Stephanie Aliaga
Global Market Strategist
Marcella Chow
Global Market Strategist
Published:
Picking winners remains difficult because AI will likely reshape industries, business models, and customer behavior, and corporate transformation may unfold more gradually than frontier advances suggest.
In Brief
- Recent AI financing concerns: Lower-rated neoclouds are more exposed to borrowing, power, and refinancing risk, while higher-rated hyperscalers are better insulated and should retain stronger access to supportive, high-quality debt demand.
- The return on investment (ROI) debate: Compute demand is strong, but capex is outpacing broad monetization—so cost of capital and timing can pressure valuations, especially for more levered, lower-quality players.
- AI adoption pace: AI is being adopted widely but monetized unevenly—infra suppliers have captured most gains so far, while the next winners will be firms that convert adoption into measurable earnings via durable moats, making selectivity and diversification across the value chain essential.
- Please visit our AI hub for more AI insights.
Recent AI financing concerns
Was last week’s neocloud selloff driven by weakening demand, or was it mainly a shift in investor sentiment?
Last week’s selloff in neocloud public debt and equity likely says more about shifting AI sentiment than a deterioration in underlying demand. That said, widening credit spreads are a clear signal that investors are increasingly focused on financing risk as the scale of the AI buildout grows.
Why have AI infrastructure concerns been building for weeks?
Concerns have been building because the debate has shifted from “is demand real?” to “can the buildout be financed and delivered at acceptable returns?” Several factors have compounded investor unease: a wave of new graphics processing unit (GPU) capacity has raised questions about future utilization, pricing, and the economics of debt-financed infrastructure, even if compute is still expected to be supply constrained. A Gartner report also projected that more than 80% of neocloud providers could cease independent operations by 2030 as cost pressures mount and business models struggle to differentiate. Meanwhile, updates from a large AI platform company were viewed as mixed, and delays around funding timelines—alongside power and construction delays—have further elevated concern.
Why has one neocloud become the focal point—and what’s being repriced?
A leading GPU-focused neocloud provider has become a focal point, but the repricing appears driven more by financing costs than any clear change in underlying demand. The model depends on the spread between fixed-price, multi-year compute contracts and the cost of financing the infrastructure required to deliver them. It buys GPUs, leases data center capacity, and finances projects through dedicated entities, with debt generally repaid during the first half of a five-year customer contract. With revenue largely fixed, higher interest rates compress new-project economics. The company has reported a substantial revenue backlog of USD 104billion and an additional USD 25billion in customer commitments, but fulfilling those contracts requires significant upfront investment and sustained access to debt financing—what investors are increasingly questioning.
Could this spill over into the broader market?
Deterioration in market sentiment and technical factors may cause volatility. In particular, higher borrowing costs weaken project economics, restricting access to capital and delaying capacity additions, which tightens financial conditions across the AI infrastructure complex. Last week’s temporary widening in hyperscaler spreads to their highest levels since July illustrates the potential spillover channel. Even if underlying demand remains firm, tighter financing and delivery constraints can still drive consolidation and earnings pressure through pricing and execution.
That said, it is important to distinguish direct credit exposure from pressure on earnings. Even without meaningful balance-sheet credit losses, a neocloud restructuring could still affect hyperscalers through operational and commercial channels—capacity delivery delays, investments, prepayments, or contractual support arrangements. Separately, a broader decline in GPU rental prices could compress returns on infrastructure spending across the ecosystem, including for larger players.
There is also a potential offset: financially stronger hyperscalers could secure cheaper capacity or acquire distressed assets, improving their relative position if weaker operators are forced to scale back. Ultimately, the outcome hinges on whether this is primarily a financing problem at individual operators or an early signal of weaker underlying AI demand.
Who is most exposed if financing tightens further?
Neoclouds typically rent GPU capacity with more concentrated customer bases and substantial borrowing, while hyperscalers have broader cloud services and more internally generated cash to fund investment. The key vulnerability for neoclouds is the combination of concentrated AI customers, debt-funded expansion, and GPU assets whose economic value could fall faster than expected. A large contracted backlog does not necessarily eliminate financing or counterparty risk because long-term contracts protect only if customers pay and projects can be completed.
Balance sheets are not equal in this race. Lower-rated neoclouds face greater exposure to borrowing costs, energy prices, and refinancing conditions. Higher-rated hyperscalers should be better positioned to absorb these pressures, with their debt issuance more likely to find support from investors seeking high-quality yields.
Why do credit markets matter more now than earlier in the cycle?
As AI capex increasingly shifts from internally generated cash flow toward external financing (Exhibit 1), credit markets will play a larger role in determining the pace and profitability of the buildout. Higher rates also lift the discount rate applied to future earnings, putting additional pressure on valuations across the ecosystem.
Why are we still constructive on underlying AI compute demand despite financing noise?
Underlying demand remains large, and the next phase of adoption is likely to be more inference-intensive as agentic workflows expand. Falling unit costs can widen the addressable market over time, even if near-term financing conditions tighten. Meanwhile, the binding constraint increasingly looks physical—land, power, and construction—so the bigger swing factor may be who can deliver capacity, not whether demand exists. If those bottlenecks persist, economics may shift toward early movers that secured land, power, financing, and infrastructure partnerships.
Investment implications
After three consecutive years of strong AI-driven market returns, investors have become more selective about the risks they are willing to take. That reinforces the view that some portfolio de-risking may be prudent, even while maintaining exposure to long-term opportunity.
The return on investment (ROI) debate
If compute demand is strong, why is ROI still being debated?
Strong demand for compute does not automatically translate into attractive returns on the capital being invested. The key issue is whether monetization scales quickly and broadly enough to justify the capital intensity and cost of capital embedded in today’s buildout.
How high is the monetization “hurdle” implied by current capex?
The hurdle for justifying AI investment continues to rise alongside capex expectations. Estimates from Empirical suggest the buildout could require blended consumer and enterprise spending of roughly USD 100–USD 300 per paying user per month by the end of the decade—well above today’s broad monetization levels. Spending is currently concentrated in the top 10% of adopters where the highest-spending 10% of enterprise subscribers already pay USD 675 per employee per month (per Ramp) (Exhibit 2), and continued adoption of higher-value applications could help narrow the gap. The challenge is expanding that willingness-to-pay from the top decile into the broader market.
Does the ROI outcome depend mainly on a handful of frontier labs?
It is important to distinguish the concentration of AI investment from the breadth of underlying infrastructure demand. Frontier AI labs are scaling rapidly, with combined revenue estimates approaching USD 200 billion this year and inference economics described as solid (roughly 70–80% gross margins). Yet, according to a leading GPU manufacturer, major AI labs and hyperscalers account for only about one-third of broader installed infrastructure demand, with another third supporting traditional workloads and the remainder serving startups, enterprises, sovereign customers, and neocloud providers. The implication is that the ROI outcome is not solely a function of two model labs, even if they drive a large share of new capacity commitments.
Investment implications
Ultimately, the challenge for investors comes down to timing and capital intensity. The AI buildout requires enormous investment today against monetization that may materialize over many years. Even if the long-term opportunity remains compelling and markets can fund the buildout, the pace of spending running ahead of demonstrated returns can keep valuations under pressure and disproportionately strain lower-quality, more levered players.
AI adoption and long-term investment implications
Is AI adoption happening at scale—or is it still mostly hype?
Adoption and monetization are distinct, and so far adoption is what is happening at scale in Corporate America. Roughly two-thirds of S&P 500 companies discuss AI on earnings calls, but only 17% have offered specific evidence from usage, and just 2% have quantified its impact on earnings.
Who has been winning financially so far?
The clearest winners have been companies selling AI infrastructure (Exhibit 3). Semiconductor stocks have climbed more than 30% this year versus roughly 12% for the S&P 500, and AI infrastructure beneficiaries are expected to deliver more than half of the index’s earnings growth this season. That said, evidence is building that companies outside tech are also benefiting, even if those gains are not always clearly visible in margins or earnings yet.
Why is it so hard to “see” AI benefits in financials?
AI adoption is broad but often shallow: nearly two-thirds of large companies are building or deploying AI agents, yet barely one in ten has scaled them across the business. Even when AI delivers productivity gains, time saved does not necessarily translate into money saved. Benefits also get blurred because AI initiatives run alongside restructuring and automation, and some gains—like quality improvements or retention—are inherently harder to quantify.
What should investors look for to identify early leaders?
Four characteristics may help distinguish companies translating adoption into durable advantage:
- As powerful models become widely available, the differentiator may become proprietary data, particularly as information that competitors cannot replicate;
- Domain and regulatory expertise can matter most in regulated sectors where outputs must be accurate, explainable, and auditable;
- A system-of-record position—authoritative transaction, credit, or benchmark data embedded in workflows—can provide distribution and trust as workflows become more agentic; and
- Finally, network effects can strengthen because data from one customer improves the product for others, as seen in shared-data models like insurance analytics.
Investment implications
Picking winners remains difficult because AI will likely reshape industries, business models, and customer behavior, and corporate transformation may unfold more gradually than frontier advances suggest. Today’s advantages may not be permanent, but they may persist long enough for certain companies to capitalize on them. For investors, that can create opportunities in less obvious AI beneficiaries converting adoption into monetization through durable advantages, while reinforcing the value of selectivity and diversification across the AI value chain.