Company Overview

Here’s the problem in one number. McKinsey estimates the US could be short about 100,000 financial advisors by 2034, with roughly 110,000 expected to retire in that window. Those retirees hold about 42% of industry assets, per the same analysis.

You can’t recruit your way out of that. You have to make each advisor dramatically more productive.

That’s the pitch behind Nevis, a London- and New York-based company building what it calls an “AI operating system for wealth management.” In plain English: one platform that combines the advisor’s CRM, meeting notes, financial planning and custodial paperwork, with AI agents that finish the operational work instead of just suggesting it. Nevis sells software to advisory firms. It doesn’t manage money or give advice itself.

The round: On October 26, Nevis announced a $70M Series B led by EQT Ventures, bringing total funding to $110M in about two years. Charles Schwab and QuantumLight, the fund of Revolut founder Nik Storonsky, came in as new strategic investors. Sequoia, ICONIQ and Ribbit Capital all increased their commitments. EQT’s Carolina Brochado joins the board.

Valuation: Not disclosed. Fortune, citing a source, reported about $500M. Bloomberg reported a $200M valuation on the $35M Series A last December. Both are outlet-reported, point-in-time figures the company hasn’t confirmed.

The founders: All three are Revolut alumni. CEO Mark Swan led global strategy and operations at Revolut. COO Ivan Chalov ran its consumer retail business. CPO Philipp Burda led AI. That’s a team that has already scaled a regulated financial product across dozens of markets. Not a bad résumé for selling into one of the most regulated industries in America.

Traction (all company-reported): Nevis says it launched in late 2024. In 2026, it says, revenue grew more than 25x, and client assets at firms on its platform went from $50B to $700B, with about $75B added per month. Its agents have completed more than 100,000 operational tasks. Named customers include Focus Financial Partners, IEQ Capital and United Capital. Headcount is about 70.

One important clarification: that $700B is client assets at the firms using Nevis. Nevis doesn’t hold or manage it. It’s a useful proxy for reach, not a revenue figure.

The Market

The incumbent here isn’t one company. It’s the stack.

According to Swan, advisors can spend up to 80% of their time on admin, spread across 8 to 10 separate software tools. Treat that as a founder’s estimate, not a study. But anyone who has watched an advisor re-enter the same client data into a CRM, a planning tool and a custodian’s account-opening form knows the direction is right.

The size of the prize:

• $176.8 trillion in assets managed by US investment advisers in 2025, up 22.3% in a year, per the IAA and Comply’s 2026 Industry Snapshot.

• 16,544 advisers serving 73.7 million clients, also per the Snapshot.

• $260 billion in US fee-based advisory revenue in 2024, up from $150 billion in 2015, per McKinsey.

• 67 to 71 million advised relationships projected by 2034, up from 53 million in 2024, per McKinsey.

Demand rises 28% to 34%. Supply shrinks. That gap is the whole opportunity.

Why the unit economics work: McKinsey says technology and generative AI could raise advisor capacity by 7% to 15%. Nevis claims its customers reclaim up to 10 hours a week and run 30% more prospect meetings. If even part of that holds, the math for a firm is simple: an advisor who can serve more households is worth far more than the software seat that enables it. That’s an easy ROI story to sell to a firm’s COO.

Why now:

1. The talent cliff is real and dated. Firms have to plan for 2034 today.

2. AI agents can finally finish tasks. Meeting summaries were the first wave. Account opening, planning and CRM updates, done end to end, is the second.

3. Consolidation is pushing standardization. Large RIA aggregators like Focus Financial want one platform across dozens of acquired firms, not 40 separate tech stacks.

4. The custodians are leaning in. Schwab serves more than 17,000 independent RIAs. When it invests in your platform, distribution conversations get easier.

Business Model and Moat

How it makes money: Software sold to advisory firms. Nevis hasn’t disclosed pricing, but the category typically charges per advisor seat or per firm. Revenue scales with seats, not with client assets, which keeps Nevis on the vendor side of the line rather than acting as an investment adviser.

The competition is crowded and well funded:

• Jump: Raised an $80M Series B led by Insight Partners in February, $105M total. It says more than 27,000 advisors use it, nearly one in ten in the US, across firms including Focus Financial, LPL and Osaic. It started with meeting notes and is now pitching itself as an “AI operating system” too.

• Savvy Wealth: Raised a $100M Series C at a reported $600M valuation, per Tech Funding News, though it’s an advisory firm with its own tech, not a pure vendor.

• Arca: Emerged from stealth with $64M.

• The legacy stack: Established CRMs, planning tools and custodian portals, each owning one slice of the workflow.

Notice something? Focus Financial appears as a customer of both Nevis and Jump. Large firms are clearly testing multiple vendors. The winner hasn’t been picked.

What’s defensible:

1. Custodial workflow integrations. Notes are easy to copy. Plugging into custodians to actually open accounts and move paperwork is slow, regulated and relationship-driven. Schwab on the cap table helps.

2. Being the system of record. If Nevis replaces the CRM, it holds the client data every other tool needs. That’s real switching cost.

3. Enterprise rollouts. Selling to aggregators means one contract can cover hundreds of advisors and become the default across acquired firms.

What’s not defensible: transcription and summaries. Every competitor does them, and foundation model vendors make them cheaper every quarter.

Subscribe now

Spence’s Take

Mental Model #1: Bundling and Unbundling

Jim Barksdale’s line: there are only two ways to make money in business, bundling and unbundling. Wealth tech spent 15 years unbundling. A tool for CRM, one for planning, one for risk, one for notes, one for onboarding. Each was best in class. Together they turned advisors into data entry clerks.

AI agents make rebundling possible because the agent is the glue. It reads the meeting, updates the CRM, drafts the plan and starts the account paperwork without a human copying fields between screens.

The bull case: Nevis becomes the bundle, and the bundle becomes the system of record across large RIAs. Point tools get pushed to the edges or absorbed.

There’s a lesson here from building /mkt, where we structure athlete offerings under Reg A+ with tZERO as trading infrastructure: in regulated finance, the workflow is the product, and every step has a compliance owner. Bundling only wins if the bundle is easier to supervise than the pieces it replaces.

Mental Model #2: Regression to the Mean

Extreme results tend to drift back toward normal. Revenue up 25x in a year is an extreme result, and it almost always comes off a small base. The next 25x doesn’t happen. The question is what normal looks like once the early adopters are signed.

What could kill it:

• The bundle war. Jump is chasing the same “operating system” position with a much larger advisor base. Two companies can’t both be the system of record at the same firm.

• Incumbent response. Big CRM and planning vendors can bolt agents onto products advisors already know. Good enough plus already installed is a hard combination to beat.

• Regulation. The SEC has already brought “AI washing” cases against advisers that overstated their AI use. Advisors also face recordkeeping and data privacy obligations, including amended Regulation S-P. Any AI that touches client communications has to fit those rules, and a single compliance failure at a customer could slow adoption across the industry.

• Agent errors. An AI that writes a meeting summary can be wrong cheaply. An AI that opens accounts or drafts proposals can be wrong expensively.

• Concentration. A few large aggregators drive the headline asset number. Losing one would show.

5. Why It Matters

(a) For investors and VCs

This is a company worth watching, not a call on anything. Sequoia, Ribbit and ICONIQ re-upping, plus Schwab joining, is a strong signal of insider conviction. Signals I’d track:

• Seats vs. assets. Platform assets are a reach metric. Paid advisor seats and revenue per seat tell you more.

• Agent task depth. The share of tasks completed fully end to end versus drafted for human review.

• Custodian integrations. Each new custodian connection widens the addressable base.

• Head-to-head wins. Which vendor large multi-vendor firms standardize on.

• Net revenue retention once growth normalizes.

(b) For potential customers

If you run operations at an RIA or a wealth management firm, here’s what’s on offer: one platform that replaces separate CRM, meeting notes, planning and onboarding tools, with agents that complete tasks rather than just drafting them. Nevis says it’s expanding agent capabilities and adding integrations across custodians and other wealth software.

Before you buy, ask the hard questions. Where does client data live? How are AI outputs logged for books-and-records purposes? Who reviews agent actions before they reach a client? Your compliance team should sign off before your advisors do.

(c) For competitors and builders

Four lessons worth stealing:

1. Sell capacity, not features. “Your advisors can serve 30% more households” beats “we have AI notes.”

2. Pick the scarce resource. Nevis isn’t selling to the fastest-growing input. It’s selling against the shrinking one: advisor time.

3. Get the infrastructure owner on your side. A custodian investor is distribution, credibility and integration access in one.

4. Hire people who’ve shipped in regulated markets. The Revolut background isn’t trivia. It’s how you get through a large firm’s compliance review.

If you’re a point-solution vendor, you have two choices: become the best integration partner for the bundle, or become the bundle. Standing still is the riskiest option.

The Bottom Line

Nevis is betting that AI agents turn wealth tech’s sprawl back into one system, at exactly the moment the industry runs short of advisors. The team, investors and early traction are strong, but Jump is chasing the same position. Watch which one large firms standardize on.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Startup Spotlight is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Thanks for reading Startup Spotlight! This post is public so feel free to share it.

Share

Subscribe now

Share Startup Spotlight

Company Overview

Here’s a number most consumers have never thought about. Chariot says the average class action payout is about $32 per person, citing a CFPB analysis, and that only about 4% of harmed consumers ever file a claim. Its own users, the company says, average roughly $460.

Those are company-reported figures. Haircut them hard and the gap is still the story.

Chariot Claims is a legal tech platform that finds consumers who may have claims against large companies (think website tracking without consent, hidden checkout fees, subscriptions that are easy to start and hard to cancel), checks whether they qualify and routes them to partner law firms. Chariot isn’t a law firm and doesn’t file cases itself. It does the messy part firms hate: finding claimants, collecting evidence and managing payouts to thousands of people.

The round: On October 15, Chariot announced a $10M Series A led by Mike Dodd at Silverton Partners. Indicator Ventures, which led a $3.6M seed in April 2025, returned, along with Lightbank, FJ Labs and others. Total raised: $15.7M.

The founder: Zim Hang, co-founder and CEO, is a three-time founder. His first company turned sleep research into a smart pillow and was sold to a medical device operator. He then worked in health tech and fintech, built a Web3 startup and landed in legal tech after helping another founder build a class action settlement tracker.

Traction (all company-reported): The company started in 2023 as ClaimClam and rebranded in 2025, around when it moved to its current model. It says it now takes in more than 200,000 pre-qualified claim submissions a month and has taken on about 80 cases since the pivot. It has around 15 employees and says it’s breaking even.

That last point matters. Breakeven before a Series A is rare.

The Market

To understand Chariot, you need to understand the trap big companies built for themselves.

For years, the playbook was simple: put a mandatory arbitration clause and a class action waiver in your terms. Customers can’t band together in court, nobody arbitrates alone over a $40 charge, and the claims disappear.

Then plaintiffs’ lawyers noticed the catch. Under consumer arbitration rules, the business pays most of the fees. One claim is cheap. Ten thousand individual claims filed at once is not.

The numbers show how fast this flips:

• DoorDash: About 6,250 couriers filed individual arbitration claims, generating roughly $12M in AAA filing fees, according to an American Bar Association analysis. DoorDash refused to pay, and a federal court held it to its own arbitration agreement.

• Amazon: Facing about 75,000 individual demands from Alexa users, Amazon dropped its arbitration clause in 2021, per the same ABA analysis.

• The AAA’s own data: 82 consumer mass arbitrations with more than 247,000 individual filings for the period ending 2025.

Even after the AAA rewrote its rules in January 2024 to blunt this, the ABA estimated that administering 500 cases through hearing could still top $5M, before a single lawyer bills an hour.

The gap Chariot fills: Mass arbitration only works at volume, and volume is an operations problem, not a legal one. As Hang put it, “It’s just really inefficient for them to try to build workflows around 10,000 claimants or 100,000 claimants.” Law firms argue cases. They don’t run ad campaigns, verify proof of purchase and pay out 50,000 people. Chariot does.

Market size: There’s no clean TAM for mass arbitration, so be skeptical of anyone who gives you one. Here’s what we do know:

• The top ten US class action settlements in each practice area totaled about $79B in 2025, per Duane Morris’s annual report, the highest in two decades. The consumer fraud slice alone was about $2.1B.

• Legal tech startups raised a record $4.6B in 2025 and more than $3B so far in 2026, per Crunchbase data.

Why now: Three things lined up. Arbitration clauses are now nearly everywhere in consumer contracts. Privacy claims over website tracking, biometrics and subscription auto-renewal have multiplied. And AI makes it far cheaper to scan for patterns of alleged misconduct, qualify claimants and organize evidence at scale.

Business Model and Moat

How it makes money: Chariot charges partner law firms flat service fees, per Crunchbase News. The firms work on contingency. The company hasn’t disclosed its fee levels.

That structure is deliberate, and it’s smart. Most US state bar rules restrict lawyers from sharing legal fees with non-lawyers. A flat service fee keeps Chariot on the vendor side of that line. A cut of recoveries would be far more lucrative and far more legally exposed.

The funnel: Hang told Crunchbase the team reviews a few hundred potential cases a month, tests ads on about 30 and pursues roughly 10 more seriously. That’s a venture-style portfolio applied to litigation.

The competition:

• Settlement finder sites. They list existing settlements. They don’t originate cases.

• Plaintiffs’ firms with in-house intake. Some firms run their own marketing. Most don’t want to run consumer software at scale.

• Big companies themselves. The real opponent. More on that below.

What’s defensible:

1. Claimant data and conversion know-how. Every campaign teaches Chariot which issues, ads and qualification flows work. That compounds.

2. Firm relationships. If Chariot becomes the default intake layer for plaintiffs’ firms running mass arbitrations, firms will route cases to it first.

3. Verification infrastructure. Courts are cracking down on unsupported claims. In Wallrich v. Samsung, the Seventh Circuit said claimants must prove each one actually agreed to arbitrate, and a list of names wasn’t enough. Whoever builds the best proof-of-claim pipeline wins. Chariot says it’s building integrations that verify claims through proof of purchase or breach notices consumers choose to share.

What’s not defensible: the legal theory. Any firm can file the same claims. Chariot’s edge is speed and scale, not secret law.

Subscribe now

Spence’s Take

Mental Model #1: Asymmetry

The best bets have limited downside and large upside. Arbitration clauses were designed to give companies exactly that: the consumer’s cost to fight is high, so the company’s risk is low.

Mass arbitration flips the asymmetry. Now the cost of each claim lands mostly on the company in the form of fees, before anyone even argues the merits. Multiply that by 10,000 claimants and the clause meant to protect the company becomes the most expensive line in its terms.

Chariot is the operations engine that makes that asymmetry real. Without volume, there’s no pressure. Without organized claimants, there’s no volume. That’s the bull case: Chariot owns the bottleneck.

I see the same pattern at /mkt, where we structure athlete offerings under Reg A+ with tZERO as trading infrastructure: the rules decide who’s allowed to play. Chariot isn’t trying to practice law. It’s built a business that sits exactly where the rules say a non-lawyer can add value.

Mental Model #2: Incentives

Show me the incentive and I’ll show you the outcome. Everyone in this system has one, and several of them point against Chariot.

What could kill it:

• Companies rewrite the rules. On August 14, 2026, Amazon reinstated arbitration with JAMS mass arbitration procedures that kick in at 25 similar demands and batch them. Expect more companies to follow with batching and bellwether clauses designed to drain the economics.

• Courts tighten proof. After Wallrich, claimant verification is a hard requirement. Per the ABA, only about 10% of AAA mass arbitration claims advanced past the initial stage in its data. Volume without quality is worthless.

• Fee fights. The Second Circuit held in 2025 that courts can’t compel a party to pay disputed fees mid-arbitration. Companies that refuse to pay buy time.

• Legal and reputational risk. Chariot says it has received cease-and-desist letters, including over its use of company names in ads. Legal tech firms that get marketing wrong draw regulator attention fast.

• Legislation. Chariot itself flags California’s SB 690, which would restrict certain private privacy lawsuits, as a threat to consumer claims.

• Partner concentration. If a handful of law firms drive most of the revenue, they hold the leverage.

Chariot’s opponents are the best-funded legal departments on earth, and they’re adapting.

Subscribe now

Why It Matters

(a) For investors and VCs

This is a company worth watching, not a call on anything. A breakeven Series A with a contrarian, regulation-shaped model is unusual. Signals I’d track:

• Claims-to-filing conversion. 200,000 monthly submissions only matter if they survive verification.

• Recovery data. Whether the reported ~$460 average holds as volume grows and defendants push back.

• Arbitration clause trends. If big platforms copy Amazon’s batching approach, per-claim economics shrink.

• Revenue mix. How concentrated fees are across partner firms.

• New categories. Chariot says it’s moving into children’s, financial and health data privacy claims.

(b) For potential customers

Chariot has two kinds of customers.

Consumers: The pitch is a way to find out whether you may have a claim against a company you already do business with, without hiring a lawyer yourself. Chariot says it’s building tools that surface possible claims automatically from data you choose to share. Understand what the partner firm takes on contingency, and remember a potential claim isn’t a guaranteed payout.

Law firms: The pitch is outsourced claimant acquisition, evidence collection and payout management, with costs covered upfront, according to LawNext. For a firm that wants to run mass arbitrations without building a consumer marketing operation, that’s the product.

For competitors and builders

Four lessons worth taking:

1. Find the cost someone else designed for you to ignore. Arbitration clauses assumed consumers wouldn’t organize. Chariot made organizing cheap.

2. Sell to the professional, not around them. Chariot didn’t try to replace lawyers. It made them more productive and stayed on the right side of fee-sharing rules.

3. Treat compliance as a product feature. Verification is now what decides whether claims count. Build it first.

4. Run a portfolio, not a single bet. Hundreds reviewed, dozens tested, a handful pursued.

Corporate side: the cheapest mass arbitration is the one you never trigger. Fix drip pricing, cancellation flows and tracking consent first.

The Bottom Line

Chariot isn’t really a legal company. It’s a growth and operations company that found the one place in American law where volume is the weapon. The traction is real, but so is the counterattack, and the next two years of arbitration clause rewrites will decide whether this asymmetry lasts.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Startup Spotlight is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Thanks for reading Startup Spotlight! This post is public so feel free to share it.

Share

Subscribe now

Share Startup Spotlight

The Austin-based company, which builds homes and other structures with large-scale robotic 3D printers, closed a $200M Series D co-led by Baillie Gifford and funds managed by BlackRock. CNBC reported a $1.4B valuation, with total funding now around $700M.

Here’s the part most people will skip past. In early 2025, local Austin outlet KXAN reported ICON’s valuation at around $2B, right as it raised a $56M Series C first close weeks after those layoffs. If both figures hold, this round prices the company roughly 30% lower.

The operating story looks very different from the valuation story, though. Per CNBC:

• ICON says it’s built 260 homes and other structures so far this year.

• Its multistory system, Titan, launched commercially in March. The company claims it cuts average build time by six weeks and overall costs by 40%.

• CEO Jason Ballard says revenue is on pace to grow 4x this year and double again next year.

• ICON built a 100-home community north of Austin with Lennar, and Wells Fargo said in May it would write mortgages on ICON-printed homes.

The backdrop: Zillow estimates the US is short as many as 4.7 million homes, while builders are squeezed by rates, land, labor and materials.

Read more

Here’s a stat that should bother every CFO in Europe. Spiko says roughly $50 trillion of cash across Europe and the US earns next to nothing. That’s the company’s math, not mine. Haircut it by half and it’s still an enormous leak

Spiko is a Paris- and London-based fintech that lets businesses park surplus cash in regulated money market funds, then automate where it goes. The twist is the plumbing: Spiko records fund shares on public blockchains and acts as its own transfer agent. Businesses use an app. Banks and fintechs embed the funds via API.

The round: On October 6, Spiko announced a $90M Series B led by NEA, bringing total funding to about $120M. Index Ventures, which led the $22M Series A in July 2025, came back, as did Bpifrance. New names include Speedinvest, Flourish Ventures and Wintermute Ventures. Angels include former Bundesbank president Axel Weber and Qonto’s founders.

Valuation: FinTech Futures, citing Bloomberg, reported an $800M valuation. Spiko didn’t disclose a valuation in its announcement, so treat that as an outlet-reported, point-in-time figure.

The founders: Paul-Adrien Hyppolite (CEO) is a Corps des Mines engineer who served as deputy head of the French Treasury’s financial markets division, the office that shapes how French securities markets are regulated. Antoine Michon was a technology adviser to France’s Minister of Public Transformation and previously worked at Palantir. One founder helped write the rules. The other built software inside government. That pairing explains this company.

Traction (all company-reported): About $400M in assets in mid-2025. $1B by February 2026. $2B by July. $2.7B by late September. Users grew from roughly 1,000 to more than 10,000 across 25+ jurisdictions, with funds in euros, dollars, sterling and Swiss francs.

Subscribe now

2. The Market

The incumbent Spiko is attacking isn’t BlackRock. It’s the bank current account.

American companies have had easy access to money market funds for decades. European SMEs mostly haven’t. Cash sits in deposits because fund access means paperwork and high minimums, and nobody in a 40-person company owns the job. Spiko calls this the “cash yield gap.” Hyppolite’s framing is simple: as long as central bank rates are positive, idle cash is a cost you’re choosing to eat.

Now the sizing, because it cuts both ways:

• European-domiciled money market funds held about $1.7 trillion as of August, per Crane Data. That money is almost entirely institutional.

• Tokenized US Treasury products sit around $15 billion in distributed value, per RWA.xyz data reported this month.

• Spiko’s $2.7B is a big number inside the tokenized niche and a rounding error against the traditional fund market.

That’s the opportunity and the problem in one line. The tokenized market is small. The pool it’s trying to drain is huge.

Why now? Four things lined up.

1. Rates are still positive. The pitch only works when cash can earn something. It can.

2. The regulatory rails exist. Spiko’s UCITS funds are authorized by France’s AMF and can be distributed in EEA countries where the share classes are registered. France was early in allowing blockchain-based registers for fund shares.

3. Stablecoins grew up. A Fipto partnership lets users move EURC and USDC straight into Spiko funds, and a Morpho and SG-Forge integration lets eligible holders borrow stablecoins against fund shares.

4. An incumbent blinked. Spiko and Amundi launched the Spiko Amundi Overnight Swap Fund (SAFO) in March. Spiko says it passed $1B in assets and 6,500+ users in roughly four months. When Europe’s largest asset manager rents your rails, that’s validation.

The unit economics angle: When the issuer is also the transfer agent, intermediaries that take a slice of every fund disappear. Spiko reached $400M in assets with nine people, per Index Ventures. That’s software margins in asset management.

3. Business Model and Moat

How it makes money: Spiko earns management fees on assets. Public fund data shows management fees of roughly 0.25% to 0.30% a year on its T-bill funds, plus a small operating expense charge.

Quick back-of-envelope, and this is my estimate, not a company figure: $2.7B at 25 basis points is roughly $6.75M a year in gross fees. The real number could be lower, since Spiko hasn’t disclosed how SAFO economics split with Amundi.

So if the reported valuation is accurate, nobody’s paying for today’s fees. They’re paying for the pipe. This is a volume business, and the bet is that the pipe gets very, very full.

Distribution: Direct (startups, medical practices, research institutes, public bodies) and embedded (partners like Memo Bank, whose clients placed more than €35M within weeks of launch in 2025). Embedded is what matters: every integration is distribution Spiko doesn’t sell one customer at a time.

The competition:

• Tokenized treasury players. Per RWA.xyz data reported October 14, Circle’s USYC sits around $2.4B, Ondo’s USDY around $2.3B and BlackRock’s BUIDL around $2.2B. Most are dollar-first. Spiko says it’s the world’s largest issuer of tokenized cash funds, a company claim that depends on how you define the category.

• Traditional fund giants. They can launch tokenized share classes whenever they want. BlackRock added two tokenized money market products in August.

• Banks. They own the deposit relationship and can simply start paying interest on balances.

• Treasury software. Tools like Agicap and Fygr forecast cash but don’t place it. Fygr partnered with Spiko instead.

What’s defensible:

1. The regulatory stack. An AMF-authorized management company, a UCITS fund umbrella and an in-house transfer agent on public blockchains. That took years to assemble. It can’t be forked on GitHub.

2. Embedded switching costs. Once a neobank wires Spiko into its sweep logic, ripping it out means rebuilding product, compliance and ops.

3. Multi-currency in Europe. Most US competitors lead with dollars. Spiko runs four currencies.

4. Composability. Fund shares that work as collateral and move over stablecoin rails do things a traditional fund statement can’t.

What’s not defensible: the yield. A T-bill is a T-bill. Nobody wins on the underlying asset. You win on access, automation and trust.

4. Spence’s Take

Mental Model #1: The Default Effect

People don’t optimize. They accept defaults. Most corporate cash isn’t idle because a CFO weighed the options and chose zero. It’s idle because nobody ever made a choice.

That’s the real product. Spiko’s stated goal is cash that earns “by default, around the clock.” Its rules let a company keep payroll money liquid and sweep the rest into funds. Set it once. Forget it.

The bull case is that Spiko becomes the default sweep layer inside dozens of European banks, neobanks and finance tools. Defaults compound quietly. Nobody revisits a setting that works.

Here’s what I’ve learned building /mkt, where we structure athlete offerings under Reg A+ with tZERO as trading infrastructure: regulation is the product spec. Spiko didn’t route around the AMF. It became a regulated manager. That path is slow and expensive, and it’s exactly why it’s hard to copy.

Mental Model #2: Base Rates

Before you get excited about “15%+ monthly growth” (Spiko’s own number), ask the boring question: what usually happens to hot cash products?

The base rate is ugly. Cash is the least loyal asset on earth. It follows yield in and out, and a $2.7B balance can shrink faster than it grew.

What could kill it:

• Falling rates. If short-term yields drop, the gap between a fund and a deposit narrows, and the urgency fades with it.

• Banks wake up. Memo Bank already pays on current accounts. If big banks follow, the pain point gets smaller.

• Partner turns rival. Amundi is a distribution partner today. It’s also a giant that could go direct tomorrow.

• Rule changes. The European Commission published expectations for money market fund resilience in May, and the UK’s FCA is consulting on MMF reforms. New rules could raise costs.

• Operational risk. Smart contracts, chain outages and redemptions have to work under stress, not just in good times.

5. Why It Matters

(a) For investors and VCs

This is a company worth watching, not a call on anything. NEA’s lead suggests it beat out what the firm described as dozens of companies it studied in this space. The signals I’d track:

• Net flows after any rate cut. That’s the real stickiness test.

• Embedded vs. direct mix. A rising API share means the default thesis is working.

• Product mix. Assets in Spiko’s own funds versus partner funds like SAFO likely carry different economics.

• Hourly yield accrual. Spiko says it’s coming but hasn’t given a timeline. Shipping it matters.

• Geographic expansion. Spiko named Germany, Italy, Spain, the Netherlands and the Nordics.

(b) For potential customers

If you run finance at a European startup, SME or institution, here’s what’s on offer: an app or API to place surplus cash in regulated money market funds across four currencies, with rules that sort cash by when you’ll need it. Spiko’s docs say instant bank withdrawals are available on certain euro products, up to €500,000 a day via SEPA Instant. Other redemptions follow each fund’s normal settlement cycle.

Coming next, per the company: more funds, more markets and continuous hourly accrual.

One important reality check. Money market funds aren’t bank deposits. They aren’t covered by deposit insurance, and their value can move. Availability depends on where share classes are registered. Read the prospectus and talk to your own advisers before changing how you hold cash.

(c) For competitors and builders

Four lessons worth stealing:

1. Pick boring money. Cash management isn’t sexy. It’s enormous.

2. Become the regulated entity. Owning the license is slower than partnering for it, but you capture the economics and the trust.

3. Partner with the giant you’ll eventually compete with. The Amundi deal gave Spiko scale and credibility in months. Just know the clock is running.

4. Ship rails, not features. An API that other platforms embed beats an app you have to market alone.

Banks: pay on balances or expect to lose them. Treasury software: add placement before a rival does. Dollar-first tokenizers: Europe’s multi-currency demand just got a $90M competitor.

6. The Bottom Line

Spiko isn’t selling yield. Yield is a commodity. It’s selling the end of idle cash as a default, wrapped in a regulatory stack that took years to build. The traction is real. The test is whether the money stays when rates fall.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Disclaimer: This post is for informational and educational purposes only and is not investment, legal, or tax advice. Nothing here is a recommendation or solicitation to buy or sell any security. Manus is a private company. Funding amounts, investors, valuations, and revenue figures are as reported by outlets including TechCrunch, Bloomberg, The Wall Street Journal, and The Information; they have not been independently verified, revenue figures are reported as unaudited, and the reported $4B valuation target was not confirmed in the closed round. Statements about possible IPO plans are reported, not confirmed, and may not occur. The author, Spencer Gareiss, is Chief Product Officer of /mkt, which is referenced in this post as an example of building in regulated markets. Past results are not indicative of future performance.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

Manus builds general-purpose AI agents that handle multistep tasks on their own. In December 2025, Meta announced it would acquire the company for roughly $2B. At the time, Manus was reportedly doing about $100M in annual recurring revenue.

Then Beijing stepped in. Chinese authorities blocked the deal, reportedly citing export control and foreign investment rules, even though Manus had moved its headquarters to Singapore. By April, the acquisition was being reversed. Early backers including Tencent, HongShan and ZhenFund reportedly bought the company back at about the same $2B price. Benchmark, a U.S. firm, took its proceeds and walked.

Now the sequel. On October 8, Manus was reported to have raised more than $500M in a round co-led by Boyu Capital and IDG Capital, with Tencent, HongShan and ZhenFund participating. Bloomberg reported in September that the company was targeting a valuation around $4B. The final valuation hasn’t been disclosed. The Information reported revenue had climbed to $400M to $500M annualized by June, figures that are unaudited.

The mental model: Second-Order Thinking

First-order thinking says: regulators blocked a deal, so the company lost. Bad outcome.

Second-order thinking asks: and then what?

The first consequence: founders and early investors got their company back at a fixed price while revenue reportedly grew several times over. A blocked exit turned into a cheap re-entry.

The second consequence is the one that matters. Every AI founder with Chinese roots just learned that a U.S. acquirer may not be a reachable exit, no matter where the company is incorporated. That reshapes behavior upstream. Cap tables get built with domestic capital from day one. Exit plans tilt toward Hong Kong listings, which TechCrunch reports Manus is already weighing. Talent stays put.

The deal didn’t just get blocked. It redrew the map of where this kind of company can end up.

Subscribe now

Spence’s take

Here’s the contrarian part: the winner in this story might be the regulator, not Manus.

Yes, Manus looks great on paper. But its buyers, its listing venue and its growth path now all run through one jurisdiction. That’s concentration risk dressed up as a comeback.

The broader lesson applies far beyond China. In regulated markets, your exit is a product decision. You have to design for who’s allowed to buy you, list you, or trade your equity before you write the first line of code. It’s something we think about constantly at /mkt, where the Reg A+ framework and tZERO’s trading infrastructure shaped the product from the start. Regulation isn’t a hurdle you clear at the end. It’s the road you build on.

Founders who treat it as an afterthought get Manus’s first chapter. Not everyone gets the sequel.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This post is for informational and educational purposes only and is not investment, legal, or tax advice. Nothing here is a recommendation or solicitation to buy or sell any security. Manus is a private company. Funding amounts, investors, valuations, and revenue figures are as reported by outlets including TechCrunch, Bloomberg, The Wall Street Journal, and The Information; they have not been independently verified, revenue figures are reported as unaudited, and the reported $4B valuation target was not confirmed in the closed round. Statements about possible IPO plans are reported, not confirmed, and may not occur. The author, Spencer Gareiss, is Chief Product Officer of /mkt, which is referenced in this post as an example of building in regulated markets. Past results are not indicative of future performance.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

That’s not a biology problem. It’s a factory problem. And factory problems are where robots win.

Company Overview

What they do: Multiply Labs builds robotic systems that manufacture complex, individualized drugs. Picture a long enclosed box packed with robot arms doing exactly what lab techs do today: pipetting, transferring bags, loading instruments, running the same validated steps. The company started with cell and gene therapies and says it’s expanding into antibodies, viral vectors and mRNA.

Stage and raise: A $75M Series B, which the company says brings total funding past $100M since its 2016 founding. Business Insider covered the round on October 6; some deal trackers date the announcement to mid-September. No valuation was disclosed.

Investors: Business Insider and Axios report the round was led by NantWorks, Patrick Soon-Shiong’s holding company. AstraZeneca, which is also a Multiply customer, participated alongside returning backers Lux Capital and Founders Fund. Other reports also list Casdin Capital, Legend Biotech, Lingotto and Strange Ventures. A few databases list a different lead syndicate, so treat the exact lead as reported, not confirmed.

Founders: CEO Fred Parietti holds a PhD in mechanical engineering from MIT, where he researched autonomous robotic systems, after robotics work at Carnegie Mellon, Politecnico di Milano and ETH Zurich. Co-founder Alice Melocchi earned a PhD in pharmaceutical technology from the University of Milan and met Parietti while she was a visiting scholar at the Novartis-MIT center. They went through Y Combinator in 2016.

Here’s the detail most coverage skips. Multiply’s first product was robots that 3D-printed personalized pill capsules. Around its 2021 Series A ($20M, led by Casdin), it pivoted to cell therapy. Different molecule, same thesis: personalized medicine only works if you can make it at scale. That’s a pivot that kept the founder-market fit intact.

Subscribe now

2. The Market

Drug discovery isn’t the choke point anymore. Per ASGCT’s Q1 2026 data, 2,132 gene therapies and 881 non-genetically modified cell therapies are in development. The pipeline is full. The factories aren’t keeping up.

The cost picture is brutal. Manufacturing some gene therapies can run past $1M per dose, according to a 2025 research paper cited by Business Insider. Multiply says the therapies it targets are priced between $300,000 and $2M per dose.

Market sizing depends on who you ask:

• Visiongain values cell and gene therapy manufacturing at $13.7B in 2026, projecting $63B by 2036 (16.5% CAGR).

• Roots Analysis puts the narrower cell therapy manufacturing slice at $7.2B in 2026, reaching $14B by 2035.

• GM Insights sizes the cell and gene therapy CDMO segment at about $3.8B in 2025, with Lonza holding over 14% share.

Those forecasts disagree by a lot, and that’s the real insight. This is a supply-constrained market. How big it gets depends on how cheap production gets. Lower the cost per dose and you don’t just win share; you expand the number of patients who can get treated at all.

The gap: Traditional drugmaking scales with bigger tanks. Autologous cell therapy doesn’t, because every batch is one patient. Scaling means more clean rooms and more trained people, and both are expensive and scarce.

Multiply’s angle: In company-reported results from August 2025, Multiply said its robotic cluster cut cost per dose by 74% and delivered up to 100x more doses per square foot of clean room. A 2024 peer-reviewed study in Cytotherapy found its robotic cell expansion matched the manual process on all critical process parameters. The key design choice: its robots operate instruments pharma already uses, from partners including Thermo Fisher, GenScript and Wilson Wolf. Customers don’t have to rebuild a validated process from scratch.

Why now: Capital is flooding into physical AI, collaborative robot arms have gotten cheaper and more capable, and regulators have shown openness to automated manufacturing. Rival Cellares’ platform received FDA’s Advanced Manufacturing Technology designation, a signal the agency wants this category to exist.

3. Business Model and Moat

Revenue model: Multiply sells the robots. Drugmakers like AstraZeneca and Legend Biotech own and operate them in their own facilities. Multiply then earns recurring revenue from cartridges used in the machines, plus service and support. Parietti told Business Insider each robot currently costs several million dollars to build, and deployments are in the “high single-digit” range. An $85M partnership with Retro Biosciences marked the company’s first commercial system sale.

The competition:

• Cellares is the heavyweight. It raised a $257M Series D in January led by BlackRock and Eclipse, bringing total funding to $612M, and signed a $380M manufacturing agreement with Bristol Myers Squibb. Its model is different: it runs its own automated “Smart Factories” and manufactures for clients. It expects commercial-scale manufacturing to begin in 2027, and NJBIZ has reported a planned IPO the same year.

• CDMOs like Lonza, Catalent and Charles River sell capacity and labor.

• Closed-system boxes from large tools vendors automate pieces of the workflow.

• In-house manual processes, which remain the default.

What’s defensible:

1. Instrument-agnostic integration. Multiply automates the process a drugmaker already validated instead of forcing a new one. That lowers adoption friction.

2. Regulatory gravity. Once a robot is part of a manufacturing process in a regulatory filing, swapping it out means revalidation. That’s a switching cost measured in years.

3. Consumables. Every run uses cartridges. The installed base pays rent.

4. Customers on the cap table. AstraZeneca and Legend aren’t just logos. They’re buyers with money in.

What isn’t defensible yet: Scale. A handful of multimillion-dollar machines with a European supply chain is a promising start, not a moat.

4. Spence’s Take

Mental Model #1: Theory of Constraints. A system’s output is capped by its tightest bottleneck. Improve anything else and you’ve wasted effort. For decades pharma optimized discovery. In cell and gene therapy, the constraint has moved to manufacturing. Multiply is attacking the constraint directly, which is why a 74% cost cut (if it holds up across customers) matters more than any new target in a pipeline.

Mental Model #2: Counter-Positioning. The best position is one incumbents can’t copy without hurting themselves. CDMOs make money renting capacity and labor. Cellares makes money running factories. Multiply sells pharma the ability to own its capacity. For a CDMO to match that, it would have to sell away the very thing it charges for. That’s real counter-positioning.

The bull case: Multiply becomes the default automation layer for advanced biologics. Razor-and-blade economics kick in as the installed base grows. Antibodies, a much bigger market than cell therapy, open up. Parietti says the goal is to bring robot costs down “from millions to tens of thousands.” That’s the company’s stated ambition, not a result, but if it gets close, the addressable market changes shape.

What could kill it:

• Hardware economics. Selling multimillion-dollar capital equipment into pharma means long sales cycles and lumpy revenue. Hardware startups die in the gap between “works” and “ships at volume.”

• Capital asymmetry. Cellares has raised roughly six times as much. If the market picks the outsourced factory model, Multiply’s “own your capacity” pitch shrinks to a niche.

• The science moves. In vivo CAR-T approaches, which engineer cells inside the patient’s body, and off-the-shelf allogeneic therapies could reduce the need for patient-by-patient manufacturing. Multiply’s push into antibodies and mRNA looks like a hedge against exactly this.

• Customer concentration. High single-digit deployments means one delayed program can dent a year.

5. Why It Matters

(a) For investors and VCs: This is a private company and not something most investors can access, so read this as a framework for tracking the category, not a recommendation. Signals worth tracking:

• Installed base growth beyond high single digits

• Consumables and service as a share of revenue, the tell for recurring economics

• Commercial (not just clinical) programs built on Multiply’s process

• A disclosed antibody or mRNA deployment

• Falling cost to build each robot

• Cellares’ IPO, if it happens, as the first public read on the category

(b) For potential customers: If you run manufacturing at a biotech or pharma company, the offer is robotic clusters that run your existing instruments, in your own facility, with you controlling capacity. The claims are strong. Ask for site-level data on contamination rates, batch success, and validation timelines before you plan around them. Patients are the downstream customer. Lower manufacturing costs could widen access over time, though pricing is set by drugmakers, not robot makers.

(c) For competitors and builders: Three lessons.

1. Integrate, don’t replace. In regulated industries, compatibility beats cleverness. Multiply won by working with validated instruments instead of asking customers to start over. I see the same dynamic at /mkt, where we build on the Reg A+ framework and tZERO’s trading infrastructure instead of inventing new market plumbing.

2. Turn customers into investors. Strategic money from buyers is the strongest form of validation, and it shortens the next sales cycle.

3. Pivot the product, keep the thesis. Multiply went from printed pills to cell therapy without abandoning its core belief. Good pivots change the what, not the why.

6. The Bottom Line

Multiply Labs is betting the next decade of medicine is a manufacturing story, and the math backs it up. The thesis is right; the open question is execution, specifically whether a handful of expensive machines can become a scalable installed base before better-funded rivals or new science change the game. It’s one of the most interesting physical AI companies almost nobody is talking about.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Disclaimer: This post is for informational and educational purposes only and is not investment, legal, tax, or medical advice. Nothing here is a recommendation or solicitation to buy or sell any security. Multiply Labs is a private company; its securities are not publicly traded. Funding figures, investor lists, and announcement dates are as reported by the company and outlets including Business Insider and Axios; some sources list a different lead investor and an earlier announcement date. Cost, throughput, deployment, and product roadmap figures are company-reported or company-stated and have not been independently verified. Forward-looking statements, including cost targets, expansion plans, and Cellares' reported IPO plans, are attributed to the companies or outlets and may not occur. Market size figures are third-party estimates that vary significantly by source. The author, Spencer Gareiss, is Chief Product Officer of /mkt, which is referenced in this post as an example of building in regulated markets. Past funding, partnerships, or reported results are not indicative of future performance.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

Company Overview

What they do, in plain English: Quartermaster makes SmartMast, a weather-hardened box of cameras and radios that bolts onto a ship’s mast. It runs AI on the vessel, watches the water around it, and sends back what it sees. Every equipped ship becomes a node in a shared ocean sensor network.

Stage and round: On September 28, TechCrunch reported a $140 million raise, preempted by Insight Partners. The split matters: roughly $100 million is Series B equity from Insight, new investor Overmatch Ventures (a defense-focused firm), and returning backers, plus a $40 million debt facility from Stifel. Tech Funding News lists First Round Capital, Quiet Capital, Steel Atlas, TMV, BoxGroup, and Operator Partners as returning investors.

Total raised: The Series B landed just four months after a $43 million Series A co-led by First Round and Quiet Capital in May. That puts disclosed funding at roughly $183 million including debt. The company hasn’t disclosed a valuation, so I won’t guess at one.

Founding team: CEO Neil Sobin helped build Hivemapper, the network that mapped streets using cameras on everyday vehicles. Quartermaster is the same playbook pointed at the sea. Chief growth officer Steve Ribaudo was previously VP of growth at Anduril, per the company’s site, and its about page describes team members who ran Navy growth at Scale AI. Preqin dates the founding to 2023.

Traction (company-reported): More than 650 vessels in 25 countries now carry a SmartMast, and 800-plus units have shipped. Sobin says the gap exists because full-fleet rollouts are about to start. Coverage has reached nearly 46 million square kilometers, up 77% since May, per Smart Maritime Network. SmartMast-equipped ships have also assisted in more than 20 rescues at sea.

Subscribe now

The Market

The gap incumbents left open. Today’s maritime intelligence stack mostly stitches together AIS pings, satellite passes, and port records. The problem is the foundation. As Sobin put it to TechCrunch in May, AIS is opt-in and self-reported, so anyone doing something shady, from petty smuggling to sanctions evasion, can go dark or spoof a position.

Satellites help, but they’re expensive and they don’t stare at one patch of water all day. A camera on a ship that’s already there does.

The numbers:

• The fleet: UNCTAD counts about 116,000 merchant vessels of 100 gross tons or more as of January 2026, including roughly 62,000 over 1,000 GT. At 650 installs, Quartermaster sits well under 1% of that fleet. That’s either a warning or a runway, depending on execution.

• The insurance pool: The International Union of Marine Insurance reported last week that global marine premiums hit $42.6 billion in 2025, up 5.5%. Cargo was $24.2 billion and ocean hull $10.5 billion.

• Surveillance spend: Precedence Research pegs maritime surveillance at $28.36 billion in 2026, growing to $50.93 billion by 2035. Research firms define this category differently, so treat it as directional.

Why now. Two words: the Gulf. TechCrunch reports that Sobin credited the conflict involving Iran and the resulting shipping chaos with sharpening investor conviction. According to the company, GPS jamming and spoofing earlier this year told ships across Gulf waters they were somewhere they weren’t, and several underwriters stopped writing cover for the region. S&P Global reported in July, citing broker Marsh, that war-risk premiums for Hormuz transits reached as much as 10% of hull value. For context, UNCTAD estimates Hormuz handles around 11% of global maritime trade volume.

When a single chokepoint that important goes dark, “we can see the water” stops being a nice-to-have.

Business Model and Moat

The model is the clever part. Per Tech Funding News, ship operators pay nothing beyond space on the mast. Quartermaster covers the hardware, installation, support, route hazard alerts, and broadband internet at sea. In exchange, the data feeds a shared network that insurers, fleets, energy operators, and government customers like coast guards and navies can use.

Sobin explained the logic in May: fleet operations are low-margin businesses, so selling a sensor to a shipowner is a brutal pitch. Giving it away is an easy one. Pricing for the data side hasn’t been disclosed.

The competitive field:

• Data aggregators. Windward, taken private by FTV Capital in 2025 at about $280 million, fuses AIS, satellite imagery, and ownership records. Kpler bought Spire’s maritime data business for $241 million in 2025. Both are software layers on top of largely shared inputs.

• New hardware at sea. Saildrone, Oshen, and Online Oceans build autonomous vessels. They add new sensors to the ocean. Quartermaster rides on ships that are already out there.

• Satellites. Insight also led a $145 million round for HawkEye 360, the RF geolocation company. That tells you how Insight views this category: layered sensing, not a single winner.

What’s defensible: Proprietary first-party data. Every mast generates tens of gigabytes a day that no aggregator can buy, and every new ship widens a moat a competitor would need years of installs to cross. The pro-mariner angle (free internet, hazard alerts, rescue support) also raises the cost of ripping a unit out.

Spence’s Take

Mental Model #1: The Map Is Not the Territory.

AIS is a map. It’s what ships say about themselves. The territory is what’s actually happening on the water. For decades the maritime world priced risk, enforced sanctions, and routed cargo off the map. When the map and territory drift apart, as they did in the Gulf this year, underwriters can’t price it, so they walk.

Quartermaster’s pitch is simple: stop trusting the map, go look at the territory. That’s a far stronger foundation than another analytics layer on the same self-reported pings.

I spent years at Robinhood building derivatives and prediction markets. Here’s what that taught me: a market is only as good as the data it settles on. If you can’t observe it, you can’t price it. Insurance is no different.

Mental Model #2: Network Effects.

One SmartMast is a camera. Ten thousand is infrastructure. Each added ship improves coverage for every customer, which makes the data more valuable, which funds more free installs. Sobin has already run this flywheel once at Hivemapper. That pattern recognition is a real asset.

The bull case: Insurers start pricing specific straits and routes off Quartermaster’s data. At that point the mast network becomes something rivals can’t copy quickly, and the company becomes a quiet utility for global trade.

What could kill it:

1. The payer problem. Tech Funding News said it plainly: if insurers don’t adopt the data, Quartermaster owns a big sensor network that still has to prove who’ll pay for it. Free hardware is a cost center until data buyers show up at scale.

2. Hardware drag. Salt, storms, installs across 25 countries, and doubled manufacturing capacity. Hardware businesses burn cash in ways software businesses don’t, and $40 million of this round is debt.

3. Geopolitical timing. Crisis-driven demand is real, but it can cool. The question is whether buyers stay when the headlines move on.

4. Trust and access. Cameras on commercial ships near contested waters raise questions for flag states, port authorities, and crews. One regulatory or privacy backlash could slow installs.

Why It Matters

(a) For investors and VCs. This is a company worth watching, not a call to action. The signals I’d track:

• Installed base velocity. Does the 650-to-800 gap close fast as fleet rollouts begin?

• Named data buyers. Especially insurers and government contracts. Paying customers on the demand side validate the whole model.

• Coverage density on key lanes. Total square kilometers is a vanity stat. Density in Hormuz, the Red Sea, and major fishing grounds is what buyers pay for.

• Debt usage. How the $40 million Stifel facility gets deployed tells you a lot about capital discipline.

(b) For potential customers. If you’re a fleet operator, the pitch is free hardware, free broadband, and hazard alerts for mast space. Read the data-sharing terms carefully and understand who can see what your vessel captures. If you’re an underwriter, energy operator, or coast guard, this is a new first-party feed that doesn’t depend on a ship choosing to report itself.

(c) For competitors and builders. Three lessons travel well:

• Pay your supply side in what they actually need. Mariners want connectivity and safety, not a sensor. Quartermaster pays in bandwidth, not cash.

• Own your inputs. Aggregators compete on models. Quartermaster competes on data nobody else has.

• Let the crisis prove the thesis, then build past it. Timing opened the door. Durable revenue has to keep it open.

The Bottom Line

Quartermaster is betting that the ocean’s biggest problem isn’t analysis, it’s observation, and that the cheapest way to watch the sea is from ships already on it. The network logic is sound and the founder has done this before. Now it has to prove that insurers and governments will pay for what the masts see, at a scale that justifies giving the hardware away.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Disclaimer: This post is for informational and educational purposes only and does not constitute investment, legal, tax, or financial advice, or an offer or solicitation to buy or sell any security. Quartermaster is a private company; its securities are not publicly traded, and nothing here should be read as a recommendation. Funding amounts, investor names, vessel counts, coverage figures, and other operating metrics are as reported by the company and by TechCrunch (September 28, 2026 and May 20, 2026), Tech Funding News (September 29, 2026), and Smart Maritime Network, and have not been independently verified. The company has not disclosed a valuation for this round. Market figures are drawn from UNCTAD, the International Union of Marine Insurance, S&P Global, and Precedence Research; third-party market estimates vary by methodology. Descriptions of geopolitical events and their effect on shipping and insurance are attributed to the cited sources and the company. Any forward-looking statements reflect the company's stated plans or the author's opinion and are subject to risks and uncertainties. Past performance and private funding rounds are not indicators of future results. The author has no position in or business relationship with Quartermaster. Always do your own research and consult a qualified professional before making any financial decision.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

A month ago, World Labs was reportedly talking about a $5 billion valuation. On Monday, AMD agreed to buy it for $8.2 billion in stock.

That’s not a startup exit. That’s a chip company deciding it can’t win the next round of AI from the outside looking in.

Subscribe now

The deal

World Labs, founded by Stanford professor and ImageNet creator Fei-Fei Li, builds “world models.” These are AI systems that generate and simulate interactive 3D environments from text, images, or video. Its Marble product creates simulated worlds used for entertainment and for training robots.

Here’s what’s been reported:

• Price: About $8.2 billion, all stock, per Axios and Quartz.

• Timing: Expected to close by the end of 2026, subject to regulatory approval.

• Leadership: Li becomes AMD’s executive vice president and chief scientist, reporting to CEO Lisa Su. Co-founders Justin Johnson and Ben Mildenhall keep running the World Labs team.

• Funding history: World Labs came out of stealth in 2024 at a valuation above $1 billion, backed by Andreessen Horowitz, NEA, and Radical Ventures. Its latest round raised $1 billion from a group including Nvidia, AMD, Autodesk, and Fidelity.

• Existing ties: The two companies already had a training and inference partnership, and World Labs was training models on AMD GPUs.

Read that investor list again. Nvidia backed the startup its biggest rival just bought.

The mental model: Circle of Competence

Know where your edge ends, and be honest about it.

AMD’s circle is silicon. Nvidia’s advantage has never been only chips. It’s the software, the developer tools, and its own world-model suite, Cosmos. Nvidia understands the workloads before they arrive, so it builds hardware for where AI is going.

Lisa Su said it plainly: building compute for the next generation of AI requires deep understanding of how models evolve. Translation: AMD knows the edge of its circle and is buying its way past it.

That’s the smart move. A company that stays inside its circle forever gets commoditized. A company that pretends its circle is bigger than it is makes expensive mistakes.

The second risk is the one I’d watch. Chipmakers have bought AI teams before. Intel acquired Nervana in 2016 and later wound down that product line in favor of Habana. Buying a lab doesn’t mean you know how to run one.

Spence’s take

The headline is the price. The story is the feedback loop.

World models are compute-hungry, and robotics is where physical AI gets real. If World Labs’ workloads shape AMD’s chip roadmap, AMD gets something Nvidia has had for years: a front-row seat to what the next models need.

Here’s the contrarian part. The biggest risk isn’t overpaying. It’s over-owning. Research labs run on independence. If World Labs becomes an internal AMD project, it could lose the external users and talent that made it valuable.

Three signals I’m tracking: whether Johnson and Mildenhall stay past the first year, whether Marble stays usable on other companies’ hardware, and whether regulators scrutinize a chip company owning a foundation model.

Buy the capability. Protect the culture. That’s the whole game.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This post is for informational and educational purposes only and is not investment, legal, tax, or financial advice. Nothing here is a recommendation or solicitation to buy, sell, or hold shares of AMD, Nvidia, or any other security. Deal terms, funding history, and valuations are as reported by TechCrunch, Axios, Quartz, and Tech Funding News (September 28 and 29, 2026) and have not been independently verified. The approximately $5 billion figure was reported by third parties and was not confirmed by World Labs. The transaction is subject to regulatory approvals and other closing conditions and may not close as announced. Statements about what the deal could mean reflect the author's opinion and are forward-looking, uncertain, and subject to change. Always do your own research and consult a qualified professional before making any financial decision.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

The headline is the size of the check.

The interesting part is everything attached to it.

SSI reportedly raised $1 billion in 2024 at a $5 billion valuation, then $2 billion in early 2025 at a $32 billion valuation. Now Nvidia is adding another $5 billion, alongside access to its next-generation Vera Rubin infrastructure.

The stated goal: roughly 10x more compute capacity.

Read that again.

This isn’t just a funding round.

It’s a bet that compute is a strategic asset.

And that distinction matters.

Subscribe now

Index Ventures just raised another $2 billion across three funds: $400 million for seed, $900 million for venture, and $700 million for growth.

Simile raised $200 million at a $2 billion valuation, just five months after its $100 million Series A.

AI capital isn’t slowing down.

But I think we’re asking the wrong question.

Everyone wants to know: Which AI company wins?

I’d ask: Which companies control the bottlenecks?

The Mental Model: Second-Order Thinking

First-order thinking says:

AI is getting more capital.

Second-order thinking says:

Capital attracts talent. Talent consumes compute. Compute creates infrastructure dependencies. Infrastructure creates bargaining power.

Eventually, the model itself may become the least differentiated part of the stack.

That’s the part I find most interesting as an operator.

At /mkt, we’re building in a regulated market around athlete tokenization. That means the constraints aren’t something you solve after the product works.

Securities regulations matter.

Offering structure matters.

Trading infrastructure matters.

Disclosure and compliance matter.

The constraint shapes the product.

AI is developing its own version of this.

Compute. Distribution. Data. Energy. Regulation. Capital.

The companies that control those constraints have leverage.

Here’s my contrarian take:

Stop looking for moats in features. Start looking for moats in bottlenecks.

A feature can be copied.

A dependency that’s expensive, regulated, scarce, or deeply integrated is much harder to copy.

That’s why I care less about who raised the biggest round.

I’d rather know what that round bought them access to.

Because in the next phase of startups, the winner may not have the best technology.

They may simply control the thing everyone else needs.

Disclosure and disclaimer: Spencer Gareiss is Chief Product Officer at /mkt, referenced above as an example of building in regulated markets. This newsletter is for information and education only. It is not investment advice, and it is not an offer or solicitation to buy or sell any security. Gatik is a private company; its valuation was not disclosed, and its revenue, delivery, on-time, and fleet figures are company-reported and have not been independently verified. Forward-looking targets, including fleet-growth plans, are the company’s statements, not facts. Market-size, driver-shortage, and consolidation figures are third-party estimates from named sources (Future Market Insights, the American Trucking Associations, and others) and will vary by methodology. Funding and operating details reflect reporting as of August 28, 2026, drawn from the company’s announcement and outlets including TechCrunch, Bloomberg, Reuters, Forbes, SiliconANGLE, and FreightWaves.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

That number gets the headline.

The structure gets my attention.

SSI previously raised $1 billion at a $5 billion valuation in 2024, then another $2 billion at a $32 billion valuation in February 2025. Now Nvidia is reportedly committing another $5 billion, alongside access to its next-generation Vera Rubin hardware. The companies said the partnership should increase SSI’s compute capacity by an order of magnitude.

That’s not just capital.

It’s capital plus scarce infrastructure.

And the same pattern is showing up across venture.

Subscribe now

Index Ventures raised $2 billion across three funds on July 31: $400 million for seed, $900 million for venture, and another $700 million for its existing growth fund. That brings Index’s available capital to $3.5 billion. The timing is notable: its portfolio company Wiz recently completed a reported $32 billion sale to Alphabet.

Meanwhile, Simile raised $200 million at a $2 billion valuation, only five months after its $100 million Series A. The company is building AI systems that simulate human behavior for product and market research.

The obvious conclusion is that AI funding is accelerating.

I think that’s incomplete.

The mental model: Second-Order Thinking

First-order thinking asks: What happened?

AI companies are raising enormous amounts of money.

Second-order thinking asks: What happens next?

When capital, compute, distribution, and infrastructure increasingly come from the same small group of companies, the competitive game changes.

The scarce resource may not be the model.

It may be access.

That matters for founders. The best startup strategy isn’t necessarily “raise the biggest round.” It’s figuring out which dependency becomes a bottleneck if your company actually works.

We’ve seen versions of this problem in regulated markets. At /mkt, building around athlete tokenization means the product can’t be separated from securities rules, offering requirements, trading infrastructure, disclosures, and compliance. The constraint isn’t something you bolt on after product-market fit. It shapes the product from day one.

AI founders face a different version of the same problem.

Compute isn’t a footnote. It’s part of the business model.

My contrarian take: the next great startup moats may be built around constraints, not features.

Everyone can see the latest model.

Far fewer people can secure the infrastructure, regulatory pathway, distribution, data, or capital structure required to turn that model into a durable business.

That’s where I’d look.

Not at who raised the most.

At what they now control that competitors can’t easily replicate.

Disclosure and disclaimer: Spencer Gareiss is Chief Product Officer at /mkt, referenced above as an example of building in regulated markets. This newsletter is for information and education only. It is not investment advice, and it is not an offer or solicitation to buy or sell any security. Gatik is a private company; its valuation was not disclosed, and its revenue, delivery, on-time, and fleet figures are company-reported and have not been independently verified. Forward-looking targets, including fleet-growth plans, are the company’s statements, not facts. Market-size, driver-shortage, and consolidation figures are third-party estimates from named sources (Future Market Insights, the American Trucking Associations, and others) and will vary by methodology. Funding and operating details reflect reporting as of August 28, 2026, drawn from the company’s announcement and outlets including TechCrunch, Bloomberg, Reuters, Forbes, SiliconANGLE, and FreightWaves.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

1. Company Overview

Gatik, founded in 2017 by veterans of the autonomous-tech world, runs driverless medium-duty box trucks on the “middle mile,” the short, repeatable regional hops between distribution centers and stores. On Aug 25 it announced a $200 million Series D led by the Qatar Investment Authority and Koch Disruptive Technologies, with Millennium Management, ARK Invest, Arca Continental, and Intact Private Capital joining. Intact tripled its prior commitment. It’s Gatik’s largest round to date and brings total funding to roughly $500 million.

The numbers Gatik put behind the raise are the part worth sitting with. The company reports more than $600 million in contracted revenue, 85,000 fully driverless orders completed, and a 99% on-time rate. Those are company-reported figures, and Gatik declined to disclose its valuation, so treat the metrics as self-reported and the price as unknown. It runs dozens of fully driverless trucks today and, per a company spokesperson, is targeting more than 100 by the end of 2026, with plans to reach thousands over the following years.

The customer list is the credential. Walmart was the first named customer, and in 2021 Gatik launched what it calls the world’s first driverless commercial delivery service on those routes. PepsiCo is now the largest public partnership: since signing a multiyear deal in June, 41 driverless Gatik trucks move Frito-Lay products from distribution centers to roughly 250 retail locations around Dallas, Phoenix, and Northwest Arkansas. Kroger, Tyson Foods, and Canada’s Loblaw are also customers. The trucks come off Isuzu assembly lines with Gatik’s “Gatik Driver” system on board, and the company partners with NVIDIA, Cummins, Ryder, and Goodyear. Operations run in Texas, Arizona, Arkansas, and Ontario.

Subscribe now

2. The Market

Gatik isn’t chasing the whole autonomous-trucking number, and that matters, because the whole number is a mess of competing definitions. The slice it actually plays in is middle-mile autonomous delivery, and Future Market Insights pegs that at about $490 million in 2026, growing to roughly $14.2 billion by 2036, a 40% compound annual rate, with Level 4 box trucks the leading vehicle type. Those are third-party estimates, not gospel, but the shape is clear: this is a small, fast-forming market, not a mature one.

The demand underneath it is structural, not hype. The American Trucking Associations estimates a driver shortage near 82,000 in 2026, projected to climb past 160,000 by 2031. Federal hours-of-service rules cap a human driver at 11 driving hours inside a 14-hour window, which means a human-driven truck sits idle for a big chunk of every day. A driverless truck doesn’t clock out. On repetitive regional routes that carriers have always struggled to staff, that’s a real economic gap, and it’s exactly the gap Gatik sells into.

One more market fact frames the raise. Autonomous middle-mile is consolidating early: by some estimates the top handful of players already account for the large majority of commercially active driverless middle-mile miles. In a market like that, the durable asset isn’t a demo. It’s certified routes and signed customers.

3. Business Model and Moat

Gatik’s model is autonomy as a service. It doesn’t build trucks, it buys them from an OEM and supplies the brain, then runs the routes around the clock. The strategic choice that makes the whole thing work is the constraint. Instead of trying to drive anywhere, Gatik locks its trucks to a tightly defined operational domain: known regional routes, mapped and validated, run over and over. Early routes were fixed and under 10 miles. Today the dynamic version stretches to as much as 400 miles with multiple pickups and drop-offs, but the principle holds. Certify a route once, and every additional shipper you layer onto it makes the route more valuable. That’s a compounding asset, not a one-off.

The moat is the stack of things that are hard to copy fast: a driver-out operational record on public roads, route certifications that took years to earn, a Fortune 50 customer base with switching costs, a manufacturing and supplier web (Isuzu, NVIDIA, Cummins, Ryder, Goodyear), and a regulatory footprint across the states with workable AV frameworks. Gatik also joined NVIDIA’s Halos AI Systems Inspection Lab in March 2026, an accreditation layer that signals it’s treating safety validation as a moat rather than a checkbox. Right now, in driver-out middle-mile box trucks specifically, Gatik largely sits alone.

4. Spence’s Take

Gatik’s edge is what it refused to build. That’s circle of competence in action, one of the 50 models in the book. The glamorous problems in autonomy, robotaxis and long-haul highway trucking, are also the hardest ones, with the widest range of situations a system has to handle perfectly. Gatik drew a tight circle around a problem it could actually master, the repeatable distribution-center-to-store run, and refused to step outside it. Staying inside your circle isn’t a limitation. It’s how you ship while everyone else is still demoing.

Pair that with a second model: margin of safety. A narrow, validated route gives you engineering margin, fewer edge cases means fewer ways to fail. And more than $600 million in contracted revenue gives you financial margin that hype-funded peers just don’t have. This is the same logic behind building in regulated markets, which is the game I play day to day as CPO at /mkt: you don’t win by moving fast and breaking things, you win by owning a defined, compliant domain so completely that speed becomes safe. Gatik took the autonomy version of that bet.

The bull case is clean. Driver-out operations are live and commercial, the revenue is contracted, the customers are household names, the tailwinds (driver shortage, hours-of-service limits) are structural, and the company now has the capital to scale.

The bear case is just as real. “Dozens” of trucks to “thousands” is a brutal jump that tests manufacturing, operations, and capital all at once. One serious safety event could reset public and regulatory trust overnight, no matter how good the on-time rate is. Customer concentration is heavy, with PepsiCo and Walmart doing a lot of the lifting. The regulatory patchwork still varies state by state, and a national framework isn’t in place yet. And with the valuation undisclosed, nobody outside the round can judge whether the price is sane. None of this is a prediction. It’s the list of things to watch.

5. Why It Matters

For investors and VCs: Gatik is a template for capital-efficient autonomy. Narrow the domain, sign real contracts, and let the revenue, not the narrative, carry the round. Notice who led it, too: a sovereign fund and strategic industrial capital taking equity in a platform they can deploy against. As middle-mile consolidates, certified routes and signed revenue are the assets that survive. That’s the pattern to pressure-test in any autonomy deal that crosses your desk.

For potential customers (retailers, grocers, CPG, and 3PLs): driverless middle-mile has moved past pilot theater and into live supply chains. The pitch is capacity and reliability on exactly the routes that are hardest to staff. If you’re evaluating it, the questions that matter are on-time performance, how much of your route network fits the operational domain, insurance terms, and how much you’re comfortable concentrating with a single provider.

For competitors and builders: the lesson is wedge selection. Pick a domain you can certify and repeat, build the operations layer rather than just the model, and partner for the hardware instead of trying to build it. Certification and operational history are becoming the real barriers to entry here, and they compound with time. Speed to a flashy demo is not the same as speed to a signed, driver-out route.

6. The Bottom Line

Gatik didn’t win by solving the hardest problem in autonomy. It won by choosing a valuable one, drawing a hard line around it, and shipping revenue while the rest of the field argued about robotaxis. The next 18 months are a straight execution test: get from dozens of trucks to hundreds to thousands without a safety event that resets the clock. Hold that 99% on-time rate through the scale-up, and Gatik doesn’t just lead the middle mile, it defines it. Fumble the scaling, and the moat starts to look like a head start. Either way, this is the clearest proof yet that boring, contracted, driver-out autonomy is a real business, not a someday story.

If this was useful, share it with someone who builds things, and if you want the full toolkit, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Disclosure and disclaimer: Spencer Gareiss is Chief Product Officer at /mkt, referenced above as an example of building in regulated markets. This newsletter is for information and education only. It is not investment advice, and it is not an offer or solicitation to buy or sell any security. Gatik is a private company; its valuation was not disclosed, and its revenue, delivery, on-time, and fleet figures are company-reported and have not been independently verified. Forward-looking targets, including fleet-growth plans, are the company’s statements, not facts. Market-size, driver-shortage, and consolidation figures are third-party estimates from named sources (Future Market Insights, the American Trucking Associations, and others) and will vary by methodology. Funding and operating details reflect reporting as of August 28, 2026, drawn from the company’s announcement and outlets including TechCrunch, Bloomberg, Reuters, Forbes, SiliconANGLE, and FreightWaves.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

Five rounds closed between Monday and Thursday that are worth your time. One values a private beta at $2.5 billion. Another only makes money when a cow gets sold. Here’s what happened, and what each one tells you.

1. Instinct raised $250M for an app you can’t even use yet

Instinct, built by Spear Street Technology, is a personal AI agent you text or call to run your life: booking travel, clearing your inbox, managing your calendar. It told the Wall Street Journal it raised a $250 million Series B co-led by Index Ventures and Benchmark, pushing total funding to $350 million at a company-reported $2.5 billion valuation. It’s still in private beta, and press reports put its valuation climb at roughly $50 million to $2.5 billion in about four months. Founder Noah Shinn is 23.

Spence’s take: A $2.5 billion price on a beta with no public revenue isn’t a product bet, it’s a bet that Shinn ships trust faster than the lawsuits ship.

Here’s where a mental model earns its keep. Apply second-order thinking, one of the 50 in the book. First-order: the agent feels like magic and early users love it, so back it fast. Second-order: the same access that makes an agent useful, your inbox, your calendar, your cards, is what turns one bad default into a privacy and regulatory mess at scale. Reporters have flagged terms that reportedly grant a perpetual license to user data, and one investor said the agent emailed on her behalf and kept reading her inbox after she cut off access. A perpetual data license reads great in a term sheet and terrible in a subpoena. That’s why regulated markets are a different sport. At /mkt, where I’m CPO, consent and disclosure aren’t features you bolt on later, they’re the starting line. The agent race won’t go to the best demo. It’ll go to whoever solves second-order trust first.

Subscribe now

2. Gatik raised $200M and already has the revenue to back it

While everyone argues about robotaxis, Gatik put driverless box trucks on middle-mile routes between distribution centers and stores. On Aug 25 it announced a $200 million Series D led by the Qatar Investment Authority and Koch Disruptive Technologies, with Millennium Management, ARK Invest, and Intact joining. It’s the company’s largest round yet and brings total funding to about $500 million. Gatik says it has more than $600 million in contracted revenue, 85,000 fully driverless orders, and a 99% on-time rate, with customers like PepsiCo and Loblaw.

Spence’s take: Everyone chased the flashy version of autonomy. Gatik built the boring one that already has signed revenue, and boring with revenue wins.

3. Emerald AI hit unicorn status by making data centers behave

The AI bottleneck stopped being chips and became megawatts. Emerald AI sells software that turns AI data centers into flexible grid assets, dialing power down when the grid is stressed without killing critical jobs. On Aug 25 it announced a $150 million oversubscribed Series A at a company-reported $1.05 billion valuation, co-led by Energize Capital and DCVC, with NVIDIA, Siemens, GE Vernova, and In-Q-Tel among the backers. Total raised now tops $220 million, and the company says flexible operation could free up more than 100 gigawatts on the existing US grid.

Spence’s take: When the constraint moves from silicon to the socket, the winner is whoever gets the utility to say yes, and Emerald’s selling exactly that.

4. Perceptron raised $21M to teach robots to see

Small check, loud signal. Perceptron, founded by former Meta researchers, is building vision models for factory and warehouse robots and just launched its Isaac 0.5 model. On Aug 26 it announced $21 million led by Bessemer Venture Partners. The bet: AI’s next decade shows up on a warehouse floor, not in a chat window.

Spence’s take: Consumer agents get the headlines, but the patient money is quietly moving into machines that see and act in the physical world.

5. Breedr raised $27M, and only gets paid when a cow does

The least flashy round of the week might be the healthiest business in it. Breedr gives every animal a digital record from birth to sale and runs a data-verified marketplace for cattle. On Aug 26 it announced a $27 million Series B led by Partech’s impact fund, taking total funding to $46.6 million. More than two million cattle are on the platform, and the company expects close to $500 million of livestock to trade through its marketplace this year, against a US herd the company notes is the smallest since 1951. The kicker: Breedr earns when an animal sells, not on a subscription.

Spence’s take: A model that only makes money when the customer does is rare and underrated. Own a painful real-world workflow, get paid on the outcome, and you’ve built something hard to kill.

Next week’s deep-dive

Next week’s Startup Spotlight goes long on the AI power crunch: whether “flexible data centers” is a real moat or a good story with NVIDIA’s logo stapled to it. Full breakdown, bull case, bear case, and where the risk actually sits, for subscribers.

If this was useful, share it with someone who builds things. And if you want the full toolkit, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclosure and disclaimer: Spencer Gareiss is Chief Product Officer at /mkt, referenced above as an example of building in regulated markets. This newsletter is for information and education only. It is not investment advice, and it is not an offer or solicitation to buy or sell any security. Private-company valuations noted here are point-in-time and company-reported. Several figures, including Instinct’s reported valuation history, each company’s revenue and usage numbers, and forward-looking targets, are self-reported by the companies or drawn from press reports and have not been independently verified. Funding details reflect reporting as of August 28, 2026, from primary releases and outlets including the Wall Street Journal, TechCrunch, Bloomberg, Reuters, Forbes, BusinessWire, and others.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

OpenAI is putting $400 million of its own money into a new venture fund.

No outside LPs.

That’s a big change from its first fund, which raised $175 million from outside investors including Microsoft. The new fund plans to make roughly 8 to 10 lead investments a year, with checks ranging from the low millions to as much as $50 million to $100 million for select companies.

And OpenAI isn’t starting from zero.

Its first fund backed 24 startups, including Harvey and Cursor. That fund is now fully deployed.

At the same time, the rest of the market is showing just how aggressive AI venture has become.

Subscribe now

Instinct is reportedly raising $250 million at a $2.5 billion valuation after launching its private beta in February. Socure raised $156 million at a $5.2 billion valuation while acquiring agentic AI startup Fravity. DeepSeek is reportedly seeking $7.4 billion at a $74 billion valuation.

So what’s actually happening?

The AI companies aren’t just raising money.

They’re increasingly becoming part of the capital allocation system around AI itself.

That’s where one mental model from Mental Models: How to Think, Act, and Win becomes useful:

The mental model: Incentives

Don’t just ask what someone says they want.

Ask what they’re structurally incentivized to do.

A traditional VC firm needs financial returns for its LPs.

OpenAI has another incentive layered on top: build an ecosystem around the technology it develops.

If a startup becomes an important customer, distribution channel, developer platform or infrastructure partner, the strategic value may extend beyond the ownership stake.

That doesn’t make the investment good or bad.

It makes the incentive structure different.

Founders should understand that difference before taking capital from anyone.

We’ve seen the same principle at /mkt. When you’re building in a regulated market, the incentives of the issuer, investors, trading infrastructure, regulators and customers all matter. You can’t design the product around one participant and assume everyone else will follow.

My contrarian take: the next generation of venture capital may look less like investors financing startups and more like companies financing ecosystems.

That’s powerful.

It’s also something founders should examine carefully.

When your investor can benefit from your success in multiple ways, understand every one of those incentives.

Capital isn’t just capital.

Who provides it matters. Why they provide it matters even more.

This is for informational and educational purposes only. It isn’t investment, financial, legal, tax, or regulatory advice, and it isn’t an offer or solicitation to buy or sell any security or financial instrument. Reported funding amounts, valuations and deal terms may change or remain unconfirmed.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

DeepSeek is reportedly raising $7.4 billion at a $74 billion valuation, up from a valuation of more than $50 billion in June. The company plans to use the capital for research, development and computing infrastructure, with a potential Shanghai IPO next year.

But here’s the number I’d pay more attention to:

$500 million.

That’s DeepSeek’s reported annual recurring revenue.

The valuation gets headlines. Revenue tells you whether there’s a business underneath the headline.

And DeepSeek isn’t operating in a vacuum. Today, AI startup Instinct is reportedly raising $250 million at a $2.5 billion valuation, bringing its total funding to roughly $350 million. Socure raised $156 million at a $5.2 billion valuation and acquired agentic AI company Fravity. OpenAI is also launching a new $400 million venture fund, using its own capital rather than outside LP money.

That’s a lot of capital chasing AI.

Subscribe now

Which brings me to a mental model from Mental Models: How to Think, Act, and Win:

The mental model: Second-Order Thinking

The first-order question is obvious:

“How big can this company get?”

The second-order question is harder:

“What has to be true for that outcome to happen?”

For an AI company, that might mean asking:

How much compute will it require?

How quickly will inference costs fall?

How durable is customer demand?

What happens when competitors offer similar capabilities for less?

And perhaps most importantly, does revenue grow faster than the cost of delivering the product?

That’s where today’s funding numbers get interesting.

A $74 billion valuation doesn’t tell you whether the business is durable. A $500 million recurring revenue base gives you a much better starting point for asking the question.

At /mkt, we’re building in a regulated market, which creates a similar discipline. You can’t just ask whether customers want the product. You have to ask what has to be true across compliance, distribution, infrastructure and market structure for the product to work.

That’s slower.

It can also force better thinking.

My contrarian take: AI founders shouldn’t be afraid of capital. They should be afraid of capital masking weak economics.

The next phase of AI won’t be won by whoever raises the most money.

It’ll be won by whoever converts extraordinary technical capability into an extraordinary business.

This is for informational and educational purposes only. It isn’t investment, financial, legal, tax, or regulatory advice, and it isn’t an offer or solicitation to buy or sell any security or financial instrument. Reported funding amounts, valuations and revenue figures may change or remain unconfirmed.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

A car company just raised almost a billion dollars for a robot business that hasn’t sold a single robot. The same day, that car company’s stock fell almost 7 percent. If those two facts feel like they cancel each other out, you’re reading it first-order.

Here’s what happened. On August 24, XPeng announced that its robotics unit had signed share-purchase agreements to raise more than $900 million, at a post-money valuation of more than $6.3 billion, roughly 43 billion yuan. XPeng calls it the largest single-round private financing in the history of China’s embodied-AI sector, so treat “largest ever” as the company’s claim. IDG Capital led it. Gaorong Ventures joined, with Tencent and Alibaba as strategic backers. Bloomberg reported the money splits roughly into $200 million from XPeng itself, about $600 million from outside investors, and around $100 million from senior executives writing their own checks.

Subscribe now

The robot in question is IRON, XPeng’s humanoid. Mass production is targeted for the end of 2026, first inside XPeng’s own stores and campuses, then a wider launch in 2027. XPeng keeps control of the unit, holding roughly 82 percent and consolidating it on the group’s books. CEO He Xiaopeng is running the robotics business himself.

And the stock? It dropped about 6.8 percent that day, partly because second-quarter results landed at the same time: a wider net loss and light third-quarter guidance. So the market got two messages at once and picked the gloomy one.

Here’s the model I’d use to read this: second-order thinking.

First-order thinking stops at the immediate consequence. Robots don’t ship yet, the parent is losing money, this raise means dilution and distraction, so sell. All true, all shallow.

Second-order thinking asks what happens next, and after that. What XPeng actually did was carve its robotics arm into a separately capitalized entity with its own cap table, its own investors, and a public price tag that didn’t exist 48 hours earlier. It funded a capital-hungry, high-risk bet with mostly outside money, put executive skin in the game of about $100 million, and kept the core EV business from absorbing the burn. The quarter got priced by the market. The decade got priced by the round.

My contrarian take: the move that looks like financial engineering is also a discipline mechanism. A standalone company with its own valuation has to earn its next round on its own results. That’s much harder to bury inside a parent’s income statement than a line item labeled “R&D.” Whether IRON hits its timeline is a real question, and humanoid timelines slip more often than they hold. But “distraction” and “smart structuring” aren’t opposites here. Sometimes the same decision is both, and the stock chart only tells you how the crowd felt for one afternoon.

The lesson for builders isn’t “spin everything out.” It’s this: when you’re funding a long bet, structure decides who carries the risk and who’s forced to prove it. Get that right and a scary line item becomes a priced option. Get it wrong and it’s just burn.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This newsletter is for informational and educational purposes only. It is not investment, financial, legal, or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Funding figures, valuations, and investor details reflect company announcements and third-party reporting (including PRNewswire and Bloomberg) around August 24, 2026, are point-in-time, and may change. XPeng’s description of the round as a record, and the $6.3 billion valuation, are company-reported. Do your own research and consult a licensed professional before making any financial decision.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

This week Castelion, a four-year-old defense startup in Torrance, announced a $1 billion Series C, $800 million in equity plus a $250 million credit facility, at a valuation the company put at $13 billion. JPMorgan’s Strategic Investment Group co-led it alongside Andreessen Horowitz and Carlyle, with T. Rowe Price joining as a new backer. The money goes toward mass-producing Blackbeard, its low-cost hypersonic strike missile, at a 1,000-acre site in New Mexico. Castelion says it’s booked more than $500 million in U.S. military contracts in 18 months, with the Navy as its main customer, and it’s targeting initial fielding next year.

Read past the hardware and the thesis is almost boring. Castelion isn’t really selling a better missile. It’s selling manufacturing.

Subscribe now

Here’s the model from my book: economies of scale. Make one of something and it’s staggeringly expensive. Make ten thousand and the price per unit collapses, partly because fixed costs spread across more units, and partly because of the learning curve, the well-documented pattern where every doubling of cumulative output pushes cost down a predictable notch as you get better at building the thing. Cars, phones, and rockets all rode that curve from bespoke to cheap. Traditional defense never did. It builds exquisite weapons in tiny numbers at brutal unit cost, which is exactly why the U.S. keeps running short on munitions the moment a real conflict burns through the stockpile.

Castelion is betting hypersonics can ride the same curve. Design the thing for manufacturability, use commercial components where you can, own the factory instead of waiting on the legacy supply chain, and the cost per unit starts falling toward something you can actually field at volume. Its CEO framed it as going back to Apollo-era industrial ambition. That’s the pitch investors just valued at $13 billion. Not a smarter warhead. A cost curve.

Spence’s take: Deterrence has never really been about owning the fanciest weapon. It’s about having enough of them at a price you can sustain, so an adversary runs the math and decides not to test you. Something you can only build ten of doesn’t deter anyone who can build a thousand cheaper ones. So the scoreboard everyone watches, the range and the speed and the specs, isn’t the one that decides this. The one that decides it is unit cost at volume. And that’s also where the bet is riskiest, because the demo is never the hard part. Making the 500th unit as good and as cheap as the first is the entire game, and it’s where manufacturing dreams usually go to die. Castelion just raised a billion dollars to prove hypersonics are a factory problem, not a physics one. Now it has to actually build the factory.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This newsletter is for informational and educational purposes only. It is not investment, financial, legal, or tax advice, and nothing here is a recommendation or solicitation to buy, sell, or hold any security or to participate in any investment strategy. Funding amounts, the $13 billion valuation, and contract figures reflect information reported by the company and by third-party news sources as of the publication date and are point-in-time figures that may change. Forward-looking statements, including production plans, cost-reduction goals, and fielding timelines, are attributed to the company and are not statements of fact. Private company figures are difficult to verify independently. Do your own research and consult a licensed professional before making any financial decision. The author may hold positions in companies or sectors mentioned and receives no compensation from any company covered here.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

On Monday, Valar Atomics announced a $1 billion Series B led by Sequoia, plus a $200 million credit facility, for $1.2 billion in new financing. Bloomberg reported the round valued the company around $6 billion, a figure Valar didn’t confirm. This is a company that raised just $130 million last October. Its pitch is simple to say and brutally hard to do: build small nuclear reactors like manufactured products instead of one-off construction projects. It says its Ward 250 reactor went critical in June, and that it has powered an Nvidia AI chip directly off a fission reaction. AI data centers are its biggest target customer.

Subscribe now

The easy read is “AI needs power, nuclear is back, money floods in.” True, but shallow. Demand for power has been around for decades. It never turned nuclear into a venture-scale bet before. Something else changed.

Here’s the model from my book: first principles. Most people reason by analogy. Nuclear is slow and absurdly expensive, so it’ll always be slow and absurdly expensive, so smart money stays away. First-principles thinking throws out the analogy and asks what’s actually true at the bottom. And at the bottom, the physics of a small reactor isn’t the problem. The problem is how we build them: as bespoke, multibillion-dollar megaprojects, each one a cathedral poured in concrete over a decade. Valar looked at that and asked a different question. What if a reactor were a product you manufacture on a line, not a monument you construct on a site? Same physics. Completely different cost curve.

That reframe is the whole thing. Valar’s own founding observation was that nuclear is one of the most abundant energy sources on earth, saddled with one of the worst cost curves in industry, and that almost nobody was actually trying to fix the curve itself. Attack it with manufacturing, vertical integration, and repetition, and a source everyone wrote off as too expensive starts marching toward the cost of simply making the thing.

Spence’s take: Sequoia didn’t write a billion-dollar check because AI is thirsty. Everyone knows AI is thirsty. It wrote the check because Valar is betting the industry’s defining number, the cost and time to stand up a reactor, was never a law of physics. It was an inherited habit. That’s the first-principles move: find the thing everyone treats as fixed and prove it’s a choice. Now comes the part the model can’t do for you. Nuclear is one of the most regulated, least forgiving arenas on earth, and manufacturing economics only matter if the reactors actually get built, approved, and run safely at scale. The insight is real. The execution is where bets this size go to be tested.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This newsletter is for informational and educational purposes only. It is not investment, financial, legal, or tax advice, and nothing here is a recommendation or solicitation to buy, sell, or hold any security or to participate in any investment strategy. Funding amounts reflect information reported by the company as of the publication date. The approximately $6 billion valuation was reported by Bloomberg and was not confirmed by the company; it is a point-in-time figure that may change and could not be independently verified here. Forward-looking statements, including plans to mass-produce reactors, drive down the cost curve, and serve AI data centers, and the company’s own descriptions of its technical milestones, are attributed to the company and are not statements of fact. Advanced nuclear energy is heavily regulated and carries substantial execution, regulatory, and safety risk. Private company figures are difficult to verify independently. Do your own research and consult a licensed professional before making any financial decision. The author may hold positions in companies or sectors mentioned and receives no compensation from any company covered here.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

The Information reported this month that Kalshi is in advanced talks to raise at least $750 million at a roughly $40 billion valuation, with Sequoia Capital and Wellington Management weighing whether to co-lead. Nothing’s closed and the company hasn’t confirmed it, so treat those figures as reported, not final. But the trajectory is real. Kalshi raised $1 billion at a reported $22 billion valuation in May, and press reports put its annualized revenue near $4 billion as of July, up from a rounding error two years ago. Sports contracts reportedly drive more than 80% of that volume.

Subscribe now

That’s the setup. Here’s the tension.

Kalshi is regulated by the CFTC, and that federal status is the whole moat. It’s why a $1.3 trillion shop like Wellington can even look at the cap table. It’s why Sequoia has publicly pegged Kalshi at roughly 95% of the U.S. prediction market. Being the compliant one is the product.

It’s also the target. More than a dozen states are challenging Kalshi’s sports contracts as unlicensed gambling, with reported actions ranging from outright bans to, in Arizona, criminal charges. The exact contracts printing the revenue are the exact contracts under legal fire.

One of the models in my book is Inversion. Instead of asking how something wins, you ask what has to be true for it to break. Flip Kalshi and the answer gets uncomfortable, because the bull case and the bear case are the same sentence. Federal regulation preempts state gambling law, right up until a court decides it doesn’t. The moat and the landmine are the same piece of ground.

Most people reading the $40 billion headline are pricing the upside: revenue, market share, an IPO the CEO has said he’s considering for 2027. Invert it and you’re really pricing one question. Does the CFTC framework hold against the states? Everything else is a footnote to that.

I’ve built inside regulated markets long enough to respect the trade. At /mkt we work under Reg A+ with tZERO as the trading infrastructure, and that framework is slow and expensive on purpose. You inherit its protection and its fights at the same time. You don’t get to keep one without the other. That’s not a bug of building in regulated markets. It’s the whole deal you sign.

So here’s the contrarian read. A $40 billion price on a company whose core product is being called illegal in a dozen states isn’t really a bet on prediction markets. It’s a bet on one legal question resolving one way. That might turn out to be a great bet. Just name the bet you’re actually making before you make it.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This post is for informational and educational purposes only and is not investment, legal, or financial advice, nor a recommendation to buy, sell, or hold any security or interest in any company mentioned. Kalshi is a private company; all valuations are point-in-time and as reported by the press. The reported $750 million raise and roughly $40 billion valuation were reported by The Information citing unnamed sources, have not been confirmed by Kalshi, and were not closed as of this writing. Revenue, volume, and market-share figures are as reported by the press and have not been independently verified. Author affiliation disclosure: Spencer Gareiss is Chief Product Officer at /mkt; the /mkt reference is included as a structural example of building in regulated markets, not as a solicitation or endorsement.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

Company Overview

Ambrook builds financial-management software for the businesses that grow, build, and haul the physical world. In plain English: it’s bookkeeping, invoicing, bill pay, payments, and cash management, except it’s designed for a cattle ranch that runs four entities, buys and sells animals that are literally born, grow, and die on the balance sheet, and collects most of its revenue in two lumpy seasons. Think of it as what QuickBooks would look like if someone rebuilt it in 2026 for people who wear boots to work.

The company just closed a $30 million Series B, announced in early August 2026. That brings total funding to about $59 million, following a $26.1 million Series A back in July 2025. The round was led by Lachy Groom, the solo capitalist who used to run Stripe Issuing, which matters more than it looks like it should, and I’ll come back to that. Thomson Reuters Ventures, Thrive Capital, Field Ventures, and Cameron Ventures joined, along with a roster of operator angels: Guillermo Rauch of Vercel, Akshay Kothari of Notion, and Tomer London of Gusto.

Subscribe now

Ambrook was founded in 2020 by Mackenzie Burnett (CEO), Dan Schlosser (product), and Jeff Anders (design), and it’s based in Denver with an office in New York. The traction is the part that made me stop scrolling. Ambrook is over 8,000 customers, up from about 2,500 a year ago. More than 1,000 are trucking businesses now, with hundreds of general contractors and property managers on top of its agricultural core. Nearly a third of users open it only on their phones each month. And here’s the number that reframes the whole company: by the CEO’s account, about half of customers are putting their finances on software for the very first time, coming straight off paper and spreadsheets.

The Market

Start with the core. USDA’s 2022 Census of Agriculture counted 1,900,487 farms in the United States, the first time the number fell below two million since before the Civil War, working 880 million acres, operated by 3.37 million producers, producing $543.1 billion in agricultural products. These aren’t hobby plots. Fewer than 6 percent of farms, about 105,000 of them, sold more than three-quarters of all output, and many are multi-million-dollar operations. A striking share of them still run on paper, on desktop software from the 1990s, or out of a shoebox of receipts.

Now the timing. American agriculture is under real financial stress. In the 2025 ABA/Farmer Mac Agricultural Lender Survey, lenders expected only about 52 percent of US farm borrowers to be profitable, a sharp drop from recent years. When margins were fat, sloppy books were survivable. When roughly half your borrowers are underwater, financial visibility stops being a luxury and becomes the difference between getting the operating loan and not getting it.

Two more tailwinds make now the moment. First, AI-native bookkeeping slashes the cost to serve a customer base that used to be too small, too seasonal, and too scattered for anyone to serve profitably. Second, embedded finance lets a startup issue cards and move money through partners instead of becoming a bank. Ambrook’s Visa commercial cards are powered by Stripe and issued by Celtic Bank, a member FDIC institution. That’s the Lachy Groom connection paying off: this is a payments-native company, not an accounting tool that bolted a card on later.

The positioning I find most interesting is that Ambrook isn’t only taking customers from QuickBooks. When half your users are first-time software buyers, you’re not fighting over a fixed pie, you’re growing the pie. That’s a very different motion from most vertical SaaS, which usually means knife-fighting to convert switchers one at a time.

Business Model and Moat

Ambrook makes money two ways, and that’s the crux of it. There’s subscription software revenue, the classic SaaS engine. And there’s the payments side: card interchange and money movement running through its Stripe-and-Celtic rails. Two revenue engines from one customer, where healthy payment economics can subsidize cheaper, stickier software.

Who else is in the market? The incumbent is Intuit’s QuickBooks, more than 30 years old and built for a generic small business, not for multi-entity operations with livestock inventory, government-program payments, and cash flow that arrives twice a year. There are legacy desktop farm-accounting tools, a handful of newer farm-finance startups, and general-purpose small-business software. None of them combine vertical depth, mobile-first design, and embedded payments the way Ambrook is trying to.

So what’s actually defensible? Four things. One, switching costs. Once your books, receipts, bill pay, cards, and audit-ready tax records all live in one place, ripping it out hurts, and it hurts double for a business that had no system before Ambrook. Two, vertical depth. Modeling a herd, multiple profit-and-loss lines, seasonal cash cycles, and federal payments is genuinely hard to replicate as a feature bolted onto a horizontal tool. Three, the embedded-finance flywheel: payment data sharpens the cash-flow features, which deepen daily usage, which grows payment volume, which funds the software. Four, distribution and trust in a word-of-mouth market where outsiders struggle to earn credibility.

That last piece connects to something I think about constantly. Issuing cards through a chartered bank, moving customer money, and keeping compliance-grade records means Ambrook is building inside a regulated financial system from day one. That’s the same reality I live in at /mkt, where regulated-markets infrastructure isn’t a feature, it’s the entire game. Building where the rules are strict is slower, but the moat is real, because the next entrant has to clear the same bar you already cleared.

Spence’s Take

Two models here.

The bull case runs on switching costs, or what I’d call the system-of-record effect. The most valuable position in any workflow software is being the place the data lives. Ambrook’s wedge is unusually strong because it’s often the first system a business ever adopts. You don’t just win the account, you define how that business keeps its books forever. Layer embedded payments on top and the account gets stickier and more valuable every quarter. If the new trucking and construction cohorts retain the way the ag base does, this compounds into a durable company that grows its revenue per customer as it moves up-market.

Now the bear case, and my favorite tool for it is inversion. Don’t ask how Ambrook wins. Ask what would guarantee it loses, then check whether the company is walking toward those things. Three ways it loses. First, focus dilution: chasing agriculture plus trucking plus construction plus property management all at once can turn a sharp vertical wedge into a shallow generalist that gets out-specialized in every category. Second, the downturn bites the hand that feeds: if farm profitability keeps sliding, core customers cancel software before they cancel diesel. Third, the interchange trade breaks: if payment economics get squeezed, or a bank-partner or regulatory hiccup disrupts the card program, one of the two revenue engines sputters. None of these is fatal today. All of them are worth watching.

The honest read is that Ambrook’s greatest strength and its biggest risk are the same decision: expansion. Customer-pulled expansion, like the Arizona cattle feeder who reportedly asked to use it for his trucking business, is a gift. Founder-pushed expansion, chasing every adjacent vertical because the deck says the market is huge, is how focused companies go soft. Which one this becomes is the whole ballgame.

Why It Matters

For investors and VCs: This is a company worth watching, and I’ll leave it there, because nothing in this newsletter is investment advice. If you’re tracking it, the signals that matter are net revenue retention, the attach rate on payments and cards, average contract value as they move up-market, and whether the trucking and construction cohorts hold through a downturn the way agriculture has. Watch whether Thomson Reuters Ventures becomes a real distribution channel into tax and compliance, not just a check on the cap table. And note that Ambrook didn’t disclose a valuation with this round, so anyone quoting one is guessing.

For potential customers: If you run a farm, a ranch, a trucking outfit, a contracting business, or a property portfolio on paper, spreadsheets, or a desktop tool from two decades ago, this is aimed squarely at you. What’s coming is AI-native bookkeeping, invoicing, bill pay, business cards, and cash management built for multiple entities and seasonal money, with about a third of users running it entirely from their phones. The problem it solves is visibility: knowing which of your enterprises actually makes money, and having the records ready when the lender or the tax preparer calls. Standard fine print applies. The cards run through Stripe and Celtic Bank, so treat it like the financial product it is.

For competitors and builders: Three lessons. One, boring verticals full of first-time software buyers can be bigger and stickier than the crowded categories everyone else fights over. Greenfield beats knife-fighting. Two, embedded finance can turn one product into two businesses, and the payments engine can fund the software engine. Three, let your expansion be customer-pulled before it’s founder-pushed. The market will usually tell you where the next vertical is if you’re listening to support tickets instead of staring at a market-size slide.

The Bottom Line

Ambrook is a sharp wedge into a huge, underserved market at exactly the moment financial stress turns its product from a want into a need, with a payments-native model most vertical-SaaS peers can’t match. The real risk isn’t the market, it’s discipline: whether a company this good at expanding can stay focused enough to keep winning each vertical it enters. Watch retention in the new categories and the attach rate on payments, and you’ll know which Ambrook you’re getting.

This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.

Disclaimer: This newsletter is for informational and educational purposes only. It is not investment, financial, legal, or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Ambrook is a private company; its $30 million Series B was announced in early August 2026, and funding figures, investor names, and customer metrics reflect company statements and third-party reporting (including FinTech Futures, Upstarts Media, FinSMEs, and agricultural trade press) available at that time. Customer outcomes and traction numbers are company-reported and point-in-time. Ambrook did not disclose a valuation. Market and farm-economy statistics are from the USDA 2022 Census of Agriculture and the 2025 ABA/Farmer Mac Agricultural Lender Survey. Do your own research and consult a licensed professional before making any financial decision. Disclosure: the author is Chief Product Officer at /mkt; /mkt is referenced only as an illustrative example of building in regulated markets and this is not a solicitation for /mkt or any security.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight

Last week the Financial Times reported that JPMorgan cut its banking relationship with Polymarket back in October, citing regulatory concerns, and told the company to find a new bank. Polymarket pushed back on the framing, telling the FT it still works closely with JPMorgan across multiple areas, and the bank reportedly kept some ties open in case it wants an underwriting role in a future IPO. Reuters confirmed the core of it through its own source. Around the same window, a judge ordered Kalshi to stop offering most wagers in Washington state, and New York City opened a probe into both platforms. On the other side of the ledger, the CFTC has told Kalshi to keep operating despite a New York lawsuit. Federal tailwind, state and banking friction, all at once.

Subscribe now

Keep the numbers in mind while you read that. Kalshi was reportedly valued around $22 billion in May and, per FT reporting, has been in talks for a round near $40 billion. Polymarket sat around $15 billion and, according to Bloomberg and CNBC, has been seeking north of $20 billion, with both reportedly clearing a billion dollars in annualized revenue. These aren’t fragile startups. And they still spent the week playing defense against courts, regulators, and a bank.

Here’s the model from my book: inversion. Most founders ask “how do we grow faster than the other guy?” Inversion flips it: “what’s the thing that kills us, and are we building to avoid it?” In consumer software the answer is usually a competitor or churn. In regulated markets, it almost never is. It’s the perimeter: a regulator, a court, a state attorney general, or a bank quietly deciding you’re too much risk to hold. None of those show up on your growth chart until the day they end you.

I built prediction markets and derivatives at Robinhood, so I’ve watched this movie before. The product can be brilliant and the demand real, and none of it matters if you lose the rails underneath it. Banking access in particular is a dependency almost nobody inverts on until it’s gone.

Full disclosure: I’m CPO at /mkt, where we tokenize athlete income and made the call early to build inside the perimeter, using a Reg A+ framework with registered trading infrastructure. I’m not holding that up as the only right answer, and I’m not knocking Kalshi or Polymarket, who’ve won plenty of their own regulatory fights. The point is narrower. In a markets company, where you choose to stand relative to the rulebook is the first design decision, not the last.

Spence’s take: The prediction-market boom is real, and I think the category keeps growing. But the scoreboard everyone watches, valuation and volume, measures the wrong risk. The one that matters doesn’t trend on a chart. It shows up as a letter from a regulator or a bank. Invert first. Build for the thing that can end you, then go chase the growth.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.

Disclaimer: This newsletter is for informational and educational purposes only. It is not investment, financial, legal, or tax advice, and nothing here is a recommendation or solicitation to buy, sell, or hold any security or to participate in any investment strategy. The regulatory, legal, and banking matters described are based on third-party reporting (Financial Times, Reuters, Bloomberg, and CNBC) as of the publication date, are evolving, and may change; certain characterizations are those of the reporting outlets, and the companies named have disputed some of that framing. Reported valuations and funding discussions are point-in-time figures, some are unconfirmed by the companies, and private-company figures are difficult to verify independently. The author is Chief Product Officer at /mkt and references it here only as an illustrative example of building in a regulated market, not as a solicitation; any securities offering by /mkt is made solely through its official offering materials. Do your own research and consult a licensed professional before making any financial decision. The author may hold positions in companies or sectors mentioned and receives no compensation from any company covered here.

If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.

Get it Here

Share Startup Spotlight