The Thread Thus Far: From Dark Fiber to the Edge of the Loop
In Part 1, we traced the thirty-year mechanical thread underlying modern capital cycles. We began not with a 1999 crash, but with the 2000–2002 telecommunications and technology bust that followed the late-1990s infrastructure boom. Telecommunications providers poured capital into fiber, switching equipment and network capacity while equipment vendors increasingly supplemented ordinary financing markets with extended payment terms, direct loans and other forms of customer financing. When the boom broke, corporate capital spending collapsed—but much of the physical network remained.
That distinction matters. The financial claims disappeared. The conduits, fiber routes, rights-of-way and communications infrastructure did not. The sunk capital of one cycle became a lower-cost input into the broadband, mobile and cloud cycles that followed.
Part 2 moved from the infrastructure to the household balance sheet. We examined the concentration created when capitalization-weighted indexes progressively concentrated ordinary retirement portfolios in a relatively small group of mega-cap technology companies, and we laid out four defensive tax maneuvers that become available during a major equity drawdown: Roth conversions at depressed valuations, direct-index tax-loss harvesting, strategic asset relocation and generational basis planning.
Part 3 brought the cycle into 2026. The essential difference from the dark-fiber era was becoming visible: fiber installed twenty-five years ago can still transmit economically useful data today. Frontier accelerators do not enjoy that luxury. Their frontier-performance value can reprice in only a few product generations, even while the data center, transmission equipment, substation, land and power interconnection surrounding them may remain economically productive.
By the summer of 2026, the repricing was no longer merely theoretical. Big Tech’s infrastructure spending continued to accelerate, free cash flow came under increasing pressure, and credit investors began paying more attention to the balance sheets financing the buildout. Goldman Sachs now estimates that Meta, Microsoft, Amazon and Alphabet alone will spend approximately $5.3 trillion in aggregate capital expenditures from fiscal 2025 through fiscal 2030. Analysis of the broader hyperscaler complex found that the projected increase in capital expenditures through 2027 materially exceeds the projected increase in operating cash flow.
The question at the end of Part 3 was therefore not whether artificial intelligence is useful. It was much narrower: Who owns the depreciation risk, who finances it, and what happens when the cost of capital rises faster than the cash yield on the equipment purchased with that capital?
Now the bond market is beginning to answer.
The Sovereign Vise: Capital Is Available—At a Price
Equity markets remain willing to capitalize extraordinary future AI earnings. The long end of the bond market is imposing a different discipline.
The Long-End Repricing
On August 13, the United States sold $25 billion of 30-year Treasury bonds at a 5.216% high yield. The auction’s 2.39 bid-to-cover ratio was not weak; primary dealers were left with only approximately 11.5% of the competitive award. This was not a failed auction or a buyer’s strike. It was something more important: buyers showed up, but they required the U.S. government to pay the highest 30-year auction borrowing cost since 2001.
That is the real signal. The market is not refusing duration. It is repricing duration.
Thirty-year U.S. real yields—the return demanded after inflation—are now around 3% and near an eighteen-year high. At the same time, Alphabet, Amazon and Meta have issued almost $220 billion of bonds so far in 2026, more than double their combined full-year 2025 total. Governments and the AI complex are increasingly bidding for the same long-duration institutional capital.
Every leveraged data-center model ultimately lives downstream from that price. A compute lease, power purchase agreement, asset-backed loan or private-credit facility may carry a very different contractual coupon than a Treasury bond, but the risk-free curve remains the foundation beneath the spread. When the long-duration baseline moves above 5%, projects conceived when money was substantially cheaper must either produce more cash, accept a lower valuation, increase leverage, obtain external support—or fail the new hurdle rate.
That is not a prediction. That is arithmetic.
The Yen Warning: Reserve Mobilization Without the Old Playbook
Japan provides a second warning about global liquidity, although the mechanism is more nuanced than a simple liquidation of Treasury reserves.
After the yen reached a forty-year low of approximately 163.99 per dollar, coordinated Japanese and U.S. intervention helped drive it as strong as roughly 155.20 before it weakened again toward the 159–160 area. Japanese officials have emphasized that intervention is not mechanically tied to a single exchange-rate line such as 160 or 162.
Under the existing FIMA repo facility, Japan may be able to obtain dollar liquidity against Treasury collateral rather than first selling those Treasury securities outright, depending on facility access and terms. That distinction matters because repo and liquidation have very different effects on the Treasury market.
But the larger message remains intact: one of the world’s largest official holders of dollar assets is being forced to mobilize balance-sheet resources to stabilize its currency.
The same global dollar-and-rate complex is producing stress elsewhere. Indonesia’s rupiah touched a record low of 18,190 per dollar on June 8, prompting a rare off-cycle Bank Indonesia rate hike the next day. India’s rupee closed at 95.4350 per dollar on August 11, with traders citing likely Reserve Bank of India dollar-selling intervention as oil rose. These currencies are not falling because Japan intervened; each country has its own fiscal, trade and political variables. They are, however, operating inside the same world of expensive dollars, elevated real rates and energy uncertainty.
The vise is global.
The Physical Floor: Energy, Fertilizer and the Cost of Reality
The AI economy is often discussed as though computation exists independently of the physical economy. It does not.
Electricity requires generation. Generation requires fuel, transmission and capital. Semiconductor fabrication requires extraordinary amounts of equipment, chemicals, water and power. Data centers require transformers, generators, cooling systems, substations and long-term utility commitments.
And geopolitical chokepoints still matter.
Before the 2026 Iran conflict, approximately one-third of globally traded urea and nearly half of seaborne sulfur moved through the Strait of Hormuz. Urea production is tightly linked to natural gas through the ammonia production chain; sulfur is a major upstream input in phosphate fertilizer production, including products such as diammonium phosphate. Hormuz disruption therefore transmitted beyond crude oil into fertilizer, food and industrial supply chains.
The World Bank’s April 2026 Commodity Markets Outlook projected fertilizer prices to rise approximately 31% in 2026, driven principally by a roughly 60% increase in urea prices.
This does not mean central banks are incapable of easing. Recent U.S. inflation data have moderated. July CPI rose 3.4% year over year; prices for items other than food and energy rose 2.5%; and energy prices rose 14.7% over the year. Policymakers remain caught between cooling core inflation and a volatile external energy shock.
The appropriate conclusion is therefore not that monetary accommodation is impossible. It is that the inflationary floor limits how aggressively central banks can rescue highly leveraged capital structures without risking another inflation cycle. And that matters profoundly when the world’s largest technology companies are embarking on the largest capital-spending cycle in modern history.
The Deflationary Wedge: When Intelligence Gets Cheaper Than the Infrastructure Built to Produce It
At the same moment that debt, electricity and physical infrastructure have become more expensive, the economic cost of model intelligence is moving in the opposite direction. That is the central contradiction.
Chinese and other open-weight model developers are forcing a dramatic repricing at the model layer. In independent testing reported in August, DeepSeek’s newly released V4-Flash model cost more than 100 times less to run on selected benchmark tests than Anthropic’s premium Claude model. Alibaba’s Qwen family and other open-weight architectures are reinforcing the same competitive pressure.
The precise multiple should not be fetishized. DeepSeek itself announced price increases ranging from 50% to 1,100% across particular V4 models, token categories and times of use. Model economics change too quickly for any fixed multiple to function as a permanent constant.
The more durable proposition is this: for some commercially relevant workloads, useful model capability can now be obtained at one to two orders of magnitude less inference cost than premium frontier offerings on selected benchmark tests.
That creates an extraordinary wedge. On one side of the transaction sits a physical infrastructure stack whose financing cost is increasing. On the other sits a software intelligence layer whose unit cost is being competed downward.
Open Weight Does Not Mean Free
Open-weight models do not eliminate compute demand. In many cases they increase the strategic value of owning or renting compute, because an enterprise can host the model itself, customize it, preserve control over prompts and data, and avoid sending sensitive information to a foreign model developer.
That migration is already visible. Ramp data reported in August showed that approximately 6.1% of businesses spending on AI were using platforms offering open-weight and Chinese-developed models in July, up from 4.5% a year earlier. U.S. cloud providers are increasingly hosting these models so customer prompts can remain within American infrastructure. The same data indicate that open-weight adoption has not yet materially reduced spending on OpenAI and Anthropic, although the adoption rates of the largest proprietary providers are slowing.
Nor does a cheaper token automatically produce a cheaper enterprise system. Private deployment introduces its own costs: compute capacity, engineering, orchestration, security, monitoring, inference optimization, storage, redundancy, compliance and cybersecurity. Open-weight systems can be dramatically cheaper at the model layer while remaining expensive at the enterprise architecture layer.
So the original conclusion must be narrowed. The $20-to-$30 software seat does not simply disappear. What disappears first is the ability to charge a premium merely for access to intelligence that has become increasingly commoditized. The economic moat must migrate upward—to proprietary workflow, data, distribution, integration, security, regulated processes and customer switching costs—or downward, into control of the physical compute infrastructure itself. The middle is being squeezed.
NVIDIA’s $500 Billion Financing Architecture: Compute Becomes an Asset Class
And now we arrive at the loop.
On August 10, NVIDIA announced memoranda of understanding with Apollo Global Management, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish compute-financing platforms intended to mobilize more than $500 billion of third-party capital for AI infrastructure over time. NVIDIA CEO Jensen Huang indicated the company has the option to backstop as much as $125 billion—25% of potential transactions.
Read that again. The world’s dominant AI accelerator supplier is no longer concerned solely with selling processors. It is helping Wall Street build the financing market capable of funding customers that need to acquire and deploy those processors.
The comparison with the telecommunications cycle is no longer rhetorical. It is becoming structural.
But precision matters here too. NVIDIA has not funded a $500 billion lending facility today. The agreements are memoranda of understanding. Individual investment commitments, final terms and deployment schedules have not yet been disclosed. What has changed is the architecture. Institutional investors are being invited to treat AI compute increasingly like an infrastructure asset capable of supporting customized loans, private credit, junior capital and potentially securitized or asset-backed structures.
An Illustrative Compute-Financing Loop
[Institutional Capital / Insurers / Private Credit]
↓
[Compute-Finance Vehicle or Project SPV]
↓
[AI Lab / Neocloud / Enterprise Customer]
↓
[NVIDIA Systems Purchased and Deployed]
↓
[Vendor Revenue Recognized on Hardware Sale]
↓
[Debt Service Depends on Future Compute Utilization and Pricing]
↓
[Collateral Value Depends on Residual Demand for Prior-Generation Hardware]
With NVIDIA potentially providing limited residual-value support around portions of the structure, the loop begins to resemble a hybrid of project finance, equipment finance and vendor-enabled credit.
That does not make it fraudulent, improper or even irrational. Automakers have financed automobiles for generations. Caterpillar finances heavy equipment. Aircraft engines, railcars, power plants and telecommunications equipment have all developed specialized financing markets.
The question is not whether equipment should be financed. The question is whether the duration of the debt and the assumed residual value of the collateral properly match the economic life of the technology. That is the fulcrum.
CoreWeave: The Counterargument That Must Be Taken Seriously
Any credible bear case must confront evidence that contradicts it. CoreWeave provides that evidence.
The specialized AI cloud company reported approximately $2.58 billion of second-quarter 2026 revenue, more than double the prior-year level, and increased its revenue and operating-profit outlook. Its revenue backlog reached approximately $104 billion, and management raised projected 2026 capital expenditures to $35 billion–$39 billion.
Those numbers matter because they rebut the simplistic version of the bubble argument. The data centers are not empty. Demand for compute is real. Customers are signing enormous contracts. Cloud revenue at Microsoft, Google and others continues to grow rapidly.
Therefore the thesis cannot honestly be: Nobody wants the compute.
The thesis is: At what cost of capital, at what utilization rate, at what contractual price, and for how long does today’s compute remain valuable enough to service the financing attached to it?
That is a far harder question. And it is the question the credit market is beginning to price.
Credit Is Not Predicting Default—Yet
Credit-default-swap markets have become an increasingly popular way to express concern about technology leverage. Hedging activity has expanded sharply as hyperscalers increase borrowing and free cash flow comes under pressure.
But the signal should not be overstated. CDS markets for many technology names remain relatively small and illiquid. Sharp spread changes can therefore exaggerate the market’s apparent judgment, and several of the strongest technology companies remain extraordinarily remote from a conventional default scenario.
The correct interpretation is not: The credit market has declared the AI complex insolvent.
It is: For the first time in this cycle, the price and structure of credit are becoming central to the AI investment thesis.
Equity asks how large the opportunity can become. Credit asks whether the borrower gets paid before the collateral becomes yesterday’s technology. Those are not the same question.
The Internal Revenue Code Playbook: How the Compute Buildout Actually Runs Through the Tax Code
The financing structure becomes still more interesting when federal tax law is placed over the top of it. And here, current 2026 law is actually more favorable to the initial compute buildout than a simple five-year depreciation model would suggest.
1. MACRS, 100% Bonus Depreciation and the Zero-Basis Problem
IRC §§168 and 1245
Computers and related peripheral equipment are generally treated as five-year property under MACRS, although an integrated data center contains many other components—electrical systems, buildings, cooling equipment, networking infrastructure and other assets—that can carry different classifications and recovery periods.
But five-year MACRS is no longer the whole story. Under current §168(k), qualifying property acquired and placed in service after January 19, 2025 can generally qualify for 100% additional first-year depreciation when the statutory requirements are satisfied. Eligible property generally includes tangible MACRS property with a recovery period of twenty years or less, meaning qualifying server and compute equipment can potentially be deducted completely in the year it is placed in service.
That flips the original tax mismatch on its head. The economic asset may still have several years of useful service remaining while its federal income-tax basis has already been reduced to zero.
That produces a different set of planning consequences. If fully depreciated equipment is subsequently sold while it still retains substantial secondary-market value, §1245 depreciation recapture can convert gain into ordinary income to the extent of prior depreciation allowed or allowable.
So the tax tension is not principally that the taxpayer is forced to depreciate a three-year economic asset over five years. It is that the tax code may front-load the deduction dramatically while the financing, collateral risk and eventual disposition remain outstanding.
Tax depreciation can occur faster than economic depreciation. Credit depreciation can occur faster than either. Those three clocks do not necessarily agree.
2. Domestic R&E Versus Foreign Development
IRC §§174A, 174, 41 and 280C
The 2025 legislation created a major dividing line between domestic and foreign research expenditures. For tax years beginning after 2024, §174A generally permits a current deduction for domestic research and experimental expenditures. Foreign R&E remains governed by §174 and generally must be capitalized and amortized ratably over 15 years, beginning at the midpoint of the taxable year in which the cost is incurred.
That creates an explicit geographic tax preference. Domestic model research, software development and semiconductor engineering may receive immediate deduction treatment when they satisfy §174A, while qualifying foreign research costs can remain trapped in a fifteen-year recovery schedule.
The §41 research credit remains separately relevant, but taxpayers must account for the coordination rules—including §280C—rather than assuming that the credit and deduction simply stack without adjustment. For multinational AI companies, therefore, the location and documentation of engineering activity is no longer an accounting footnote. It directly influences current cash tax.
3. The Interest Deduction Constraint
IRC §163(j)
The business-interest limitation remains a genuine risk for highly leveraged compute structures, but the law changed materially. Section 163(j) generally limits deductible business interest to the sum of business interest income, 30% of adjusted taxable income, and qualifying floor-plan financing interest, subject to statutory exceptions and special rules.
Beginning with tax years after 2024, depreciation, amortization and depletion are again added back in calculating ATI. In practical financial shorthand, the calculation moved back toward an EBITDA-like base, rather than the EBIT-like regime that applied immediately before the legislative change.
That matters enormously for data-center finance. Massive depreciation deductions do not, by themselves, crush the §163(j) limitation in the manner they would have under the prior EBIT-oriented computation. The constraint now appears elsewhere. If the debt load is sufficiently large, interest expense can still exceed the taxpayer’s available §163(j) capacity. The result is deferred deductibility rather than immediate cash-tax relief, with additional special rules applying to partnerships and other pass-through entities.
The financing question therefore becomes: How much interest-bearing leverage can the project carry before the tax shield ceases to arrive at the same time as the interest bill? That is the real §163(j) pressure point.
4. Obsolescence, Worthlessness and Abandonment
IRC §165 and Rev. Rul. 2004-58
Technological obsolescence does not automatically create a tax deduction. Under §165 and the principles reflected in Rev. Rul. 2004-58, an abandonment loss generally requires both an intent to abandon and an affirmative act evidencing that abandonment. Worthlessness similarly requires an identifiable event establishing a closed and completed transaction.
That creates a potentially important divide between economic impairment and tax disposition. Suppose a frontier cluster can no longer command premium training economics after a new accelerator generation arrives. If the operator continues using that equipment for inference, internal workloads, batch processing or lower-value customers, the asset may have suffered a severe economic markdown—but the taxpayer has not necessarily abandoned it.
A commercial downgrade is not the same thing as a tax abandonment. The taxpayer therefore needs a defensible record of what actually happened: continued use, sale, retirement, physical disposition, abandonment, or another identifiable event establishing worthlessness. A third-party valuation can be powerful evidentiary support, but the Code does not impose a universal appraisal requirement for every §165 claim.
This distinction could become extremely important if lenders mark collateral down long before taxpayers have a tax event allowing them to recognize the same economic loss.
5. True Lease Versus Conditional Sale
Rev. Rul. 55-540
As compute financing grows more sophisticated, another old tax doctrine returns to center stage. Calling a contract a “lease” does not make it a lease for federal income-tax purposes.
The IRS applies a facts-and-circumstances analysis under authorities including Rev. Rul. 55-540. Financing-like indicators can include payments that economically build equity, nominal or bargain purchase options, rent approximating the purchase price plus financing return, and other terms transferring the practical benefits and burdens of ownership to the purported lessee.
A five-year contractual term attached to rapidly evolving hardware does not, standing alone, cause automatic recharacterization. But residual value matters. If the “lessor” has little meaningful residual economic interest, the customer bears substantially all downside, and contractual economics function principally as installment financing, tax ownership can become a serious issue.
That matters because ownership determines who claims depreciation. And when 100% bonus depreciation may be available, who owns the asset for tax purposes at the placed-in-service date can be worth an extraordinary amount of cash.
6. The Foreign-Derived Deduction After 2025
IRC §250 — FDDEI, Not FDII
The international provision has also changed. For tax years beginning after December 31, 2025, the statutory terminology moved from foreign-derived intangible income—FDII—to foreign-derived deduction eligible income, or FDDEI. The §250 deduction percentage applicable to qualifying FDDEI is now 33.34%, rather than the prior 37.5%.
At a 21% corporate tax rate, a fully qualifying dollar receiving the 33.34% deduction produces a theoretical federal rate of approximately 14% before other limitations, interactions and tax attributes. That replaces the old 13.125% headline rate.
For U.S. semiconductor and technology companies, qualifying sales or services to foreign customers for foreign use can therefore continue to receive a meaningful federal tax preference—but physical export revenue does not automatically qualify merely because the purchaser is overseas. The statutory foreign-use and deduction-eligible-income requirements still have to be satisfied.
The policy signal remains unmistakable: Domestic research is favored. Domestic investment is favored. Qualifying foreign-market income produced from the U.S. is favored. The tax code is not standing in the way of the AI capital cycle. In several important respects, it is accelerating the front end of it.
The Snake Eating Its Tail
Put the pieces together.
$5.3 trillion of projected hyperscaler capital expenditure.
Almost $220 billion of bond issuance this year from Alphabet, Amazon and Meta alone.
A 30-year Treasury auction clearing at 5.216%.
Thirty-year real borrowing costs near 3%.
Open-weight models delivering selected workloads at more than 100× below premium-model inference costs on selected benchmark tests.
NVIDIA and six of the world’s largest financial institutions preparing platforms intended to mobilize more than $500 billion of third-party compute finance.
And a tax code permitting qualifying compute hardware to be written off 100% in its first year.
This is no longer merely a technology story. It is a capital-structure story.
The physical cost of building compute is rising at the same time that the unit cost of intelligence is falling. The supplier of the dominant hardware is helping organize financing for the customers purchasing that hardware. The financial sector is increasingly treating the hardware as financeable collateral. The federal tax system accelerates the initial deduction. The customer then has to produce sufficient economic value from that hardware—year after year—to service financing negotiated against assumptions about utilization, pricing and residual value. And each new silicon generation tests those assumptions again.
That is the snake.
What Breaks—and What Survives
The most likely failure mode is not that artificial intelligence suddenly becomes worthless. Nor is it necessarily that the data centers go dark.
The more plausible repricing occurs inside the capital stack.
A data center can remain full while the equity beneath it is impaired.
A GPU can remain busy while the price per unit of inference collapses.
A borrower can report rising revenue while free cash flow remains inadequate for the amount of capital required to sustain that growth.
A lender can remain money-good while earning a far lower return than originally underwritten.
And silicon can remain technically functional while losing the premium economic role on which its original residual-value assumption depended.
The durable assets may therefore be different from the assets that receive most of today’s attention: land near transmission, interconnection rights, substations, fiber, cooling infrastructure, generation capacity, utility agreements, permitted data-center shells. Those assets do not become obsolete every time NVIDIA changes an architecture. The silicon sitting inside them might.
That is why the terminal phase of this cycle may resemble the telecom bust only in part. After telecommunications collapsed, distressed buyers acquired infrastructure whose usefulness increased as the internet matured. After an AI-credit repricing, buyers may again acquire valuable physical infrastructure at distressed prices—but this time they may have to separate the long-duration value of power and connectivity from the short-duration value of the processors installed inside it.
That distinction will determine who survives the clearing process. It will also determine who has the tax basis, losses, depreciation recapture, interest carryforwards and restructuring flexibility when the clearing price finally arrives.
Next: The Buyers Waiting at the Bottom
In the subscriber edition, we will move from the financing loop to the other side of the trade. We will examine how sovereign wealth funds, infrastructure managers, private-credit firms and strategic operators are positioning around AI infrastructure; which assets retain durable value after a compute repricing; how distressed data-center transactions should be analyzed from a tax-basis and depreciation perspective; and where the next generation of buyers may find the equivalent of yesterday’s dark fiber.
Because every capital cycle eventually reaches the same final question:
Not what did it cost to build?
What is it worth when the forced seller has to sell it?