1. Menlo Ventures Just Cashed In a $500M Bet Nobody Else Wanted to Make

Menlo Ventures raised $3 billion in new capital, the largest fundraise in the firm’s 50-year history, after its early investment in Anthropic, now worth nearly $14 billion as Anthropic’s valuation climbed past $900 billion. Menlo led Anthropic’s Series D in 2024 by structuring roughly $500 million of that round as a special purpose vehicle, at a time when no major firm was writing three-quarter-billion-dollar checks. AI Funding TrackerTech Funding News

Spence’s take: Everyone calls these bets “contrarian” after they pay off. The actual contrarian move happened in 2024, when nobody knew if it would work.

2. XCures Doubled Its Valuation Cleaning Up Medical Records With AI

XCures landed a $46 million Series B led by Innovius Capital, bringing total funding to more than $76 million and valuing the company at $127 million post-money, more than double its prior round. The company grew from roughly $3 million to $10 million in annualized recurring revenue in 2025 and is on track to break $20 million in 2026, while achieving cash-flow breakeven last year. Calcali Tech

Spence’s take: A profitable company intentionally entering a burn phase to grow faster is one of the most underrated signals in venture. It means the founders chose growth, not the other way around.

3. Healthcare AI Just Crossed an $8.5 Billion Threshold This Year Alone

Investors have put an estimated $8.5 billion into seed- to growth-stage funding for AI-powered health tech companies in 2026 as of late June. Calcali Tech

Spence’s take: Healthcare moves slower than every other sector until it doesn’t. This number says we’re past the “if” stage and into the “how fast” stage.

4. Twenty’s $1B Cyber Warfare Bet Is Still the Story Nobody’s Fully Priced In

Twenty, the Arlington-based offensive cyber startup for the U.S. military, became America’s first venture-backed cyber warfare unicorn earlier this month after raising a $100 million Series B at a $1 billion valuation, less than two years after founding.

Spence’s take: A billion-dollar valuation with no disclosed revenue isn’t proof of a great company. It’s proof that investors believe the category just became fundable. Those are very different things.

5. Menlo’s Anthology Fund Shows the New VC Playbook: Be the Insider, Not Just the Investor

Menlo launched a $100 million fund called Anthology in partnership with Anthropic, which has since grown to roughly $250 million deployed, backed more than 60 companies, and produced exits including Graphite (acquired by Cursor) and Astrix Security (acquired by Cisco). Tech Funding News

Spence’s take: The smartest funds aren’t just writing checks anymore. They’re building proprietary pipelines into the company they bet the firm on. That’s a moat most LPs don’t fully appreciate yet.

Why Menlo Is the One to Watch

I want to run this through asymmetric bets, a model I cover in my book, Mental Models: How to Think, Act, and Win. Menlo’s 2024 Anthropic investment had a defined downside: the capital they put in. The upside, if Anthropic became one of the most valuable private companies in the world, was essentially unbounded. That’s the entire logic of venture capital distilled into one decision. Most investors talk about asymmetric bets. Menlo actually structured a $500 million SPV around one, at a moment when the broader market had gone quiet on big checks. The lesson isn’t “go invest in AI labs.” It’s that real conviction shows up as structuring your bet correctly when consensus is against you, not just picking the right company.

——————————————————————————————————————

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, right now.

Get it Here

——————————————————————————————————————

Next week’s deep-dive:

——————————————————————————————————————

Two ex-Palantir guys built a search bar for data the government already owns. Investors just put a $6.8 billion price tag on it.

That’s Peregrine Technologies, and on Monday it announced a $250 million Series D at a $6.8 billion valuation. The round was led by existing investors, including Fifth Down Capital, Sequoia Capital, OG Venture Partners, Goldcrest Capital, XYZ Ventures, and Godfrey Capital. Here’s the number that matters most: 15 months ago, Peregrine’s Series C valued it at $2.5 billion. So this is close to a 3x markup in barely over a year.

What does it actually do? It connects the data a city already has, police records, 911 logs, permit databases, sensor feeds, and makes it searchable in real time, with role-based access and a full audit trail baked in. The company says it doesn’t create or collect new data and doesn’t use facial recognition. It now serves more than 400 agencies covering roughly 125 million people across North America, doubled its customer base in a year, and is running security fusion centers for eight of the eleven U.S. host cities for the World Cup this summer.

Subscribe now

The model from my book that explains this is Circle of Competence.

Founders Nick Noone and Ben Rudolph didn’t chase the flashiest AI demo. Noone ran Palantir’s Special Operations business. Rudolph built data infrastructure for the UN’s refugee agency. They went straight into the hardest, most trust-sensitive arena there is, public safety, where a wrong answer carries real consequences, and they earned competence and credibility there first. Now they’re expanding outward from strength: federal, enterprise, international, even pilots in financial services and travel. That’s how a circle of competence is supposed to grow. You start where you’re genuinely excellent, then push the edges deliberately. You don’t start at the edge.

There’s a second-layer lesson here, and it’s the one builders should sit with. The AI model isn’t the moat. Models keep getting cheaper. The hard, scarce part is the context: knowing where the data lives, what the rules are, and who’s allowed to see what. As the CTO put it, the model isn’t enough on its own, it needs context. Peregrine owns the unglamorous layer everyone else skipped past.

I see the same thing building in regulated markets. At /mkt, the compliance and structure aren’t friction sitting on top of the product. They’re the product. The boring layer is the moat.

My contrarian flag, kept conservative because I’m not telling anyone what to buy. Two things. First, a 3x markup in 15 months is investors repricing a category, not a business literally tripling, and Peregrine won’t disclose revenue, so watch that gap. Second, the real risk isn’t the tech, it’s the politics. Rival Flock Safety learned what happens when public backlash catches up to a surveillance-adjacent product. Peregrine’s restraint may be its smartest design choice, because in this market, trust is the thing that’s hardest to rebuild once it breaks.

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, right now.

Get it Here

Startup Spotlight is for informational and educational purposes only. It is not investment advice, an offer, or a solicitation to buy or sell any security. Company metrics are self-reported, and funding details are drawn from public reporting and company statements. Figures may change.

For a century, the house always won because the house wrote the rules. A startup out of New York just got Washington to change them.

The Commodity Futures Trading Commission granted Novig approval to operate as a federally regulated Designated Contract Market, which is the formal term for an exchange. In plain English, Novig now has a path to offer sports trading in all 50 states, not as a sportsbook, but as a venue where people trade contracts against each other. A competitor, ProphetX, cleared the same bar about a week earlier. The race to be the regulated sports exchange is officially on.

Quick context on the company. Novig was founded in late 2024 by Jacob Fortinsky and Kelechi Ukah. It runs a commission-free, peer-to-peer exchange with an order book, meaning there’s no house taking the other side of your trade. Back in February it raised a $75 million Series B led by Pantera Capital, pushing total funding past $105 million at a reported valuation around $500 million. The company says its annualized trading volume has run into the billions after roughly 10x growth in 2025. It also claims a far higher share of profitable users than traditional sportsbooks. Treat that as a company figure, because it is one.

Subscribe now

The model from my book that cuts to the heart of this is Follow the Incentives.

A traditional sportsbook is your counterparty. When you win, it loses. So its incentives are exactly what you’d expect: shade the odds, limit the sharp players, discourage anyone who wins too often. That isn’t a bug in a few bad operators. It’s the structure. As Charlie Munger liked to say, show me the incentive and I’ll show you the outcome.

An exchange flips that. It doesn’t take the other side of your trade. It makes money on volume and fees regardless of who wins, so its incentive is a deep, active, trusted market with lots of participants. Align the incentives, and you get better pricing and a platform that doesn’t punish you for being good. That structural alignment, not the app design, is the real threat to the incumbents.

I’ve watched this movie before from the building side. At Robinhood I worked on prediction markets and derivatives, and at /mkt we build on regulated trading infrastructure. The pattern repeats: take a market run by a middleman, financialize it, and move it onto a regulated exchange. It happened to stock brokers. It’s now happening to the sportsbook.

Here’s my contrarian take, and I’ll keep it conservative. A license is not a victory. It’s a permission slip. Novig is walking into a brutal field where Kalshi already dominates and Robinhood, DraftKings, and FanDuel have enormous distribution. The exchange model is structurally better, but better structure loses to better distribution all the time. And there’s a real societal question underneath all of this: turning every game into a 24/7 tradable asset is exactly why some states are pushing back. The incentives that make this a great business are the same ones worth watching closely.

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, right now.

Get it Here

Startup Spotlight is for informational and educational purposes only. It is not investment advice, an offer, or a solicitation to buy or sell any security. Company metrics are self-reported, and funding details are drawn from public reporting and company statements. Figures may change.

That’s the headline that should make you stop scrolling. Stark Defence, founded in 2024, just closed a €500 million round (about $570 million) led by Sequoia Capital and Peter Thiel’s Founders Fund, with the NATO Innovation Fund, Project A, Air Street Capital, 201 Ventures and Döpfner Capital joining. Reported valuations differ by outlet, but every source puts it north of €3 billion, roughly triple where the company sat back in February. Total raised now runs around €640 million, up from roughly €140 million before this round.

Here’s the part the headlines skip. According to Financial Times reporting, Stark’s Virtus drones failed to hit a single target across four attempts in trials with British and German forces last November. Germany approved a €269 million Bundeswehr contract anyway, with options stretching toward €2.86 billion. More than 80% of the new money is earmarked for manufacturing and R&D, not go-to-market. And Stark is now on the clock to equip a full German brigade in Lithuania by the end of 2026.

So why would two of the sharpest funds on earth write that check?

Subscribe now

This is where one of the models from my book earns its keep: Second-Order Thinking.

First-order thinking reads the trial result and stops. The drone missed, so it’s a bad product, so it’s a bad bet. Clean, fast, wrong.

Second-order thinking asks what happens next, and then what happens after that. The thing being priced here isn’t this quarter’s hit rate. It’s European strike-drone manufacturing capacity during a rearmament cycle, sitting on top of a procurement position inside contracts that are already signed. In a war of attrition, a good-enough system you can build by the thousand beats a perfect system you can’t ship. Capacity is the scarce asset. Stark is being financed like a factory, because a factory is what the buyers actually need.

I’ve spent enough time around regulated, infrastructure-heavy markets to recognize the shape of this. At /mkt, we build inside systems where the demo is the easy part and compliant execution at scale is the hard part. Defense procurement rhymes with that: signed contracts, certifications, and hitting a delivery deadline carry more weight than a single flashy test.

Now the contrarian flag, and I’ll keep it conservative, because I’m not here to tell anyone what to buy. When capital races to “be in the room,” price can run out ahead of proof. A signed contract isn’t a battlefield result, and €3 billion-plus is a lot of expectation to load onto a two-year-old company that hasn’t yet shown it can manufacture at the volumes it’s promising. The second-order logic is sound. The valuation is where the risk actually lives.

Watch the factories, not the press releases. The next eighteen months are a single question: can Stark build what it already sold?

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, right now.

Get it Here

Startup Spotlight is for informational and educational purposes only. It is not investment advice, an offer, or a solicitation to buy or sell any security. Company metrics are self-reported, and funding details are drawn from public reporting and company statements. Figures may change.

Subscribe now

That single decision is why this company is worth your attention this week.

Company Overview

In plain English: Adaptive Innovations is a home health care provider that runs on its own AI operating system. When a hospital discharges an elderly patient who needs care at home, a traditional agency handles the intake, scheduling, clinical documentation, and compliance billing through a tangle of legacy software and human coordinators. Adaptive automates that back office with AI and sends its own clinicians into the home. It isn’t selling the tooling to someone else. It’s the licensed provider getting paid to deliver the care.

The company is at the Series A stage. It came out of stealth in early June with $60 million raised to date: a $50 million Series A led by Felicis, plus a previously undisclosed $10 million seed round that Bain Capital Ventures led. The Series A drew in Bain Capital Ventures, Optum Ventures, Sunflower Capital, Conviction, BoxGroup, SV Angel, Dorm Room Fund, and Constellation, along with angels from healthcare services and frontier AI labs.

The founding team runs deep on both sides of the problem. Co-CEOs Alex Wendland and Logan Stinson lead alongside CTO Ryan Tolsma and COO Hunter Stinson, working out of New York City and Dallas. The company says its bench includes engineers from Palantir and Jane Street and operators from McKinsey, Harvard, Stanford, and the U.S. Army Rangers. That’s an unusual mix: people who can ship hard software sitting next to people who’ve actually run operations under pressure.

What have they built so far? According to the company, 200 clinicians have completed more than 100,000 home visits in 18 months. Adaptive reports a rehospitalization rate below 5% against an industry average around 11%, and says its AI workflows have cut clinician documentation time by roughly 80%. Treat those as company-reported figures, because they are. They’re still the kind of early numbers that make sharp investors lean in.

Read more

On June 23, Cadence, the clinical AI company managing chronic care for older adults, announced a $100 million Series C led by Spark Capital. The syndicate is worth reading closely. Thrive Capital, General Catalyst, Coatue, and B Capital came in alongside three health-system investors: Corewell Health Ventures, Memorial Hermann, and Duke Health. Cadence also announced a new affiliation with Duke Health the same day.

The product fits the moment. Cadence supports more than 90,000 patients across 20-plus health systems, using AI to handle the high-volume routine work of chronic care, monitoring vitals and surfacing risk, so clinicians spend their time on the decisions that need a human.

Here’s the part the headline skips. Medicare’s innovation center rolled out an experimental program called ACCESS that pays providers a set amount per patient, roughly $180 to $420 in year one depending on the condition, and pays the full amount only if patients hit outcome targets. Compare that to a standard remote-monitoring setup, where a provider can bill around $100 per patient per month for doing the work regardless of results. Cadence said it would be among the first to sign up.

This is where I’d reach for Skin in the Game, one of the models in my book. The idea is simple: you carry real downside, not just upside, and your incentives are wired to the result instead of the activity. ACCESS flips healthcare’s default. The old model pays for motion. ACCESS pays for outcomes. Cadence is volunteering for a structure where it eats the risk if patients don’t get better.

That also explains the cap table. When three health systems write checks into the company they’re buying from, they’ve got skin in the game in its success too. That’s not a logo grab. That’s alignment.

The contrarian read: the $100 million is the least interesting number in this story. The interesting numbers are $180 to $420 per patient paid on results, against $100 a month paid no matter what. Cadence is betting it can use AI to make the outcomes math work where labor-heavy competitors can’t. If patients don’t improve, the model doesn’t pay. Most companies run from a constraint like that. The good ones build toward it.

I think about this at /mkt, where we operate inside a Reg A+ structure with tZERO as our trading infrastructure. In regulated markets you don’t write the rules, you build to fit them. The structures that force you to win only when your users win can look like constraints. They usually turn into moats.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

This newsletter is for informational and educational purposes only. It is not investment, legal, tax, or financial advice, and it is not an offer or solicitation to buy or sell any security. Figures are drawn from publicly available reports as of June 23, 2026, and are subject to change. Do your own research and consult a licensed professional before making any decision.

Read more

They’re building agents that do actual work. Sierra is raising billions to automate customer support workflows. Harvey is raising hundreds of millions to write legal documents. Cognition, Adept, and Poolside are racing to build coding agents that handle real development tasks. Meanwhile, every startup that pitched itself as a “ChatGPT for X” is either pivoting or dying.

The difference isn’t subtle. And it completely changes how you should think about building an AI company right now.

A chatbot improves information retrieval. An agent completes tasks. That one word, “complete,” is the difference between a feature and a business.

Why task completion is eating the market.

Here’s the brutal math: if an AI chatbot saves you 30 minutes per day but requires a human to make the final decision, you’ve built a tool. If an AI agent saves you 3 hours per day and handles the decision autonomously with human review gates, you’ve built a replacement for work. One is a nice-to-have. The other is a must-have.

Enterprises don’t pay $10K to $100K per month for something that saves a human 30 minutes. They pay for something that reduces headcount or saves significant operational cost. That’s why task-completion agents are raising while chat interfaces are sitting in acquihires.

Let me show you the capital concentration: the top 10 agentic AI startups account for roughly 49% of all funding in the entire market, even though there are 99 funded companies. That’s not distributed hype. That’s concentrated capital moving toward the things that actually work.

The breakdown tells you everything:

Customer service agents (Sierra, Parloa, Decagon, Moveworks) have raised several billion combined because they handle inbound requests autonomously and route exceptions to humans. They measure success in “tickets resolved without human,” not “conversations completed.”

Legal AI (Harvey, Legora, EvenUp, Luminance) is raising hundreds of millions because these agents actually draft documents, conduct discovery, and flag risk. They sit in workflows that cost law firms six-figure retainers. The ROI is undeniable.

Coding agents (Cognition, Poolside, Magic, Adept) are winning because they handle pull request reviews, refactor code, and complete features. Developers measure the agent not by how helpful its suggestions are, but by how many lines of code it actually writes.

This is the flip. When the AI system is advising a human, the human stays in control and the metric is “time saved.” When the AI system is doing the work autonomously, the metric is “work completed,” and the leverage is asymmetric.

Why conversational AI hit a ceiling.

The problem with chatbots is that they’re advisory, not autonomous. You ask it a question. It gives you an answer. You still have to interpret the answer, fact-check it, and act on it. The human cognitive load doesn’t actually go down. You’re just replacing one tool (Google) with another (ChatGPT).

Enterprises realized this in 2024. By 2025, they started funding agents instead. By 2026, the funding gap is obvious.

The better chatbot doesn’t scale. It plateaus. Every marginal improvement in instruction-following or reasoning takes more training, more tokens, more infrastructure. Meanwhile, agents that automate specific workflows hit exponential returns. Do the same task 1,000 times, the agent gets cheaper and faster. Do the same conversation 1,000 times, the chatbot just handles the same easy questions.

The architecture that matters.

Real agentic AI systems have three things chatbots don’t:

Autonomous decision-making. The agent makes choices (with guardrails, with human review gates, but autonomously). It doesn’t propose and wait. It acts.

Multi-step workflow execution. The agent breaks complex tasks into subtasks, executes them across systems, and tracks context. A chatbot answers one question per turn. An agent plans across an entire workflow.

Integration with business systems. The agent reads from your CRM, writes to your legal database, executes transactions, updates your spreadsheets. A chatbot talks to you. An agent talks to your infrastructure.

That’s why platforms like Adept (which positions itself as an “autonomous collaborator” that navigates software and handles workflows), UiPath (which merged RPA with agentic AI to automate cross-system workflows), and CrewAI (which lets you assemble teams of specialized agents) are winning. They’re not building better conversationalists. They’re building systems that do work.

Here’s how it applies to how we think about /mkt and regulated markets.

In a trading infrastructure, agents need to understand regulatory rules, execute transactions autonomously, and maintain audit trails. That’s not a chatbot use case. That’s an agentic AI use case. The agent needs to know the rules, make decisions within them, and prove its work. That’s why agentic systems matter in regulated markets. They’re not advisory. They’re operational.

The mental model is “Asymmetric Leverage.”

When you use AI to advise humans, the leverage is symmetric. You’re both doing cognitive work. You’re just dividing it differently. When you use AI to replace work, the leverage is asymmetric. The AI does the work, humans do the review. That’s a 10x efficiency gain, not a 1.2x gain. Capital flows toward asymmetric leverage.

My contrarian take: if you’re building “ChatGPT for X” in 2026, you’re already late.

The winners aren’t building conversational AI. They’re building task-completion agents. They’re not asking “how do we make better suggestions?” They’re asking “how do we do the entire workflow autonomously with the right safety gates?”

The chatbot graveyard is real. And it’s full of startups that could have been worth billions if they’d recognized the shift from “advise” to “automate.”

If you’re building with AI right now, don’t optimize for better answers. Optimize for autonomous task completion. That’s where the capital is. That’s where the defensible moat is. That’s where the 100x returns are.

Everything else is a chatbot. And nobody’s funding chatbots anymore.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

Ramp raised $500 million in Series E. Tennr closed $101 million in Series C. Gecko Robotics is one of the most watched startups in the market. They didn’t win because they had the best narrative. They won because they solve something expensive that companies already have budgets to pay for. They show real retention numbers. Real unit economics. Real customer demand.

Ask an investor today what they care about, and they’ll tell you it’s not your story. It’s your burn rate.

That sounds counterintuitive until you realize what’s happened. In 2025, pitch decks became spray and pray. Every founder could slap “AI-powered” on a problem and raise capital. The narrative won. The hype won. The result? A graveyard of startups that couldn’t explain why their gross margin was negative or why their CAC never paid back.

Now investors are done. They’re looking at burn rate, CAC, LTV, payback period, churn, and runway. If your unit economics don’t check out, your story doesn’t matter.

Here’s the thing that’s changed: in 2026, VCs openly tell founders to raise only what they need to hit the next milestone. Not a $20 million war chest because you’re “in a race.” Just enough to hit proof points. Each milestone steps up your valuation as execution is proven.

This is actually better for founders. It means you’re not expected to grow at any cost. It means the investor gets nervous if your burn rate is too high relative to your revenue. It means you can’t hide bad unit economics under the cover of “we’re in growth mode.”

The mental model here is Incentive Alignment. When investors only write checks for what you can prove you’ll spend well, you’re suddenly aligned. You’re both betting on discipline, not growth at all costs.

Here’s my contrarian take: the founders getting funded in 2026 aren’t the ones with the best pitch. They’re the ones with the best spreadsheets.

This is actually easier than the hype era. You don’t need to convince someone that your technology is the future. You need to convince someone that your business won’t run out of money before you hit the next inflection point. Those are radically different problems.

At /mkt, we don’t pitch on the tokenization thesis or the future of athlete equity. We pitch on unit economics, regulatory defensibility, and trading volume. We show the burn math. We show the moat. We show the path to profitability. That’s what wins capital in 2026.

The founders who are still trying to win with narrative are already behind. The ones moving fast are the ones who understand: in 2026, your burn rate is your brand.

Build something that makes sense on a spreadsheet first. The story will follow.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

Eclipse Ventures just disclosed a $1.3 billion raise. $720 million for early-stage, $591 million for growth. Their focus? Not large language models. Not AI apps. Not SaaS wrappers around APIs. They’re backing physical industries: robotics, semiconductors, defense tech, manufacturing, and AI infrastructure.

Cerebras IPO’d two months ago. Market cap: $95 billion on day one. They make AI chips. Not software. Chips.

Figure just received orders for 70,000 humanoid robots. They’re production-constrained, not demand-constrained, and their revenue pipeline through 2029 is $14 billion+.

The pattern isn’t subtle. Capital is flowing from apps to atoms. From software to silicon. From SaaS to supply chains.

This isn’t a niche trend. In Q1 2026 alone, venture investors deployed $297 billion globally. Nearly half went into infrastructure, hardware, and manufacturing plays. The other half was split between frontier labs and everything else.

Here’s what happened: when AI made software cheap and easy to build, it destroyed software margins. Every idea can now be implemented in two weeks. That means competition is instant. Defensibility is near zero. Margin compression is real.

But you can’t build a frontier chip fab in two weeks. You can’t manufacture a humanoid robot in a sprint. You can’t scale power infrastructure with a prompt. That’s why capital is moving toward the things that are hard to build and hard to copy.

The mental model is Picking and Shoveling. When the California Gold Rush hit, the smart money didn’t pan for gold. They sold picks and shovels. Today’s picks and shovels are chips, robots, power systems, and the infrastructure that AI companies need to exist.

/mkt is a real-world test of this principle. We’re building in a regulated market (Reg A+) with infrastructure that’s difficult to replicate (tZERO for trading). We’re not trying to out-engineer a LLM. We’re building defensible infrastructure in a space where regulation is actually a moat, not a headwind.

Here’s the contrarian part that most founders won’t admit: if you’re building a pure-play AI software company right now, you’re already late. The capital has moved on. Your pitch will compete against a thousand other pitches. Your margins will be worse. Your funding will be harder.

But if you’re building infrastructure, manufacturing, or physical products that leverage AI as a tool (not your product), you’re swimming downstream with the capital.

The venture market is realigning. The question is whether you’re building the pan or the pick.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

Let that sink in. A company valued at $75 billion with recurring Starlink revenue still can’t write the check for AI infrastructure from equity alone. They’re borrowing money they’ll have to repay.

The numbers tell you everything. That $20 billion is financing xAI’s compute cluster, power infrastructure, and data center hardware. This isn’t a feature release or a new product line. It’s the basic cost of playing the game in frontier AI right now.

Here’s the uncomfortable truth: the AI race has shifted from a margin game to a capital structure game. You can’t win by being smarter anymore. You win by having enough capital to afford the electricity, the cooling, the silicon, and the space to run models at scale. SpaceX can afford debt. Most startups can’t.

This is where I’d normally say “but it’s not all bad for startups.” Except it kind of is. Here’s the pattern I’m seeing:

Frontier AI labs (OpenAI, Anthropic, xAI, etc.) are now capital-raising at scales that only sovereign wealth funds can support. They’re not getting Series A and B rounds. They’re getting $30 billion, $50 billion, $122 billion checks from Singapore, Saudi Arabia, and Abu Dhabi.

Meanwhile, traditional VC firms are shifting capital down-market to picks-and-shovels plays. Databricks, CoreWeave, and the compute/data infrastructure layer are eating disproportionate funding. Why? Because if the frontier labs own the compute, the picks-and-shovels players own the picks.

The mental model here is Asymmetric Risk. When one player can absorb risk that competitors can’t, the market structure changes. SpaceX raising $20 billion in bonds doesn’t mean the AI race is expensive. It means the financing game has become asymmetric. Frontier labs backed by sovereign wealth funds can borrow against future revenue. Startups can’t.

Here’s my contrarian take: The SpaceX bond sale is actually a warning shot for founders. If you’re building an AI company that requires massive compute infrastructure, you’re competing in a market where the barriers aren’t innovation or talent anymore. They’re balance sheet and capital access. That’s not a venture game. That’s a capital markets game.

For most startups, the lesson is brutal. You’re not raising money to build something. You’re raising money to survive long enough to either (a) get acquired by a larger player, or (b) find a use case that doesn’t require frontier compute at all.

The frontier labs won. The question now is what layer of the stack is still available for new players to own.

If you’re building something, make sure you’re not competing on the thing that costs the most to scale.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

A founder with sloppy compliance gets a discount. A founder with compliance baked into product design gets a premium. The difference isn’t 10%. It’s often 20% to 40% on funding terms, and that’s before you count the enterprise customers who won’t buy from you without it.

The “compliance premium” is real. At seed stage, founders now face 15 to 20% higher legal expenses just to get the infrastructure in place. That’s not a bug. That’s the market’s way of pricing in regulatory risk. But here’s what makes 2026 different from 2020: it’s not just about legal defense anymore. It’s about competitive advantage.

Let me show you why.

The procurement revolution is happening right now.

Enterprise security teams and procurement departments are asking AI startups three questions before they even look at product features: How was your model trained? What data goes into your prompts? Who checks the outputs, and how can we challenge bad results?

Those questions used to come from compliance lawyers. Now they come from CISOs, CFOs, and procurement directors who are already spending $500K to $2M on AI vendors. They’re not buying demos. They’re buying defensibility.

The EU AI Act is a formal, risk-based regime that ties obligations to use case, model category, and role in the system. The US is a patchwork of state laws, federal guidance, and agency enforcement. The UK is principle-led, which sounds looser until you realize it means you’re responsible for interpreting the principles yourself. China maintains state control.

For most founders, this sounds like chaos. In reality, it’s clarity. The chaos was 2020 to 2023, when you could build anything and apologize later. That era is dead.

Now here’s the part that matters for your startup: the founder who understands these regimes doesn’t see them as constraints. They see them as a moat.

Why? Because 90% of competing startups will ignore compliance until fundraising breaks down or a customer refuses to sign. At that point, you’re scrambling to retrofit governance into a product that wasn’t designed for it. You’re 6 months behind. Your valuation discount is permanent.

But the 10% of founders who treat compliance as a product constraint from day one end up with:

First, a cleaner product. When you design for governance, data handling, and auditability from the start, you end up with fewer exploits, clearer workflows, and easier customer onboarding. You’re not adding compliance to something fragile. You’re building something resilient.

Second, a competitive edge in sales. Your customer procurement process becomes your strongest sales tool. While competitors are hiring lawyers to defend their infrastructure, you’re showing audits, certifications, and transparent decision logs. You close deals faster and on better terms.

Third, a stronger fundraising position. When a VC asks about compliance risk in 2026, the founder who says “we have ISO 42001 certification and we’ve mapped our training data and we’ve designed for audit” doesn’t sound cautious. They sound smart. The founder who says “we’re thinking about that” sounds naive.

Here’s how I think about it, and how we build at /mkt.

We’re operating in a regulated market (Reg A+ offerings with tZERO trading infrastructure). Our competitors might say that’s a constraint. We treat it as an asymmetric advantage. Because we had to build compliance into the core from the beginning, we ended up with a cleaner infrastructure, better data governance, and a product that enterprise partners actually trust.

The result? Regulators don’t see us as a risk. They see us as a partner. Customers trust us faster. And when a competitor tries to copy our product, they can’t copy the compliance scaffolding. That’s the moat.

The mental model here is “First-Principles Thinking.”

Most founders start with a product idea and retrofit compliance later. First-principles thinking means starting with “what’s the regulatory reality?” and designing product around it. It’s harder upfront. It’s slower at the beginning. But it’s the only way to build something that survives contact with enterprise customers, regulators, and serious capital.

My contrarian take: compliance-first founders aren’t being cautious. They’re being aggressive.

They’re building a moat that’s expensive for competitors to copy. They’re capturing customer trust at a time when trust is the scarcest resource. They’re moving into markets where compliance is the barrier to entry, not the barrier to profit.

The founders complaining about “regulatory burden” in 2026 are losing. The founders building regulatory resilience into product are winning.

If you’re raising capital for an AI startup right now, the single best question to get asked in a board meeting isn’t “how fast can you grow?” It’s “how did you design for compliance?” If you can answer that clearly, you’ve just signaled something most founders can’t: that you think like a builder, not a hustler.

The era of move-fast-and-break-things in regulated markets is over. The era of move-thoughtfully-and-build-trust has started. The founders who recognize that shift first are the ones building the defensible companies.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

1. Odyssey Just Raised $310M to Teach AI How the Physical World Actually Works

Odyssey, a world model AI startup founded by self-driving vehicle veterans Oliver Cameron and Jeff Hawke, raised a $310 million Series B at a $1.45 billion valuation, led by Natural Capital with Amazon, AMD Ventures, GV, and others participating. Instead of predicting the next word in a sentence, the company’s models learn to predict how the physical world behaves, including how objects move and what happens after something is touched or pushed. Odyssey’s CEO says the field is approaching the “GPT-3 moment” for world models, the point where the technology shifts from promising research into a breakthrough foundational layer.

Spence’s take: Everyone’s been racing to make AI talk better. This round is a bet that the next unlock is making AI understand physics better. Different race, same stakes.

2. SiFive’s $400M Says the Chip Wars Aren’t Just an Nvidia Story

Semiconductor startup SiFive raised a $400 million Series G round led by Atreides Management, challenging incumbent Arm Holdings with chip designs built on an open rather than proprietary standard.

Spence’s take: Open standards versus proprietary lock-in is one of the oldest fights in tech. SiFive just got $400 million more ammunition to fight it in chips.

3. Capital Is Concentrating Around Problems Nobody Can Commoditize

This week’s funding tape shows investors concentrating capital around a narrow set of hard-to-commoditize problems: sovereign cyber defense, enterprise workflow automation with measurable ROI, custom silicon design, and infrastructure sitting close to the AI compute bottleneck. Deals like Dream’s $260 million round, Gradial’s $65 million Series C, and Architect Labs’ $24 million seed all point to the same investor instinct: back teams attacking expensive, mission-critical systems where incumbents are slow to adapt.

Spence’s take: When investors stop chasing novelty and start chasing bottlenecks, that’s usually a sign the easy money phase of a cycle is ending.

4. Odyssey’s Amazon Deal Is the Real Story Buried in the Headline

Alongside the funding, AWS became Odyssey’s preferred cloud provider, with the startup committing to optimize its models for Amazon’s Trainium AI chips, positioning them as an alternative to Nvidia’s hardware.

Spence’s take: Funding rounds get headlines. Cloud and chip commitments get ignored. This one’s worth watching, because it’s Amazon picking a side in the AI infrastructure fight before most people noticed there was a fight.

5. The Smartest Money in the Room Showed Up as Individual Investors, Not Funds

Odyssey’s investor list includes existing backers like Jeff Dean, Google’s chief scientist, Elad Gil, Qasar Younis, co-founder and CEO of Applied Intuition, Garry Tan, president and CEO of Y Combinator, Guillermo Rauch, founder and CEO of Vercel, and Kyle Vogt, founder of Cruise.

Spence’s take: When the people building competing AI companies personally write checks into a rival, that’s not just a vote of confidence. That’s insiders hedging their own thesis.

Why Odyssey Is the One to Watch

I want to run this through second-order thinking, a model I cover in my book, Mental Models: How to Think, Act, and Win. The first-order story is “AI startup raises $310 million.” The second-order story is what happens if world models actually work: robotics, gaming, autonomous systems, and scientific simulation all get access to AI that understands cause and effect in the physical world, not just language patterns. That’s a bigger shift than a better chatbot. It’s AI gaining a working model of reality itself, and the company that gets there first doesn’t just win a product category. It potentially becomes the infrastructure layer everyone else builds on top of.

——————————————————————————————————————

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, right now.

Get it Here

——————————————————————————————————————

Next week’s deep-dive:

——————————————————————————————————————

1. Bezos Bets $12B That AI Can Replace the Slowest Part of Engineering

Prometheus raised $12 billion in a Series B from JPMorgan Chase, Goldman Sachs, BlackRock, and Bezos himself, pushing total funding past $18 billion after launching just seven months ago. That puts the company’s valuation at roughly $41 billion, placing it among the top five most valuable AI startups on the planet. The 150-person company is building an “artificial general engineer” designed to compress the design-to-manufacturing cycle for physical products from years down to months.

Spence’s take: Seven months, $18 billion, zero disclosed revenue. This is either the smartest bet in physical AI or the most expensive science project in history. There’s no third option.

2. Robinhood’s Prediction Markets Playbook Just Got a $1.4B Validation in Germany

NEURA Robotics, a German physical AI robotics company, raised up to $1.4 billion in Series C funding led by Tether, alongside Qualcomm, Amazon, and NVIDIA, marking the largest round ever for a full-stack robotics company.

Spence’s take: When crypto-native capital and chip giants write the same check, that’s not hype. That’s two completely different risk appetites agreeing on the same thesis.

3. The Gaming Industry Just Bought a Piece of Its Own Plumbing

Interchecks, an instant payments platform serving sportsbooks and fintechs, raised $50 million in a Series C led by Bettor Capital, a growth fund focused entirely on real-money online gaming, with Commerce Ventures, Decades Holdings, and Thayer Street Partners also joining. The company has processed more than $50 billion in transactions over a decade, with triple-digit revenue growth for seven straight years and profitability since 2023.

Spence’s take: A profitable, decade-old company raising venture money on its own terms is rarer than the unicorn headlines. Pay attention when boring and profitable decide to raise anyway.

4. Voice AI Is Crowded, But Investors Still See Room for a Winner

Bland closed a $50 million Series C led by Dell Technologies Capital, with HubSpot Ventures and others joining. The company says it handles more than 3.5 million calls per week and has processed over 175 million AI phone calls total.

Spence’s take: Crowded markets don’t scare good investors. They just raise the bar for who gets funded next.

5. Wall Street’s Biggest Names Just Quietly Bet on Tokenized Assets

Receipts Depositary Corporation announced a $7 million oversubscribed round led by LiveOak Ventures, with Hivemind Capital, OTC Markets Group, and GTS also participating, to issue depositary receipts on digital and alternative assets.

Spence’s take: Small check, big signal. Traditional market infrastructure players are building the plumbing for tokenized assets before retail even notices. This is the kind of deal that looks irrelevant until it isn’t.

Why Prometheus Is the One to Watch

I want to run the Prometheus deal through second-order thinking, one of the models in my book. The first-order story is “AI startup raises a lot of money.” The second-order story is what happens if this actually works: entire categories of physical engineering jobs get compressed into a software workflow, and the bottleneck for inventing new jet engines, drug compounds, or hardware shifts from human design cycles to compute and data access. That’s not an incremental shift. That’s a complete rewrite of who controls the pace of physical innovation, and right now Bezos and a handful of banks are betting $41 billion that they’ll own that rewrite.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is out now.

Get it Here

Next week’s deep-dive:

Here’s the number that should stop you: Interchecks has processed more than $50 billion in transactions over its 10-year history, achieved triple-digit net revenue growth year-over-year for the past seven years, and operated profitably since 2023. That’s not a typical venture story. Most companies raising Series C rounds are burning cash to chase growth. This one’s been profitable for three years and just decided to raise anyway. EnkiAI

So what does Interchecks actually do? It runs a payment infrastructure platform that connects banks, fintechs, and sportsbooks to various instant deposit and payout networks via a single API, routing real-time transactions across multiple payment rails while handling embedded compliance, fraud monitoring, and risk management. In plain English: when you deposit money into a sportsbook app or pull your winnings out, there’s a decent chance Interchecks is the invisible plumbing making that happen instantly instead of in three business days. VC News Daily

Alongside the raise, the company launched Account Funding Transactions, letting businesses fund accounts using debit credentials in real time, with built-in fraud protections covering account verification, duplicate card detection, and customizable velocity limits. That product now sits next to their existing Pay-by-Bank option on a single RESTful API, so clients can offer multiple deposit methods without juggling separate integrations. EnkiAIEnkiAI

I want to apply moat through compliance, a mental model I dig into in the book, because this deal is a perfect case study. Most fintech founders treat compliance as a tax they have to pay before they get to build the fun stuff. Interchecks treated it as the actual product. The company is a certified PCI Level 1 Service Provider and SOC 2 Type 2 certified, and works with major global networks including Visa and Mastercard. Those aren’t marketing badges. They’re the reason sportsbooks, banks, and regulated fintechs trust this company to touch their money flow at all. EnkiAI

I spent years at Robinhood building inside one of the most heavily regulated corners of fintech, and now I’m doing the same thing at /mkt with Reg A+ offerings running through tZERO’s infrastructure. The lesson repeats everywhere: in regulated markets, the company that wins isn’t the one with the slickest app. It’s the one that made compliance so airtight that everyone else has to build on top of them instead of around them.

Here’s the contrarian read. Bettor Capital, a growth equity firm focused entirely on the real-money online gaming market, led this round. That’s not a generalist fund dabbling in fintech. That’s the gaming industry’s own capital deciding that payment rails, not better odds or flashier apps, are the next place real value gets created in sports betting. Everyone’s watching the front end of gaming apps. The smart money just quietly bought a piece of the plumbing underneath all of them. EnkiAI

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

The round was led by Evolution Equity Partners, with participation from Accel, AT&T Ventures, Blackstone, Coatue, Spark Capital, Cyberstarts, and Temasek. Total funding now exceeds $2 billion, making it one of the most valuable privately held cybersecurity firms in the world. Yahoo!

Here’s the number that actually explains why investors are throwing this much money at a data security company: 68% of organizations cannot tell the difference between human activity and AI agent activity inside their own systems. Read that again. Most companies racing to deploy AI agents have no idea what those agents can see, touch, or do once they’re turned loose on internal data. FinSMEs

The AI industry built out chips, models, and compute. What’s still missing is the layer that governs what AI can actually see and do, and how to handle the risk that creates. Cyera is betting the entire $12 billion valuation on being that layer. FinSMEs

The growth numbers back up the bet. Annual recurring revenue has tripled for three consecutive years, and the company has expanded to more than 1,500 employees across 18 countries in the past 18 months, completing five acquisitions along the way. Customers reportedly include Paramount, Chipotle, and Valvoline, using Cyera to control exactly what data their AI systems can reach. PR Newswire

I want to apply constraint identification here, one of the models in my book. Every founder eventually has to answer one question correctly: what’s actually blocking growth in this market? Most AI infrastructure plays have chased compute, chips, or model performance. Cyera’s bet is that none of that matters if a company doesn’t trust its own data enough to let AI near it. They identified the real constraint wasn’t capability. It was permission.

That’s a sharper read than it sounds. A lot of founders solve impressive technical problems nobody was blocked on. Cyera found the boring, unglamorous gate that every single enterprise has to walk through before AI adoption becomes real, and they built the gatekeeper.

Here’s my contrarian take: a $12 billion valuation for a company at this stage isn’t really a bet on Cyera’s product. It’s a bet that data security posture management becomes one of the most valuable layers of the entire AI economy, and right now Cyera is simply the company furthest along in claiming that territory. That’s a real opportunity, but it’s also a crowded one. The company competes with established security vendors and startups spanning cloud security, data discovery, access governance, and AI protection, and a premium valuation increases the pressure to expand revenue and prove the products work together as one system, not just a pile of acquisitions. MannpublicationsMannpublications

I’ve spent my career building inside regulated markets, first at Robinhood, now at /mkt, where every product decision runs through a layer of compliance and trust before it ever reaches a user. The companies that win in regulated spaces aren’t the ones with the flashiest tech. They’re the ones customers trust enough to hand over the keys. Cyera is making the same wager at enterprise scale: trust is the actual product, and everything else is just the delivery mechanism.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

Interchecks doesn’t fit any of that. They’re a New York-based instant payments infrastructure company that was founded in 2016, has about 50 employees, processed more than $50 billion in transactions over ten years, achieved triple-digit net revenue growth every single year for seven straight years, and has been profitable since 2023. On June 15, they closed a $50 million Series C led by Bettor Capital, Commerce Ventures, Decades Holdings, and Thayer Street Partners.

No billion-dollar valuation announcement. No charismatic founder on CNBC. Just a company that quietly built the plumbing that makes money move inside some of the most compliance-intensive industries in fintech — sportsbooks, earned wage access platforms, neobanks, digital asset wallets.

Their core product routes instant deposits and payouts across multiple payment rails — Pay-by-Bank, Push-to-Card, Visa Direct, Mastercard — through a single RESTful API, with fraud monitoring, account verification, and 1099 tax compliance built in. Their new Account Funding Transactions product, launched alongside the round, adds real-time debit card funding with layered risk controls: duplicate card detection, velocity limits, and proactive monitoring of suspicious activity before funds move.

Fifty employees. Fifty billion dollars processed. Think about that ratio for a second.

Mental Model: The Quiet Compounder

There’s a pattern in infrastructure businesses that rarely gets the coverage it deserves. I call it the Quiet Compounder.

It works like this: you find a workflow that’s painful, regulated, and handled differently by every company that touches it. You build a compliant, reliable solution that abstracts the pain away. You charge a transaction fee or a SaaS margin. And then you grow — not because you’re pitching a vision, but because every new customer is adding to a base of revenue that compounds quietly, year after year, with low churn and high switching costs.

Interchecks is a textbook example. Sportsbooks, fintechs, and financial institutions don’t rip out their payment infrastructure once it’s embedded. The cost of switching is high. The cost of a failed transaction in a regulated environment is higher. And every year Interchecks stays compliant, processes volume, and adds rail options, the gap between them and a new entrant trying to replicate the stack grows wider.

Triple-digit revenue growth for seven consecutive years is not luck. That’s a flywheel that’s been running since 2019, hidden in plain sight because the company never needed a flashy round to fuel it.

The Contrarian Take

The $50M Series C is interesting for a specific reason that isn’t in the press release.

Interchecks has been profitable since 2023. They didn’t need this capital to survive or to grow. They raised it. That’s a different kind of fundraise — one where the company has leverage in the negotiation and is choosing to accelerate from a position of strength, not necessity.

Compare that to the raise-to-survive rounds that dominated 2022 and 2023. The difference in founder leverage, valuation discipline, and investor dynamics is significant.

The businesses that didn’t need a round but chose to raise one are often the most interesting companies in a given funding cycle. They’ve already answered the hardest question — can this work without venture capital — and the answer was yes. The new capital is acceleration, not oxygen.

Seven years of triple-digit growth. Profitable. 50 employees. One clean API.

That’s the startup story that gets ignored because it doesn’t fit on a slide deck cover. It’s also the story that tends to look obvious in retrospect.

Podium Automation makes industrial control panels. Not software. Not AI agents. Actual physical boxes full of wires, switches, and circuit breakers that tell factory equipment when to start, stop, and how to communicate safely with everything around it.

If that sounds unglamorous, that’s the point.

The company just closed an $18 million Se…

Read more

A stellarator is a type of fusion reactor that uses twisted magnetic fields to contain superheated plasma long enough to generate power. It’s one of the oldest fusion concepts in physics, and it’s historically been considered too mechanically complex to build at scale, because traditional stellarator designs require highly warped, three-dimensional magnetic coils that are extraordinarily difficult to manufacture.

Thea just raised $100 million in an oversubscribed Series B, led by Thomas Tull’s US Innovative Technology Fund, with General Innovation Capital Partners, Linse Capital, Calm Ventures, Climate Capital, Divergent Capital, Emerald Technology Ventures, Gaingels, Idemitsu Kosan, Overlay Capital, Timescale Ventures, and Whatif Ventures all participating. That brings total private investment in the company to roughly $130 million, following a $20 million Series A closed in early 2024.

The company spun out of Princeton University and the Princeton Plasma Physics Laboratory in 2022, originally under the name Princeton Stellarators before rebranding. Co-founder and CEO Brian Berzin brings a background in electrical engineering and private equity. Co-founder and CTO David Gates previously led stellarator magnet array research at PPPL, meaning the technical leadership here isn’t a software team that pivoted into energy. It’s a team that spent years inside the exact research institution that’s been working on this problem for decades.

That pedigree matters in fusion more than almost any other sector. This is a field littered with ambitious claims and very few working prototypes. A founding team with direct lineage to the national lab doing the underlying physics research is a meaningfully different signal than a founder with a deck and a vision.

The Market

Fusion has spent the last several years going from a punchline to a serious line item on institutional balance sheets. Cumulative investment in the fusion sector surpassed $15 billion by January 2026, reflecting strong investor confidence in the technology’s potential to address growing demand for clean, firm power, a need amplified by the energy requirements of AI and data centers.

That last point is the real story. This isn’t climate-driven capital chasing a feel-good thesis. Surging electricity demand is spurring interest in fusion technology, much of it driven directly by the power needs of AI infrastructure. Data centers need baseload power that doesn’t fluctuate with weather the way solar or wind does. Fusion, if it works, is exactly that: constant, clean, scalable power.

The competitive field is heating up fast. In June 2026 alone, Helion raised $465 million for its Field-Reversed Configuration approach, while Focused Energy secured a $240 million Series A for laser-based inertial fusion, showing investor appetite spreading across multiple fusion architectures rather than concentrating in a single approach the way it did in the 2021 to 2024 period.

What makes Thea’s timing specifically interesting is the technical shift underlying its raise. The core shift in the stellarator field is the move from physically complex hardware to software-defined magnetic control, which fundamentally changes the technology’s readiness level for commercial manufacturing. In plain terms, the bottleneck that made stellarators impractical for decades was a manufacturing problem, not a physics problem. If Thea has genuinely solved the manufacturing side, they’re not just one more entrant in a hot sector. They’re unlocking an architecture that was sitting on the shelf for a reason that had nothing to do with whether it worked.

Business Model and Moat

Thea doesn’t have a product on the market yet, and won’t for years. This is deep tech in the truest sense: the business model right now is building toward a demonstration system, not generating revenue from customers.

The new funding will help Thea expand manufacturing for its smaller magnets and begin construction of its Eos demonstration device, with construction starting next year. The company has utilized arrays of mass-manufacturable magnets paired with dynamic software controls to reinvent the stellarator design.

The defensibility argument here is unusual compared to a typical startup moat. It’s not a network effect or a brand. It’s a combination of deep technical IP around magnet array manufacturing, a direct talent pipeline from the lab where stellarator research originated, and government backing that’s hard for a new entrant to replicate quickly. Thea was selected as an inaugural awardee of the Department of Energy’s Milestone-Based Fusion Development Program and is supported by six separate Department of Energy INFUSE awards. That’s not capital, but it is validation, and in a field this technically demanding, validation from the agency that runs the national fusion research program is a moat in its own right.

Competitively, Thea is one of several stellarator-focused companies globally, including Germany’s Proxima Fusion, which spun out of the Max Planck Institute for Plasma Physics. But within the stellarator approach specifically, Thea’s manufacturing-first thesis, built around mass-producible magnets rather than the historically complex twisted coil designs, appears to be a genuine differentiation rather than a marketing angle.

Spencer’s Take

Two mental models apply directly here.

The first is constraint identification. For decades, people assumed the stellarator’s problem was a physics problem. Thea’s founders, sitting inside the actual lab doing this research, correctly identified that the real constraint was manufacturability, not the underlying science. That’s a subtle but critical distinction. Founders who misdiagnose the actual bottleneck waste years solving the wrong problem. Founders who correctly identify the constraint can unlock value that’s been sitting untouched for decades, simply because nobody else framed the problem correctly.

The second is the power of credentialed conviction. In binary, high-uncertainty fields like fusion, investor conviction often comes down to who’s in the room more than what’s in the deck. A CTO who spent years running the exact research program this company is now commercializing is a different kind of signal than industry experience alone. It’s worth noting how often deep tech rounds get led by funds and individuals, like Thomas Tull here, who are making a long-horizon bet on a specific technical thesis rather than chasing near-term returns.

The bull case: this is a genuinely massive market opportunity, riding a real demand tailwind from AI infrastructure power needs, backed by a technical team with direct institutional lineage to the underlying science, and validated by serious government funding programs. If the manufacturing thesis holds, Thea could be sitting on an architecture advantage that’s extremely difficult for competitors to replicate quickly.

The bull case has a serious counterweight, though. Fusion has a multi-decade history of well-funded, well-credentialed teams missing timelines by years, sometimes by a decade or more. The Eos demonstration device hasn’t been built yet. Every fusion company at this stage says the breakthrough is close. Most of them are right about the science and wrong about the calendar.

Why It Matters

For investors and operators evaluating this space: the signal to track isn’t the funding round size, it’s the Eos demonstration device timeline next year. Deep tech rounds are easy to raise on a compelling thesis. The actual test is whether construction milestones hit on schedule. If Eos breaks ground and shows progress on schedule, that’s a far stronger signal than any amount of additional capital raised in the meantime.

For potential customers, meaning utilities, data center operators, and large industrial energy buyers: there’s nothing to evaluate or purchase yet, and that’s an important distinction from the Podium Automation story in this same publication. This is a multi-year technology bet, not a near-term vendor decision. The relevant action for energy buyers right now is awareness, not procurement. Worth tracking which utilities or hyperscalers eventually sign early offtake agreements with fusion companies, since that will be the real tell for which architecture the market believes in first.

For competitors and builders thinking about adjacent spaces: the broader lesson is about where deep tech founders should look for unlocked value. Thea’s advantage didn’t come from a flashy new discovery. It came from correctly identifying that an old, well-understood physics concept had been blocked by a manufacturing constraint that modern software and materials science could finally solve. There are other corners of physics and engineering sitting in the same state, technically sound but commercially stalled because of a manufacturing or tooling gap nobody’s revisited with current technology. That’s a more repeatable playbook than waiting for a brand-new scientific breakthrough.

The Bottom Line

Thea Energy is making a credible, well-capitalized bet that the stellarator’s real obstacle was always manufacturing, not physics. The team’s direct lineage to Princeton’s fusion research, the scale of government validation, and the demand tailwind from AI power consumption all support the thesis. The real test starts next year, when the Eos demonstration device either breaks ground on schedule or joins a long list of fusion promises that slipped.

This newsletter does not provide investment advice or recommendations to buy or sell any security. The information presented is for educational and informational purposes only and reflects publicly available data at the time of writing. Private company valuations, fundraising terms, and projections are inherently uncertain, and this is especially true in early-stage deep tech sectors like fusion energy, where commercialization timelines carry significant risk. Always conduct your own due diligence and consult a licensed financial advisor before making investment decisions.

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

The company is Prometheus, the industrial-AI startup co-led by Jeff Bezos and Vik Bajaj, former co-founder of Alphabet’s Verily. It announced a $12 billion Series B at about a $41 billion valuation, backed by JPMorgan, Goldman Sachs, BlackRock, DST Global, Arch Venture Partners, and Bezos himself. It launched in November 2025 with $6.2 billion, so total funding sits north of $18 billion in roughly seven months. The team is spread across San Francisco, London, and Zurich. It’s Bezos’s first operating role since he left the Amazon CEO seat in 2021.

What it’s building isn’t a chatbot. It’s what the founders call an “artificial general engineer,” AI meant to design and help manufacture complex physical products, from jet engines to drug compounds. Bezos’s framing to Axios: ask a jet-engine maker for the same engine with 10% more thrust and you might be staring at a ten-year program, not because anyone’s lazy, but because the physical world is brutally hard.

The model: Second-Order Thinking

First-order thinking stops at the obvious, immediate consequence. Second-order thinking asks the harder question: and then what? What do the consequences cause?

First-order on this news is easy and loud. $41 billion for 150 people and no revenue? Bubble. AI hype. The emperor has no thrust. But run it second-order. The bet isn’t on what Prometheus sells today, which is close to nothing. It’s on one thesis: the slowest, most expensive step in the entire physical economy is the engineering design loop, and software is about to bolt a flywheel onto it. And then what? Compress that loop and small teams attempt projects that used to need thousands of engineers and a decade. Whoever owns that tool sits upstream of trillions in industrial output. The banks and Bezos aren’t pricing a product. They’re pricing a position.

There’s a second-order tell most coverage skipped. Bezos has reportedly explored a roughly $100 billion fund to buy manufacturers outright and run them on this AI. First-order, that’s a side rumor. Second-order, it might be the whole plan: own the tool and the factories it runs.

My take

Here’s where I’d push back on the bulls and the bears.

The bears’ “it’s a bubble” reflex misses the upstream position. But the bulls have a second-order crack that physics won’t forgive. Software iterates because you can ship, break, and patch in a day. Jet engines can’t. You can’t A/B test a drug compound or hot-fire your way past certification (Bezos knows the cost; Blue Origin just lost a New Glenn in a test).

So the question that decides whether $41 billion is genius or fantasy: does an artificial general engineer collapse the whole design loop, or just the cheap front half, while the expensive, regulated back half (testing, certification, tooling, supply chain) stays exactly as slow? Compress the front, leave the back, and you’ve sped up the part that was never the real bottleneck.

That’s the bet under the bet, and it’s why I’d watch the regulated back half closest. I see the same truth at /mkt, building in a heavily regulated market: software speeds the front of the loop, but the durable moat is mastering the slow, regulated part nobody gets to skip.

If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, my book is coming soon.

This newsletter is for informational and educational purposes only. It isn’t investment advice, nor an offer or solicitation to buy or sell any security. References to companies, including /mkt, are illustrative and not recommendations. Figures cited are drawn from public disclosures and reporting, and the reported manufacturing-fund plan is unconfirmed. Do your own research and consult a licensed professional before making any financial decision.