When Code Eats Capital
What happens to debt vs. equity when the predictability horizon collapses?
A research paper examining what happens to the capital structure - the relationship between debt and equity - when AI disruption compresses the time horizon over which business viability can be assessed. The approach is evidence-first: start with what’s actually happening in credit markets, in software economics, and in the physical infrastructure of AI, then reason from first principles to the implications.
I. The Question
The entire architecture of corporate finance rests on one assumption: that you can distinguish between debt and equity risk because you can forecast business viability over the life of a loan. Debt holders accept lower returns for priority claims. Equity holders absorb the uncertainty. The system works because the predictability horizon - the window over which you can reasonably assess whether a business survives - is longer than the duration of most credit instruments.
This paper asks a simple question: what happens to that distinction when the predictability horizon collapses?
I’m not starting with a thesis. I’m starting with three sets of evidence that, taken together, suggest something structural may be shifting in the relationship between debt and equity for technology-exposed companies. The evidence comes from credit markets, from the economics of AI capability, and from the physical infrastructure being built to support it. The question is whether these are isolated data points or symptoms of a deeper change in the capital structure itself.
II. The Evidence - What’s Actually Happening
A. Private credit is cracking in ways that don’t fit the normal cycle
HARD DATA
BlackRock TCP Capital marked a $25 million loan to Infinite Commerce Holdings - an Amazon aggregator - from 100 cents on the dollar to zero in a single quarter (Q3 to Q4 2025). This was the second consecutive quarter with a par-to-zero writedown, following Renovo Home Partners. As of December 31, 2025, 14 portfolio companies were on non-accrual status, representing 9.7% of the portfolio at cost, up from nine companies at end of September. (BlackRock TCP Capital 10-K; Bloomberg, March 5, 2026)
Blackstone’s $82 billion Private Credit Fund (BCRED) saw record redemption requests of 7.9% of assets in Q1 2026 - up from 4.5% in Q4 2025 and 1.8% in Q3 2025. Blackstone raised its redemption cap to 7% and, together with 25+ senior executives, injected $400 million of personal capital to meet all requests. Net withdrawals: $1.7 billion. JPMorgan called it “a significant expression of souring investor sentiment on direct lending.” (SEC filing; JPMorgan; CNBC, March 3–5, 2026)
Blue Owl Capital ended regular quarterly redemptions from its retail-oriented private credit vehicle and sold $1.4 billion in loan assets to manage liquidity. (InvestmentNews; CNBC, February–March 2026)
BlackRock’s private credit fund received $1.2 billion in withdrawal requests representing 9.3% of fund assets. Redemptions were capped at 5%, leaving roughly half of requesting investors locked in. (European Business Magazine, March 2026)
THE STRUCTURAL POINT
The conventional explanation is cyclical: rising rates, tightening conditions, a correction after years of easy money. That’s part of it. But three things don’t fit neatly into a cyclical frame:
Speed of repricing. Par to zero in one quarter is not normal credit deterioration. It’s the market discovering that what it classified as a lending relationship was actually an equity-like bet on a business model’s survival.
Sector exposure. CNBC reported that redemption pressure is being influenced by exposure to SaaS companies, as investors worry that AI tools could erode traditional software business models. Blue Owl is a significant direct lender to the software sector. The Cliffwater Corporate Lending Fund ($29.7B) carries 24% exposure to information technology.
Redemption dynamics. In a social media world, as Silicon Valley Bank demonstrated, fear spreads faster than funds can liquidate illiquid positions. The rational move for any individual investor is to redeem first. The illiquidity premium they were paid to accept becomes a trap when the underlying assets can be disrupted faster than the redemption windows allow.
First-principles question: If the assets backing private credit loans can lose their competitive position in months rather than years, is the credit market pricing these as debt instruments or as equity instruments with a coupon?
B. The cost of building competitive software has collapsed by orders of magnitude
HARD DATA
Stanford HAI’s 2025 AI Index Report: inference costs for GPT-3.5 - level performance fell 280-fold between November 2022 and October 2024 ($20/million tokens to $0.07). Hardware costs declining 30% annually. Energy efficiency improving 40% annually. Open-weight models have closed the performance gap with proprietary models from 8% to 1.7% in a single year. (Stanford HAI, April 2025)
OpenClaw, NVIDIA’s open-source autonomous agent framework, became the most downloaded open-source software in history within three weeks of release. (Jensen Huang; Jordi Visser/22V Research)
Alex Finn, an independent developer, operates five autonomous AI agents across three Mac Studios and a Mac Mini ($25,000 total hardware, $200-$250/month subscriptions), producing continuous software output 24/7. He replicated a feature that took a funded startup (Cursor) weeks and presumably millions to develop - in five minutes. (Moonshots Podcast #237)
Mac Mini sales are growing at exponential rates as consumers discover local AI. Apple’s unified memory architecture enables running frontier-class open-source models on consumer devices. (Moonshots Podcast; multiple industry sources)
THE STRUCTURAL POINT
The cost of competing with an incumbent software company has dropped from millions of dollars and years of engineering to a consumer hardware purchase and a monthly subscription. This isn’t theoretical; it’s operational today.
Think about what this means from first principles. A lender extending a five-year term loan to a software company is making an implicit bet that the company’s revenue stream will persist over the loan’s duration. That bet rests on competitive moats: accumulated feature complexity, enterprise relationships, switching costs. When any mid-sized company - or a single developer with a Mac Mini - can generate a custom alternative tailored to exact workflows, the competitive moat isn’t a permanent structure. It’s a temporary feature of a market that hasn’t yet adjusted.
First-principles question: If the moats that justified the credit are dissolving, and the lender’s risk profile now depends on whether the borrower can adapt to AI disruption faster than the loan matures, how is that functionally different from equity risk?
C. The infrastructure bet is straining under its own weight
HARD DATA
Oracle: $106 billion in total debt. Negative free cash flow of $13.2 billion on a trailing four-quarter basis as of Q2 fiscal 2026. Stock down 54% from September 2025 highs, erasing $463 billion in market value. Took on $58 billion in new loans in 60 days to finance data centers. Bonds now trading like junk-grade credit. Planning to cut 20,000–30,000 jobs (12–18% of workforce). Free cash flow not expected to turn positive until approximately 2029–2030. (Oracle 10-Q/SEC filings; Bloomberg; CNBC; Fortune, December 2025–March 2026)
Oracle’s gross margin projected to fall from 77% (fiscal 2021) to approximately 49% by 2030 as revenue mix shifts from high-margin legacy software to lower-margin AI infrastructure. Analysts project roughly $34 billion in cumulative negative free cash flow over the next five years. (FactSet analyst consensus; CNBC, December 2025)
Oracle and OpenAI scrapped plans to expand a Texas data center -negotiations collapsed over financing and OpenAI’s changing needs. OpenAI reset total industry CapEx expectations from $1.4 trillion to $600 billion. (22V Research; Bloomberg)
Jamie Dimon warned that bankruptcies like Tricolor and First Brands may signal broader private credit stress. (JPMorgan; multiple sources)
THE STRUCTURAL POINT
Oracle is the clearest case study of the capital structure question in real time. The company’s $106 billion debt load was accumulated to finance an AI infrastructure buildout that assumes enterprise adoption will generate returns by 2027–2028. But the enterprise adoption is lagging. Models are ready; organizations are not. And the competitive landscape is shifting beneath the bet: open-source models running on consumer hardware, inference moving to the edge, startups adopting faster than enterprises.
The second-order effect: if inference shifts from centralized data centers to edge devices, the infrastructure being financed with debt becomes less valuable than projected. Not worthless - but worth less. And in credit markets, “worth less than projected” on a levered balance sheet can cascade quickly.
The third-order effect: Oracle’s bonds are already trading like junk. If the market begins to reprice AI infrastructure debt broadly, it tightens financing for the entire buildout - which in turn slows the enterprise adoption that the debt was supposed to finance. A reflexive loop.
First-principles question: Is Oracle’s debt a credit instrument backed by forecastable cash flows, or a leveraged equity bet on whether centralized AI infrastructure remains the dominant architecture? And if lenders can’t answer that question with confidence, what is the appropriate risk premium?
III. First Principles -Why the Predictability Horizon Matters
Before drawing conclusions, I want to lay out the first-principles framework that connects these three evidence sets. The argument isn’t that AI is bad for business. It’s that AI compresses the time horizon over which business outcomes can be forecast, and that compression has specific, measurable consequences for the capital structure.
The debt-equity distinction is a function of predictability
At its core, the difference between lending to a company and owning a company is the lender’s ability to assess, with reasonable confidence, that the business will generate enough cash to service the loan over its duration. When that confidence is high, debt behaves like debt: lower risk, lower return, priority in the capital structure. When that confidence collapses, debt starts behaving like equity: the lender is no longer senior to uncertainty. They’re embedded in it.
AI is compressing the horizon
Traditional credit analysis assumed industry structure was relatively stable over a 3–7 year loan maturity window. The evidence above suggests that AI is restructuring competitive dynamics on much shorter timescales. A 280-fold decline in inference costs over 18 months. A competitive feature replicated in 5 minutes. Open-source models closing the performance gap from 8% to 1.7% in one year. These are not gradual shifts. They compress the window within which a business model can be confidently assessed.
Second-order effects
Duration mismatch: If disruption cycles are 18–24 months and loan durations are 3–7 years, the lender is effectively making a multi-cycle bet without multi-cycle visibility. A five-year term loan to a software company is no longer a credit instrument. It’s a structured equity bet with a coupon.
Illiquidity premium inversion: Private credit sold investors an illiquidity premium. When the underlying businesses can be disrupted faster than redemption windows allow, illiquidity becomes a cost, not a compensated feature. The premium inverts.
Cost of capital convergence: If lenders begin pricing AI disruption risk into credit, the cost of debt for AI-exposed businesses rises toward the cost of equity. This changes capital structure decisions, M&A financing, and leveraged buyout economics.
Third-order effects
The funding asymmetry: OpenAI raised $110 billion in private equity - four times the largest IPO in history. Anthropic’s annual run rate is $25 billion and growing at hyper speed. These disruptors are equity-funded and largely immune to credit tightening. The companies they’re disrupting are leveraged and exposed. The disruptors don’t need the credit markets. The disrupted depend on them.
The zero-capital competitor: Alex Finn’s autonomous organization costs $25,000 in hardware. When he says a displaced worker could build a $5 million company for $200, he’s describing a world where the capital required to compete has collapsed. The credit underpinning the incumbents was priced for a world where competition required capital. That world is ending.
The talent reflexivity trap: Software companies facing multiple compression lose their ability to retain talent through stock-based compensation. Talent leaves. Competitive position weakens. Credit quality deteriorates. The stock-based compensation model that made SaaS companies attractive borrowers becomes the mechanism of their decline.
The buyback disappearance: Hyperscalers are spending so aggressively on AI infrastructure that cash return as a percentage of net income has fallen to historically low levels. Buybacks - which have been a major support for stock prices - are decelerating. This removes a floor from equity prices, which in turn affects the collateral underpinning credit.
IV. Historical Parallel - LTCM and the Lesson of Crowded Models
There’s a historical precedent worth examining. In 1998, Long-Term Capital Management - a hedge fund employing leveraged convergence trades across global bond markets - nearly collapsed when a liquidity shock (Russia’s default) caused spreads to diverge instead of converge. LTCM’s positions had been correct from a valuation standpoint but failed because of financing constraints: margin calls forced deleveraging at distressed prices, and the size of their positions meant there were no buyers.
The parallel is not that private credit in 2025-2026 is identical to LTCM. The parallel is structural:
Crowded positioning: LTCM’s trades were crowded - multiple funds running similar models on similar positions. Today, private credit has concentrated heavily in the same sectors (24% IT exposure in Cliffwater; significant SaaS exposure at Blue Owl) using similar underwriting assumptions about recurring revenue predictability.
Leverage against a single variable: LTCM was leveraged against the assumption that spreads converge. Private credit is leveraged against the assumption that software business models remain defensible over loan durations. Both assumptions are correct in normal conditions and catastrophically wrong when the regime shifts.
The policy response: LTCM required a Fed-coordinated $3.65 billion recapitalization. The S&P didn’t bottom until the Fed cut rates. This is the pattern Jordi Visser at 22V Research highlights: historically, every major drawdown coinciding with the financial sector’s 200-day moving average turning down has required a policy intervention to find a floor -the 2015 Shanghai Accord, the 2018 Fed pivot, COVID stimulus, the 2022 rate pause.
The critical difference: LTCM’s problem was temporary. Spreads did eventually converge. The question for private credit lenders exposed to software companies is whether AI disruption is temporary or structural. If it’s structural, the convergence never comes. The loan doesn’t recover. And the credit instrument was equity all along.
Visser also makes a point worth sitting with: after both the LTCM crisis and the 2007 Quant Quake, markets went to higher levels. The deleveraging was painful but not terminal. The question is whether this time is the same - painful but ultimately a buying opportunity - or whether something has fundamentally changed about the assets being delevered.
V. Where I Might Be Wrong
I’ve assembled evidence suggesting that AI disruption is compressing the predictability horizon faster than credit markets can adjust. Here’s where that reasoning is most vulnerable.
Enterprise stickiness may outlast the disruption cycle
Enterprise software has survived predictions of its demise before. Switching costs are real. Integration complexity is real. Regulatory requirements lock in existing systems. If enterprise adoption of AI-native alternatives lags long enough - and the evidence suggests it’s lagging - existing credit instruments may mature before the disruption arrives in full force. The Ford analogy cuts both ways: Ford survived 36 years at the same stock price. The equity went nowhere, but the debt got paid.
This could be cyclical, not structural
Rising rates, oil price shocks, and tightening conditions are sufficient to explain private credit stress without invoking AI at all. Blackstone’s fund has a 9.8% annualized return since inception. BlackRock’s TCP Capital noted that 91% of its writedowns came from deals underwritten in 2021 or earlier, hit by prolonged high rates. I need to be careful not to attribute to AI what the macro environment explains.
The edge computing thesis may be premature
Running five AI agents on consumer hardware is impressive but narrow. Enterprise workloads - large-scale training, multi-modal inference, mission-critical applications - still require centralized compute. If data center demand holds, Oracle’s bet looks less like a stranded asset and more like well-timed infrastructure investment. The “your iPhone runs a frontier model in three years” timeline could be off by a factor of two or three.
Sample size
The disruption evidence draws heavily on one developer’s workflow, one macro commentator’s weekly update, and one podcast. The full paper needs broader cross-referencing: VC funding data, software company earnings revisions, actual default rates in private credit, historical precedents for technology-driven credit events.
If the structural thesis is correct, the repricing has barely begun. If the cyclical thesis is correct, this is a buying opportunity in disguise. The honest answer is that I don’t know yet which it is. But the evidence is tilting structural, and the market is still pricing cyclical.
VI. Implications - If the Evidence Holds
I want to be clear: this section is conditional. These implications follow only if the structural thesis survives further scrutiny. I’m laying them out not as predictions but as the logical consequences of the evidence assembled above.
For credit markets
Credit duration assumptions need revision for any AI-exposed sector. A five-year term loan to a software company should probably be stress-tested against 18–24 month disruption cycles, not 5–7 year default assumptions.
The illiquidity premium for tech-exposed private credit may need to become an illiquidity discount. Investors are being paid for a risk (illiquidity) that may be smaller than the risk they’re actually taking (business model obsolescence).
CDX high yield has barely moved. Breadth hasn’t seen a true panic day. If the structural thesis is right, the credit market hasn’t repriced yet. The adjustment is ahead.
For the broader market
If financial conditions continue to tighten -financials below the 200-day moving average, VIX trending higher, credit spreads widening - historical precedent says markets don’t bottom without a policy response. The question is what triggers that response and how much damage accumulates before it arrives.
The commodity supercycle thesis may be the mirror image of the software repricing. Energy ($2.4T market cap) and materials ($1.5T) need to grow relative to software ($5.4T) if AI drives a structural supply shock for physical infrastructure while deflating digital goods.
For bitcoin and scarce assets
This connects to the framework I’ve been developing across my previous papers. If AI deflation eventually compels a monetary policy response (the slingshot thesis from Bitcoin Investor Week), and if the credit markets are simultaneously repricing tech-exposed debt toward equity-like risk, then capital seeking safety flows toward genuinely scarce assets - the gravitational basin Dan Hillery and Mason Foard described at True North.
The digital credit instruments I wrote about in The Seasoning Clock - STRC, SATA, and their successors - are building track records through exactly the kind of stress that validates the three-year seasoning window. The instruments designed for the new world are seasoning while the instruments designed for the old world are cracking.
The connection: the same AI disruption that is breaking the traditional capital structure is strengthening the case for assets that don’t depend on it.
VII. What I’m Watching
High conviction: The cost of competing with incumbent software companies has collapsed structurally, not cyclically. A 280-fold decline in inference costs, open-source models at near-parity with proprietary ones, and autonomous agents operating on consumer hardware represent a permanent shift in the economics of software. The credit market has not yet repriced for this.
High conviction: Private credit stress is more than cyclical. The sector concentration in IT and SaaS, combined with AI-driven competitive disruption, introduces a risk that standard credit models weren’t built to assess. Par-to-zero in one quarter is a regime signal, not a one-off.
Probable but timing-uncertain: The edge computing shift threatens centralized infrastructure bets. The trajectory is clear but the timeline for enterprise-relevant disruption is genuinely uncertain. Oracle is the test case.
Worth monitoring: Whether any major private credit fund defaults or gates permanently. Oracle’s ability to refinance its $106 billion debt stack. Bitcoin’s correlation with software stocks - if it breaks to the upside, the market is beginning to distinguish between tech proxies and scarcity assets. The Fed’s response function as credit conditions tighten.
VIII. Conclusion
The conclusion should not overstate the case. The evidence is suggestive, not conclusive. The honest framing: something may be shifting in the relationship between debt and equity for technology-exposed companies, and the market is not pricing for it.
Return to the question from Section I. The answer isn’t that debt and equity are the same thing. The answer is that the boundary between them - which is a function of predictability - is moving. And when boundaries move in credit markets, the adjustment is rarely smooth.
Connect to the broader arc: position for the outcome, size for the delay. The delay, as I wrote at True North, is where the information advantage lives. That applies to bitcoin. It also applies here. The market is pricing the credit cycle. If what’s happening is structural, the repricing hasn’t fully started.
The market is pricing the cycle. The evidence suggests the capital structure is shifting. The gap between those two things is where the risk - and the opportunity - lives.
Research Sources
Primary data cited in this paper
BlackRock TCP Capital Corp. 10-K filing (Q4 2025); Bloomberg, March 5–6, 2026 (Infinite Commerce, Renovo Home Partners writedowns)
Blackstone Private Credit Fund SEC filing, March 2026 (BCRED redemptions, 7.9% of assets)
Blue Owl Capital redemption halt and $1.4B asset sale (InvestmentNews; CNBC, Feb–March 2026)
Oracle Corp. 10-Q, SEC filings, Q2 fiscal 2026 ($13.2B negative FCF trailing; $106B debt; $50B capex guidance)
Stanford HAI 2025 AI Index Report (280-fold inference cost decline; 30% annual hardware cost decline; open-weight model parity)
LTCM crisis: Federal Reserve History; President’s Working Group Report (1999); PBS Frontline
Primary analysis and commentary
22V Research (Jordi Visser) - weekly commentary on financial conditions, turbulence model, software repricing, LTCM parallel
Moonshots Podcast #237 (Alex Finn, Peter Diamandis, AWG) - OpenClaw workflows, hybrid economics, SaaS replication
Author’s previous papers: “The Supersonic Tsunami Meets the Hardest Asset” (BIW 2026) and “The Seasoning Clock” (True North 2026)
Additional sources for the full paper
Cliffwater Corporate Lending Fund composition (SEC filings; cliffwaterfunds.com)
PIMCO, Apollo (Mark Rowan), Lloyd Blankfein public statements on private credit
Steve Eisman podcast on insurance – private credit – retail chain
Jim Chenos analysis of Oracle
OpenAI funding rounds ($40B March 2025; $110B subsequent); Anthropic ARR data
CDX high yield spread data (current vs. historical stress periods)
All-In Podcast: SaaS structural risk discussion (Chamath, Sacks)
Andre Karpathy “Software 2.0” essay
Peter Murdoch / Insight Partners podcast on agents, ASICs, edge computing
Apple Mac Mini/Studio sales data; M5 chip marketing materials
VC funding trends and software company earnings revision data
Historical precedents: 2007 Quant Quake; 2015 Shanghai Accord; 2018 Fed pivot
Methodology and Disclosure
This paper is based on the author's analysis of primary source transcripts, SEC filings, and publicly available market data as cited throughout. Empirical claims were cross-referenced against original filings and reporting. Research, structural editing, and drafting were conducted with the assistance of Claude Opus 4.6 (Anthropic). Images were generated using Nano Banana AI.