For more than two years, the artificial intelligence trade has been the single most powerful force in American equity markets—a relentless upward march that minted trillions in market capitalization and transformed companies like Nvidia from chipmakers into cultural phenomena. That era of unchallenged dominance is now facing its most serious stress test, and the fault lines are exposing a stark divide between the companies that build AI infrastructure and those that stand to profit from deploying it.
The catalyst arrived in late January from an unlikely source: a Chinese AI laboratory called DeepSeek, which released an open-source large language model that appeared to rival the capabilities of leading American systems—at a fraction of the cost. The implications rippled through global markets with startling speed, triggering what MSN described as a dramatic reassessment of “winners and losers” across the technology sector.
A $600 Billion Monday and the Questions That Followed
On January 27, Nvidia suffered its worst single-day loss in history, shedding roughly $600 billion in market value as investors scrambled to recalculate the economics of AI infrastructure spending. The reasoning was straightforward, if brutal: if a startup operating under U.S. export restrictions could build a competitive model using fewer and less advanced chips, then perhaps the massive capital expenditure plans of hyperscale cloud providers—the very spending that has powered Nvidia’s extraordinary revenue growth—were built on shakier assumptions than Wall Street believed.
The sell-off was not confined to Nvidia. Broadcom, which supplies custom AI accelerators and networking chips, fell sharply. Power infrastructure companies that had surged on expectations of electricity-hungry data center buildouts—names like Vistra Energy, Constellation Energy, and Vertiv Holdings—saw steep declines. The entire supply chain that had been bid up on the thesis of virtually unlimited AI capital expenditure came under pressure simultaneously.
The Other Side of the Ledger: Software and Services Rally
But what made the January sell-off particularly instructive was not just the damage—it was the divergence. While hardware and infrastructure names plunged, a different cohort of technology companies rallied. As reported by MSN, companies positioned as AI adopters rather than AI suppliers saw their shares climb. The logic was the mirror image of the infrastructure sell-off: if AI becomes cheaper to build and run, then the companies that integrate AI into their products and services—enterprise software firms, application providers, and platforms with large user bases—stand to see their margins expand rather than contract.
Salesforce, ServiceNow, and Palantir were among the names that benefited from this rotation. The market was effectively saying that cheaper AI is deflationary for infrastructure providers but inflationary for the value of AI-enabled applications. Morgan Stanley analysts noted at the time that the DeepSeek development, if validated, could accelerate enterprise AI adoption by lowering the barrier to entry for companies that had been priced out of large-scale AI deployments.
The Capex Debate Intensifies
The tension at the heart of this market reassessment centers on one question: how much will the world’s largest technology companies spend on AI infrastructure, and for how long? In recent earnings calls, the answers from Big Tech have been emphatic. Meta Platforms announced plans to spend between $60 billion and $65 billion on capital expenditures in 2025, with the vast majority directed toward AI infrastructure. Microsoft disclosed plans for more than $80 billion in data center spending. Alphabet signaled similarly aggressive investment.
These numbers initially reassured investors who had been rattled by the DeepSeek news. If the customers buying Nvidia’s chips were not pulling back, the thesis remained intact—or so the argument went. But a more nuanced reading suggests the picture is more complicated. Wall Street analysts have begun to distinguish between “Phase 1” AI spending, which is dominated by training large foundation models and requires enormous GPU clusters, and “Phase 2” spending, which focuses on inference—the process of actually running AI models in production at scale. Inference workloads may require different hardware configurations, potentially favoring custom silicon from companies like Broadcom and Marvell Technology over Nvidia’s general-purpose GPUs.
Nvidia’s Moat Under the Microscope
Nvidia remains, by virtually any measure, the dominant force in AI semiconductors. Its CUDA software platform, which developers use to write code for its GPUs, represents a formidable competitive advantage that has taken more than a decade to build. The company’s upcoming Blackwell architecture promises significant performance improvements, and CEO Jensen Huang has repeatedly argued that demand for AI compute is essentially insatiable—that as models become more efficient, developers will simply use the freed-up capacity to build more capable systems.
This argument, sometimes called the “Jevons Paradox” of AI, holds that efficiency gains do not reduce total demand for compute but instead expand it by making new applications economically viable. There is historical precedent for this view: the dramatic decline in the cost of computing over the past half-century has been accompanied by an equally dramatic increase in total computing consumption. But skeptics point out that the Jevons Paradox is not a law of physics. It describes a tendency, not a guarantee, and the specific economics of AI training versus inference may not follow the same pattern as general-purpose computing.
Power Plays and the Infrastructure Premium
Perhaps no corner of the market illustrates the AI infrastructure trade’s excesses and vulnerabilities better than the power sector. Utilities and independent power producers had become unlikely Wall Street darlings in 2024, with companies like Vistra Energy more than tripling in value on expectations that AI data centers would drive unprecedented electricity demand. Constellation Energy’s proposed acquisition of Calpine Corp. for $26.6 billion underscored the thesis that reliable, large-scale power generation was becoming a scarce and valuable resource.
The DeepSeek-triggered sell-off hit these names hard, but many have since recovered a significant portion of their losses. The underlying demand picture for data center power remains strong—even if individual AI models become more efficient, the sheer number of models being trained and deployed continues to grow. According to the International Energy Agency, data center electricity consumption is projected to more than double by 2030. Grid constraints, permitting delays, and the long lead times required to build new generation capacity mean that companies with existing power assets near major data center markets retain significant strategic value.
What the Bond Market Is Saying
Equity investors are not the only ones paying attention to the shifting AI narrative. Credit markets have begun to differentiate more sharply between technology companies based on their AI positioning. Companies with strong free cash flow generation and clear paths to AI monetization are finding favorable borrowing conditions, while more speculative AI-adjacent firms are seeing spreads widen. The broader fixed-income market is also grappling with the macroeconomic implications of AI-driven productivity gains, which could theoretically reduce inflationary pressures and influence the Federal Reserve’s rate path—though such effects remain highly uncertain and likely years away from materializing in economic data.
Meanwhile, the venture capital market has shown no signs of cooling its enthusiasm for AI startups. Private funding rounds for AI companies reached record levels in early 2025, with investors betting that the technology’s long-term transformative potential justifies current valuations even as public market investors grow more discriminating. The disconnect between private market exuberance and public market selectivity is itself a signal worth watching—historically, such divergences have resolved in one direction or the other, often painfully.
The Sorting Mechanism Wall Street Needed
For institutional investors who have spent the past two years debating whether AI represents a generational investment opportunity or a speculative bubble, the recent market turbulence has provided something valuable: differentiation. The monolithic “AI trade” that lifted virtually all related stocks in 2023 and 2024 is giving way to a more granular assessment of which companies will capture lasting value and which were simply riding the wave of sentiment.
The winners in this new phase are likely to be companies that can demonstrate concrete returns on AI investment—measurable productivity gains, new revenue streams, or defensible competitive advantages derived from proprietary data and distribution. The losers may include companies whose valuations were predicated on an assumption of ever-increasing infrastructure spending without a corresponding focus on the efficiency and economic return of that spending. As one senior portfolio manager at a major asset management firm told colleagues in a recent internal note, the AI trade is not over—but the era of buying everything with “AI” in the pitch deck almost certainly is.
The market is entering a phase that demands more precision, more skepticism, and a willingness to distinguish between the picks and shovels of AI and the gold they are meant to unearth. For Wall Street, that kind of analytical rigor has been a long time coming.
WebProNews is an iEntry Publication