Thoughts on Artificial Intelligence (AI)

This summer, my family and I were driving from Vancouver to Calgary. Somewhere near Revelstoke, a cluster of warning lights lit up on the dashboard and the car suddenly refused to go into reverse. Not exactly the development one hopes for in the middle of the mountains!

I snapped a photo of the warning display and showed it to an AI assistant. It suggested a short sequence of steps. I followed them and—voilà—the car started working again. Without AI, the next move would likely have involved a mechanic, a tow truck, a long delay, or some combination of the three. A problem that might easily have cost hundreds of dollars was solved with a $27 /month subscription.

That small experience captures the magic of AI better than most grand predictions do. It put a competent, patient adviser in my pocket at the exact moment I needed one. It didn’t manufacture a car, replace the mechanic forever, or eliminate uncertainty. It simply converted information into useful action, instantly, and at almost no incremental cost.

That is real value. The harder question for investors is who will capture it, and whether the returns will justify the extraordinary capital now being deployed.

The tool is real

The debate around AI has become oddly binary. One camp is racing toward Artificial General Intelligence (AGI) and treats almost any investment as justified. The other dismisses it as a bubble built on clever chatbots and circular financing. As usual, reality sits somewhere in the middle.

AI fits into a long progression of tools. The printing press lowered the cost of reproducing knowledge. Steam power and electricity amplified physical effort. The internet collapsed the cost of finding and distributing information. AI lowers the cost of turning information into a first draft of thought or action.

A few commercial use cases are already past the demonstration stage. In a controlled experiment with 95 professional developers, GitHub found that engineers using Copilot completed a programming task 55% faster. Coding has the attributes AI favors: abundant structured data, rapid feedback loops and output that can be tested. AI use cases now also span customer service, where models draft responses and summarize interactions; operations and compliance, through automated report drafting and policy checks; HR and training, by generating onboarding materials and answering routine policy questions; and finance and accounting, where they reconcile transactions, flag anomalies and prepare first‑draft analyses.

AI is far from perfect, but with some of the world’s brightest minds working to improve it, we have to assume that its capabilities will only become more powerful.

The open question is how large the end market becomes. AI excels where tasks are language‑heavy, repetitive enough to learn and easy to verify. Judgment, accountability and checking remain human responsibilities. But if the first draft becomes dramatically faster, the economics can still be profound.

Hyperscaler returns

The scale of the infrastructure build is difficult to overstate. Across the major U.S. platforms, capital spending is now running at roughly $600 billion this year, though not all of it is strictly AI. Amazon alone expects about $220 billion of cash capital expenditures in 2026, spanning long‑lived data centres, power and networking, as well as shorter‑lived servers and accelerators.

This spend is currently leading to an acceleration in revenues. In the second quarter, AWS revenue grew 37% year over year to a $169 billion annualized run rate, while AWS operating income reached $16.6 billion at a 39% margin. More importantly, Andy Jassy gave unusually specific unit economics on the Q2 call: “For servers and networking equipment, on average, it takes a little less than 3 years to break even on that investment.” The data-centre economics are even more interesting. Jassy said, “the buildings have lives of more than 30 years and should support at least five or six generations of servers”. Later generations can earn better returns because Amazon does not have to repeat the cost of the building, land and power.

Microsoft’s CEO, Satya Nadella, effectively corroborated Jassy’s numbers when he validated a Morgan Stanley analysis showing that hyperscale AI infrastructure can generate ROIC in the high‑20s, noting that software continually improves the productivity of existing hardware.

The Sensitivity of GPU Economics

Consider a deliberately simple $1.5 million investment in 25 deployed GPUs at an cost of $40,000 each. Assume a five-year useful life, 90% billable utilization, a base rental rate of $3.99 per GPU-hour and cash operating costs of $0.75 per utilized GPU-hour. The returns look attractive, including in the downside case. But the table also shows how sensitive the economics are to price. A 20% reduction in rental rates cuts the unlevered IRR from almost 25% to approximately 17%.

Scenario	Billable rate	Annual revenue	Cash payback	Five-year IRR
Rate –20%	$3.19	$629,000	3.5 years	16.7%
Base case	$3.99	$786,000	2.8 years	25.6%

However, the danger in any shortage, like the one being experienced today, is that peak pricing can masquerade as normalized economics. This is the crux of the debate. Bears argue that supply will inevitably catch up, pushing prices down and compressing returns. Bulls counter that falling prices will unlock even more demand, expanding the total addressable market rather than shrinking it. A 20% drop in price won’t translate into a 20% drop in revenue if usage surges enough to overwhelm the decline. Both sides may ultimately be right but timing this shift is very difficult.

Negative free cash flow is not automatically bad

After nearly $600 billion of combined AI‑related capex, the major U.S. hyperscalers are now producing only marginal free cash flow, and consensus expects free cash flow to turn negative next year as the build‑out accelerates. This is exactly the kind of headline that worries investors as AI spending compresses free cash flow across large technology companies.

We think about this differently. Capital allocation shouldn’t be judged solely through the lens of near‑term free cash flow. When a company spends $1 today on capacity that pays back quickly and then throws off cash for years, free cash flow dips precisely because management is pursuing a high‑return opportunity. As shareholders, we would rather see companies deploy capital at high rates of return than hand it back to us through dividends or buybacks. Growth capex that compounds value is fundamentally different from maintenance spending — and the accelerating revenue growth of the hyperscalers is evidence of that dynamic playing out in real time.

But financing is becoming increasingly creative

The financing behind the AI buildout is more complicated than reported capex alone suggests and leverage cuts both ways. Rather than fund every data center directly, hyperscalers increasingly use third-party developers, joint ventures, private credit and long-term leases, often supplemented by residual-value guarantees or financial backstops that make the projects financeable. This reduces upfront cash requirements and transfers construction funding to outside investors, but it does not eliminate the economic obligation.

CreditSights calculates that Amazon, Microsoft, Alphabet, Meta and Oracle already have approximately $287 billion of lease liabilities on their balance sheets and another $1.16 trillion of signed leases that have not yet commenced or are off-balance sheet. The concern is that many lease payments begin only when facilities are completed, creating a significant commencement ramp from 2027 onward.

The ultimate test is whether profitable end‑customer demand emerges at sufficient scale to justify trillions of dollars in investment. If it does, these structures are sensible infrastructure finance. If AI demand or pricing disappoints, model companies, and ultimately hyperscalers could still be obligated to pay for capacity, turning today’s aggressive buildout into future impairment and write-offs.

AI is now part of the macro cycle

AI spending is no longer only a technology-sector story. It supports demand for construction, electrical equipment, power generation, networking, cooling, software and skilled labour. Fed model estimates datacentre investment could add roughly 0.3 to 0.5 percentage points to 2026 GDP growth.

Goldman Sachs estimates that companies connected to the AI trade now represent roughly 45% of the S&P 500! The S&P 500’s dependence on AI is hardly Buffett’s idea of a diversified portfolio. More importantly, Goldman expects AI-infrastructure beneficiaries - including semiconductors, hardware, industrials and utilities - to generate roughly half of all S&P 500 earnings growth in both 2026 and 2027. However, leverage and tightly correlated portfolios could turn a slowdown in AI spending into a broader market shock, pressuring earnings and valuations well beyond the obvious beneficiaries.

In essence, the U.S. economy and markets have become increasingly exposed to the AI capital cycle, so evidence that returns are weakening can have outsized impact on the overall market and economy.

What We Are Watching as Investors

A sustainable ecosystem ultimately requires consumers, governments and ordinary businesses to pay for AI. AI agents must deliver reliable results without creating an unacceptable number of false positives. Across the portfolio, we regularly ask management teams how they are measuring these trade-offs and where they see tangible returns. Beyond coding and customer service, scalable use cases are still emerging.

It is important to note that the usefulness of AI does not guarantee attractive returns for participants. Technology determines what is possible. Industry structure and competitive advantages determine who makes money.

At White Falcon, we’re keeping a close eye on both sides of the AI debate. The landscape is shifting fast and understanding whether the bull or bear forces are gaining ground demands constant attention and discipline. We do not need to predict the ultimate future of artificial intelligence today. We need to identify the evidence that would cause us to become more bullish or more bearish.

How White Falcon Uses AI

White Falcon is materially more productive today than it was two, three or four years ago. We can review transcripts faster, compare disclosures across years, stress-test a financial model, and generate better questions before a management meeting.

But the investment process has not been outsourced. AI does not know which management team deserves trust, whether an accounting adjustment is economically honest, or when an industry is over-earning. It can help assemble the mosaic. Conviction still must be earned. At White Falcon, AI is a tireless junior analyst - not the portfolio manager.

We live in a remarkable time. A photograph taken beside an old car in Revelstoke can become useful mechanical guidance in seconds. Work that took days can take hours; work that took hours can take minutes. That is magic in the everyday sense of the word.

PS: In the interest of full disclosure, the first draft of this post was written entirely by AI.