The Capital Trends Behind the AI Computing Power Rebound: JPMorgan Fund Flows Reveal Retail Buy-In "Shrinking," Pouring Into Nvidia, SanDisk and Other Computing Power Core Companies

According to Investing.com, the latest "Retail Radar" research report released by Wall Street financial giant J.P. Morgan shows that after the Federal Reserve's rate hike was finally implemented, retail investors did not fully exit AI, but instead, under macroeconomic pressure, their overall market entry pace has slowed significantly, and their stock selection has become more concentrated on AI computing power leaders. For the FOMC's unanimous decision on Wednesday, Eastern Time, to raise rates by 25 basis points—setting the policy rate at 3.75%—4.00%, J.P. Morgan's baseline view is: if the cycle is only a rollback of last year's "insurance-like cuts" into a shallow rate hike, while corporate profits remain strong and the Middle East situation doesn’t spiral out of control, the stock market is still able to absorb rising interest rates and long-term U.S. Treasury yields.

When the pressure on valuations from oil prices and long-end yields eases, the core assets of the AI computing power theme, which have continued to attract buying interest, have a foundation for participating in the recovery of risk appetite, and the market focus can conditionally shift back from rate hike shocks to orders, revenue, and profit realization. This is why the latest data compiled by J.P. Morgan shows that from September 10 to 16, retail investor capital inflows helped the seven mega-cap stocks achieve a total net purchase of $1.601 billion, with Nvidia alone accounting for $1.196 billion. Leaders across the AI computing power supply chain, such as SanDisk, ASML, and Oracle, also received strong inflows. This “total contraction, core concentration” structure better explains why AI computing power theme stocks have a foundation for rebound after the Fed's hike compared to a broad “retail exit from technology.”

J.P. Morgan’s data-driven historical analysis suggests that under persistent high-growth economic conditions, the equity market can withstand the 10-year U.S. bond yield gradually approaching 6%. However, this is an endurance assessment premised on AI-driven strong profit expansion, not a yield forecast of 6%.

Similarly, two other Wall Street giants, Goldman Sachs and Jefferies, have recently issued similar positive signals. Last weekend, Goldman Sachs published a research note stating that since ChatGPT took the globe by storm in 2022, the U.S. stock market’s long-term bull run is likely to continue under a “profits above all” bullish logic—predicting S&P 500 EPS to reach $340 by 2026, marking a sharp 24% year-on-year growth from an already high base, and a further rise to $385 (up 13%) by 2027.

Meanwhile, the forward P/E ratio has fallen from 22 at the start of the year to 19, indicating that the headwind of rates has been reflected through valuation compression. Historical samples show that three months after the start of a rate hike cycle, the S&P 500 averages a 2% drop, but after twelve months, the average gain is 9%. These data do not support the view that “the Fed’s rate hikes will necessarily end the bull market,” nor can they be used as 100% proof that future returns will replicate history; Goldman Sachs emphasizes in the research that the real key is whether the trend in profit delivery can offset any further valuation declines.

Jefferies recently commented that, fueled by the dual engines of the AI investment frenzy and AI-related companies’ earnings surprising to the upside, the S&P 500 is expected to soar to 8,000 by the end of 2026 and further reach 9,000 in 2027. Jefferies' core logic is clear and strong: in a cycle where AI-driven earnings growth is more than double the historical average, fighting the profit trend is dangerous. Jefferies' baseline S&P 500 forecast of 8,000 in 2026 is based on an EPS of $373 (up 35% YoY, well above the consensus 29%) and a 21.5x P/E ratio.

Fed rate hike implemented, retail capital appears to contract buying: funds concentrate on core AI computing power assets

The latest retail fund flow data compiled by J.P. Morgan shows that, from September 10 to 16, net retail inflows totaled only $2.5 billion—about 63.2% lower than the 12-month average of $6.8 billion per week; ETF net inflows were $3.1 billion, with individual stocks seeing a net outflow of $600 million. The report ranks the total inflow intensity at the 2nd percentile in history, ETF inflows dropped to a one-year low, and individual stock capital flows were in the 12th percentile. Thus, the crucial market message is “incremental buying has decreased and funds are pickier:” High oil prices, long-end rates, and AI safety discussions are suppressing short-term risk appetite, but core assets such as Nvidia still manage to attract net buying.

On Thursday (September 17), the Philadelphia Semiconductor Index surged 3.1% after a brief dip, coinciding with a sharp drop in oil prices and the 10-year Treasury yield falling to 4.93%. This shows the marginal easing of yield curve pressure and the strong recovery of investment confidence in AI computing power after the Fed’s rate hike and the central bank’s anti-inflation credibility expectations were repaired.

Though the J.P. Morgan "Retail Radar" was published just before the global tech stock rally, it still provides vital clues for investors—overall market entry has slowed, but selective buying of core AI assets hasn’t stopped. Notably, the S&P 500’s “Magnificent Seven” (including Nvidia and Google) saw a combined $1.601 billion net buying, with Nvidia at $1.196 billion, and SanDisk, ASML, and Oracle netting $202 million, $119 million, and $102 million, respectively.

This “total contraction, core concentration” structure better explains the rebound basis for the AI theme than the generic view of "retail exiting tech:” Funds are shrinking exposures to non-Magnificent Seven tech stocks, while continuing to bet on select names in accelerated computing, storage, semiconductor equipment, and cloud infrastructure. The report also clearly notes that AI data center core infrastructure chains, electrification, major AI beneficiaries, and commercialized AI software remain favored by retail buyers.

Oracle, in particular, demonstrates the logic of “profit and order support underpinning dips”: its revenue grew 30% YoY, cloud-related infrastructure revenue saw triple-digit growth, and its remaining performance obligations increased by $26 billion quarter-on-quarter. Subsequently, on September 11, retail net buying of Oracle reached $87 million in a single day, highlighting that verifiable business growth still attracts capital inflows.

Combining J.P. Morgan’s latest assessment that “so long as profits remain strong and geopolitical risks are under control, shallow hike cycles and persistently high long-end Treasury yields can still be digested by equities,” one infers a possible investment strategy: when the pressure on valuations from oil and long yields eases, AI computing power core assets with previous buyer support have a foundation to participate in the risk appetite rebound, providing a backdrop for Thursday's rally. The market focus can shift from the shock of rate hikes back to orders, revenues, and realized profits.

Trading pace and sector flows together imply that retail is engaging in both defensive allocation and selective tech investment, not simply rotating from technology as a whole to defensive sectors. In the past four sessions, retail was still a net buyer of individual stocks at midday, but shifted to net selling in the afternoons, mainly due to shifting from buying to selling in tech stocks outside the Magnificent Seven and accelerated industrials selling. Excluding the Magnificent Seven, only consumer staples posted a net buy ($11 million); industrials, communications, and tech saw net outflows of $555 million, $493 million, and $405 million, respectively, while utilities and real estate posted the smallest net sales.

On the ETF side, sector ETFs experienced the third-largest weekly net outflow in history, mainly led by technology products; but broad large-cap ETFs still saw $1.3 billion in net buys, precious metal ETFs brought in $188 million, and medium-to-long duration bond ETF demand slightly improved. Notably, the market exposure of high-dividend strategies is near a three-year high, but retail inflows to dividend ETFs have actually fallen.

Meanwhile, J.P. Morgan’s compilation shows that retail option trading as a share of the total (one-month rolling average) is still at the 96.8th percentile historically, indicating participation remains high—even though spot buying willingness and trading direction have diverged.

From fund flows to profit realization: AI computing power infrastructure investment shifts from broad chase to select mega-cap configuration

The specific purchase list shows that capital not only continued to support the Magnificent Seven but also maintained clear preferences in storage, semiconductor equipment, and cloud infrastructure. J.P. Morgan’s retail fund flow data for the week of September 10–16 is as follows:

All Magnificent Seven names saw net buying: Nvidia $1.196 billion, Tesla $201 million, Amazon $87 million, Apple $63 million, Google $25 million, Microsoft $19 million, and Meta $10 million—a total of $1.601 billion, with Nvidia making up about 74.7%. Outside of the Magnificent Seven, SanDisk (SNDK.US) was net bought by $202 million, ASML (ASML.US) $119 million, and Oracle (ORCL.US) $102 million. Together with Nvidia and Tesla, they constitute the top five weekly net buys. The top five net sells were SpaceX (SPCX.US) $249 million, Marvell (MRVL.US) $95 million, Intel (INTC.US) $76 million, NuScale Power (SMR.US) $60 million, and AMD (AMD.US) $52 million.

The report clearly notes, retail investors continue buying the AI data center electrification supply chain, the most core and primary beneficiaries of AI and data center infrastructure, the Magnificent Seven, growth stocks, as well as themes like AI software and AI application commercialization—fully demonstrating that retail investors have not abandoned AI, but have instead focused their buying on the Magnificent Seven and the most core, capital-favored companies in storage, equipment, and cloud platforms. Non-Magnificent Seven tech names saw net outflows, but this does not mean every stock was sold off.

The team of analysts at J.P. Morgan upgraded Meta to “Overweight,” optimistic about its advanced models, the Muse agents, and model API services. They believe AI commercialization will spread from advertising into new products and revenue sources. In cybersecurity, Okta, Palo Alto Networks, CrowdStrike, Varonis, and Zscaler are especially favored—the core reasoning being that expanding AI applications widen the attack surface, making security companies foundational partners for model developers and ramping up security spend. It should be noted, however, that the retail investor base remains a net seller of this security software portfolio, indicating a divergence between institutional fundamental views and short-term retail trades.

AI cloud computing infrastructure leader Oracle is a key case of positive flows: J.P. Morgan’s report cites its 30% YoY revenue growth, triple-digit growth in cloud infrastructure revenue, and remaining performance obligations (RPO) rising by $26 billion quarter-on-quarter, arguing that these results address concerns about ongoing order growth, order-to-revenue conversion, and subsequent financing. On September 11, retail net buying of Oracle reached as high as $87 million in a single day, seen as post-earnings dip buying. Further, on September 15, Skyworks and Qorvo received net buys of $7.5 million and $400,000, respectively, following news of their planned merger—these are not directly attributed to AI hardware buying.

Financing, applications, and delivery data in the AI sector provide fundamental support for continued core asset attention, although these are supplemental to fund flow observations. Media reports say OpenAI is discussing new funding at a valuation exceeding $1.2 trillion; Reuters reports that Anthropic is prepping a potential IPO valued at around $2 trillion, aiming to raise up to $100 billion, which could make it the largest initial public offering in history—both firms are at the funding discussion or preparation phase.

On the application front, Astra has enhanced programming, browsing, computer operation, and complex workflow capabilities, further expanding the scope of AI-executable tasks; OpenAI has confirmed it is suspending new subscriptions and upgrades for its $200/month Pro 20X plan effective September 10, with existing subscriptions unaffected. On the procurement side, Anthropic has disclosed agreements with Amazon for up to 5 GW, Google and Broadcom for 5 GW, $30 billion in Azure capacity, and $50 billion in U.S. AI infrastructure investment tied to Fluidstack.

On the delivery side for AI computing power infrastructure resources, Nvidia’s second fiscal quarter 2027 data center revenue reached $89 billion, up 117% YoY; South Korea’s exports in August rose 68.7% YoY to $98.26 billion, with semiconductor exports at $46.65 billion, and year-to-date exports as of early September hitting $709.4 billion, surpassing the full-year record set in 2025. From an investment logic perspective, these data points ground the “expanding AI demand” narrative in application usage, capacity procurement, and hardware revenue, supporting the long-term growth for the likes of Nvidia, storage, and cloud infrastructure firms—a key reason why short-term retail fund flows have focused on selective buying, continuing to choose some core AI investments.

The most noteworthy technical change is that AI computing power resource demand is expanding from concentrated pretraining, to post-training, continuous inference, and agent execution—helping explain why capital is reassessing hardware vs. software positions internally.

J.P. Morgan, citing SemiAnalysis, noted that the share of pretraining in overall computing power allocation has dropped to below 15%, with terminal token demand expanding. Post-training technology is also becoming a key capability-expansion direction for models; the description of declining share refers to structural shift, not an absolute decreased demand for compute. According to Jevons Paradox, total AI demand will continue to expand as compute costs drop. From a system architecture’s perspective, user count, task frequency, number of calls, and context length together affect total token processing: prefill (input), decode (output), and KV cache place demands on GPU memory and bandwidth; multi-tool agents require CPUs for program execution, storage chips for data and cache, high-speed ethernet for connecting compute resources, and optical interconnect/optical communication in AI data centers for fast data transmission, with cloud platforms facilitating orchestration and delivery.

This essentially forms a complete demand chain of "accelerated computing—memory & storage—infrastructure delivery—software commercialization," with J.P. Morgan’s fund flow data for Nvidia, SanDisk, ASML, and Oracle corresponding to different links in the chain. Retail investor flows still selectively favor AI computing power and tech themes, but the chase is no longer indiscriminate for every AI-labeled stock—when overall buying cools, capital continues to overweight core computing and crucial bottleneck AI industry links.