White Line | NVIDIA bets its credit on AI, while Microsoft and Meta face scrutiny over AI returns

Source | White Line

Finding direction before change arrives

"White Line" is produced by the Wu Shuo team, focusing on trend shifts and trading opportunities in the AI era.

Full text below:

Core view: The market hasn’t abandoned AI; instead, it’s re-pricing the AI supply chain in layers. In the past, rising capital expenditures alone were enough to convince investors to pay for future growth. Now, the market is asking who will provide financing, who bears the credit risk, and when these investments will translate into orders, revenue, and free cash flow.

In the last week of July, major technology companies released earnings reports that are difficult to categorize simply as 'AI-positive' or 'AI-negative.'

Both Microsoft and Meta are accelerating their investments, yet market sentiment toward them has diverged; Apple, by contrast, has refrained from building data centers on a comparable scale. Its lighter capital expenditures and robust cash flow have made it an 'anti-AI CapEx' play in the eyes of some investors. Meanwhile, NVIDIA may be providing financing guarantees for OpenAI’s data center project, signaling that the AI infrastructure race is no longer limited to chip procurement—credit risk is now entering the supply chain.

AI investment is shifting from procurement to financing

According to The Wall Street Journal, NVIDIA is discussing providing approximately $250 billion in financing guarantees for OpenAI’s leased data center in Ohio. The campus, developed by SB Energy—an energy subsidiary of SoftBank—is planned to have a capacity of 10 gigawatts, with total costs, including chips, potentially exceeding $500 billion. NVIDIA may also offer financing support for OpenAI’s purchase of up to $350 billion worth of chips. The deal remains under negotiation and may not materialize at its current scale. Nevertheless, even entering discussions reflects an evolving approach to financing AI infrastructure.

Previously, NVIDIA’s role was relatively straightforward: cloud providers and model developers raised capital, built data centers, and then purchased GPUs from NVIDIA. Now, downstream projects have grown so large that tenants’ own creditworthiness alone can no longer secure sufficiently cheap funding. To keep these projects moving forward, chip suppliers may also need to provide guarantees to lenders. If such arrangements are finalized, NVIDIA could lock in more long-term orders—but it would also assume risks related to customer defaults, project delays, and weaker-than-expected demand for computing power. AI-related risks are beginning to propagate upstream along the supply chain.

These discussions are unfolding in a less accommodative financing environment. On July 29, the Federal Reserve kept its target range for the federal funds rate at 3.50%–3.75%, though three voting members advocated for a 25-basis-point rate hike. Funding costs remain elevated, and whether data centers can generate sufficient revenue to cover rent, interest, and equipment depreciation is becoming a new valuation constraint.

Azure grows 43%: Microsoft’s AI investments are already backed by orders

Microsoft lowered its expected capital expenditures (CapEx) for calendar year 2026 from approximately $190 billion to $175 billion, but this does not signify a corresponding reduction in investment plans.

Starting in fiscal year 2027, Microsoft will extend the estimated useful life of its data centers and office buildings from 15 years to 25 years. Due to this change in estimated useful life, more future data center leases will be classified as operating leases and excluded from reported CapEx. Microsoft emphasized that its construction and procurement plans remain unchanged, and FY2027 CapEx will still increase year-over-year. In short, what’s declining is CapEx on financial statements—not actual spending on data centers or computing capacity. The market has accepted this explanation largely because Microsoft has already demonstrated tangible revenue and order commitments.

This quarter, Azure cloud revenue grew by 43%; Microsoft expects Azure’s constant-currency growth rate to be around 45% next quarter. Commercial remaining performance obligations (RPO)—contracts signed but not yet recognized as revenue—reached $678 billion, growing 25% even after excluding OpenAI. Paid seats for Microsoft 365 Copilot have also surpassed 30 million.

Microsoft reported operating cash flow of $55.4 billion and free cash flow of $19.6 billion for the quarter. Free cash flow declined 23% year-over-year, confirming that investments in computing capacity are indeed consuming cash, though newly added capacity is already contributing more quickly to Azure revenue. Of course, the $678 billion RPO (remaining performance obligation) cannot all be recognized as near-term revenue. The weighted average contract duration is 2.3 years, with approximately 30% expected to be recognized over the next 12 months. Microsoft’s cloud gross margin also fell to 65%, indicating that cost pressures from AI infrastructure have not abated.

Microsoft’s current advantage isn’t that it spends less, but rather that investors can more clearly see how this spending translates into orders and then into revenue.

Capital expenditures reached $31.08 billion, leaving Meta with only $784 million in free cash flow for the quarter.

Meta’s second-quarter revenue grew 28% year-over-year, driven by a 14% increase in ad impressions and a 12% rise in average ad pricing. Its core advertising business remains robust, and AI-driven recommendations may already be enhancing user engagement and ad efficiency. However, costs are rising even faster. Total expenses for the quarter reached $42.03 billion, up 55% year-over-year, including $2.4 billion in legal expenses and $1.18 billion in severance costs. Capital expenditures hit $31.08 billion, operating cash flow was $31.86 billion, and free cash flow amounted to just $784 million. The company also narrowed its 2026 CapEx guidance to $130–145 billion.

Microsoft can separately disclose Azure growth rates, cloud backlog, and Copilot seat adoption, whereas Meta’s AI returns are more embedded within metrics like recommendation efficiency, ad impressions, and ad pricing. While its advertising business may already be benefiting from AI, external investors find it difficult to isolate those gains and directly compare them against investments in data centers, models, and talent. This is why, despite both companies making massive AI investments, the market finds Microsoft’s narrative easier to understand.

Why Apple is viewed as an 'anti-AI CapEx' stock

The label 'anti-AI CapEx' is one the market has applied to Apple—it doesn’t mean Apple isn’t pursuing AI.

Apple continues to ramp up R&D spending but has not replicated the data center scale of Microsoft, Meta, or Google. Instead, it channels more resources into in-house chips, operating systems, end-user devices, and its services ecosystem. It can also reduce the infrastructure investment needed for independently training and deploying large models by partnering with external model providers.

Apple’s latest reported Q3 fiscal 2026 results show quarterly services revenue of $30.74 billion and R&D expenses of $11.73 billion. Operating cash flow for the first nine months was approximately $117 billion, with capital expenditures totaling about $6.8 billion.

Apple isn’t avoiding investment—it’s simply allocating most of its spending to R&D and product ecosystems rather than building its own computing infrastructure. In a high-interest-rate environment, this model exerts relatively less pressure on free cash flow and is more likely to command a valuation premium. However, this asset-light approach comes with trade-offs. Apple must rely on external models and cloud-based compute, and whether it can master core model capabilities and ensure product experience remains to be seen. For now, the market is primarily rewarding its balance sheet strength, not a proven leadership position in AI.

From refining margins to AI computing power: what’s scarce may be delivery capacity

This week, the refining sector offered an apparently unrelated but highly instructive comparison.

In mid-July, the U.S. 3-2-1 crack spread briefly surged to a record $69.66 per barrel. This metric estimates the theoretical profit a refiner earns by processing three barrels of crude oil into two barrels of gasoline and one barrel of distillate. Even after the risk premium on crude oil receded, refining margins remained elevated, due to low refined product inventories and refineries already operating near full capacity.

HF Sinclair reported an adjusted refining margin of $25.95 per barrel in the second quarter, up approximately 57% year-over-year. Feedstock isn’t scarce—the bottleneck lies in short-term processing capacity. The AI supply chain faces a similar issue. Even as GPU supply increases, new bottlenecks may emerge in power infrastructure, data center construction, advanced packaging, optical communications, memory, and testing capacity. Total capital expenditure only indicates the scale of investment; it’s the constrained segments that determine how quickly this spending translates into usable computing power. This analogy need not be stretched too far—it simply reminds investors that the most profitable part of a supply chain isn’t always the priciest raw material, but could instead be the processing or delivery capacity that can’t be expanded quickly.

After the semiconductor rebound, watch how orders are confirmed

Some oversold semiconductor stocks have recently staged a recovery, but price rebounds don’t substitute for earnings validation. Recent announcements of partnerships and orders from several companies illustrate the various stages through which demand evolves from intent to revenue. Amkor Technology, a U.S.-based semiconductor assembly and test services provider, signed a multi-year agreement with NVIDIA valued at approximately $1.5 billion. NVIDIA will provide an upfront payment to support Amkor’s expansion of advanced packaging capacity in the U.S., and the two companies will jointly develop packaging and testing technologies for AI and accelerated computing platforms. This deal demonstrates NVIDIA’s willingness to secure U.S. packaging capacity in advance, but neither the contract value nor the prepayment should be treated as confirmed revenue for Amkor. Investors should monitor factory construction progress, customer acceptance, actual shipments, and gross margins going forward.

AXT, a compound semiconductor substrate manufacturer, signed a long-term indium phosphide (InP) substrate supply and capacity reservation agreement with optical communications component maker Lumentum. Lumentum will pay a $43.5 million deposit upfront, which will later be offset against future purchase payments. The customer’s willingness to pay in advance signals scarcity in indium phosphide capacity. However, the deposit still represents funds tied to future deliveries and can only be recognized as revenue once products are actually shipped.

Semiconductor test equipment company Aehr Test Systems is closer to the order confirmation stage. Its fourth fiscal quarter orders reached a record $60.7 million; including new orders received after the fiscal year-end, its effective backlog stands at approximately $100.6 million. Nevertheless, Aehr’s revenue for the prior fiscal year declined from $59 million to $50 million, and it reported negative operating cash flow. While the growing backlog improves revenue visibility for the next fiscal year, whether the company can achieve a turnaround depends on when these orders actually convert into revenue and profits.

From cooperation agreements and prepayments to backlogs, then to shipments, revenue, and free cash flow—there’s still a long way to go. The market is pricing the AI supply chain with increasing granularity and will increasingly differentiate where each company stands along this path.

Capital expenditures continue to expand, and the market is starting to ask about returns

The final week of July did not see a unified AI-driven market rally.

NVIDIA may be providing guarantees for OpenAI’s financing, signaling that AI infrastructure is still expanding—though it now requires increasingly substantial credit support. Both Microsoft and Meta are ramping up investments, with Microsoft benefiting from Azure revenue and long-term contracts that afford it greater patience. Apple, which has not incurred data center expenditures on the same scale, appears leaner in a high-interest-rate environment.

The criteria for evaluating companies along the supply chain are also evolving. Advance payments demonstrate customers’ willingness to lock in capacity, and order backlogs enhance revenue visibility—but only upon delivery do orders translate into revenue, gross margin, and cash flow.

What matters next is not how much additional capital expenditure a few companies announce, but whether cloud revenue can continue growing, whether free cash flow can halt its decline, and whether supply chain firms can deliver on their existing agreements and orders on schedule.

The AI arms race hasn’t stopped—it’s just that, starting this week, being willing to spend money is no longer enough.

This article is for informational and market structure analysis purposes only and does not constitute investment advice.

Source:The Wall Street Journal, Gavin Baker, Microsoft FY2026 Q4 earnings report and earnings call, Meta Q2 2026 earnings report, Apple Q3 FY2026 financial statements, Federal Reserve FOMC statement dated July 29, 2026, EIA Weekly Petroleum Status Report, Reuters, HF Sinclair Q2 2026 results, AXT and Lumentum long-term supply agreement, Aehr Test Systems FY2026 Q4 results

79K Views