Big Tech's AI trillions: what quality of agent is actually being built?

Investing.com - Meta Platforms is indicated down roughly 10% in premarket trading Thursday, pointing toward $526.50, a level just above its 52-week low of $520.26, as investors challenge whether the company's massive AI capital expenditure program is producing returns that justify its scale.

NVIDIA (NASDAQ: NVDA) and Palantir (NASDAQ: PLTR) are directly in the crosshairs of the same repricing: NVIDIA's credit default swaps have been rising sharply, and Palantir is indicated down about 0.87% to $121.93 in premarket trade Thursday, according to Investing.com data.

Oracle (NYSE: ORCL) is adding another dimension to the pressure. The company's credit default swaps have drawn attention in credit markets as its stock has also come under heavy selling pressure, reflecting investor concern about the debt Oracle has taken on to build out its cloud and AI data-center infrastructure. Oracle has aggressively financed data-center capacity through bond issuances, and its long-term debt load has climbed sharply in recent years as it races to compete with hyperscalers. That leverage now looks precarious to credit markets if AI revenue ramps more slowly than projected. The dynamic illustrates a broader pattern across the sector: data-center debt is priced on the assumption that AI workloads will fill capacity quickly, and when that assumption wavers, credit spreads widen and equity valuations compress simultaneously. Oracle is not alone. Across the industry, companies have issued billions in corporate bonds to finance land, power agreements, and hardware — obligations that do not shrink if enterprise AI adoption lags.

The selloff is spreading globally. South Korean stocks slumped more than 9% Thursday as AI spending fears hammered chip stocks, a sign that markets are beginning to ask a harder question: what, exactly, is all this infrastructure for?

The answer that keeps surfacing is not the consumer-facing AI agent that was supposed to be the payoff. It is the plumbing underneath: data ingestion, normalization, storage, and processing at scale. And the data being processed is not Gemini chats or ego-driven chatbot interactions. It is the relentless torrent of real-world signals — location pings from phones, food delivery orders, RFID chip reads at warehouses and logistics hubs, license plate scans from Flock cameras, flight tracking data, and thousands of other streams that modern commerce and surveillance infrastructure generate every second. Big Tech's trillion-dollar buildout has cascaded well beyond GPU clusters. A recent SEC filing from renewable energy company Greenbacker Renewable Energy explicitly cites "rising data center demand" and "lack of power" as core strategic positioning pillars, illustrating how the AI capital cycle has pulled in power infrastructure financing at multiple removes from any consumer product.

There is also a question that almost no one in the industry is answering publicly: how good are the agents being built on the back of all this capital? Companies are competing aggressively to acquire land, lock up power agreements, and procure GPU clusters, yet almost nothing is disclosed about the actual capability or quality of the AI systems being trained on this infrastructure. The buildout is proceeding at a scale that implies transformative results, but those results are not being demonstrated in any systematic or public way. Hundreds of billions in annual capital expenditure flows into data centers, power capacity, and chips — and at the other end, the agents and assistants companies publicly tout remain, in most enterprise deployments, early-stage tools with narrow reliability and little independent verification of their quality. For investors, this opacity is increasingly difficult to ignore: the capital expenditure is visible and measurable; the output is not.

Analysts covering Asia's semiconductor selloff have characterized it as a market repricing of AI expectations, with the central concern being whether enterprise data-collection and processing infrastructure is the actual end product being constructed, rather than the agents and assistants that companies have marketed publicly.

NVIDIA sits at the pressure point of this debate. Its credit default swaps have drawn attention on Wall Street amid what analysts describe as circular financing concerns: AI companies borrow to buy NVIDIA chips, but those chips' valuations depend on continued AI investment by the very same companies. NVIDIA last traded at $190.01 on Wednesday, off more than 3.5% on the day per Investing.com data, though premarket Thursday showed a partial recovery to around $193.93. The company is scheduled to report Q2 fiscal 2027 earnings on August 26, which will be the first major opportunity for management to address whether data-center revenue growth is accelerating or plateauing and whether the circular-financing narrative has any structural merit.

Palantir offers a different lens. Its core business has always been large-scale government and enterprise data integration and analytics, a mission that predates the current AI boom by more than a decade. If the broader AI buildout is effectively constructing a substrate for institutional data aggregation rather than consumer convenience, Palantir's original thesis looks less like a niche government contractor story and more like a preview of where the entire industry was heading. Oppenheimer recently reiterated its Outperform rating on Palantir ahead of the company's upcoming earnings report, a signal that the data-analytics framing remains central to how the Street values the stock. Palantir shares closed Wednesday at $123.00, per Investing.com.

For investors trying to limit exposure during the current AI ETF turbulence, WarrenAI analysis published by Investing.com flagged KOMP and AIQ as offering the least risk in AI ETF selloffs, reflecting the preference for broader, less concentrated AI exposure over single-stock bets.

What makes the current moment different from prior AI skepticism cycles is the breadth of the repricing. Meta's premarket move, if it holds into the regular session, would push the stock toward a test of its 52-week low. Combined with the South Korean chip rout, rising CDS spreads on NVIDIA and Oracle, and debt-market anxiety about data-center leverage across the sector, the market is no longer just questioning individual company valuations; it is questioning the architecture of the entire investment thesis. NVIDIA's August 26 earnings call will be a critical stress test of whether data-center revenue can sustain the narratives that have supported the sector's valuation through two years of aggressive capital deployment.

You May Also Be Interested In

• VSee Health signs LOI to acquire healthcare commerce platform for $42M
• Coastal Financial Corporation (CCB) Misses Q2 EPS by 376c
• Mizuho upgrades Bloom Energy to Outperform on stronger execution, margin expansion