Mind the Gap
Funding 30-Month Assets with 30-Year Paper
TL;DR
Hyperscaler capex for the four largest U.S. cloud platforms is tracking close to $725 billion in 2026, with Oracle adding another $50 billion. The build is increasingly debt-financed: about $108 billion of AI-linked IG issuance in 2025, Meta’s $25 billion six-tranche sale this month, Alphabet’s 100-year Sterling note in February, Amazon’s $37 billion blockbuster in March. A second pool of capital is now stepping in alongside the bond market: private credit, with Blackstone, Blue Owl, Brookfield, KKR and Pimco financing data centers and GPU fleets through SPVs that keep the debt off hyperscaler balance sheets entirely. The Meta-Blue Owl Hyperion structure ($27B in 2049 secured notes anchored by Pimco at 6.58%) is the prototype. Frontier compute has an economic half-life I would put near thirty months. The paper funding it runs out to a hundred years. That is the duration mismatch that defined the 1990s telecom build, the 2010s shale cycle, and dry bulk in the mid-2000s. The asset side is what the equity market discusses. The liability side is what the credit market is starting to. Watch Oracle CDS, the long end of the hyperscaler curve, and the language of the next set of capex guides.
Hyperscaler capital expenditure for the four largest U.S. cloud platforms is now tracking close to $725 billion for 2026, with Oracle adding roughly another $50 billion on top. Meta priced a six-tranche bond deal of as much as $25 billion on May 1, the same week it lifted full-year capex guidance to as much as $145 billion. Alphabet sold a £1 billion 100-year Sterling note in February, the first century bond from a tech issuer since Motorola in 1997. Oracle’s five-year CDS roughly tripled between July and December 2025. The headline framing is a race to AGI. The mechanics are something else.
I have spent a chunk of my working life on the financing and balance-sheet side of cycles like this. My read is that the equity market is being asked to underwrite a build whose asset duration and liability duration have come unmoored, and whose output is commoditizing on a curve that the depreciation schedule does not reflect. That is the conundrum. It is not a question about whether AI matters. It is a question about who pays for the gap between economic life and accounting life, and on what terms.
I. Two Templates: The Patent Vault and the Capacity Trap
Industrial “borrow-to-build” cycles tend to resolve into one of two templates. The first is the patent vault: capital deployed against a legally protected, non-fungible product where entry is barred for a defined period. The second is the capacity trap: capital deployed against fungible output where the firm holds no pricing power and where competing supply enters faster than demand can absorb it.
The patent vault is the friendlier template, but it is not as friendly as the marketing suggests. Eli Lilly today, riding the GLP-1 cycle, looks like the archetype: high incremental returns on capital, demand inelastic, supply legally constrained. AbbVie sat in the same chair a few years ago with Humira, then watched U.S. biosimilar entry from January 2023 compress that franchise on a schedule the equity market had known about for years and still managed to misprice. Even legally protected monopolies face an exclusivity cliff that resets economics violently. The duration of the moat is the duration of the patent, not the duration of the asset.
The capacity trap is the less friendly template, and it has recurred almost on schedule across the last forty years.
• Shale, 2014–2020. Medium-tenor debt funded wells with eighteen-month decline curves. The mismatch did the damage.
• Dry bulk shipping, mid-2000s. A wave of vessel orders met a softening trade cycle. Charter rates collapsed roughly 90%. Highly leveraged owners ended up with hulls worth less than the debt against them.
• U.S. Class A office, post-2020. Long-duration assets financed against historical occupancy assumptions, repriced by a structural shift the underwriting did not contemplate.
• Late-1990s telecom. This is the comparison worth dwelling on. WorldCom, Global Crossing, 360networks and Williams Communications laid the dark fiber. U.S. telecoms issued more than $500 billion of debt between 1996 and 2001. By 2001, roughly 95% of installed fiber was unlit. Bandwidth prices fell about 90% in the early 2000s. Two dozen telecom companies went bankrupt in 2001–2002 alone. Around $1 trillion of industry debt was written off. The fiber itself was not wrong. Over the following decade and a half, it was lit, leased and absorbed. But the original financiers were wiped out long before the asset became economic. The capital cycle does not require the build to be useless. It only requires the financing to outrun the absorption.
That last distinction is the one I want to keep on the table for the rest of this note.
II. The 2026 Pivot: From Software Margins to Industrial Capital Intensity
For roughly fifteen years the public market priced the platform franchises as asset-light software businesses: gross margins north of 70%, capital intensity below 15% of revenue, free cash flow that funded buybacks. That profile is changing in front of us. CreditSights has top-five hyperscaler capital intensity at 45–57% of revenue in the most recent quarter. Microsoft 45%. Oracle 57%. These are not software-company numbers. They are utility-company numbers, on software-company multiples.
The free-cash-flow line tells the same story. Amazon’s trailing twelve-month free cash flow has fallen from roughly $26 billion a year ago to a little above $1 billion. Microsoft is down 22%. Alphabet is down 38%. Meta is the only one of the four still growing FCF, and even Meta now sits with about $12 billion of free cash flow against $145 billion of planned capex. Operating cash flow is no longer enough to fund the build. The bond market is being asked to fill the gap, and increasingly, so is the private credit market.
1. The Depreciation Lag
GPU clusters are being capitalized on five- to six-year depreciation schedules. The economic half-life of frontier compute is shorter than that, and the gap is the most important unpriced variable on these balance sheets.
DeepSeek’s V3/R1 release in January 2025 was the moment the gap became visible. The headline number, around $5.6 million for the final pre-training run, was always a partial figure. DeepSeek’s own paper said as much, and SemiAnalysis later put total infrastructure cost closer to $1.6 billion once R&D and prior model work were included. The market briefly conflated the two, and Nvidia lost $589 billion of market capitalization in a single session on January 27, 2025, the largest one-day loss in U.S. stock-market history. The conflation was wrong on the small number. It was directionally right on the larger point: the cost-efficiency frontier had moved sharply, and a meaningful share of installed Hopper-generation capacity had become economically older than its book life implied. Blackwell, Rubin and the merchant-silicon roadmap from AMD and the hyperscalers’ own ASICs have continued to compress that half-life since.
The bull rebuttal is that demand is rising faster than supply. Token volumes, agentic workloads, video generation. Utilization stays high. Depreciation gets spread across a much larger revenue base. There is real evidence for this. Microsoft’s AI business is now at a $37 billion annualized run rate, up 123% year over year. I take that seriously. My counter is narrower: high utilization does not save you if the unit price of the output is falling toward marginal cost faster than you can amortize the kit.
2. The Commoditization of Inference
The pharma analogy for AI rests on the assumption that proprietary models can sustain pricing power. The evidence so far runs the other way. Open-weights releases from Meta, DeepSeek, Mistral and Alibaba have repeatedly closed the capability gap to within a quarter or two of the frontier, and inference pricing on equivalent capability tiers has fallen by roughly an order of magnitude over the past eighteen months.
The right way to phrase it is not that inference price trends to the cost of electricity. That is too cute. It trends toward the cash marginal cost of serving a token: power, networking, cooling, plus the unrecovered amortization of the cluster the token ran on. In a competitive equilibrium where capability is broadly fungible, the unrecovered amortization is what gets squeezed first, because it is the only line where pricing power exists. That is the regulated-utility outcome dressed in different clothes.
The question, then, is not whether AI is useful. It plainly is. The question is whether the operators of the infrastructure capture the rents, or whether those rents flow through to power producers, equipment vendors and end users. History has a clear majority view on that.
III. The Liability Side: Funding 30-Month Assets with 30-Year Paper
Note to the reader: this section runs long because the liability side now has two stories rather than one — the public bond market and the private credit market. Readers familiar with the bond data can skip directly to “The Private Credit Pivot, and the Hyperion Prototype.”
This is the part of the story that is missing from most of the equity research I have read on the capex cycle, and it is the part I want to spend the most time on. Through the mid-2020s the hyperscaler build was overwhelmingly self-funded. That has changed. The aggregate 2026 capex plan for the top four hyperscalers sits at roughly $725 billion, and the bond market is now being asked to do the heavy lifting.
A few data points worth sitting with.
• AI-linked U.S. investment-grade issuance reached roughly $108 billion in 2025, about four times the prior five-year average for tech, per Mellon. The combined IG index weight of Meta, Alphabet, Amazon and Oracle has nearly doubled in twelve months, from 2.2% to 4.1% of the Bloomberg U.S. Corporate IG Index, per Breckinridge. Morgan Stanley has gross U.S. IG supply rising about 25% to a record $2.25 trillion in 2026, with hyperscaler-and-related issuance reaching $400 billion, roughly ten times the 2024 figure.
• Meta has come back to the well twice in seven months. $30 billion in October 2025, then a six-tranche deal of up to $25 billion this month, with the longest tranche, a 2066 maturity, marketed at initial price talk of as much as 180 basis points over Treasuries. Meta’s five-year CDS hit a record high the day the new deal was launched.
• Alphabet priced a $20 billion seven-part deal on February 9, 2026, upsized from $15 billion after drawing more than $100 billion of orders. The deal included a 40-year tranche that compressed 25 basis points during book-building and a £1 billion 100-year Sterling tranche at a 6.125% coupon, the first century bond from a tech issuer since Motorola in 1997. A $17.5 billion November 2025 deal had already brought the longest U.S. dollar tech corporate of last year, a 50-year note, which has tightened in secondary.
• Amazon raised $37 billion in U.S. dollars across eleven tranches on March 10, alongside a €14.5 billion (~$16.8 billion) European debut the following day, taking the combined deal to roughly $54 billion. The dollar book hit $126 billion in orders. The structure ran out to a 50-year tranche. The combined size is the fourth-largest U.S. corporate bond sale on record and the largest non-M&A deal since Verizon’s $49 billion in 2013.
• Oracle is laying out a 2026 funding plan of $45–50 billion, split between an early-year IG bond, a $20 billion at-the-market equity program, and mandatory convertible preferreds. A separate $38 billion JPMorgan/MUFG-led debt package is being put in place against the Texas and Wisconsin data centers. Total Oracle debt is now reported at $153 billion, up roughly 60% in twelve months.
• Oracle is the credit to watch. S&P-rated BBB+ with a negative outlook. The 5-year CDS widened from roughly 40 basis points in July 2025 to 151.3 basis points on December 12, the highest since the global financial crisis and a tripling in five months. CDS trading volume in Oracle hit $9.2 billion in a ten-week window, against $410 million in the comparable period a year earlier. OpenAI now represents about 58% of Oracle’s contracted backlog, and the late-July agreement to build up to 4.5 GW of capacity is reported at more than $300 billion over five years. On April 28, the WSJ reported that OpenAI had missed internal user and revenue targets. Oracle traded down on the print. Q2 FY2026 free cash flow was negative $10.3 billion, the worst since 1992. Morgan Stanley and JPMorgan together project something like $1.5 trillion of incremental tech debt issuance over the next several years to fund this build.
1. The Private Credit Pivot, and the Hyperion Prototype
The next leg of the funding stack is not the bond market. It is private credit. Through the second half of 2025 and the first part of 2026, the largest non-bank lenders, Blackstone, Blue Owl, Brookfield, KKR, and the credit arms of Pimco and BlackRock, have stepped from the periphery of this cycle to the center of it. KKR has now committed roughly $34 billion of equity into digital infrastructure across 23 investments, alongside a separate $50 billion strategic partnership with Energy Capital Partners targeting data centers and the power generation behind them. Blue Owl’s credit platform stands above $145 billion. Blackstone has called digital infrastructure one of its highest-conviction themes.
The October 2025 Meta-Blue Owl Hyperion transaction is the prototype, and it is worth walking through in some detail because the structure tells you how the next $500 billion gets funded.
Hyperion is a 2,250-acre, 2–5 GW data center campus in Richland Parish, Louisiana, scheduled to come online in 2029. Total project size is roughly $30 billion. The capital stack is built around a special purpose vehicle named, with some humor, Beignet Investor LLC. Blue Owl-managed funds own 80% of Beignet. Meta owns 20%, against a $1.3 billion equity check. The other $2.5 billion of equity comes from Blue Owl. The remaining $27 billion is debt, raised through Pimco-issued, A+-rated, fully amortizing senior secured notes due 2049. Pimco anchored $18 billion of that paper at a 6.58% coupon. BlackRock took roughly $3 billion. The structure was reportedly cleared by an SEC private letter.
The mechanics are worth understanding. Beignet owns the campus. Meta leases it back. Lease payments service the bonds and the equity coupon to Blue Owl. Critically, Meta’s lease was engineered into four-year increments specifically so that the rating agencies do not classify the obligation as long-term debt on Meta’s books. The $27 billion is real. The cash-flow obligation is real. The campus is real. None of it appears as Meta debt. From a creditor’s perspective the result is a 24-year amortizing claim against future Meta lease payments, dressed as project finance. From Meta’s perspective it is operating expense.
Hyperion was the largest private-credit transaction ever executed at signing. It is also a template. Microsoft, BlackRock, GIP, MGX and NVIDIA established the AI Infrastructure Partnership in September 2024, targeting $30 billion of equity and up to $100 billion of total mobilized capital including debt; AIP’s first transaction was the October 2025 acquisition of Aligned Data Centers at a roughly $40 billion enterprise value. Oracle’s Stargate project is built on adjacent architecture with Blue Owl in the consortium. CoreWeave, the pure-play AI cloud, has now stacked $9.8 billion of private credit secured directly against its GPU fleet across two Blackstone-and-Magnetar-led facilities, the second of which, at $7.5 billion, was described by Blackstone as one of the largest private-credit financings in history.
Three things follow. First, the headline capex number understates the true commitment. Meta’s contractual obligations rose roughly $107 billion in a single quarter; Alphabet disclosed $232.7 billion of non-cancelable supply, content and energy commitments at March 31, 2026. The on-balance-sheet bonds I described above are the visible portion of a funding stack that runs significantly deeper. Second, the duration mismatch I am tracking on the bond side gets sharper, not softer, when you include the SPV layer. Beignet’s notes mature in 2049, twenty-four years out, against a campus full of GPUs that may or may not be the right asset class in 2031. Third, and this is the part that should make a credit investor pause, the structure transfers the technology-obsolescence risk from Meta’s rating-sensitive balance sheet to Pimco’s and BlackRock’s and Blue Owl’s, and ultimately to the insurance-company and pension end-investors who own those funds. The risk has not disappeared. It has changed addresses.
GPU-as-collateral is a particular feature of the private-credit layer worth flagging. The CoreWeave facilities are secured against Nvidia hardware whose secondary-market value is tied to a fast-moving silicon roadmap. A facility lent against H100s in 2024 is, by 2026, lent against an asset class whose newest peers are Blackwell and Rubin. The collateral does not literally vanish, but its value relative to the loan principal compresses faster than any traditional asset class an institutional credit investor is used to. The closest historical analogue is vendor financing in the late-1990s telecom cycle, when Lucent and Nortel lent into the build to keep their order books moving. We know how that ended.
2. Step Back
The hyperscalers are issuing thirty-, forty- and hundred-year paper, and now twenty-four-year SPV notes secured against four-year-life hardware, to fund assets whose economic half-life I would put at thirty months. That is the duration mismatch. It is the same mismatch that defined the late-1990s telecom build, the 2010s shale cycle, and dry bulk in the mid-2000s. The asset side of the capex conundrum is the part the equity market discusses. The liability side is the part the credit market is starting to.
The defense is that four of the five hyperscalers carry stellar ratings, that demand for long-dated investment-grade paper is structurally enormous (insurance balance sheets, sovereign reserves, pension liability matching), and that even at current spreads the all-in cost of capital is well below the expected return on the build. That defense is not unreasonable. It is the same defense that was made, with similar credit ratings and a similar investor base, in 1999. What changes is not the rationale at issuance. What changes is the gap between the maturity of the bond and the useful life of the asset against which it was raised. That gap is now wider than at any point in the modern history of the technology sector. In one Mirabaud portfolio manager’s words to CNBC in February: “what if, in three years, these Nvidia chips get outstripped by a Chinese competitor, and I’m lending for five or eight years, and in year three, my data center is obsolete?” That is the question the bond and private-credit markets are now beginning to ask out loud.
IV. The Antitrust Misread
I want to flag a counter-narrative that is becoming consensus and that I think is wrong, or at least overstated. The argument goes: the antitrust cases against Big Tech are dismantling the monopoly rents that have been subsidizing the AI build, and the cycle therefore becomes self-financing or it collapses.
The actual rulings cut the other direction. Judge Mehta’s September 2025 remedies opinion in U.S. v. Google, finalized in December, rejected the Chrome divestiture the DOJ had requested, declined to mandate choice screens, and explicitly cited generative AI competition as a reason to choose narrow behavioral remedies over structural ones. Wall Street read it as a win for Google. The analyst community used the word “home run.” The FTC’s actual antitrust case against Amazon, the marketplace and logistics integration matter and not the Prime dark-patterns settlement, has been pushed to a February 2027 bench trial. There has been no structural remedy against any of the platforms.
If anything, the regulatory backdrop is doing the opposite of what the bear thesis assumes. Courts are openly invoking AI competition as a reason not to break up the incumbents. That gives the platforms a regulatory incentive to keep building, because demonstrable competitive intensity in AI is now itself a defense against structural antitrust action. Cash flow from the legacy monopolies remains intact for at least the duration of the appeals cycle, and the build continues unconstrained. The regulatory risk to this capex cycle is real, but it sits in the future, not in the present.
V. What the Regret Phase Looks Like
Every capex cycle has a regret phase. It begins not when the assets stop being useful, but when the marginal return on the next dollar of capex falls below the weighted average cost of capital. For a software business with 70% gross margins and minimal capital intensity, that threshold is far away. For an industrial business with 45–57% capital intensity and falling output prices, it is much closer than the equity market currently believes.
Two things would tell me the regret phase has begun. The first is a hyperscaler missing on free cash flow not because of revenue, but because depreciation and interest expense outrun the business case for incremental capacity, and management lowering rather than raising the next year’s capex guide in response. We have now seen guides raised for eight quarters running. A cut would be the signal. The second is credit-spread widening that decouples from rates: AI-issuer spreads moving wider while the rest of IG holds firm. The early signs are already visible in Oracle’s CDS and in the term premium being demanded on the longest hyperscaler tranches. It is not yet a trend. It is a thing to watch.
Inside the cycle, the firms that win are not the ones that build the most. They are the ones that recognize when the build has out-run absorption and pivot to capital preservation before the credit market forces the pivot on them. That has been true in shale, in shipping, in fiber, and in commercial real estate. I do not see why it would be different here.
VI. Practitioner Conclusion
The capex conundrum is not a question about the utility of artificial intelligence. The technology is real and the demand is real. The question is whether the capital structure that has been put in place to fund the buildout is durable across the cycle that follows.
On the asset side, hyperscalers are capitalizing thirty-month assets on six-year schedules. On the liability side, they are funding those assets with paper that runs out to a hundred years on the public side and twenty-four years through SPVs on the private side. On the output side, the price of inference is converging toward cash marginal cost. On the regulatory side, the antitrust environment that the bears expect to constrain the build is, for now, doing the opposite. The bondholder is being asked to absorb the duration mismatch. The private-credit lender is being asked to absorb the technology-obsolescence risk. The equity holder is being asked to assume the rents are durable. All three assumptions deserve more scrutiny than they are currently receiving.
I do not think this ends in 2026. It does not need to. Cycles like this end when the marginal financing turns more expensive than the marginal return, and we are not there yet. We are at the point where the gap between asset life and liability life has become the most important variable on the page, and where the smartest thing a practitioner can do is read the credit market, not the equity market, for the first signal that the gap is starting to be priced.
That is the watch I am keeping. Oracle CDS, the long end of the hyperscaler curve, the spreads on the next set of SPV-securitized data-center deals, and the language of the next set of capex guides. The conundrum will be resolved on those four dials before it shows up in the equity prints.
Sources: company filings and earnings releases (Alphabet, Amazon, Meta, Microsoft, Oracle, Blue Owl, CoreWeave); SEC Form FWP filings; Bloomberg; Reuters; CNBC; Wall Street Journal; Financial Times; Fortune; CreditSights; SemiAnalysis; Mellon Investments; Breckinridge Capital Advisors; Cambridge Associates; MUFG Americas; Barclays; Morgan Stanley; JPMorgan; Janus Henderson; Nuveen; Mirabaud Asset Management. Antitrust references: U.S. v. Google LLC, remedies opinion (Mehta, J., D.D.C., Sept. 2 and Dec. 5, 2025); FTC v. Amazon.com, Inc. (W.D. Wash., trial scheduled Feb. 9, 2027).
The Macro Fireside
Macrofireside.com · @Macrofireside on X
The Macro Fireside is a practitioner’s publication — written at the intersection of markets, policy, and geopolitics by an experienced hand who has spent decades managing money through moments the world would only later recognize as inflection points. Analysis here is earned, not assembled. This piece does not constitute investment advice.
For professional enquiries: gs@macrofireside.com