Before the 1890s, a factory that wanted electricity generally bought its own generator. Then, Samuel Insull, in Chicago, helped turn power from a piece of equipment into a service purchased by the kilowatt-hour. Central stations could spread their fixed costs across as much demand as possible. And to finance those stations, Insull pushed utility bonds and holding-company structures to their limits, bringing long-duration capital from the financial markets into heavy infrastructure on an unprecedented scale.

The 2026 AI earnings season feels, in some ways, like history replaying itself.

Capital spending is accelerating. Optical module demand is surging. Software results are diverging sharply. And beneath all of this, we believe a new financial architecture is taking shape, one that is turning AI from an investment sustained largely by technological optimism into an asset class capable of supporting long-duration debt and increasingly complex financing structures.

Large infrastructure companies may have a decisive advantage in this transition: stronger balance sheets, cheaper capital, and more sophisticated asset-management capabilities. For that reason, we remain firmly constructive on the infrastructure holdings in the AGIX portfolio.

We will examine this earnings season in a three-part series of Portfolio Manager (PM) Notes. The first question is the most fundamental: How does compute move from being merely a capital expenditure to becoming collateral?

Many investors have looked at hyperscalers' free cash flow turning negative, and at their growing reliance on debt financing, and concluded that AI capital expenditure (capex) is becoming unsustainable. We disagree.

Alphabet generated $39.1 billion of operating cash flow this quarter and spent $44.9 billion on capital expenditures, producing free cash flow of -$5.9 billion.¹ To read that figure as evidence that the investment itself is unsustainable is like declaring a restaurant unprofitable because it used the cash from its first location to open a second. It confuses the cash-flow profile of growth with the quality of the underlying asset.

The real question is: Can an individual asset or project generate a sufficiently high return on invested capital? And does the balance sheet have sufficient capacity to fund capex at scale?

On the first question, Nebius' latest earnings call offered a striking answer. Nebius is an Amsterdam-based neocloud provider. A neocloud is an independent cloud compute provider that builds and operates Graphics Processing Unit (GPU) and Tensor Processing Unit (TPU) infrastructure for AI workloads, unlike major hyperscalers such as Amazon and Google. The payback period on Nebius' GPUs has fallen to roughly one-third of their depreciation life. GPU assets are generally depreciated over five to six years. A 22-month payback period against a 60-to-72-month useful life implies that the asset can recover its initial cost roughly three times over during its economic life.²

That is the metric that matters. Not whether aggregate free cash flow is temporarily negative, but how quickly the underlying asset earns back the capital invested in it.

The economics are even more attractive because customers are effectively funding a substantial portion of the project before the asset is completed. In some cases, more than half of the construction cost is covered in advance. The neocloud contributes less of its own capital, while the project-level return rises.

The second question is whether balance sheets can continue to support capex at this scale. CoreWeave's earnings call provided an answer there as well. CoreWeave is widely considered the flagship and largest neocloud.

Over the past year, the company's weighted-average cost of debt has fallen by nearly 300 basis points. Applied to its debt balance at the end of the second quarter, that decline translates into approximately $1.1 billion of annual interest savings.³ That tells us something important. The market is not merely repricing CoreWeave as a company. It is re-pricing compute as an asset, hence the robustness of the balance sheet.

Compute is also becoming increasingly financialized. NVIDIA has signed memoranda of understanding with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR to create dedicated pools of capital for customers, with potential financing capacity exceeding $500 billion.⁴ Look closely at the structure, and the logic becomes clear. The compute infrastructure itself secures the debt. Financing is raised through private placements and special-purpose entities. A single vehicle can issue tens of billions of dollars of debt, acquire the infrastructure, and lease the resulting compute capacity to NVIDIA's customers.

The most consequential detail is NVIDIA's potential role in the capital stack. Jensen Huang has said that NVIDIA could provide up to 25 percent financing for an individual opportunity.⁵ In practical terms, NVIDIA would become the subordinated investor beneath the senior lenders.

That subordinated layer changes everything. Without it, the entire asset pool must be priced against the downside case. That is how companies such as CoreWeave and Nebius originally ended up borrowing at costs above 10 percent. But once a subordinated tranche absorbs the first losses, the same stream of contractual cash flows can be divided according to risk. Senior creditors receive greater protection. Their claims can then be distributed through the much larger and much cheaper mainstream bond market.

For NVIDIA, the economics are especially unusual. The effective cost of providing subordinated capital may be negative because the risk capital is paid for by the gross profit generated from incremental GPU sales enabled by the financing.

The structure is not entirely new. Google's next major strategic priority is to turn the TPU into a third-party platform and compete more directly with NVIDIA in the inference market. Fluidstack, in effect, the Google ecosystem's answer to CoreWeave, has already entered into a similar financing arrangement.

Fluidstack builds data centers equipped with Google TPUs and leases the capacity to Anthropic. Google backstops Fluidstack's lease obligations. Broadcom provides residual value support (meaning protection against a shortfall in the TPUs' resale value) for the senior notes. If Anthropic fails to make its lease payments and the TPUs are subsequently sold for less than the amount needed to repay senior noteholders, Broadcom will cover the resulting shortfall. Together, these guarantees reduce the risk borne by the senior tranche of the special-purpose vehicle and push its financing cost toward quasi-investment-grade levels.

The difference is already visible in pricing. The TPU ecosystem has been able to finance at roughly 7.1 percent, compared with about 9.3 percent in the NVIDIA ecosystem. That gap has ultimately pushed NVIDIA toward a similar structure.

We believe the message from this earnings season, and from the financing arrangements now being developed by NVIDIA, Google, and Meta, is increasingly difficult to ignore: Compute is becoming an asset class with contracts, collateral, layered risk, differentiated pricing, and growing access to the mainstream capital markets.

This is the foundation of our view that AI capex remains sustainable. It no longer depends solely on optimism about the future of artificial intelligence. It increasingly rests on project-level economics and a financing architecture that is already working. And sustainable capex does not remain trapped on the balance sheet. It travels through the value chain.

A lower cost of capital raises a project's Internal Rate of Return (IRR). A broader financing frontier allows more customers to build. Longer-duration contracts improve revenue visibility. Better asset utilization strengthens margins. Over time, those improvements should appear in revenue growth, earnings, and shareholder returns.

Large infrastructure companies may be best positioned for this transition because they possess what smaller participants often lack: financing capacity, operating expertise, and the ability to manage assets across cycles. We therefore remain confident and intend to continue our overweight holding in AI infrastructure companies within the KraneShares Public-Private AI & Technology ETF (Ticker: AGIX) portfolio.

---

Holdings are subject to change.

For AGIX standard performance, top 10 holdings, risks, and other fund information, please click here.

Citations:

1. Alphabet Earnings Report as of 6/30/2026.
2. Nebius Earnings Report as of 6/30/2026.
3. CoreWeave Earnings Report as of 6/30/2026.
4. "NVIDIA Partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital," NVIDIA Newsroom, August 11, 2026.
5. "NVIDIA just soothed a major market fear about AI, analysts say," Morningstar, August 11, 2026.

Term Definitions:

Free cash flow: The cash a company generates from its operations after subtracting capital expenditures (spending on property, equipment, and other long-term assets).

Operating cash flow: The cash a business generates from its core day-to-day business activities, adjusted for non-cash items like depreciation and changes in working capital.

GPU (Graphics Processing Unit): A specialized processor originally designed for rendering graphics, now widely used to run the parallel computations required to train and deploy AI models.

TPU (Tensor Processing Unit): A custom chip developed by Google specifically to accelerate machine learning workloads, designed as an alternative to GPUs for AI inference and training at scale.

Weighted average cost of debt: The blended interest rate a company pays across all of its debt obligations, weighted by the size of each debt relative to total debt, often calculated on an after-tax basis since interest is tax-deductible.

Pools of capital: Distinct sources of funding, such as equity, debt, retained earnings, or private capital, that investors or companies draw from to finance investments, each with its own cost, risk profile, and return expectations.

A subordinated tranche: A layer of debt or securities in a structured deal that ranks below senior tranches in priority, meaning it absorbs losses first but typically offers higher returns to compensate investors for that added risk.