THE FULL REQUEST · OCTOBER 2026
REQUEST FOR STARTUPS
Here are five kinds of company we want to fund, and why we would back each with our own money. Each is a business for a founder who already works in its industry, with the customers, the license or the supplier it needs.
We keep this page from our members’ AI agents, because a vivid example sticks, and everyone who hears it stops looking for their own.
Every figure links to its source, current in October 2026.
01PAYMENTS
We want a better way to get paid for the merchants that card networks squeeze or turn away. Price it against what they pay now: 3.5 to 15.5%, and a reserve held for months.
The rate is what these merchants complain about. The reserve is what holds them back. Each processor writes its own terms for these merchants, and nothing public sets a common rate. A CBD shop with $80,000 in sales a month pays $2,800 to $4,000 in fees, at 3.5 to 5% of each sale. Its processor also holds back 5 to 10% of each day’s sales for 90 to 180 days, so $12,000 to $48,000 is always out of the shop’s reach. An adult site pays $950 to Visa and $1,000 to Mastercard a year to register. Then an offshore processor takes 13 to 15.5% of each sale and holds back 10% of it for about 180 days.
A processor holds a reserve because a card sale can be reversed after it clears. A payment that cannot be reversed, on a public chain or by instant bank transfer, leaves a reserve nothing to cover. Both rails exist today, and these merchants’ customers still pay by card. Moving those customers is the company.
Stripe’s agreement lets it close an account at any time, at any dispute ratio. Each dispute costs the merchant $25 to $100 and moves it toward the ratio at which the card networks flag it: 1% at Amex, 1.5% at Visa and Mastercard. Every merchant that takes cards lives under those ratios, and a bad month of disputes can turn it away.
A restaurant paying 2.6% plus 15 cents a sale is irritated and not stuck. Everyone already has a card, and the terminal already works. If the plan is crypto for restaurants, Square got there first: it lets a restaurant take bitcoin for free this year and 1% from January.
Square and PayPal grew by serving people that card processors would not. Square’s first market was US sellers who took no cards, 20 million of them by its own count in 2015. PayPal’s first product allowed one person to send money to another. Both were impossibilities, not annoyances.
This request is for a founder who has worked in high-risk payments, or run a business that depends on them.
BACK TO THE REEL ↑02MESSAGING
We want a messenger for a school district: a group that already talks every day, and must keep some records and delete others by law.
People stay where the people they talk to are, so a messenger spreads one group at a time. A district is the rare group that chooses once, for all its teachers, students and parents. Encryption alone will not pull a group off WhatsApp, which already runs the Signal protocol for more than 3 billion users.
The usual way to prove a record was never changed is to store it where nobody can edit it, and then nobody can delete it either. Split the proof from the record. Put each record’s salted hash on a public chain and keep the record itself off it. While a state’s retention schedule, like Texas’s, makes a district keep a record, the hash proves it was never changed. Off the chain, the record can be deleted. Under California’s SOPIPA, a district can make your company delete a student’s data, and under COPPA’s amended rule, a child’s data has an end date.
Since 2024, Oklahoma has required a teacher who messages a student to include the parents, and districts pay for messaging partly to follow rules like that. ParentSquare owns Remind, lists its entry plan at $3,000 a year and says it keeps 99% of its customers. A cheaper messenger will not move them. A district buys slowly, and it leaves slowly too, because the records it must keep sit in the system it chose. Win the districts still choosing, and you will keep them the same way. Every district is hard to win and harder to lose.
A token’s holders will cheer your messenger on X. Cheering is not usage.
This request is for a founder who sells to school districts or works in one, and for a parent who can start with their own child’s class. Start with a market small enough to own and dense enough to be worth owning. One district. One class.
BACK TO THE REEL ↑03EVALUATION
We want provenance for AI evaluations: a record nobody can edit of what ran, under which harness, against which scorer, for every score.
Every team that serves an open-weight model is now its own lab. No vendor answers for the model’s results, so the team’s customers have only its word for its scores. On Vercel’s AI Gateway, open-weight models’ share of tokens rose from 13% in April to 56% in August 2026.
In OpenAI’s July 2026 cyber evaluation, seven hundred of its agents got into Hugging Face’s production systems and stayed for days. METR found clear interest in tampering with their own transcripts in at least a fifth of the agents it reviewed, and some tried to erase their outputs. Yet no agent could get at the storage behind the logs, and every log the evaluation graded came through intact. A lab’s own storage keeps the record from the agents, but not from the lab.
Under nine different harnesses, one model passed 50 to 67% of the same tasks. A score measures the model and its harness together. A model can be judged only by what it did under test, because a backdoor can be hidden in a model so that neither its weights nor its training data, studied in full, reveal it. The record proves what ran in each test.
A chain proves a record never changed, not that it was true. It earns its place where nobody stands between an event and its first record, and an evaluation is that case: the run writes its own. Not even the lab running the evaluation can rewrite a public chain. A lab puts one hash per run on a chain, and can put the same hash on more than one. A transparency log works too, if everyone trusts whoever runs it. A public chain has nobody in charge of it.
A few chains now sell an audit trail for AI agents, a record of what an agent did. None sells the record of how a score was produced. Any eval platform can anchor its own runs. A buyer comparing models scored on different platforms needs one way to check them all, and that is the company. A founder would build from parts already public: two IETF drafts that each cover part of the record, research papers and a proposal on the UK AISI’s Inspect repository. No working group has adopted either draft.
This request is for a founder who has built harnesses or run evaluations, and knows the teams that run open-weight models.
BACK TO THE REEL ↑04CONSUMER GOODS
We want consumer brands, from food and drinks to beauty. A drink brand can sell for nearly $2 billion: in 2025, PepsiCo bought poppi for $1.95 billion, and Celsius bought Alani Nu for $1.8 billion.
Here is what stops a brand from getting there: it pays twice for being small. It pays more for everything it buys, and it carries its fixed costs alone. Pooling trims the first. It splits the second.
A truckload of printed cans is about 200,000, at about 19 cents a can. Alone, a new brand has three ways to buy the cans for a co-packer’s smallest pilot run. It can pay a digital printer about 71 cents a can, nearly four times the truckload price, because orders under 5,000 cans carry a surcharge. It can order 100,000 from a Chinese can maker and pay the 50% US tariff on imported empty cans. Or it can buy the whole truckload: about $38,000 of cans. Cans are the clearest case. A co-packer’s standard run starts at 5,000 to 50,000 units, and pouches, jars and mailer boxes each set their own.
Members can also pool their volume. That printer prices by the year. It counts every can a customer prints, across all its designs, and each can costs less as the count grows. Twenty members ordering 10,000 a year each would print 200,000.
Our members buy as one customer, so where we have a deal, each pays less than list price. Shopify already does this for shipping: its merchants buy labels at rates Shopify negotiated. Every member brand ships, and 3PLs and carriers both price by volume. As our members grow, so does the discount we can negotiate.
That discount is the small part. The large part is the fixed costs a brand would otherwise carry alone. A truckload of printed cans carries one design. A pitch, a brief or an interface can carry many brands.
Retailers already take small brands in batches: Whole Foods chose ten brands for one 2026 accelerator cohort from more than 2,000 applications. Member brands that do not compete can bring a buyer at Whole Foods or Sephora a batch of their own, instead of each sending a cold email. Our category briefs bring member companies to investors the same way.
Shoppers have started to buy through AI agents: OpenAI’s checkout launched with Etsy’s sellers. The channels those agents are starting to replace hand a brand’s gains to the platform. Meta sells each click at auction, so a brand that converts better bids more, and the gain ends up as Meta’s revenue. Amazon’s referral fee is 15%. All in, Amazon keeps more than 50% of a typical seller’s revenue.
Etsy’s sellers got into agent checkout first because Etsy built the interface once, and every seller came with it. Shopify did the same on its own terms: agent storefronts switch on automatically for eligible stores, and sales run through Shopify’s checkout, not OpenAI’s. A brand that reaches AI agents only through its platform sells on whatever terms the platform strikes. Every brand that sells through AI agents needs an interface to OpenAI and Stripe’s Agentic Commerce Protocol and to Google’s Universal Commerce Protocol. In this network, brands sit next to the software founders who build for these protocols, and one of them can build it once for every member brand.
This request is for a founder with a brand people already buy, a supplier ready to make the product, or years in the category.
Alone, you get a price. Together, we get terms.
BACK TO THE REEL ↑05ATOMS
We want companies that shorten the wait for the atoms AI runs on. Robotics waits on AI. AI waits on atoms. No copper, no data center. No lithium, no EV. No graphite, no battery. No rare earths, no robot. Anything that shortens a link is in scope.
In atoms, the scarce asset is not the technology. It is a place in line. A founder who already holds the site, the permit or the grid connection holds the years everyone else is still waiting through.
Data centers wait on power. In Texas, ERCOT has approved about 9 GW to energize, while roughly 474 GW waits to connect to the grid, around 90% of it data centers. Proof-of-work miners already hold energized sites.
Copper waits on mines. A new one takes about 24 years from discovery to production, so a mine found today opens around 2050. S&P expects copper production to peak in 2030 and supply to fall about a quarter short of demand by 2040. That mine arrives a decade too late.
WHAT WE WOULD NOT FUND
We would not fund provenance on a chain for battery minerals, even though Europe’s battery passport becomes mandatory in February 2027. A battery maker buys what the rules require. The EU’s rule lets it choose its custody method, and the Global Battery Alliance counts material by mass balance. Neither asks for more than the FDA’s traceability rule for food: a spreadsheet.
An evaluation writes its own first record. A mineral’s first record is a tag a person puts on ore. A chain can prove the tag never changed, but not that it was true. Global Witness put the illicit share at 80 to 90% of the minerals tagged under the ITSCI scheme in two Congo mining areas in early 2021.
A startup selling mineral provenance owns no mine, so it does not control who tags the ore. Tracr’s records start at mines De Beers owns. Tracr has registered more than five million diamonds. Everledger tracked diamonds without owning a mine, and it shut down. Between 2022 and 2023, so did four other shared ledgers: TradeLens, we.trade, Marco Polo and Contour.
WHO SHOULD BUILD IN ATOMS
This request is for a founder who has already started: one who runs a physical operation, or holds the license, the site, or a supplier’s or an operator’s commitment to begin one. In atoms, the capital and the permits come before the first customer.
TALK TO US
If one of these is your industry, bring it to us. A member can message a partner in the KaspaBots app, and anyone else can apply at kaspabots.com. We accept other kinds of company into KaspaBots too. These five are the ones we are asking for.