent one layer up the stack. The team includes a founder who has built tactile data capture before at a leading robotics company, and the hardware works today.
What has to go right: 6thSense needs to become the standard tactile format before large customers build gloves in-house or open-source designs commoditize the rig.
Agency Tool Company
What they do: Agency Tool Company builds infrastructure for shipping software to real-world robots. ATC Deploy pushes over-the-air updates to robot fleets by sending only the bytes that changed and resuming when the network drops, which the company says is about 20x faster than a Docker pull. ATC Build offers CI runners on real embedded hardware such as NVIDIA Jetson, so teams test on the computer the robot actually uses. Website: agencytool.com YC Profile: ycombinator.com/companies/agency-tool-company
Why we invested: Jack Morrison and Davis Foster spent eight years at Scythe Robotics, took an autonomous commercial mower from concept to hundreds of machines running daily, and sold the company to ASI in 2026. They hand-rolled this tooling for years and are now building the version they wish they had had. Deploy launched with partners across agriculture, construction, and logistics robotics. Our long-term view is simple: if there are going to be billions of robots, every one of them needs a CI/CD pipeline. Whoever owns robot DevOps sees every release, every rollback, and every fleet.
What has to go right: Device OTA and fleet management has historically produced good businesses rather than venture-scale ones. ATC needs pricing that scales with machine count, and it needs Build to become where robot software is validated, not just shipped.
Agnost AI
What they do: Agnost monitors AI agents in production. It analyzes real conversations to detect silent failures and user frustration, clusters them into named issues, and proposes fixes a team can approve. It is also building toward training smaller specialized models from those production traces. Website: agnost.ai YC Profile: ycombinator.com/companies/agnost-ai
Why we invested: Every company shipping an agent is flying partially blind. Traditional product analytics tracks clicks, but agents fail quietly inside conversations, and nobody files a ticket. CEO Shubham Palriwala was the first engineer at Formbricks, an open-source product-feedback platform, so he has built this category before. CTO Parth Ajmera built large-scale data pipelines at Microsoft. The strategic bet is that owning labeled failure data gives a head start in distilling cheaper custom models, because you already know exactly where the frontier model breaks.
What has to go right: This is the most crowded corner of the batch, with LangChain and several batchmates shipping adjacent loops. Agnost has to turn its failure dataset into switching costs before observability platforms reframe for the same buyer.
Belvedir
What they do: Belvedir is a custom model factory. It captures a company's agent and tool-use traces, trains and benchmarks private models on that data, deploys them on infrastructure the customer controls, and keeps improving them from production usage. The stated ambition is to make private models cheap enough that the world can have trillions of them. Website: belvedir.ai YC Profile: ycombinator.com/companies/belvedir
Why we invested: The economics of AI are shifting from renting one giant model to owning many small ones. Open-weight models plus automated fine-tuning make that viable for ordinary companies, not just labs. Founder Zachary Yu has worked this problem from several angles, including reinforcement-learning data at Mercor and RL environments at his previous company. Early pilots reported large cost reductions against frontier models on the same tasks.
What has to go right: Belvedir competes with well-capitalized inference clouds and several batchmates. It has to own the loop from traces to deployed model tightly enough that nobody wants to reassemble it from parts.
COACH
What they do: COACH gives every field sales rep personalized coaching after every in-person meeting, helps managers see where to spend their coaching time, and turns what top performers do into company knowledge. Website: getcoach.com YC Profile: ycombinator.com/companies/ai-coach
Why we invested: Inside sales has been instrumented for a decade. Every call recorded, every deal scored. Field sales, meaning the reps knocking on doors, walking showrooms, and closing in customers' living rooms, is still mostly a black box. Mathieu Perez, Thomas Perez, and Nicolas Sebag built fast traction in Europe, a market the US category leaders largely ignored, and are now moving into the US. We liked the founder-market fit, the speed of execution, and a business that looked structurally healthy for its stage.
What has to go right: There are well-funded US incumbents in conversation coaching. COACH has to win US accounts on coaching quality rather than geography, and build a compounding data asset out of field conversations.
Decawork
What they do: Decawork trains small, company-specific models that approve or reject AI agent actions in real time, using company policy, task context, and past actions. Around that sits the control plane: agent identity, scoped credentials, approval gates before production access, full audit trails, and retirement of agents nobody uses. Website: decawork.ai YC Profile: ycombinator.com/companies/decawork
Why we invested: Agent sprawl is the new shadow IT. Employees ship internal agents with coding tools faster than IT can inventory them, and the last year produced several public security incidents that started with an AI tool holding more access than anyone realized. Aman Raj built AI compliance tooling at Barclays and previously founded a fintech company. Sarthak Aggarwal built AI systems at NVIDIA. The reason we wrote the check is what could go right: if an independent company becomes the authority layer for the AI workforce, the way Okta became that layer for humans, the outcome is enormous.
What has to go right: The field is crowded and Microsoft, Okta, and AWS all ship native agent governance. Decawork has to win on the authority model itself, not on being another credential vault.
DeepReach
What they do: DeepReach recruits local entrepreneurs around the world who take on the company's stereo capture devices, hire local workers, and record real-world physical work. That data is sold to frontier AI and robotics labs. Website: deepreach.ai YC Profile: ycombinator.com/companies/deepreach-inc
Why we invested: Physical AI has a data diversity problem. A model trained on footage from one lab in one city does not generalize to a kitchen in Lagos or a warehouse in Manila. DeepReach's answer is a franchise-style network instead of a centralized collection lab, which makes distribution capital-light and genuinely global. The founder fit is unusually tight. CEO Tim Li ran a global staffing company for years, which is exactly the hard part of this business. Co-founder Chris Liu holds a USC PhD, was a research scientist at Meta, and published on crowdsourced 3D capture of the physical world. At launch the company reported more than 150 entrepreneurs, over 1,000 workers, seven countries, and 500,000+ clips within roughly three months.
What has to go right: Better-capitalized data vendors can copy a recruiting mechanic. DeepReach has to convert its head start into network liquidity and quality advantages that compound faster than a rival can outspend it.
Evergrove
What they do: Evergrove is an AI workforce for workers' compensation. Its voice agents make and receive the care-coordination calls that run the industry: intake, triage, scheduling, dispatch, records retrieval, and billing, for managed care organizations, TPAs, and provider networks. Website: evergrovelabs.com YC Profile: ycombinator.com/companies/evergrove
Why we invested: Workers' comp is permanent, regulated, non-discretionary, and still runs on phone, fax, and hold music. That is exactly the operationally ugly market we wrote about in Spring, where the pain is real and nobody else wants to finish the diligence. Sid Unnithan and Jacob Chia both came from Newfront, where they built voice AI for insurance workflows, so they have shipped close to this product before in the same domain. The company already reports hundreds of thousands of calls handled in production, which is a different thing from a pilot. This is software as labor in the most literal form in our portfolio.
What has to go right: Evergrove needs account-level expansion and real switching costs through workflow depth and compliance, ahead of same-vintage competitors and claims platforms building AI in-house.
Fabraix
What they do: Fabraix builds red-teaming AI agents that continuously find security vulnerabilities in customer-facing AI. Its agent attacks chat, voice, browser, and coding agents black-box, with no integration and no source code access. Fabraix also runs Playground, a public arena where anyone can try to break live AI agents for weekly prize money, with the system prompts published so you can read what the agent was told. Website: fabraix.com YC Profile: ycombinator.com/companies/fabraix
Why we invested: Every customer-facing agent is an attack surface, and most companies deploy one without ever seriously trying to break it. Ahmed Aly came out of fraud data science at Monzo and Ibrahim Abdu from engineering at Meta, so they think about adversaries for a living. Playground is the interesting strategic piece. If it works, it becomes a crowdsourced engine for novel attacks, and a live attack library is a moat in a way that a static benchmark never is.
What has to go right: AI security is consolidating fast. Fabraix has to show that its attack data compounds, and extend from finding problems into preventing them at runtime.
hardware intelligence
What they do: hardware intelligence builds AI tools for the chips AI runs on. Wave is a terminal-native, agentic waveform debugger: engineers open enormous simulation traces instantly and ask, in plain English, why a signal is wrong at a given moment. WaveZip is a compressed format for capturing those waveforms. Website: hardwareintelligence.ai YC Profile: ycombinator.com/companies/hardware-intelligence
Why we invested: Verification eats the majority of chip design time, and a bug that reaches silicon can be a nine-figure mistake. Every AI accelerator being built needs this work done, over and over. CTO Rishov Sarkar's Georgia Tech PhD is on making hardware simulation dramatically faster, which is precisely the bottleneck the product depends on. CEO Athreya Anand was a software engineer at Google. Building for how verification engineers actually work, in a terminal rather than a license-gated GUI from twenty years ago, is a real insight about the buyer.
What has to go right: A much better-funded AI chip-design competitor and the EDA incumbents are all moving on this. hi needs paid conversions and proof its capture format holds up on commercial simulators, not just open ones.
Marengo
What they do: Marengo is an AI-native engineering firm for data centers. It takes a site from due diligence through feasibility, concept, and permit-ready design, using proprietary internal tooling, with the stated goal of roughly half the time at roughly half the cost of a traditional firm. Website: marengox.com YC Profile: ycombinator.com/companies/marengo
Why we invested: Speed-to-power is the scarcest resource in AI, and pre-construction design and permitting is a genuine bottleneck in front of hundreds of billions of dollars of capex. Marengo sells the finished engineering work rather than software to engineering firms. That is the model we argued for in Spring: the companies that take responsibility for outcomes instead of selling tools. Emil Ares, a Cambridge physicist, and Gad Marconi, a TU Delft aerospace engineer, are elite systems engineers who surrounded themselves with senior advisors from the top of the data center EPC world.
What has to go right: Services economics, professional liability, and licensed engineering sign-off all have to scale without Marengo becoming as headcount-bound as the firms it is displacing.
NebuSec
What they do: NebuSec builds an AI security platform for continuous code and cloud protection. It models threats, audits code, triages and verifies findings, and generates patches. It is model-agnostic and can run on frontier or self-hosted models, including on-premises and air-gapped. Website: nebusec.ai YC Profile: https://www.ycombinator.com/companies/nebusec
Why we invested: This was the most externally verified technical team we met in the batch. The founders come out of r3kapig, which finished as the number one CTF hacking team in the world in 2025, and the team's public record includes zero-days in Chrome and the Linux kernel plus a long list of real CVEs and six-figure bug bounty earnings. The why-now is structural. Frontier models can now find exploitable vulnerabilities at a pace defenders cannot match by hand, while access to the most capable models is gated. A model-agnostic pipeline that deploys anywhere is built precisely for that gap.
What has to go right: NebuSec has to carry a proven edge in low-level systems code into the application and cloud code most enterprises actually buy protection for, and convert expert audit work into recurring software revenue.
OpenVector
What they do: OpenVector connects existing cameras to vision language action systems. You describe in plain English what matters, it builds a task-specific model, watches the feed continuously, and then acts inside your business software when the event occurs. No new hardware required. Website: openvector.com YC Profile: ycombinator.com/companies/openvector
Why we invested: There are more than a billion installed cameras in the world and nearly all of them only record. CEO Andrey Gizdov is first author on a CVPR 2025 spotlight paper on compute-efficient, human-like vision models, work done with one of the most respected figures in computational vision, and he walked away from a fully funded PhD to build this. Cheap task-specific models, rather than heavy general-purpose ones, is his actual research specialty, and that could translate into a real cost and edge-deployment advantage. The demo was among the best we saw all summer.
What has to go right: Well-funded players in video security and physical AI already own camera relationships at scale. OpenVector has to pick a beachhead and win it on cost, latency, and accuracy.
Palisade
What they do: Palisade puts an AI sales agent on niche online marketplaces through a single script tag. Every visitor gets an agent that learns what they want, remembers them, guides them from discovery through checkout, and follows up over text and email. Website: palisade.run YC Profile: ycombinator.com/companies/palisade
Why we invested: Amazon can build its own shopping agent. The long tail of vertical marketplaces, meaning trading cards, vintage goods, vacation rentals, and collectibles, cannot. Solo founder Jonathan Salama built influencer storefront tooling at Amazon and was a founding engineer at an agentic shopping venture, so he has been building toward this product for years. The company publicly reported millions of dollars of agent-touched GMV within its first months in the batch, and part of its model is performance-based, which aligns its revenue with sales it actually drives.
What has to go right: Palisade has to broaden beyond its largest accounts, stay ahead as the big platforms roll out agentic commerce protocols, and build a team around a solo founder.
Pennant
What they do: Pennant is a corporate governance operating system. It gives institutional investors, and the companies and advisors they engage with, voting and governance intelligence they can act on, with citation-backed reasoning. It also publishes Governance Arena, a benchmark for how AI models handle governance questions. Website: getpennant.ai YC Profile: ycombinator.com/companies/pennant
Why we invested: Proxy advice has been a two-firm market for forty years, and that duopoly is under real regulatory, political, and commercial pressure right now, with major banks publicly moving away from the incumbents. CEO Ryan Nowicki Stewart spent years on the buyer side in investment stewardship and board advisory at firms including State Street and BlackRock. He is automating the job he used to do by hand, for the buyer he used to be. That is the kind of founder-market fit we pay up for. Pennant is also a Palantir Startup Fellowship company.
What has to go right: Pennant has to become the system of record for governance decisions before the incumbents' own AI efforts, or in-house builds at the largest asset managers, close the window.
Stoa
What they do: Stoa is an institutional marketplace to buy and sell GPUs and AI hardware, with verified counterparties, firm quotes, and price discovery. It handles verification, contracts, shipping, and settlement, and publishes a GPU price benchmark alongside the marketplace. Website: stoaexchange.com YC Profile: ycombinator.com/companies/stoa
Why we invested: GPUs have become a trillion-dollar asset class with almost no transparent price discovery, and they are increasingly pledged as loan collateral. Whoever becomes the reference price for AI hardware sits at the center of that market. The founders come from exactly the right world: Eren Berke Saglam traded interest-rate derivatives at Citi, Kaan Yigit built commodity pricing systems at Uniper, and CEO Berat Celik has been building trading systems since he was an undergraduate. The company publicly reported hundreds of millions of dollars of RFQ volume in its first month.
What has to go right: Becoming the reference price is a liquidity race, and at least one competitor has a head start on dealer relationships. Stoa has to turn RFQ volume into settled trades and make its benchmark the number the market quotes.
Vernius Systems
What they do: Vernius builds low-cost active radar seekers for counter-drone interceptors, letting any interceptor detect and lock onto Shahed-class targets beyond visual range and in all weather. The company uses an automated engineering pipeline to design hardware in weeks rather than years. Website: vernius.systems YC Profile: ycombinator.com/companies/vernius-systems-inc
Why we invested: Cheap one-way attack drones broke the economics of air defense. Shooting down a $20,000 drone with a million-dollar missile is a losing trade at scale. Low-cost interceptors fix the arithmetic only if they can find the target, and terminal guidance is the hard part. Vernius sells a component into every interceptor platform instead of competing with all of them, which is the better position in a fragmenting market. The team came out of frontline hardware work through Defense Tech for Ukraine and Sandia National Laboratories, and the company counts former National Security Advisor H.R. McMaster among its early backers. Paired with Arlo from Spring, this is a deliberate counter-drone position built across two batches.
What has to go right: Vernius has to cross the gap from fielded prototypes to production volume and repeat defense procurement. That gap has ended plenty of promising hardware companies.
Part Three: What This Portfolio Says
Summer 2026 by theme:
- Physical AI and robotics (4): 6thSense, Agency Tool Company, DeepReach, OpenVector
- AI compute supply chain (3): hardware intelligence, Marengo, Stoa
- Security and governance for the AI workforce (3): Decawork, Fabraix, NebuSec
- Model and agent infrastructure (2): Agnost AI, Belvedir
- Vertical AI that sells the outcome (4): COACH, Evergrove, Palisade, Pennant
- Defense (1): Vernius Systems
Against Spring, we moved away from applied workflow software and toward the physical world, the compute layer, and the trust layer. Our bet is that the next two years of AI value creation look less like chat interfaces and more like buildouts: data centers, chips, robots, drones, and the systems that let autonomous software act without breaking something expensive. Summer 2026 was the first YC batch that looked like it agreed.
We are clear-eyed about what we bought. This batch was more expensive, more crowded in its fastest-growing categories, and harder to underwrite on fundamentals than any batch we have seen. We responded the way we always do: stay early, stay small on the first check, use structure where structure is available, and be useful enough afterward that we earn the right to go bigger. Some of these companies will be absorbed by the platforms beneath them. A few, we think, become the platforms.
We're Raising Fund IV
Team Ignite was ranked the third most active early-stage venture firm in the world in PitchBook's Q2 2026 Global League Tables, behind only a16z and Y Combinator. That velocity is the point. YC batches are large, fast, and noisy. Allocation disappears quickly and consensus arrives after the price has moved. Our edge is seeing the whole batch, moving before Demo Day, and being useful enough that founders want us on the cap table.
Fund IV is how we scale that model: larger initial checks into the companies we believe in most, and the capacity to keep supporting the ones that break out.
We're raising our next fund. See the deck here: tr.ee/fund-iv-deck
If you are an accredited investor and want to talk, reply here or reach out directly. And to the seventeen teams in this post: thank you for letting us in early. Now go build.