마진 위의 AI – OpenAI를 기반으로 구축된 1,000억 달러 규모의 부채 타워

AI on Margin – The $100 Billion Debt Tower Built Under OpenAI THE DAILY --- H.E.A.T. Torching Wall Street’s Obsolete Playbook. Delivered straight to your inbox 5 days a week. Subscribe Home Posts AI on Margin – The $100 Billion Debt Tower Built Under OpenAI AI on Margin – The $100 Billion Debt Tower Built Under OpenAI Dec 1, 2025 I’ve been a trader and investor for 44 years. I left Wall Street long ago—-once I understood that their obsolete advice is designed to profit them, not you. Today, my firm manages around $4 billion in ETFs, and I don’t answer to anybody. I tell the truth because trying to fool investors doesn’t help them, or me. In Daily --- H.E.A.T. , I show you how to Hedge against disaster, find your Edge, exploit Asymmetric opportunities, and ride major Themes before Wall Street catches on. ❝ Today is Payable Date: BITK: .12/Share MSTK: .28/Share ❝ I’m hosting a webinar entitled “Why Covered Call ETFs Suck and What to Do Instead” (More Info Below) December 9 2-3pm. Sign Up Here Table of Contents --- H.E.A.T.--AI Margin Call --- News vs. Noise: What’s Moving Markets Today --- A Stock I’m Watching Why Covered Call ETFs Suck-And What To Do Instead --- H.E.A.T. The ORCL/Open AI story is interesting and I went back into ORCL last week. I think the massive pop was unjustified, but so was the massive selloff….. Today we talk about the relationship in more detail….. AI on Margin – The $100 Billion Debt Tower Built Under OpenAI If you want to understand what late‑cycle AI risk actually looks like, don’t stare at $NVDA’s P/E. Look at the balance sheets wrapped around OpenAI. According to new FT reporting, OpenAI’s partners and lenders are on track to pile up almost $100 billion of debt that is effectively tied to OpenAI’s growth – while OpenAI itself carries almost no debt at all. ( Financial Times ) The basic structure: SoftBank, Oracle, CoreWeave and others have already borrowed $30B+ to invest in OpenAI or build its data centers. Blue Owl + Crusoe + other infra players have another ~$28B of loans whose economics rely on OpenAI‑linked deals. A bank syndicate is now working on ~$38B more for Oracle and Vantage to fund new OpenAI sites in Texas and Wisconsin. Stack that up and you get ~ $100B of bonds, bank loans, and private credit whose repayment ultimately depends on OpenAI using a massive amount of compute over the next decade. FT notes that’s on the order of the net debt of the six largest corporate borrowers in the world – VW, Toyota, AT&T, Comcast, etc. Meanwhile, OpenAI itself is reportedly running with minimal balance‑sheet debt , beyond a ~$4B credit facility it hasn’t touched. It has signed roughly $1.4 trillion of compute and data‑center commitments over eight years, versus an annualized revenue run‑rate around $20B , but the actual borrowing sits on other people’s books. ❝ OpenAI’s strategy, in their own executive’s words: “How do we leverage other people’s balance sheets?” This is “AI on margin” in the most literal sense. How the structure really works There are three main pillars: Big sponsors levering up Oracle has already sold about $18B of bonds to fund OpenAI‑related infrastructure, and one sell‑side desk thinks it may need up to $100B in new borrowing over four years to deliver on these contracts. SoftBank has raised ~ $20B this year for AI, with OpenAI as its largest single bet. CoreWeave borrowed $10B+ to lease GPU‑dense data‑center capacity for Microsoft and OpenAI workloads. Special‑purpose vehicles (SPVs) doing the building Data‑center projects in Texas, New Mexico and other sites are being financed in SPVs and variable‑interest entities , often with non‑recourse debt . If things go wrong, lenders take the land and buildings, not the sponsors. Example: Blue Owl + Crusoe’s Abilene, TX site took out ~$10B from JPMorgan to build a facility that Oracle is leasing for 17 years. That loan reportedly has no recourse to Blue Owl or Crusoe: if Oracle stops paying, the bank effectively owns a giant empty hyperscale shell. Long‑dated leases and take‑or‑pay style contracts Oracle is signing long‑term leases on these SPV facilities, which it then uses to fulfill OpenAI’s compute contracts. Oracle’s debt is serviced by those lease payments; the SPVs’ debt is serviced by Oracle. Other infra players like Blue Owl and Crusoe are effectively equity wrappers around these leases: their return comes from the spread between what they can borrow at and what Oracle pays. The common thread: everyone is levering up to build the AI grid, and OpenAI is the demand anchor – but almost none of the leverage is on OpenAI itself. This is structurally very similar to the late‑1990s/early‑2000s telecom build‑out: carrier commitments, vendor financing, and asset‑heavy SPVs funded by cheap debt… all predicated on a long runway of traffic growth. What could go right? If the AI thesis plays out anything like the believers think, this structure is insanely powerful for OpenAI: OpenAI locks in access to custom, AI‑optimized infrastructure (compute, power, networking) without loading its own balance sheet with $100B+ of recourse debt. Partners like Oracle, SoftBank, CoreWeave, Blue Owl, Vantage and Crusoe turn AI demand into bond‑like cash flows via long‑term leases and contracts. Nvidia and other chip vendors still get paid upfront — this debt is, in large part, just a way to finance massive GPU purchases and data‑center builds . If AI workloads and pricing power grow as hoped, these assets could end up under‑built , not over‑built: a world where $100B of debt looks small relative to the cash being produced. In that world, OpenAI has effectively: shifted the capex and credit risk to big infra partners, kept its own equity “option” extremely clean, and reserved the right to capture a disproportionate share of the economics if AGI‑ish systems emerge. What could go wrong? There are three main failure modes I’d worry about: Demand risk: OpenAI doesn’t monetize fast enough If AI usage and monetization disappoint – slower enterprise adoption, pricing pressure, regulatory brakes – then $1.4T of compute commitments over eight years starts to look very heavy vs a $20B revenue base. OpenAI can try to renegotiate, but the debt is sitting at Oracle, SoftBank, CoreWeave, Blue Owl, and in SPVs held by lenders. Their lenders don’t care about the dream of AGI; they care about debt service . Spread risk: rates or credit spreads blow out A big chunk of this borrowing is long‑dated, but not all of it. If rates back up or spreads widen, refinancing in the late 2020s could be painful . Think “telecom 2001”: once capital markets lose faith in the return on all those fibers in the ground, you go from cheap leverage to distressed refinancing very quickly. Concentration and counterparty risk Many of these deals are essentially single‑tenant, single‑theme assets : if Oracle or OpenAI walked away, there isn’t a deep pool of alternative tenants who need 100MW+ of GPU‑optimized capacity in Abilene tomorrow. SPVs and non‑recourse structures protect sponsors, but concentrate risk at the bank and private‑credit level . JPMorgan’s $10B Abilene loan is ultimately a bet on Oracle and OpenAI and the durability of AI capex. This is not a 2008‑style systemic bomb — $100B is big but not “blow up the global banking system” big — but it is absolutely the kind of thematic leverage tower that can hurt a lot of balance sheets if the story cracks. Winners and Losers Likely Winners (if the AI capex cycle stays intact) 1. OpenAI (and similar model providers) Gets exclusive‑ish access to premium infrastructure with minimal on‑balance‑sheet leverage. Keeps equity upside if the models truly become foundational infrastructure. Essentially runs an asset‑light AI monopoly attempt while others own the concrete and copper. 2. Nvidia and high‑end chip vendors Much of this debt is being used to buy their hardware . They get paid early in the chain; even if infra owners struggle later, the chips are bought and installed. 3. Oracle and hyperscaler‑adjacent infra players… if they execute Oracle is taking real risk here, but if OpenAI usage scales and margins hold, it becomes a quasi‑hyperscaler with embedded AI tenants and long‑duration revenue. Data‑center developers like Vantage, and capital providers like Blue Owl, get leveraged exposure to AI infra cash flows —especially where SPVs are structured carefully and tenant quality is strong. 4. Private credit and infra investors who underwrite this stuff correctly AI data‑center loans with long leases to Oracle/Microsoft‑tier credits can be fantastic assets if structured conservatively. They offer yield, growth, and embedded call options on AI demand without having to pick specific models. Likely Losers / Tougher Lane 1. Over‑levered infra sponsors with weak recourse or weak tenants If you’re levering 10–15x on a single hyperscale tenant whose economics hinge on one unprofitable AI customer, you’re effectively writing a very large, very concentrated AI call option with debt. SPVs protect the sponsor equity to a point, but once you’ve put real capital in the ground, a vacant 100MW GPU barn is not fun. 2. Oracle’s bondholders, if expectations prove too rosy On paper, Oracle is turning into a high‑growth AI infra landlord , with OpenAI contracts as proof. In a bear case where AI infra is overbuilt or pricing compresses, the equity can re‑rate and the debt stack gets heavier . Oracle is not a risk‑free utility; it’s becoming a leveraged AI infra play. 3. Banks and private‑credit funds late to the party Early lenders with tight structures and good covenants will probably do fine. Latecomers chasing yield at tight spreads into speculative AI‑themed SPVs are potential bag‑holders if refinancing risk shows up around 2028–2030. 4. Investors who think “OpenAI risk” = only OpenAI’s equity The real financial risk is distributed through the ecosystem : Oracle’s bonds, SoftBank’s leverage, infra sponsors’ equity, banks’ loan books, private‑credit vehicles. If you’re long all of these at peak AI optimism, you’re more exposed to the AI credit cycle than you might think. How I’m thinking about it Big picture, this is what late‑stage AI looks like financially: OpenAI is effectively running AI on margin, but the margin calls will land on its partners, not on its own balance sheet. The $100B debt tower is both a vote of confidence in AI demand and the first clear sign of a leveraged infra bubble forming beneath the model layer . I don’t think this headline means “short AI and go home.” But it does change how I think about where to be long: I’d rather own picks‑and‑shovels with diversified tenants and strong sponsors than the most levered, single‑tenant SPVs. I’d be very careful with anyone whose AI infra strategy hinges on one or two big contracts with lossmaking AI unicorns. And I’d treat OpenAI‑driven infra debt as a new macro risk factor to watch as we get deeper into this cycle – right alongside Fed policy, labor, and whether AI revenues start to catch up with AI promises. This is what an AI supercycle feels like in real time: exponential model progress on the front end, and a slowly rising wall of debt on the back end. --- News vs. Noise: What’s Moving Markets Today In a holiday‑thinned tape, nothing looked dramatic: bonds and equities chopped in a narrow range, Eurostoxx was flat, spreads barely moved. Under the surface, though, the story is a lot cleaner than it was a month ago. Equities have retraced most of the November sell‑off, and Fed pricing has round‑tripped with them – we started the month with ~17 bps priced for a December cut, we’re sitting near ~20 bps now. What’s actually changed is positioning . At the end of October, everyone was maxed out in risk – long AI, long beta, long crypto. After the flush, investors are basically flat on our positioning gauges, and you can see it in the price action: high‑octane retail momentum names and crypto are still well below their peaks even as the index recovers. That’s exactly what you want to see if you’re looking for a base, not a blow‑off top. That brings us back to the question everyone keeps asking: is AI in a bubble? November felt like a perfect storm – stretched AI valuations, a wobble in global liquidity, hawkish Fed soundbites, and a jobs report that gave the Fed “permission” not to cut. That was enough to hit the AI complex and bitcoin hard and resurrect every “this is 1999” headline. But step back and nothing fundamental has broken in the AI story. The capex is being driven by absurdly cash‑rich giants (GOOGL, MSFT, META, AMZN, etc.) who are still only spending ~50–60% of free cash flow on AI infrastructure and can turn the tap if they have to. Adoption is real – ChatGPT alone touches more than 10% of the planet, AI is already outperforming humans on more and more tasks, and it’s the control layer for everything from humanoid robots to autonomous vehicles to industrial automation. The market is increasingly valuing companies on their role in that ecosystem – chips, data centers, cloud, power – not on this quarter’s earnings. That can morph into a bubble a couple of years down the road, but trying to top‑tick “the AI bubble” today is a great way to miss the bulk of the S‑curve. For now, we think this is a violent correction inside an ongoing AI cycle , not the end of it. On the macro side, the real risk isn’t that AI capex suddenly stops – it’s that the labor market cracks faster than the Fed expects . Our base case has U.S. unemployment drifting toward 4.5–4.6% as SMEs hit a refinancing wall in 2026 and pull back hiring; that’s a modest slowdown, not a collapse. But if we’re wrong and the labor picture deteriorates sharply, we’d have to revisit the whole macro stance. Inflation and fiscal worries haven’t gone away either, especially with governments leaning into expansionary budgets. The Fed, for its part, is already on an easing path – whether they cut in December or wait for January is mostly noise. The direction of travel is still toward ~3% policy rates into 2026, and global liquidity (on a combined M2 basis) is likely to rise, not fall. So where does that leave us? With a healthier starting point than we had a month ago. Positioning in risky assets has been cleaned up, some froth has come out of the AI/crypto complex, and the Fed is leaning easier into what’s still, for now, a slow‑motion labor slowdown rather than a sudden stop. We’re not in the “AI is a bubble, run for the hills” camp. If there is a bubble, it’s probably a couple of years away and there’s little edge in trying to short it in Q4 2025. Instead, our focus for 2026 is on the broadening of the AI trade – reallocating away from just the high‑end chipmakers into the second layer (data centers, power, networking, robotics, non‑U.S. tech) and using any renewed “AI bubble” scare as an opportunity to upgrade quality, not abandon the theme. Key Takeaways Positioning, not a broken narrative, drove November’s damage. Crowded longs in AI, beta, and crypto were forced out; investors are now close to flat, which is a healthy base for the next leg. We do not think we’re in a full AI bubble yet. Capex is funded by huge cash flows, leverage isn’t the issue, and adoption + infrastructure build‑out are real. A true bubble is more likely a 2027 problem than a 2025 trade. The Fed is in easing mode; the calendar is noise. Whether the cut comes in December or January, the path is lower rates toward ~3% into 2026, with global liquidity set to rise as fiscal policy stays loose. Labor is the real macro swing factor. Our base case is a modest slowdown (unemployment in the mid‑4s). A sharper deterioration would force us to rethink the constructive stance on risk. Play the AI cycle by broadening, not bailing. Use cleaned‑up positioning and any renewed “AI bubble” panic to rotate from the most crowded AI winners into second‑layer beneficiaries and infra, rather than trying to call the top of the entire theme. --- A Stock I’m Watching Today stock is SiTime (SITM)…. SiTime (SITM) is one of those boring‑sounding companies sitting in a very exciting place. They don’t make GPUs or data centers – they make the tiny timing chips (MEMS oscillators and clocks) that act like the “heartbeat” inside high‑speed AI hardware: 1.6 terabit optical modules, switches, NICs, and accelerator boards. As AI networks upgrade from today’s 800G links to next‑gen 1.6T speeds, each module needs more and higher‑performance timing content, so SiTime’s dollar content per system roughly doubles at the same time that unit volumes are ramping. On top of that, they’ve won key sockets in Apple’s in‑house 5G modem and are positioned for the next wave of on‑device AI in phones, wearables, and sensors. In plain English: if AI data centers and AI‑enabled devices keep scaling, SiTime quietly collects a toll on every faster connection, without having to bet on which GPU, TPU, or model “wins” the AI war. How Else I Can Help You Beat Wall Street at Its Own Game Inside --- H.E.A.T. is our monthly webinar series, sign up for this month’s webinar below…. Why Covered Call ETFs Suck-And What To Do Instead Tuesday December 9, 2-3PM EST Covered call ETFs are everywhere — and everyone thinks they’ve found a “safe” way to collect yield in a sideways market. The truth? Most of them suck. They cap your upside, mislead investors with “yield” that’s really your own money coming back , and often trail just owning the stock by a mile. Join me for a brutally honest breakdown of how these funds actually work — and what you should be doing instead. What You’ll Learn: 🔥 Why “high yield” covered call ETFs are often just returning your own capital 📉 How most call-writing strategies quietly destroy compounding 🚫 Why owning covered calls in bull markets is like running a marathon in a weighted vest 💡 The simple structure that can fix these problems — and where the real daily income opportunities are hiding Sign Up Here All rights reserved. Keep Reading No posts found View more caret-right 0