Circular Capital: How AI Investment Might Be Building Its Own House of Cards

The 2023-2026 AI build-out is being propelled by a small set of extraordinarily profitable upstream vendors, an expanding constellation of downstream firms that are not yet profitable, and increasingly circular financing structures. When you run the numbers (capex, depreciation horizons, utilization, power, and realistic end-market demand), the revenue and margin required to make today's investments pencil look far beyond what the paying economy currently supports. That mismatch doesn't make AI "fake." But it does make the prevailing financing model fragile.

Executive summary:

• Profit concentration upstream. Nvidia is the profit center of the ecosystem, doing on the order of $160B in revenue and ~$100B in EBIT, while relying on a small number of very large customers and on a single critical foundry supplier (TSMC).
• Circular financing is real. Public "order books" tied to AI (e.g., Oracle) surged on enormous future bookings - reportedly hundreds of billions from OpenAI - even as vendor financing for GPUs and pre-booked compute flowed around the loop through firms like CoreWeave. At each step the "numbers go up," while no end customer has paid a dollar yet.
• The break-even hurdle has moved up sharply. With 3-5-year effective asset lives rather than a decade, industry revenue of roughly $320B-480B per year would be required just to break even on the 2025 capex wave. Current revenue is a small fraction of that.
• Capex is outpacing monetization. The sector is spending more than $30B per month (~$400B in 2025) yet taking in ~$15-20B per year on current estimates; combining 2025-2026 implies ~$1T of revenue needed merely to get back to breakeven.
• Macro spillovers are large. The AI build-out plausibly accounts for ~1.5-2% of U.S. GDP via direct capex, with an additional, uncertain wealth-effect kicker from equity appreciation. That's supportive on the way up - and destabilizing if it stalls.

1. Follow the profits: one node dominates

Strip the story back to where cash profits are actually being earned. Nvidia is the fulcrum. In a recent MacroVoices interview, Matt Barrie noted that Nvidia is running at ~$160B in revenue and ~$100B in EBIT, an outlier profitability anchor for the entire AI value chain. He also underlined extreme customer concentration (top two about 40% of revenue, next four about 40%) and a foundry single-point dependence on TSMC.

Why does that matter? Because the rest of the "AI compute" universe doesn't look anything like that. Barrie's back-of-envelope framing of the AI compute market as "sub-$40B" today hints at how small downstream monetization remains relative to the capital flowing in. If you accept the giant "bookings" figures being touted, a single hyperscaler or cloud challenger would need to expand the paid compute market - in just a few years.

That juxtaposition (one vendor with extraordinary profitability vs. a patchwork of aspirants that are still revenue-small and cash-burning) is the first tell that financing rather than earnings is propelling the machine.

2. How the circular money loop works

Let's walk the loop, using the recent "infinite money glitch" narrative as a guide:

• Big future bookings juice the stock. Oracle's shares ripped after it presented an eye-popping hockey-stick order book for its cloud buildout; subsequent reporting and commentary highlighted that roughly $300B of those bookings were tied to OpenAI. That figure dwarfs OpenAI's current revenues.
• Vendor credit fills the gap. In parallel, Nvidia is described as offering roughly $100B of GPUs on vendor-financing terms: take now, pay later.
• Pre-booked compute creates "bankable" demand. OpenAI takes those GPUs and pre-books ~$22B of compute with CoreWeave on a take-or-pay basis; CoreWeave turns around and raises private credit against those bookings. Proceeds buy - what else - more Nvidia hardware.
• Round-tripping before revenue. At each hop, the appearance of future demand supports higher equity marks, more capacity orders, and more credit. Yet "at no point has a single customer spent a dollar."
  

This is not (necessarily) fraud. It is a financing architecture that turns future intentions into present-day capital - and it's not new. As hedge-fund manager Harris "Kuppy" Kupperman notes, we've seen similar patterns in late-1990s/2000 fiber-optic telecom: contracts, vendor financing, and bookings round-tripped to meet targets until the funding window narrowed, with painful consequences for names like Lucent and Nortel.

Barrie extends the analogy to the rise of "neo clouds" - CoreWeave, Lambda, Nebius - which he characterizes as the "WeWorks of GPUs." The model: collect pre-bookings ... raise debt/equity ... buy more gear ... repeat. It works while credit is easy; it is unkind when it tightens.

3. What the paying economy would need to deliver

Even if you accept the financing loop as a bridge to a bright future, the destination has to exist: paying customers at scale and with margins. Kupperman revisited his own model after speaking with data-center operators, lenders, and engineers. Their consensus critique was simple and devastating: the math doesn't work on a 10-year asset life. Buildings, racks, cooling, and GPUs age out fast; practical asset lives cluster at 3-5 years. That one change doubles to triples the annual depreciation charge. On that basis, he argues his prior break-even revenue estimate (~$160B) for the 2025 capex wave was "woefully inadequate." A more realistic break-even range is $320B-$480B of annual revenue - just to cover the capital being deployed this year.

Meanwhile, run-rate revenue today is nowhere near that. Kupperman's tally: more than $30B/month (~$400B in 2025) of capex going out, vs ~$15B-$20B/year of revenue coming in. Add the 2026 builds and "you would need ~$1T in cumulative revenue across 2025-2026 to hit break-even, with "many trillions more" to earn an attractive return.

Barrie's "sub-$40B compute market today" framing dovetails here. If you believe the bookings narratives, one counterparty (Oracle) tied to one customer (OpenAI) would have to grow the paid compute market by about 5 times in five years, leapfrog AWS by 2029, and do it with real cash - despite the customer's limited current revenues and the vendor's own cash constraints.

Implication: even generously assuming strong end-market adoption, the revenue ramp and margin structure required are extreme. The economics likely only close at very high utilization, minimal price compression, and favorable power costs - three assumptions that rarely coexist for long in computing.

4. A simple model you can sanity-check

To make the break-even math concrete, consider a stylized single-campus model:

• Capex: $10B for land, build, power distribution, racks, GPUs, networking.
• Useful life: 4 years (midpoint of 3-5 years). -> Depreciation: $2.5B/yr.
• Operating costs: power + O&M $0.9B/yr (this scales with utilization and local tariffs).
• Required return on capital (WACC proxy): 10% on average invested capital (assume $7.5B average) -> $0.75B/yr "capital charge."
• Target EBIT margin: 20% (so earnings after opex and depreciation must still be positive).
  

Under those (arguably generous) assumptions, required annual revenue is roughly:

• Cover depreciation + opex + capital charge = $2.5 + 0.9 + 0.75 = $4.15B, before any EBIT.
• For a 20% EBIT margin, revenue needs to be $4.15B / (1 - 0.20), or about $5.2B.

Scale that across $400B of 2025 capex industry-wide (about 40 such "$10B campuses"), and you're back near the $200B+ of annual revenue just to tread water, with plausible ranges that align with the $320-$480B Kupperman posits once you add the realities of idle time, price compression, and overheads. (This calculator is illustrative; the order of magnitude is the point.)

5. The physical-world brake: power, parts, and time

Money can be printed; transformers, turbines, and electrons can't. Even if the financing "works," power and grid constraints will meter how quickly capacity can be monetized, while assets depreciate rapidly.

• Power share: U.S. data centers are estimated around ~4.5% of national electricity demand today, plausibly reaching ~9% by 2030 - a doubling that must fight through interconnection queues and local resistance.
• Local price spikes: Wholesale prices near data-center hubs have reportedly climbed sharply in recent years, stressing the notion that cheap marginal power is endlessly available.
• Long lead-times: Critical balance-of-plant items (high-capacity switchgear, transformers, etc.) are on multi-year lead times, effectively gating delivery of usable compute. (Industry anecdotes abound; the research takeaway is simple: time-to-power is slow.)
  

Even if a CFO is comfortable with an aggressive revenue forecast, a slip of 12-18 months in energizing a campus can destroy the NPV on a 3-5-year asset-life assumption. That's before you consider the annual GPU refresh cadence that compels further spend just to remain competitive.

6. Macro linkages: why this matters beyond tech

Kupperman's macro thought experiment is worth sitting with. Start with ~$400B of 2025 capex tied to AI and adjacent infrastructure. Layer in associated R&D, grid upgrades, and ancillary buildout, and you plausibly get ~1.5% of U.S. GDP directly from AI-linked spending, maybe ~2% with multipliers. Now add a wealth-effect boost from equity gains in the AI complex. On the way up, that pads consumption; on the way down, it reverses.

His conclusion is unambiguous: if funding slows, buildout slows, beneficiary equities fall, consumption softens, and you risk a feedback loop that spills into the broader economy. Unlike railroads (long-lived assets), AI data centers risk going technologically obsolete fast, making the eventual write-downs larger and swifter.

7. "We've seen this movie" historical rhyme, not copy-paste

Skeptics aren't claiming a carbon-copy of 2000. Rather, they recognize rhymes: late-stage booms often feature big round numbers, mutual investments, vendor financing, and bookings that later fail to translate into cash. Kupperman's warning is measured: when the economics don't work, doing it at massive scale doesn't fix them, but rather it amplifies the eventual adjustment. Timing is the hard part.

8. What would have to go right (the bull case as a checklist)

From a skeptical lens, it is helpful to articulate what success requires - not to dunk on it, but to make the hurdle transparent:

1. Demand must multiply 10-20 times quickly and profitably. The compute market would need to scale from sub-$40B today to $320-$480B+ annually with strong gross margins just to validate 2025 investment, and higher still for acceptable returns.
2. Pricing can't compress too fast. In past compute cycles, unit prices fell as capacity ramped. If token/inference prices compress faster than utilization rises, margins vanish.
3. Utilization must be very high. Idle capacity kills this model under 3-5-year asset lives. Even short delays in application adoption cascade into missed depreciation windows.
4. Power must be abundant and cheap. Without long-term, low-cost power and swift interconnects, revenue ramps slip while assets age.
5. Financing must normalize from vendor credit to customer cash. The loop should evolve from bookings and vendor financing to billings and cash; otherwise growth is mark-to-model, not mark-to-market.
   

Is that impossible? No. Is it heroic? Yes.

9. Plausible bust paths (and why they rhyme with past cycles)

• Credit window narrows. If private credit, structured vendor financing, or equity appetite recede, the round-trip stalls. The same structures that made growth look effortless render it brittle.
• Bookings fail to translate to cash. If counterparties stretch, renegotiate, or de-book capacity, the "future demand" that underwrote debt vanishes. The feedback into share prices can be swift.
• Physical bottlenecks delay monetization. Interconnects, switchgear, and transformers push energization timelines out quarters or years; depreciation clocks do not stop.
• Single-point risk at the profit center. Nvidia's customer concentration and dependence on TSMC magnify any supply hiccup or geopolitical event into systemwide volatility.
• Confidence shock from the inside. After canvassing dozens of senior industry participants, Kupperman reports a shared, unspoken conclusion: "no one understands how the financial math is supposed to work." If insiders blink, the market will follow.

10. A skeptic's monitoring list

If you're writing for investors or operators, close with an empirical checklist:

• Bookings -> billings -> cash. Track disclosure that splits "bookings" from recognized revenue and cash receipts by counterparty. Does the mix improve quarter to quarter?
• Capex-to-earnings ratios. Are hyperscalers front-loading capex against flat or slowing core businesses? Watch for re-guides or pauses.
• Power contracts and queue positions. Are campuses actually securing firm, multi-year power at stable prices? How long to energize?
• Utilization and price per unit. Follow realized utilization and price per inference/token; the product of the two must rise faster than cost per watt and depreciation per seat.
• Supply chain cadence. What is the GPU refresh interval? Are customers forced to repurchase more often than business value would justify?
• Concentration & substitution. Do customers diversify away from a single upstream vendor? Are in-house accelerators or software efficiency gains (better compilers, sparsity, caching) lowering required capex per workload?

11. What skepticism is - and isn't - saying

None of this argues that AI is a mirage. The use-cases are real; the technology is improving; the need for compute is rising. Skepticism is about finance and pacing: are we building capacity faster than paying demand can absorb, and are we using financing to bridge a gap that is wider than most models admit? Today, the profit center is effectively one vendor upstream; the downstream services remain small relative to the sums discussed. Financing round-trips future intention into today's capex; that's a feature of late-cycle booms.

Meanwhile, depreciation clocks are short, power is scarce, and macro linkages are significant. If the funding impulse slows, the gear keeps aging while revenues lag. That combination has an unhappy history in railroads, power plant booms, and especially fiber/telecom. The mechanics rhyme; the timeline is unknowable.

Bottom line: for the current trajectory to work without a painful reset, paying demand must explode in both scale and quality (margin), quickly. Until we see the mix shift from bookings and vendor credit to billings and cash, prudence argues for treating the "AI ouroboros" as a financing phenomenon first and a monetization story second. If the math doesn't start working in cash, the eventual adjustment will be as much a macro event as a tech one.

Source notes:

• MacroVoices: Erik Townsend with Matt Barrie - on Nvidia's profit concentration, customer dependence, the Oracle/OpenAI bookings narrative, vendor financing, and the "infinite money glitch."
• Praetorian Capital (Kuppy): "An AI Addendum." - Revised depreciation (3-5 years), the $320-$480B break-even revenue band, $400B 2025 capex vs $15-$20B revenue, and macro spillovers/wealth effects.