The AI Capital Stack: How AI Gets Funded
Nvidia just recruited Wall Street to help its customers borrow more than half a trillion dollars. The plan, announced with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aims to mobilize over 500 billion dollars in third-party capital so hyperscalers, labs, and startups can build data centers and buy chips. Jensen Huang described the shift plainly: AI compute is turning into an investable asset class, “AI factories” that get financed the way power plants and toll roads do. The same week, Cognition, the company behind the coding agent Devin, entered talks at a 40 billion dollar valuation, up from 26 billion three months earlier, on a revenue run rate nearing a billion.
Those headlines are the disposable part. The half-trillion number will be stale by the time you finish this sentence, and the valuation will have moved again by next quarter. What does not go stale is the structure underneath them: the way money moves through the AI economy, who funds whom, and where a small founder actually sits in that machine.
Most builders read these announcements as either a reason to panic or a reason to raise a bigger round. Both reactions miss the point. The financing arms race is not a scoreboard you are losing. It is a map. Read it correctly and it tells you exactly which layers to avoid, which dependencies to stress-test, and where a single dollar of your own capital buys the most durable thing. This is a piece about reading that map. Call it the AI capital stack.
What this covers
- Why founders misread the financing arms race
- The framework: the AI capital stack
- The reflux loop and the round-trip dollar
- The gravity rule: where value actually accrues
- The capital-light lane
- The round-trip test: trace your revenue
- Two balance sheets, same ten million
- How much should you actually raise
- The contrarian take
- What to do Monday morning
- FAQ
Why founders misread the financing arms race
When a founder reads that Nvidia and six of the largest asset managers on earth are assembling 500 billion dollars, the gut reaction is a mix of awe and inadequacy. The numbers feel like a different sport. And the quiet message the press coverage sends is that AI is a game of capital now, so if you are not raising nine figures you are already behind.
That is the wrong lesson, and acting on it is expensive. The founders who internalize it go raise more than they need, hire ahead of revenue, and sign multi-year compute commitments to look serious. They confuse the capital requirements of one layer of the stack with the capital requirements of their layer. A company that manufactures chips or builds data centers genuinely needs billions. A company that sells a workflow to a few thousand customers does not, and pretending otherwise burns the one advantage a small team has.
Here is the number that should reframe the whole thing. AI-native software companies are averaging about 1.13 million dollars of annual recurring revenue per employee, against roughly 283,000 for the median public software company. The standouts are not close to average. Midjourney runs around 200 million dollars in revenue with roughly 11 people, near 18 million per head. Lovable crossed 400 million in recurring revenue with 146 employees. Telegram reportedly runs a billion in revenue with about 30. These are not companies that borrowed half a trillion dollars. They are companies that understood which layer they were in and refused to buy the wrong assets.
The financing story and the capital-light story are the same story told from two ends of the stack. To see why, you have to see the stack itself.
The framework: the AI capital stack
Every AI product, from a solo founder’s side project to OpenAI, sits on top of the same five layers. Money is poured into the bottom. Revenue has to climb up from the top. The whole question of who wins and who gets wiped out comes down to how those two flows meet.
Read from the bottom up, the layers are physical, then get more abstract as you rise. Energy and land sit at the floor: power contracts, grid interconnects, and the real estate a data center occupies. This is slow, scarce, and quietly the hardest thing to acquire in 2026. Silicon sits above it: the chips, dominated by one company that holds around 80 percent of AI accelerators by revenue. Clouds and iron come next: the data centers and neoclouds that buy the chips, wrap them in cooling and networking, and rent the capacity out. Models sit above the iron: the labs that train frontier systems on rented or owned compute. And applications sit at the top, closest to an actual customer with an actual budget. That is where most founders reading this live.
Now watch the two flows. Investor capital pours into the bottom of the stack, because that is where the expensive physical assets are. Chips, buildings, power. Revenue, meanwhile, only enters the stack at the very top, when a human or a business pays for something an application does. That revenue is supposed to flow down the stack: the customer pays the app, the app pays the model, the model pays the cloud, the cloud pays the chip maker, the chip maker pays for silicon and power. In a healthy stack, the money climbing down from real customers is enough to service the capital that was poured into the base.
The entire anxiety of 2026 is a single question: is there enough revenue at the top to pay for the build at the bottom? Current AI-attributable revenue is estimated somewhere between 50 and 150 billion dollars a year, even when you generously credit all incremental cloud growth to AI. Hyperscalers alone plan to spend roughly 700 billion on capital projects this year, about three quarters of it on AI. That is a gap of several times over. The base of the stack has been financed on the belief that the top will grow into it. And when belief has to bridge a gap that revenue cannot yet cover, the layers do something that should make every founder pay attention. They start financing each other.
The reflux loop and the round-trip dollar
Here is the move that has drawn scrutiny from banks, regulators, and short sellers alike. When the top of the stack cannot yet produce enough revenue to pay for the bottom, the players near the top start putting capital back into the bottom to keep the whole thing funded. Nvidia invests in or guarantees financing for a lab. The lab signs enormous cloud contracts. The cloud provider borrows to buy more chips from Nvidia. The dollar that left Nvidia comes back to Nvidia, having done a lap of the stack and generated revenue on the books at every stop. I call this the reflux loop, and the dollar that makes the trip the round-trip dollar.
The scale is not hypothetical. By mid-2026 there was well over 800 billion dollars in circular arrangements across the sector, according to reporting on the sector’s money flows. Nvidia was weighing a financing guarantee reported near 250 billion dollars tied to a lab leasing part of a planned data center. One lab locked in a 300 billion dollar cloud commitment with one provider, a 90 billion deal with a chip designer, and a 38 billion contract with another cloud. One large neocloud reported a widening net loss, more than 500 million dollars of interest expense in a period, and negative free cash flow measured in billions, against total liabilities above 50 billion. The cascade risk is real: if one large buyer’s AI revenue disappoints, it trims cloud spending, which hits chip revenue, which dents the neocloud’s valuation, which circles back to a lab’s ability to raise. Analysts keep reaching for three historical rhymes at once: telecom vendor financing in 2000, off-balance-sheet counterparty opacity in 2008, and physical overbuilding before a price crash in the mid-2010s.
I want to be precise about what this means for you, because it is easy to slide into doom, and doom is not useful. This is not the same claim I made in the piece on building on a subsidy. There the point was about your unit costs: the price you pay for tokens is being held down by someone else’s losses, so your margin is partly a loan. Here the point is one level up and structural. It is not just that the input is cheap because it is subsidized. It is that the revenue holding up your suppliers, and sometimes the revenue on your own dashboard, may be financed rather than earned. Cheap input is a cost question. The reflux loop is a solvency question. They rhyme, but they are different tests, and a careful founder runs both.
The reflux loop is not a reason to quit. It is a signal to read. It tells you that the deeper you build into the capital-heavy layers, the more your survival depends on money that is circling rather than money that a customer chose to spend. And it points to a rule about where value actually ends up.
The gravity rule: where value actually accrues
There is a tidy story people tell about technology stacks: value flows to the top, to the application closest to the user, because that is where the experience lives. In AI so far, that story has been almost exactly backward. Value has been gravitating down the stack, toward the layer that controls the scarcest input.
The numbers are stark. Semiconductors capture roughly 79 percent of all gross profit dollars in the AI stack. The dominant chip maker posted data-center revenue north of 75 billion dollars in a single recent quarter, growing more than 90 percent year over year, at gross margins around 75 percent, on something like 80 percent share. That is not the economics of a competitive commodity market. That is the economics of a control point. Meanwhile the application layer, the asset-light top of the stack where most software founders build, runs at average gross margins near 33 percent and earns a fraction of the profit dollars. The model labs sit in between and are climbing fast, with at least one lab reportedly moving inference gross margins from the high 30s to over 70 percent as scale kicked in.
Call this the gravity rule: value accrues to the layer with the tightest constraint, not to the layer that does the most visible work. Chips are hard to make and one company is years ahead, so profit pools there. Applications are easy to spin up, thousands exist, and most are thin, so margin gets competed away. This is the same force I described in the commoditization clock, seen from the money side. If your layer has no constraint, price falls to the cost of the next competitor, and right now the next competitor is a weekend project built on the same model API you use.
A naive reader draws a naive conclusion here: value is at the chip layer, so I should go compete near the chips. That is a trap, and it is the trap the financing headlines set. You cannot out-capital the firms assembling half a trillion dollars. You will not build a better data center than a consortium of the world’s largest asset managers. The gravity rule tells you where the profit is; it does not tell you to go stand there and get crushed. What you need is the layer where a small amount of your capital buys a constraint that the giants cannot simply fund their way past. That layer exists, and it is not where you would guess.
The capital-light lane
Put two things on two axes. On the horizontal axis, how capital-intensive is what you are building: does it take a little money or a mountain of it. On the vertical axis, what kind of asset does that money buy: a depreciating one that loses value on a clock, or a compounding one that gains value the longer you hold it. Four quadrants fall out, and only one of them is a place a small founder should stand.
Start with the two quadrants you avoid. Asset-heavy, top right, is the infrastructure game. The margins are excellent and the assets can be durable, but the price of entry is measured in billions, and it is the game the 500 billion dollar consortium is built to win. You are not invited, and you should not want to be. Melting ice, bottom right, is worse than it looks. That is heavy capital buying a depreciating asset: borrowing to buy GPUs that lose most of their value in three to five years while the debt behind them runs for decades. This is where the duration mismatch bites, and it is why some of the biggest names in the buildout carry the most fragile balance sheets. A data center is a 25 to 40 year building, but the chips inside it are obsolete in a few years, and the financing does not care.
Now the two quadrants on the left, the capital-light half. Most AI apps land in the commodity trap, bottom left: cheap to build, but with no constraint of their own, so they are the thin wrappers I wrote about in the wrapper trap. Little capital in, but little defense, and margins get competed to the floor. The one quadrant worth standing in is the top left. The capital-light lane: a small amount of money that buys a compounding asset. Proprietary data you cause to exist. A workflow customers rebuild their day around. A distribution channel or a brand that a competitor cannot clone with a bigger model. None of these depreciate on a hardware clock. All of them get more valuable the longer you run.
The revenue-per-employee receipts are proof this lane is not a consolation prize. Look at the capital-light standouts again, and notice what they own. It is never a data center. It is a product experience, a community, a distribution habit, a proprietary dataset. This is the same argument as the incompressible core and what to build when building is free, now grounded in the capital map: the giants are pouring money into the layers that depreciate, which leaves the layer that compounds wide open for someone who does not need permission or a nine-figure round to enter it.
| Layer | Who funds it | Value capture | Founder exposure |
|---|---|---|---|
| Applications | Venture, revenue, your savings | Low on average (near 33 percent), high if you own a constraint | Your home. Win with data, workflow, distribution |
| Models | Mega-rounds, chip-maker guarantees | Rising fast (60 to 70 percent for leaders) | A dependency, not a place to compete head-on |
| Clouds and iron | Debt, off-balance-sheet vehicles | Thin, and strained by interest | A vendor whose solvency you should track |
| Silicon | Its own cash, plus the financing loop | The control point (near 79 percent of profit) | Sets your floor cost; you do not compete here |
| Energy and land | Project finance, sovereigns, utilities | Slow, scarce, increasingly the real bottleneck | Irrelevant to build, relevant as a systemic risk |
The round-trip test: trace your revenue
The capital stack is not only a map for choosing where to build. It is a diagnostic you can run on a business you already have. I call it the round-trip test, and it has two halves: trace your revenue up, and trace your dependencies down.
First, trace your revenue to its ultimate source. For every meaningful customer, ask one question: whose budget does this money come from. If it comes from a company selling a real product to real end users, a bank, a retailer, a manufacturer, a hospital, then your revenue is earned and it will survive a financing winter. If it comes from another AI company that is itself burning venture money and has no profits, then your revenue is a slice of a fundraise, and it evaporates the moment that company’s next round gets harder. Plenty of 2026 startups have beautiful growth charts made almost entirely of other startups’ seed money. That is a round-trip dollar wearing a customer’s clothes. It counts on the dashboard and disappears in a downturn. This is why I keep pushing founders toward revenue models anchored to an outcome a real business pays for, not to another startup’s runway.
Second, trace your dependencies down the stack to find where you are exposed. Your app runs on a model, the model runs on a cloud, the cloud is financed by debt. Somewhere in that chain sits a vendor whose survival depends on the reflux loop staying intact. If your product cannot function without a specific neocloud that is burning billions, or a specific lab whose economics only work if the next mega-round closes, then you have imported that fragility into your own business. This is the capital version of the dependency risk I covered in platform risk and vendor lock-in. The fix is not paranoia. It is portability: an architecture where you can move the model or the compute if a supplier wobbles, so a solvency problem two layers down does not become your outage.
Run both halves and you get an honest picture that no growth chart shows you: how much of what you have built rests on earned money and durable ground, and how much rests on the loop. You do not need the loop to fail for this to matter. You only need it to slow down, and every serious observer expects it to be tested this year.
Two balance sheets, same ten million
Make it concrete. Two founders each raise or set aside ten million dollars. One builds heavy, one builds light. Watch what each dollar becomes three years later.
| Question | The heavy build | The light build |
|---|---|---|
| Where the money goes | Owned GPUs, long compute commitments, big team | People, product, distribution, proprietary data |
| What it buys | A depreciating asset on a 3 to 5 year clock | A compounding asset with no clock |
| Value in year three | Hardware mostly obsolete, commitments still due | Data deeper, workflow stickier, brand stronger |
| What a downturn does | Fixed costs stay, revenue can fall out from under them | Costs flex down with usage; the asset survives |
| Who has to keep believing | Lenders, and the next round of buyers | Your customers, and only your customers |
The heavy build looks more impressive on day one and more fragile every day after. Its costs are fixed and its main asset is melting. The light build looks modest on day one and gets sturdier: its costs flex with usage, and the thing the money bought, the data, the workflow, the audience, is worth more in year three than in year one. The lesson from the pricing side reinforces this. As I argued in pricing under cheap inference, the input cost is collapsing, which means the durable value has to sit in what you own, not in the compute you rent. The capital map says the same thing from the balance sheet: own the compounding thing, rent the depreciating thing, and never borrow to buy ice.
How much should you actually raise
The capital map has a direct answer to the question every founder eventually asks: how big should my round be. The instinct, fed by the headlines, is bigger. The map says the opposite for almost everyone at the application layer.
Look at where venture money actually went. In the first half of 2026, United States venture funding hit about 412 billion dollars, but 81 percent of it went to rounds of 100 million or more. The top five labs alone absorbed roughly three quarters of all deal value. This is not a market spreading capital across thousands of promising teams. It is a market making a handful of enormous bets on the capital-heavy layers and running a slower, higher-bar game for everyone else. Seed dollars actually rose, but the number of seed deals fell about 30 percent, which means investors are writing larger checks to fewer companies and holding the rest to a stricter standard.
Read through the capital stack, that split makes sense. The money is chasing the layers that need money: chips, clouds, frontier models. Those are the businesses that genuinely cannot exist without billions, so billions flow to them. The application layer does not have that excuse, and increasingly investors know it. Raising a giant round to build a capital-light business does not make you look serious. It saddles you with a valuation you have to grow into, dilution you did not need, and a burn rate that turns your greatest structural advantage, low fixed costs, into a liability.
The capital-light move is to raise what buys the compounding asset and no more. Enough to hire the few people who deepen the data, sharpen the product, and build the distribution, then let revenue from real customers fund the rest. A team doing millions in revenue per employee does not need a war chest. It needs runway to keep compounding. The founders who win the capital-light lane treat a fundraise as fuel for an asset that is already appreciating, not as a substitute for one. When infrastructure is being financed by the largest institutions on earth, the scarce and valuable thing is a small team that turns a little money into a lot of durable value. Price your raise to protect that, not to imitate a layer you were never meant to compete in.
The contrarian take
Here is what most founders get exactly backward. They treat the financing arms race as a reason to feel behind, and they treat capital as the thing they most need more of. Both are wrong, and the capital stack shows why.
The half-trillion-dollar financing alliance is not a threat to a small founder. It is the clearest possible map of where not to compete. When the largest pools of capital on the planet organize to fund a layer, that layer is now a capital contest, and capital contests are won by whoever has the most capital, which is not you and never will be. The giants are effectively marking the squares of the board you should never step on. Every dollar they commit to chips, data centers, and power is a dollar making those layers more efficient and more commoditized for you to rent, while making them more crowded and more dangerous for you to build in. Their capital intensity is your operating advantage. You get to buy the output of a 500 billion dollar build as a metered service and point your own scarce money at the one thing their money cannot buy: a constraint at the application layer that belongs to you.
And the deeper reframe: in a world where compute is being turned into an investable, borrowable, commoditized asset, being capital-light stops being a weakness and becomes the whole advantage. When infrastructure is abundant and financed by strangers, scarcity moves to the things that cannot be financed into existence, taste, trust, a workflow customers love, a proprietary dataset, a distribution channel you built by hand. The best time to be capital-light is precisely when everyone else is going capital-heavy, because they are competing to own the abundant thing and leaving the scarce thing unattended.
I owe you the honest counterweight, because a contrarian take that only flatters the small founder is just a different flavor of hype. Capital-light is not the same as free, and it is not risk-free. Some businesses genuinely require heavy iron, and if you are building one of those, wishing you were light will not make it so. The application layer’s average margins really are thin, and the capital-light lane is narrow: you have to actually own a constraint, or you are just sitting in the commodity trap with a nicer story. And the reflux loop cuts both ways. Its abundance is what makes your cheap compute possible, so if the loop seizes, your costs can jump even as your capital-light structure protects your balance sheet. Light protects you from the debt. It does not make you immune to the weather. The point is not that capital does not matter. It is that a dollar of your capital, aimed at the compounding layer, is worth more than a dollar of theirs aimed at the melting one.
What to do Monday morning
This framework is only worth anything if it changes what you do this week. Here is the capital audit, concrete enough to run on your own business by Monday afternoon.
Draw your stack. Write down the five layers and mark exactly where your product sits. For almost everyone reading this, the honest answer is the application layer. Good. That is the layer with the open lane. Stop feeling like you should be lower in the stack than you are.
Name your quadrant. Are you capital-light and compounding, or capital-light and commodity. Be brutal. If a competitor could rebuild your core in a weekend on the same model API, you are in the commodity trap, and no amount of funding fixes that, only a real constraint does. Write down the one asset you own that gets more valuable over time. If you cannot name it, that is this quarter’s most important problem.
Run the round-trip test on your revenue. Take your top ten customers and trace each one’s budget to its source. Put a mark next to every dollar that ultimately comes from another venture-funded, unprofitable AI company. That marked number is your exposure to the loop. If it is most of your revenue, your growth chart is more fragile than it looks, and diversifying into customers with real budgets is more urgent than adding features.
Trace your dependencies down. List every vendor your product cannot run without, and check which of them are burning cash on borrowed money to stay alive. For each single point of failure, ask whether you could move within a week if they wobbled. Where the answer is no, buy yourself portability now, while it is cheap and calm, not during a scramble.
Aim your next dollar at the compounding layer. Before you spend on more compute, more headcount, or a bigger raise, ask the capital-map question: does this dollar buy something that depreciates or something that compounds. Point it at the data, the workflow, the distribution, the brand. Let the giants finance the ice. You are building the thing that gets more valuable while you sleep. That is the whole game, and it is one a small team can win. The rest of the operating system for building this way is in the AI-native founder playbook and the AI opportunity map.
Frequently asked questions
What is the AI capital stack?
The AI capital stack is the five layers every AI product sits on: energy and land at the base, then silicon (chips), then clouds and data centers, then model labs, then applications at the top. Investor capital pours into the bottom, where the expensive physical assets are, and revenue only enters at the top when a real customer pays for an application. Reading the stack tells a founder which layers are capital contests to avoid and which layer, the application layer, is where a small team can actually own something.
What is circular financing in AI, and should founders worry about it?
Circular financing is when players near the top of the stack fund the players below them to keep the build going, before end-customer revenue is large enough to pay for it. A chip maker finances a lab, the lab pays a cloud provider, the cloud provider borrows to buy chips from the chip maker, and the dollar completes a loop, booking revenue at every stop. By mid-2026 there were well over 800 billion dollars in such arrangements. Founders should not panic, but they should run the round-trip test: trace whether their own revenue and their vendors’ solvency rest on earned money or on the loop.
Why do applications have the worst margins if they are closest to the customer?
Because value accrues to the layer with the tightest constraint, not the layer that does the most visible work. Chips are hard to make and one company dominates, so profit pools there, capturing around 79 percent of the stack’s gross profit. Applications are easy to build, so thousands exist and most are thin, and competition pushes average gross margins to about 33 percent. The exception is an application that owns a real constraint, proprietary data, a deep workflow, or unique distribution, which escapes the commodity average.
What is the capital-light lane?
It is the one quadrant a small founder should build in: little capital in, buying a compounding asset. On one axis, capital intensity from light to heavy; on the other, whether the money buys a depreciating asset or a compounding one. Heavy plus depreciating is the worst spot (debt-funded chips on a short clock). Light plus commodity is the thin wrapper. Light plus compounding, data, workflow, distribution, brand, is the capital-light lane, and it is wide open precisely because the giants are pouring money into the layers that depreciate.
Does the Nvidia 500 billion dollar financing plan help small startups?
Partly, and carefully. The stated aim includes making it easier for smaller AI companies to borrow to buy compute, and turning AI compute into an investable asset class does push down the price of the compute you rent. But borrowing to buy or commit to depreciating compute is exactly the melting-ice quadrant a small founder should avoid. The benefit to take is cheaper metered compute. The trap to skip is taking on fixed obligations against an asset that loses most of its value in a few years.
How is this different from saying AI compute is a subsidy?
The subsidy point is about cost: the price you pay for tokens is held down by suppliers running at a loss, so part of your margin is effectively a loan. The capital-stack point is one level up and structural: the revenue holding up your suppliers, and sometimes the revenue on your own dashboard, may be financed by the circular loop rather than earned from real customers. Cheap input is a cost question. The reflux loop is a solvency question. A careful founder runs both tests, because they can fail independently.
How do I run the round-trip test on my own business?
Two steps. Trace your revenue up: for your top customers, identify whose budget the money ultimately comes from, and flag any that come from other venture-funded, unprofitable AI companies, because that revenue is a slice of a fundraise and can vanish in a downturn. Then trace your dependencies down: list the vendors you cannot run without, note which are burning borrowed money to survive, and buy portability wherever you could not switch within a week. The result is an honest read of how much of your business rests on earned money versus the loop.
If there is an AI bubble, what actually protects a founder?
Structure, not prediction. You cannot time a correction, but you can be built to survive one. Being capital-light protects your balance sheet, because your costs flex down with usage instead of sitting as fixed debt against depreciating hardware. Owning a compounding asset protects your value, because data, workflow, distribution, and brand keep gaining worth even when funding gets tight. And earned revenue from customers with real budgets protects your top line, because it does not depend on anyone’s next round closing. The founders who get hurt are the ones who borrowed heavy to buy ice and sold mostly to other startups.
This is part of the AI-native founder playbook. For the cost side of the same coin, see building on a subsidy, AI gross margins, and the efficiency trap.