A structural rotation inside the AI trade — long telegraphed by institutional strategists but now impossible for executives to ignore — has broken the Magnificent Seven’s stranglehold on market returns and handed the baton to chipmakers and memory hardware suppliers, raising a question that will define capital allocation decisions across the technology sector for the rest of this decade: how much AI infrastructure spending is too much before markets demand proof of return?
What Happened
The numbers are stark. The Philadelphia Semiconductor Index — the benchmark for chip stocks — recently posted its best-ever quarterly performance, surging 88% in the second quarter. The Roundhill Magnificent Seven ETF, which tracks Amazon, Alphabet, Apple, Meta, Microsoft, Nvidia, and Tesla, has tumbled from its May peak. Every one of the seven mega-cap tech stocks has lagged the Nasdaq 100’s 16% year-to-date gain, with several down double digits.
The divergence is not accidental. Amazon, Alphabet, Meta, and Microsoft are collectively on track to spend approximately $725 billion on capital expenditure this year, according to figures cited by JPMorgan — a figure that represents roughly 100% year-over-year growth in AI-related capex. Goldman Sachs has projected that by the end of this decade, the cumulative AI infrastructure spending by those four companies alone will likely exceed the GDP of major economies, including Japan.
JPMorgan analysts identified a “spectacular revision in the 2026 plans by hyperscalers” as a major trigger for the chip sector’s outperformance, noting that markets have begun to price in hardware suppliers as the primary beneficiaries of the spending wave — while simultaneously questioning whether the companies writing those enormous cheques will ever recover their investment through AI monetization.
The Gold Rush Analogy Wall Street Is Using
Art Hogan, chief market strategist at B. Riley Wealth Management, offered the framing that has begun circulating among institutional strategists: the current moment resembles the classic gold rush pattern, in which those selling shovels and picks captured the wealth, while prospectors actually panning for gold largely came up empty. “It’s certainly more of an evolution,” Hogan told Business Insider, describing the market rotation as a natural, if uncomfortable, phase transition.
Memory stocks have been the most vivid expression of that thesis. Micron Technology recently crossed a trillion-dollar market capitalisation, and South Korea’s SK Hynix has posted a blistering rally over the same period. For Bret Kenwell, investment and options analyst at eToro, Micron’s milestone is evidence that the AI trade can sustain itself without the Magnificent Seven at the front. “From the perspective of if the trade can keep going without the participation of the Mag Seven — without question, it could keep going,” Kenwell said.
The Reading
Why the Divergence Has Institutional Weight
What elevates this rotation beyond typical sector churn is the analytical pedigree of the voices flagging it. JPMorgan — whose macro and quantitative research teams carry significant weight with institutional allocators — described the divergence as “somewhat unsustainable” and explicitly drew a parallel to the months leading up to the dot-com crash. A team led by Nikolaos Panigirtzoglou wrote that “the semiconductor trade could come under severe pressure, inducing a more significant and sustained correction in the AI trade,” though the bank simultaneously stated it was leaning toward the more bullish resolution.
That dual framing is itself significant. When a tier-one investment bank puts a bearish scenario on the record while leaning bullish, it is performing risk management for its own institutional clients — flagging the tail risk without abandoning the base case. Executives at companies with heavy AI exposure should read JPMorgan’s note not as a market call, but as a signal that the window for demonstrating AI monetization is narrowing.
Combining JPMorgan’s capex revision data with Goldman Sachs’ GDP-scale projection produces an observation neither institution stated directly: the four largest hyperscalers are, in aggregate, behaving less like technology companies optimising for return on investment and more like sovereign infrastructure builders — committing capital at a scale and timeline that can only be justified if AI becomes a utility-layer technology embedded in every commercial workflow. Markets appear to be pricing in the possibility that this thesis is correct for the hardware layer but unproven for the application layer, which is precisely where hyperscaler revenues must eventually materialize.
This reading is consistent with the broader institutional shift already reshaping how AI investment is evaluated: the easy narrative of AI as a growth story is giving way to harder questions about which companies in the value chain actually capture the economics.
Two Scenarios, One Feedback Loop
JPMorgan outlined two paths to resolving the divergence. In the bullish scenario, hyperscalers begin demonstrating meaningful AI monetization — through advertising efficiency gains, cloud revenue acceleration, or enterprise software attach rates — leading their stocks to “catch up” with chip valuations. In the bearish scenario, hyperscalers blink first, pulling back capex commitments in response to investor pressure. That pullback would trigger a feedback loop: reduced hardware demand would hit chip stocks, spreading the correction from the application layer back into the infrastructure layer that has so far been insulated from it.
The feedback loop risk is not hypothetical. Analysts who have been tracking the structural dynamics of the AI super-cycle have noted that the entire chip rally is predicated on the assumption that hyperscaler capex commitments are real and durable. Any credible signal that Amazon, Alphabet, Meta, or Microsoft is slowing its build-out would reprice that assumption immediately.
Who Is Actually Winning Right Now
The current beneficiaries of the rotation are concentrated in two hardware sub-segments: GPU and accelerator manufacturers, led by Nvidia, and memory chip producers, led by Micron and SK Hynix. The distinction matters. Nvidia’s position is well-understood — its H100 and successor architectures are the computational substrate of virtually every frontier AI model. But the memory story is newer and, in some ways, more instructive. High-bandwidth memory demand is a direct function of model size and inference volume, meaning memory chip performance is a real-time proxy for how much AI computation is actually being run. Micron’s trillion-dollar valuation is, in this reading, the market’s cleanest measure of AI workload growth.
For context on how analyst credibility functions in this environment, it is worth noting that even executives with strong AI conviction — such as Palantir’s Alex Karp — have warned that AI labs have oversold their models to enterprise customers, creating an expectation gap that eventually lands on hyperscalers’ balance sheets as underutilized infrastructure.
What to Watch
The next quarterly earnings cycles for Amazon, Alphabet, Meta, and Microsoft will be the most consequential data points in resolving JPMorgan’s two-scenario framework. Institutional investors will be scrutinizing not just revenue growth but the ratio of AI-attributable revenue to AI-related capex — a ratio that, for most hyperscalers, remains deeply unfavourable by conventional investment standards.
Signals of AI monetization are beginning to emerge in specific product lines: Microsoft’s Copilot attach rates in enterprise subscriptions, Meta’s advertising efficiency gains attributed to AI ranking models, and Google’s AI Overviews impact on search revenue. But none of these signals has yet been presented in a form that directly closes the capex justification gap. Until that happens, the rotation toward hardware suppliers is likely to persist, and possibly accelerate.
Memory chip dynamics also bear watching independent of the hyperscaler narrative. The memory chip supply cycle has historically been characterized by boom-bust dynamics that can move faster than AI infrastructure build-out timelines. If inference volumes continue growing — and there is little evidence they are not — memory demand should remain structurally supported. But the market’s pricing of that demand has become aggressive enough that even modest downward revisions to hyperscaler capex guidance could trigger outsized corrections in memory valuations.
The interoperability and governance questions surrounding AI infrastructure are also becoming more relevant to capital allocation decisions. As internet pioneers like Vint Cerf have argued, the infrastructure decisions made in the early phases of a technology transition tend to lock in competitive structures for decades — a point that should focus the attention of every hyperscaler CFO currently reviewing capex commitments.
What This Means for the Industry
The AI capex divergence is not merely a market rotation story — it is a credibility test for the hyperscaler business model in the AI era. Amazon, Alphabet, Meta, and Microsoft have collectively staked their capital allocation strategies on the premise that AI infrastructure investment will generate compounding returns through services, cloud attach, and advertising. Markets are now demanding a timeline. The fact that chip stocks have continued rallying while hyperscaler stocks lag is not a vote of no confidence in AI — it is a vote of no confidence in the current distribution of AI economics, and a signal that investors believe hardware suppliers are capturing value that has not yet reached the application layer.
For CFOs and boards at the four major hyperscalers, the strategic response is not straightforward. Pulling back capex to appease investors risks ceding infrastructure position to competitors — a move that would be nearly impossible to reverse given the lead times on data center construction and chip procurement. Maintaining current spend levels requires a credible and near-term monetization narrative that most of these companies have not yet fully articulated to institutional audiences.
The companies that must respond most urgently are those in the middle of the value chain: enterprise software vendors, cloud-native application builders, and AI platform providers whose revenues depend on both hyperscaler infrastructure investment and end-customer willingness to pay for AI features. If the hyperscaler capex story shows any signs of deceleration, the repricing will not stop at the Magnificent Seven — it will cascade through every company whose growth assumptions are built on continued AI infrastructure expansion.
Micron’s crossing of the trillion-dollar threshold is, in the end, the most precise institutional signal in this story. It means the market has decided, at least provisionally, that the AI infrastructure build-out is real, durable, and supply-constrained. What it has not decided — and what the next twelve months will likely resolve — is whether the companies building on top of that infrastructure can generate the returns their valuations require.
