Everything About AI
NextFin News - The AI trade is not breaking — it is digesting. That is the message from Sitara Sundar, head of alternative investment strategy at JPMorgan Private Bank, who argues that investors should be diversified across infrastructure, applications, geography and sector rather than chasing the narrow group of chipmakers that led the rally. Her assessment lands as the market itself splits in two: on Monday, Nvidia fell more than 3% while Intel, AMD and Marvell dropped between 5% and 6%, dragging the Philadelphia Semiconductor Index down almost 6%, even as Alphabet rose nearly 2%, Microsoft added 1.6% and Meta gained roughly 1.4%.
The divergence is the story. Chipmakers — the picks and shovels of the AI buildout — sold off on fears that the frontier-model race may slow. The hyperscalers buying that equipment held up, and in some cases rose, on the logic that a slower race would let them stop spending while harvesting the capacity they already built. Sundar's "digestion" framing sits between those two readings: a pause in a structural shift, not the end of it.
What the Market Is Actually Pricing
The trigger for the latest volatility was a weekend call for restraint from the AI labs themselves. Anthropic chief executive Dario Amodei called for stronger safety measures and even a slowdown in model development after demonstrations of models' cyber capabilities; OpenAI chief executive Sam Altman and other AI leaders quickly backed him. The market's reaction, however, was far from a uniform "sell AI."
Nvidia, Intel, AMD and Marvell led the decline, with the SOX index of semiconductor stocks falling almost 6% at one point on Monday before recovering ground. Amazon fell about 1.6% but held up considerably better than the chipmakers. Brent crude climbed above $108 a barrel after Houthi attacks in the Middle East intensified, the 10-year Treasury yield touched 5%, and investors began pricing another Federal Reserve rate hike. In other words, the AI names were hit by a macro crosswind at the same moment as the safety debate — which makes the chip-versus-hyperscaler split all the more informative.
By mid-September, the SOX was trading around 11,887, having closed the prior session at 11,146 after the Monday pullback. The index had set an all-time high of 9,556 in mid-April, meaning even after the selloff it sits well above its recent peak — a draw within a larger uptrend rather than an obvious trend reversal on its face.
The positioning behind the move matters as much as the price. The infrastructure trade had become one of the most crowded on Wall Street: spending from the AI leaders — Alphabet, Amazon, Meta, Microsoft, Oracle and Nvidia — now accounts for almost 25% of total U.S. corporate capital expenditure, according to JPMorgan Private Bank's 2026 outlook. When a single theme absorbs a quarter of all corporate capex, even a pause in the growth rate of that spending is enough to shake out marginal holders.
Why Digestion Is Not the Same as Deflation
Sundar's argument rests on a three-phase view of the AI evolution that JPMorgan Private Bank has laid out. The first phase produced the large-language models that power generative AI. The second phase drew capital into the hardware and infrastructure needed to run those models — chipmakers and the hyperscale cloud and datacenter providers. The third phase, which Sundar says is now beginning, is defined by what the bank calls "services as a software": models that evolve from probabilistic systems producing plausible text into deterministic systems that can reason and act, integrated into industry-specific workflows.
"The next wave of AI value creation, spanning agentic AI systems, vertical industry applications and AI-enabled horizontal software, is still in its formative stages and largely unfolding within private markets," Sundar said.
That sequencing matters for how to read a slowdown in chip demand. If the second phase — the infrastructure buildout — pauses while the third phase — enterprise application adoption — accelerates, the trade does not die. It rotates. Revenue and profit growth migrate from the companies selling the capacity to the companies using it.
The scale of what has already been spent helps explain why a pause is plausible without being fatal. Large U.S. technology companies have tripled their annual capital investment, from $150 billion in 2023 to a projected $500 billion or more in 2026, according to JPMorgan Private Bank's 2026 outlook. AI-related investment is contributing more to U.S. GDP growth this year than consumer spending. Yet the bank notes that AI investment still accounts for less than 1% of GDP — a figure it uses to argue that the cycle has room to run even after the most aggressive buildout in the sector's history.
The bank's conclusion is explicit: "We believe the rally in AI-related investments is justified and sustainable. Capex is massive, and adoption is accelerating."
Sundar has repeated the theme across multiple appearances this year. In January she said the AI capital expenditure cycle would continue in 2026 and expected a broadening of earnings growth. In June she pointed to memory-chip stocks such as Micron, which had gained as much as 1,000% over the prior year, as evidence that bottlenecks in chip supply were opening opportunities beyond the best-known names. In August she estimated that roughly 25% of the multi-trillion-dollar investment needed over the next five years for AI infrastructure — data centers, compute power and grid enhancements — would come from private capital rather than public markets alone. The through-line is consistent: the buildout has years to run, but the beneficiaries will broaden.
The Mechanism: Who Wins When Capex Digests
The transmission channel is straightforward, and it is why the market split the way it did. Nvidia's revenue depends on customers continuing to spend. Google, Microsoft, Meta and Amazon are the customers. If the frontier race slows, the hyperscalers can stop adding capacity and simply harvest the returns from what they have already built — while the chipmakers lose their order flow.
Gil Luria, head of technology research at D.A. Davidson, put it plainly: "They'll just all stop building data centers and just digest what they have." He compared the dynamic to Amazon's post-pandemic pullback in warehouse construction, which paradoxically allowed cash flow to surge as the company kept using the capacity it had already built. Revenue and profits could keep rising while capital expenditure falls.
That is the first-order effect. The second-order effect is less discussed: a digestion phase is where the application layer gets its margin. When infrastructure spending decelerates but model capability keeps improving, the cost of intelligence per task falls. Companies that were priced out of frontier models during the buildout sprint can now afford to embed them in workflows. That is the economic logic behind the "services as a software" thesis — and behind Sundar's push to diversify beyond the infrastructure names that dominated the first two phases. JPMorgan Private Bank puts the services-as-a-software opportunity at $3 trillion to $5 trillion, a range that would exceed the addressable market of both the cloud and mobile transitions because the ultimate target is labor-cost reduction across entire industries.
There is also a geographic dimension to the rotation. While U.S. labs debate restraint, China is not slowing: one of its top AI labs, Z.AI, raised another $5 billion over the weekend. If frontier development continues abroad, the hyperscalers cannot afford to stop spending — they would simply lose the capability race. That dynamic puts a floor under the long-term capex thesis even if near-term growth rates moderate.
The Counter-Thesis: What If This Is the Capex Peak?
The strongest case against the digestion reading is that the infrastructure buildout was the entire value proposition, and that what follows is not a rotation but a contraction. If frontier progress genuinely slows — because of safety constraints, because the low-hanging fruit is gone, or because the economics of training ever-larger models break down — then the hyperscalers do not just pause capex; they face a future in which the capacity they built earns a lower return than expected. In that world, the chipmakers fall first and the hyperscalers follow.
Citigroup has warned that AI safety concerns could weigh on earnings and stocks, and sees a 3% to 5% pullback in the S&P 500 as a level that might attract dip buyers rather than a full reversal. Bank of America analyst Vivek Arya, by contrast, called the warnings "noise" in what he described as a "secular bull market" for semiconductors. The two views are not reconcilable; one of them is right about the cycle.
There is a historical analog that should give pause. The March 2000 peak in the communications-equipment capital investment cycle followed a similar pattern: a parabolic buildout of fiber-optic capacity that ultimately far exceeded demand, leaving the infrastructure builders crushed and the equipment buyers sitting on stranded assets for years. The difference this time, according to digestion-thesis advocates, is that the hyperscalers are also the model owners and the application distributors — they capture value on both sides of the infrastructure, so excess capacity can be monetized through services rather than written off. Whether that difference is enough is the central question of the cycle.
The falsifying signal is specific: if hyperscaler capital expenditure growth decelerates for two consecutive quarters while AI-driven revenue contribution fails to accelerate, the "digestion before the next leg" thesis is wrong. That would indicate a capex-cycle peak, not a pause. Investors should also watch whether model capability improvements continue to compound — if the labs actually halt training rather than merely calling for restraint, the application-layer thesis loses its fuel.
What Comes Next
In the short term, sentiment and positioning will dominate. The SOX's 6% intraday drop shows how crowded the infrastructure trade remains, and a 5% Treasury yield plus oil above $108 a barrel leaves little room for multiple expansion in any rate-sensitive growth name. A digestion phase is typically marked by elevated volatility and sharp, news-driven swings — exactly what the market delivered on Monday. JPMorgan's technical strategists have flagged the risk explicitly: Jason Hunter warned in late August that "the medium-term trend deceleration near longer-term channel resistance, the recent switch in market leadership, and vulnerable setups for the former leadership in the AI segment of the market could favor trimming exposure in early-September and as seasonality turns negative," if the risks persist. The firm nonetheless raised its year-end S&P 500 target to 8,000, underscoring that the warning is about timing and positioning rather than a call for a bear market.
Over the medium term, the earnings reports will separate the thesis from the narrative. Nvidia's fiscal 2026 revenue reached $215.9 billion, up 65% from $130.5 billion in fiscal 2025, with gross margin holding near 75%. Consensus still expects revenue of $108.4 billion for the quarter ending in October, revised upward by more than 5% over the prior month. If those numbers hold while hyperscalers guide capex growth lower but not negative, the rotation thesis is confirmed. If both fall together, the digestion frame fails.
Over the long term, the structural question is whether deterministic, agentic AI actually reaches the enterprise workflows Sundar describes. The $3 trillion to $5 trillion services-as-a-software opportunity is a claim about labor-cost displacement at scale. If that materializes, the current pullback will look like a pause inside a regime shift. If it does not, the 2023-2026 buildout will be remembered as the cycle itself.
The base case is digestion: a cyclical consolidation inside a structural transition, with leadership rotating from infrastructure sellers to application adopters. The upside case is that safety-driven restraint buys time for competitors to close the gap — and with Chinese labs still funding at pace, the global race prevents any true pause. The downside case is a genuine capex peak, confirmed by two quarters of decelerating hyperscaler spending without offsetting AI revenue growth, in which case the communications-equipment analog from 2000 becomes the relevant template rather than the post-pandemic warehouse analogy.
The AI trade is not dying; it is growing up. The question is no longer whether the technology works, but who captures the value once the buildout sprint ends — and digestion is the phase where that answer gets decided.
Explore more exclusive insights at nextfin.ai.
