The parallels to 1999 are unmistakable. Then, the Internet promised to change the world – and it did – but not before vaporizing trillions of dollars of investor capital. The exuberance of that period was not built on ignorance but on conviction: everyone knew the Internet would matter; they simply misjudged how much, how soon, and at what cost. Today’s AI boom carries the same blend of insight and excess. The technology is real and the applications are profound, but the economics remain unproven. Faith in AI’s inevitability has replaced analysis of its profitability.

AI-related spending has become one of the principal drivers of U.S. economic resilience. According to multiple estimates, it added roughly one percentage point to U.S. GDP growth in the first half of 2025.¹ In the second quarter alone, the ten largest U.S. technology firms – Alphabet, Meta, Amazon, Microsoft, Tesla, Apple, Nvidia, Oracle, Broadcom, and IBM – spent at an annualised rate of US$426 billion on capital investment, up 73 percent from a year earlier.² Economy-wide technology spending as a share of GDP has now surpassed the peak seen during the late 1990s.

That surge in investment has powered not only GDP but also stock markets. Since the launch of ChatGPT, the combined market value of these ten companies has increased by nearly US$10 trillion, equivalent to roughly one-third of annual U.S. GDP. Historically, the early stages of repricing around transformative technologies are rational; capital chases genuine innovation. But enthusiasm has a way of turning to excess. Whether the current stampede will prove wise depends on the economic return on investment, a question the data have yet to answer.

The near-term math is sobering. Total revenue for the “big ten” reached about US$2.4 trillion (seasonally adjusted annual rate) in the second quarter, up 15 percent from a year earlier.³ Revenue growth, however, remains well below the pace of investment, lifting the group’s capex-to-revenue ratio from 10 percent in early 2022 to roughly 18 percent today.⁴ Most profits still derive from established franchises—advertising for Meta, iPhone sales for Apple—rather than from AI itself. In other words, the AI story is driving valuations while legacy businesses are funding the experiment.

History also suggests that genuine productivity benefits from new technologies take longer to materialise than investors expect. While there are promising micro-level efficiencies, U.S. labour productivity has not yet broken decisively above its multi-decade trend. Despite the fanfare, the macro concern across advanced economies remains sluggish productivity growth. A recent MIT study found that 95 percent of organisations experimenting with generative AI are earning zero return on their investments.⁵

The problem is not intent but arithmetic. The AI buildout is exceptionally capital-intensive and depreciates at an extraordinary pace. Unlike traditional infrastructure such a poles, wires, and transmission lines that last decades, the useful life of GPUs and related processing units is measured in years, not decades. Each generation of chips is rendered obsolete by the next within three to four years, compressing the window for earning an adequate return. The physical data centres and power connections will endure; the computational core will not. As a result, the hurdle rate for profitability is far higher than most models imply. It is increasingly difficult to see how current market enthusiasm can be reconciled with the investment’s short economic life.

The AI boom’s initial financing has come largely from equity, venture capital, and retained earnings. Big Tech’s profitability and deep cash reserves gave it an extraordinary ability to self-fund early stages of the revolution. But as investment needs swell, particularly for data centres, semiconductor fabrication, and power generation, companies are turning decisively toward credit markets. Debt, not equity, will fund the next leg of AI’s expansion.

The scope has already shifted. According to Bank of America, AI-linked technology firms issued more than US$75 billion in U.S. investment-grade debt during September and October 2025 alone, more than double the sector’s average annual issuance of US$32 billion over the past decade.⁶ The total included US$30 billion from Meta and US$18 billion from Oracle, plus new borrowing from Alphabet and others. Barclays now identifies AI-related issuance as the key determinant of U.S. investment-grade supply in 2026,⁷ while J.P. Morgan estimates that AI-linked companies account for 14 percent of its investment-grade index, surpassing U.S. banks as the largest sector exposure.⁸

These are not speculative borrowers in the traditional sense; they are profitable, well-capitalized enterprises with legitimate business cases. Yet the sheer scale and complexity of their financing structures are giving investors pause. Meta’s US$27 billion off-balance-sheet financing with Blue Owl Capital – the largest private-credit deal ever recorded – keeps debt off Meta’s books but transfers it into opaque vehicles owned by yield-hungry investors. Such structures may provide flexibility, but they also disperse risk into corners of the market that are illiquid and difficult to price. It is this migration of leverage from transparent to opaque balance sheets that prompted the Bank of England to warn of “pockets of vulnerability” within the global financial system.⁹

The deeper the industry digs into infrastructure, the more circular its financing becomes. OpenAI owns a warrant to buy up to a 10 percent stake in AMD. Nvidia has invested roughly US$100 billion in OpenAI. Microsoft, OpenAI’s largest shareholder, is also one of Nvidia’s biggest customers, accounting for nearly 20 percent of its revenue, and a major client of CoreWeave, another Nvidia-backed venture. Oracle, for its part, is financing OpenAI’s US$300 billion data-centre buildout, which OpenAI will pay for using capital from investors who rely on Nvidia’s chips.

It is an ecosystem in which suppliers finance customers, competitors invest in one another, and equity stakes blur the distinction between partnership and exposure. The arrangement may be efficient in theory, but in practice it resembles the reflexive financing that defined the late-1990s telecom boom, when equipment makers lent to carriers who borrowed to buy more equipment. When funding tightened, the illusion of demand evaporated. The same pattern could repeat: when the flow of capital slows, valuations and balance sheets will adjust together.

The credit footprint of AI now extends well beyond the investment-grade market. High-yield issuance tied to AI has surged, with TeraWulf, a bitcoin miner turned data-centre operator, raising US$3.2 billion in BB-rated bonds, and CoreWeave issuing US$2 billion in high-yield debt earlier this year. These deals are small relative to the broader market, but they mark the familiar progression of optimism down the credit spectrum.

Private credit is also becoming a significant source of financing. UBS estimates that private-credit loans to AI-related borrowers nearly doubled over the twelve months through early 2025.¹⁰ Morgan Stanley projects that non-bank lenders could supply over half of the US$1.5 trillion required for the global data-centre buildout through 2028 (Figure 1).

Securitization is re-emerging as well. Digital-infrastructure asset-backed securities (ABS) – bonds backed by long-term rent payments from data-centre tenants – have expanded eightfold in five years to about US$80 billion. Bank of America expects that figure to reach US$115 billion by the end of 2026, with roughly two-thirds of issuance tied directly to data-centre construction.¹¹ These instruments are standard financing tools, but their proliferation in a short span and their dependence on illiquid assets echo the layering that preceded the global financial crisis.

Figure 1: Morgan Stanley (MS) estimates for how the global data centre buildout will be financed (October 2025)

In this way, the logic of the cycle is self-reinforcing. Companies that spend the most on infrastructure drive the narrative that justifies further investment by others. As in past technological manias, optimism migrates from valuation to financing. The danger is not that the technology fails, but that the return profile cannot keep pace with the capital committed to it. The wisdom of this stampede will ultimately be judged by whether the economic return on investment matches its financial cost.

None of this diminishes AI’s transformative potential. Like the Internet before it, the technology will ultimately reshape how economies function and how capital is deployed. But revolutions of this scale follow a familiar pattern: optimism, overinvestment, correction, and consolidation. The productivity gains outlast the financial losses, but the investors who finance the early excess seldom own the eventual winners. Railroads in the 19th century, electricity in the early 20th, and fiber-optic networks at the turn of this century all followed this trajectory. AI will be no exception.

For credit markets, the implications are clear. As capex rises faster than cash flow, companies will lean more heavily on debt. Investment-grade issuance will expand further, private credit will deepen its exposure, and securitization will spread risk into less transparent corners of the system. The temptation for lenders will be to chase yield by funding complexity. The wiser course will be to underwrite what is real (i.e., cash-flowing, collateralised, and verifiable) rather than what is fashionable.

From a lending standpoint, AI’s credit dimension is still young but growing fast. For now, the borrowers are large, liquid, and well-rated. But as the ecosystem broadens, the risk will migrate to smaller operators and second-tier suppliers, where disclosure is weaker and covenants thinner. That is where underwriting discipline will matter most. The current environment rewards patience. We prefer to lend into dislocation, not euphoria. And to finance the infrastructure that survives the shake-out, not the speculation that precedes it.

For the moment, AI investment continues to underpin growth, earnings, and sentiment. But beneath the surface, leverage is building, complexity is rising, and transparency is declining – the same combination that has marked every great market inflection. The credit cycle is always quieter than the equity cycle until the very end. When the adjustment comes, it will not disprove the promise of AI; it will simply reprice the capital that made it possible.

[1] Macrobond, September 2025
[2] Macquarie Economics, September 24, 2025
[3] Bloomberg
[4] Macquarie Economics, September 24, 2025
[5] https://nanda.media.mit.edu/
[6] BofA Global Research, October 15, 2025
[7] Barclays Research, October 2, 2025
[8] “At $1.2 Trillion, More High-Grade Debt Now Tied to AI Than Banks,” October 7, 2025, Bloomberg
[9] “Bank of England warns of growing risk that AI bubble could burst,” October 8, 2025, The Guardian
[10] “Private Credit-Powered AI Boom at Risk of Overheating, UBS Says,” August 18, 2025, Bloomberg
[11] BofA Global Research, October 15, 2025

The post Echoes Of 1999: Equity Euphoria, Credit Consequences (Q3-25) appeared first on Third Eye Capital.

Consider Canada’s largest supermarket chains, home-grown Loblaw and U.S. retail giant Walmart. They emerged from the pandemic with record earnings, fortified balance sheets, and even greater negotiating leverage over their supply chains. Their size allows them to dictate pricing, shelf placement, and promotional calendars to co-packers and smaller food producers. These dominant players can lock in multi-year retail pricing with their customers and pass most cost increases upstream, while simultaneously imposing “vendor programs” that reduce the supplier’s take even further: early-payment discounts, slotting fees, promotional chargebacks, and penalties for missing on-time-in-full delivery targets.

For a smaller producer, these terms are not optional; refusing them risks losing the contract entirely. The result is a self-reinforcing dynamic: the retailer preserves its own gross margins and cash conversion cycle by pushing working capital and volatility onto the supplier. The supplier’s margin erosion reduces its ability to invest in automation, product innovation, or marketing, further cementing its dependence on the buyer. Over time, this dependence makes it even harder to negotiate better terms, ensuring the imbalance perpetuates itself.

It is not simply that the big players negotiate harder. They set the price to the end customer, and then cascade the economics backwards, telling suppliers to make it work. In concentrated markets like Canada’s, where a handful of companies can make or break an industry, these terms are not just aggressive, they are de facto standard.

In the supply agreements and vendor guides that govern these relationships, there is no visible malice, only a quiet but relentless transfer of margin. A co-packer for a national grocery chain will find its gross price whittled away by promotional chargebacks and service-level penalties, even as it is told to hold extra stock to ensure on-time delivery. A precision manufacturer in aerospace will invest in new machining capacity at a customer’s request, only to watch volumes lag and annual “cost-down” clauses eat away at unit prices. An energy services contractor may win a coveted “preferred vendor” slot with a resource major, then discover that it is expected to maintain crews and equipment on call, unpaid, until the next job. The common element is a mismatch of leverage. The customer can diversify or reshore or simply move to the next bidder; the supplier has fewer options and less time.

In good years, these pressures are masked by growth, cheap credit, or both. A thin-margin supplier can refinance or roll over its bank line, paying today’s bills with tomorrow’s sales. But when inflation runs through materials and freight, when interest rates reset upward on floating debt, when borrowing bases shrink because a single large customer now accounts for too much of receivables, the mathematics change. The erosion of a few hundred basis points of margin becomes a liquidity crisis. Debt service that was easily covered at six percent profit margins cannot be met at one percent. The default is not a shock; it is the predictable end state of a system in which the stronger party captures the upside and shifts the downside.

In Canada, the problem is sharper than it is in the U.S. Our corporate landscape is more concentrated, and our banking sector more unified in its approach to risk. When a borrower breaches a covenant, it is not just one lender pulling back; the whole sector tends to tighten in unison. Asset-based lending availability is re-margined with new reserves, cutting liquidity at the worst moment. For middle-market companies that have already stretched their vendors and drawn down their lines to meet customer demands, the withdrawal of bank support can be terminal. They arrive in formal restructuring proceedings – NOIs and CCAAs – later than they should, having exhausted their cash to keep supplying a customer that cannot or will not adjust the terms.

We have seen it play out in every sector we touch. In one case, a Western Canadian industrial supplier saw its largest OEM customer cut prices annually under a long-term agreement, levy new quality penalties, and insist on expensive retooling without guaranteeing volume. Margins went from healthy to barely positive in a year; the senior lender, wary of concentration and foreign receivables, slashed the borrowing base. In another, a national food distributor’s revenue grew in step with a major retailer’s promotions, but cash drained away through deductions and extended terms. The facility that funded its inputs was suddenly reduced because too much of its receivables sat with one debtor. In both cases, the companies were not failing because they had lost their markets or their competence. They were failing because the contractual architecture of their business relationships left them unable to capture enough value to pay their bills.

The most dangerous shocks in this environment are those that give the stronger party cover to hold the line or push harder: tariffs, for example. Large public companies often manage to turn trade disruptions to their advantage, re-engineering supply chains, pushing price increases downstream, and using the disruption to consolidate share. Their smaller suppliers, lacking the same strategic options, see costs rise and orders fluctuate without any compensating change in terms. A ten percent tariff on key inputs can be absorbed, in theory; but, in practice, for a supplier on thin margins with variable-rate debt, it becomes a debt service problem within months.

When the squeeze comes, traditional credit often makes things worse. Loan agreements are designed to protect lenders from loss, not to give borrowers time and flexibility to adapt. Covenants are based on trailing numbers; borrowing bases are sensitive to reserves for slow-moving stock or concentrated receivables; cross-defaults can shut off new money overnight. Customers with their own working-capital targets to hit may stretch payables further or impose new demands on suppliers, knowing the bank, not they, will be left holding the risk. The result is a narrowing corridor in which management can operate, and a rising probability that the situation will tip into formal default.

It is in these corridors that we operate. The businesses we step into are almost always in the grip of this corporate class system – caught between a demanding customer and an inflexible lender, with margins too thin to carry the load. Solving the problem means more than injecting capital. It means renegotiating the terms that caused the problem, diversifying the customer base even at the cost of short-term revenue, and changing operations so that cash is generated by every unit of output, not just booked as throughput. It often means using court-supervised processes not as a threat but as a way to organize competing interests, freeze unhelpful behaviour, and push through changes that would be impossible in bilateral talks.

These are granular, unglamorous fixes: inserting indexation clauses into supply contracts; capping aggregate penalties and chargebacks; securing take-or-pay minimums in exchange for investment; reclaiming early-payment discounts in price; aligning production schedules with actual, profitable demand. They are not negotiated in a single meeting; they are won over months, with credible alternatives in hand and the discipline to walk away from unprofitable volume. They are paired with changes on the floor and in the yard – shorter production runs to reduce rework, stricter WIP control, maintenance to improve uptime, freight planning to end premium shipments.

None of it works without capital that is anchored in the assets and structured to survive the drama that comes with change. Fancy layered financings collapse when a single customer dispute consumes the cash budget. We favour simple, senior positions underwritten to the value of what can be sold or collected, with equity earned through milestones like contract resets or customer diversification. Sometimes we acquire the existing debt to control the process; sometimes we provide super-priority facilities to fund the fix. In every case, the capital is there to buy time for operational and contractual change, not to subsidize structural unprofitability.

The corporate class system is not a passing condition. It will keep reallocating margin to those with brand, channel, and balance-sheet power, and it will keep pushing volatility onto those without it. In Canada’s concentrated markets, the effects are magnified; in a world of higher rates, they are accelerating. For most lenders and investors, this is a reason to pull back from the middle market. For us, it is a signal of where the next complex situations will arise, and a reminder that the defaults, the distress, and the inevitable drama of these relationships are not anomalies, they are features of the system itself. The companies caught in it can survive and even prosper again, but only if someone is willing to change not just their capital structure, but the rules of the game they have been forced to play.

The post Corporate Class Warfare (Q2-25) appeared first on Third Eye Capital.

In our world, each investment is a unique situation, shaped by a highly specific set of circumstances that defy algorithmic generalization. The analysis is neither formulaic nor mechanical. It demands something else entirely: discretion informed by context, a capacity to synthesize the idiosyncratic, and judgment honed through the repetition of non-repeatable events. The very nature of special situations investing means we are engaging where others have stepped away: with companies in transition, in distress, or at an inflection point where conventional financing is no longer accessible. There is no dataset large enough to encode how a founder’s or management team’s motivations will evolve when facing a cash crunch, or how a key supplier’s trust will influence a turnaround. Nor can AI models reliably predict the behavioral dynamics of teams under duress, or the strategic thinking of regulatory authorities, creditors, and stakeholders in moments of dislocation.

The process of assessing credit risk in these situations is deeply interpretive. It is grounded in a nuanced understanding of capital structure, asset value, and downside protection – but equally in an appreciation of context: What are the incentives of the owners? Is the management team capable of adapting under stress? Is the path to recovery defensible not just contractually, but practically? We often encounter incomplete information, contested narratives, and legal ambiguity. These are not obstacles to be sidestepped but rather the terrain. Our methodology is not about extrapolating from clean data but about evaluating under uncertainty, diagnosing root causes of business stress, and identifying what is salvageable and what is not.

Recent commentary from venture capitalist Marc Andreessen captures this distinction well¹. He notes that the most valuable skills in his business are “psychological” – understanding how people behave under pressure, anticipating their reactions, and guiding them through ambiguity. He describes venture investing not merely as capital allocation, but as psychological navigation. The same holds true in our domain. What separates a successful special situations investor is not access to a better model, but a more accurate reading of motives, incentives, and real-world constraints.

It would be misguided to suggest that AI cannot augment this work at the margins. We already employ technology to support many of the mechanical aspects of due diligence – for example, in processing large document sets, modeling downside cases, and analyzing comparative trends. But these are tools in service of a fundamentally human exercise: constructing a view of reality from fragmented, often conflicting information, and deciding what action, if any, has merit. The essence of our strategy remains interpretive and interventionist, not predictive, because the investment decisions we face are not reducible to spreadsheet logic or sentiment analysis. Consider, for example, the restructuring of a complex capital stack in a jurisdiction where creditor protections are uncertain, asset values are contested, and management alignment is fragile. Or the assessment of a family-owned enterprise undergoing succession amidst balance sheet distress and strategic dislocation. These are decisions where capital flows not on the basis of data completeness, but on trust, control levers, and an ability to navigate what is left unsaid.

Some have asked whether AI’s ascendancy in direct lending suggests a similar trajectory for the broader credit market. We see it differently. In fact, the expanding use of AI in conventional credit may well increase the value of human discretion in complex investing. As more capital flows toward standardized risk, guided by AI-enhanced underwriting tools, the opportunity set in our segment only widens. What we offer is not speed or low-cost execution, but a willingness to engage in the difficult, the mispriced, and the misunderstood. That willingness is underpinned by specialization more than scale and by the kind of insight that comes only from having worked through past cycles, restructurings, and recoveries, often where information was limited and pressure was acute.

We see this dynamic already playing out in the divergence between sponsored and non-sponsored lending. In sponsored transactions, loan structures are increasingly pro forma, reliant on sponsor-calibrated EBITDA adjustments and covenant-lite packages. These deals lend themselves to AI underwriting, because the inputs are normalized and the outputs – primarily spread and loss assumptions – are modeled from a wide base of historical data. In contrast, in the non-sponsored, special situations market where we operate, the capital structure is often bespoke, the documentation negotiated not off a template but a strategy, and the exit dependent on active intervention.

We do not believe AI will replace what we do. Not because we resist change, but because the work itself is, by its nature, resistant to commodification. Our value lies in discernment, in negotiation, in the orchestration of outcomes that would not occur without deliberate human engagement. These are not skills that can be learned from data alone. They are learned through experience.

Importantly, the real risk in complex credit is not modelable risk, but unpriced fragility: shifting collateral realities, hidden intercreditor disputes, or counterparty behavior under stress. It is worth recalling that in the lead-up to the 2008 financial crisis, it was the structured products – AAA rated, algorithmically priced – that misfired most catastrophically. The lesson was not that models were useless, but that they were blind to emergent behavior. In the same way, AI may replicate credit ratings and improve underwriting throughput, but it cannot substitute for the practitioner’s judgment in assessing unstated risks, opaque governance structures, or adversarial restructurings.

This is not just theoretical. Recent default episodes, such as the Serta Simmons capital structure controversy or the Envision Healthcare uptiering dispute, reveal a market increasingly characterized by litigation risk, document arbitrage, and first-mover advantage among lenders. These scenarios require investors to think not only about cash flows, but about process, control, and recovery paths. As Harvard Law Professor Jared Ellias and Duke Law Professor Elisabeth de Fontenay recently chronicled in their article about the meteoric rise of private credit, credit investing today is as much legal strategy as it is financial analysis.² This is precisely the space we inhabit.

In a financial world increasingly shaped by automation, we believe our strategy represents a durable and necessary counterpoint: one grounded in the conviction that complexity still requires judgment, and that outcomes still depend on the quality of decisions made by people with the experience to navigate what others overlook. As parts of the credit market become increasingly commoditized, we remain focused on opportunities that cannot be packaged, traded, or automated. The value we create is not found in higher leverage or broader distribution. It is found in being able to identify and resolve complexity that others cannot or will not engage. That is not something artificial intelligence is equipped to do. It is something only experience can.

[1] https://fortune.com/article/mark-andreessen-venture-capitalism-ai-automation-a16z/
[2] Ellias, Jared A. and Elisabeth de Fontenay. The Credit Markets Go Dark. The Yale Law Journal. January 2025.

The post Artificial Experience (Q1-25) appeared first on Third Eye Capital.