The capital trends behind the AI computing power counteroffensive: JPMorgan fund flows reveal retail buying "narrowing," pouring into AI computing power core names such as Nvidia and SanDisk.

Title context: The capital trends behind the AI computing power counteroffensive: JPMorgan fund flows reveal retail buying "narrowing," pouring into AI computing power core names such as Nvidia and SanDisk. Text: According to the latest "Retail Radar" research report released by Wall Street financial giant JPMorgan, as the Federal Reserve's interest rate hike finally landed, retail investors have not fully withdrawn from AI. Instead, under macroeconomic pressure, their overall pace of entering the market has slowed significantly, and stock selection has become more concentrated in AI computing power leaders. Regarding the Fed's FOMC monetary policy committee members unanimously approving a 25 basis point rate hike on Wednesday Eastern Time, raising the policy rate to 3.75%-4.00%, JPMorgan's baseline judgment is that if this is merely a shallow rate hike cycle that retracts part of last year's "insurance-style rate cuts," while corporate earnings remain strong and the Middle East situation does not further spiral out of control, the stock market still has the capacity to digest rising interest rates and long-dated U.S. Treasury yields. When the pressure on valuations from oil prices and long-end yields eases somewhat, the core AI computing power theme assets that had previously still received buying support will have the foundation to participate in a recovery in risk appetite, and the market's focus will also be positioned to shift from the rate hike shock back to orders, revenue, and earnings realization. This is also why the latest data compiled by JPMorgan shows that from September 10 to 16, retail investors' fund inflows helped the Magnificent Seven collectively obtain $1.601 billion in net buying, of which NVIDIA Corporation alone accounted for $1.196 billion. Leaders in the AI computing power industry chain such as SanDisk, ASML Holding NV ADR, and Oracle Corporation also received strong inflows. This structure of "overall contraction, core absorption" explains better than a broad claim that "retail investors are withdrawing from technology" why the AI computing power theme still has a basis for rebound after the Fed's rate hike landed. JPMorgan's analysis report combining historical data even argues that under the current positive conditions of a sustained high-growth economic environment, the stock market can withstand the 10-year U.S. Treasury yield gradually approaching 6%. However, this is a tolerance judgment premised on the resilience of strong AI-driven earnings expansion, not a forecast of a 6% yield. Coincidentally, the other two Wall Street giants, Goldman Sachs Group, Inc. and Jefferies Financial Group Inc., have also recently released similar positive signals. In a research report released last weekend, Goldman Sachs Group, Inc. laid out a "profits trump everything" bullish logic that the long-term bull market in U.S. stocks since ChatGPT swept the globe in 2022 will continue stronglyprojecting that S&P 500 earnings per share will reach $340 in 2026, implying a substantial year-over-year increase of 24% from an already high base; in 2027, it is expected to further reach $385, a year-over-year increase of 13%. At the same time, the forward P/E ratio has fallen from 22x at the beginning of the year to 19x, indicating that interest rate headwinds have already been reflected through valuation compression. Its historical sample shows that three months after the start of seven rate hike cycles, the S&P 500 fell by an average of 2%, but twelve months later it rose by an average of 9%. These data do not yet support the idea that "the Fed's initiation of rate hikes will inevitably end the bull market trajectory," but they cannot be used 100% to prove that future investment returns will necessarily replicate history; Goldman Sachs Group, Inc. emphasized in its research report that what is truly key is whether the earnings realization trend can offset a further decline in valuation factors. Jefferies Financial Group Inc. recently stated that, driven by the dual engines of the AI investment frenzy and AI-related corporate earnings rising above expectations, the S&P 500 is expected to surge to 8,000 by the end of 2026 and further reach 9,000 in 2027. The core logic of Jefferies Financial Group Inc. is clear and powerful: in a cycle where AI-driven earnings growth exceeds the historical average by more than twice, fighting the earnings trend is dangerous. Jefferies Financial Group Inc.'s baseline forecast of 8,000 for the S&P 500 in 2026 is based on earnings per share (EPS) reaching $373 (a year-over-year increase of 35%, far above the market consensus of 29%) and a P/E ratio of 21.5x. The Fed's rate hike lands, and retail funds appear to be shrinking buying: capital concentrates in core AI computing power assets The latest retail fund flow data compiled by JPMorgan shows that from September 10 to 16, net retail inflows were only $2.5 billion, about 63.2% lower than the average of $6.8 billion per week over the past 12 months; among this, ETFs saw $3.1 billion in net inflows, while single stocks saw $600 million in net outflows. The report places overall inflow intensity at the 2nd historical percentile, with ETF inflows falling to a one-year low and single-stock fund flows at the 12th percentile. Therefore, the most important market signal is "reduced incremental buying, more selective capital": high oil prices, long-end rates, and discussions about AI safety suppress short-term risk appetite, but core assets such as NVIDIA Corporation can still attract net buying. On Thursday (September 17), the Philadelphia Semiconductor Index, after a short-term decline, surged 3.1%, appearing in sync with a sharp pullback in oil prices and the 10-year U.S. Treasury yield falling to 4.93%, reflecting a resonance between the marginal easing of pressure from interest rates/long-dated U.S. Treasury yield curve and a strong recovery in investment confidence related to AI computing power after the Fed's rate hike restored expectations for the central bank's credibility in fighting inflation. Although in terms of timing JPMorgan's "Retail Radar" was published on the eve of a major rebound in global market technology stocks, it provides investors with crucial fund flow clues for understanding this reboundnamely, overall market entry is slowing, but selective buying of core AI assets has not been interrupted. In particular, the Magnificent Seven U.S. stocks, including NVIDIA Corporation and Alphabet Inc. Class C, still collectively obtained $1.601 billion in net buying, of which NVIDIA Corporation alone accounted for $1.196 billion. Leaders in the AI computing power industry chain such as SanDisk, ASML Holding NV ADR, and Oracle Corporation also received net buying of $202 million, $119 million, and $102 million, respectively. This structure of "overall contraction, core absorption" explains better than a broad claim that "retail investors are withdrawing from technology" why the AI theme still has a basis for rebound: while compressing overall exposure to non-Magnificent Seven technology stocks, capital continues to select certain companies in accelerated computing, storage, semiconductor equipment, and cloud infrastructure; the report also clearly records that themes such as the core infrastructure chain of AI data centers, electrification, major AI beneficiaries, and AI software commercialization continued to receive retail buying. Oracle Corporation in particular embodies the logic of "buying the dip supported by earnings and orders": its revenue grew 30% year over year, cloud computing-related infrastructure revenue grew triple digits, and remaining performance obligations increased by $26 billion quarter over quarter. It then received $87 million in single-day net retail buying on September 11, highlighting that verifiable business growth can still attract capital entry. Combining JPMorgan's latest judgment that "when earnings remain strong and GEO Group Inc risk is controlled, a shallow rate hike cycle and persistently elevated long-dated U.S. Treasury yields can still be digested by the stock market," one investment strategy that can be inferred is thiswhen the pressure on valuations from oil prices and long-end yields eases somewhat, the core AI computing power assets that had previously still received buying support will have the foundation to participate in a recovery in risk appetite, providing a capital preference backdrop for Thursday's rebound, and the market's focus will also be positioned to shift from the rate hike shock back to orders, revenue, and earnings realization. Trading rhythm and sector flows together indicate that retail investors are simultaneously making defensive allocations and selective technology investments, rather than simply switching from technology overall to defensive sectors. Over the most recent four trading days, retail investors were still net buyers of single stocks at midday but turned net sellers in the afternoon. The main change came from technology stocks outside the Magnificent Seven shifting from buying to selling, as well as accelerated selling in industrial stocks. In sector statistics excluding the Magnificent Seven, consumer staples was the only sector with net buying, at $11 million; industrials, communications, and technology saw net outflows of $555 million, $493 million, and $405 million, respectively, while utilities and real estate were among the sectors with relatively small net selling. On the ETF side, sector ETFs saw the third-largest weekly net outflow in history, mainly driven by technology products; however, broad-based large-cap ETFs still received $1.3 billion in net buying, precious metals ETFs received $188 million, and demand for medium- and long-duration bond ETFs also improved slightly. One phenomenon worth distinguishing is that crowding in market-wide high-dividend strategies is near a three-year high, but retail inflows into dividend ETFs have instead declined somewhat. At the same time, JPMorgan-compiled data shows that the one-month rolling average of retail options trading share remains at the 96.8th historical percentile, indicating that trading participation is still very high, but willingness to buy in the spot market and trading direction have diverged. From fund flows to earnings realization: AI computing power infrastructure investment shifts from broad pursuit to selected super-leader allocation The specific buy list shows that capital not only continued to support the Magnificent Seven, but also retained clear preferences in storage, semiconductor equipment, and cloud infrastructure. JPMorgan-compiled retail fund flow data is as follows (weekly retail fund flow summary for September 10 to 16): All of the Magnificent Seven received net buying: NVIDIA Corporation $1.196 billion, Tesla, Inc. $201 million, Amazon.com, Inc. $87 million, Apple Inc. $63 million, Alphabet Inc. Class C $25 million, Microsoft Corporation $19 million, Meta $10 million, totaling $1.601 billion, with NVIDIA Corporation accounting for about 74.7%. Outside the Magnificent Seven, SanDisk (SNDK.US) received $202 million in net buying, ASML Holding NV ADR (ASML.US) $119 million, and Oracle Corporation (ORCL.US) $102 million, together with NVIDIA Corporation and Tesla, Inc. forming the top five net buys for the week; the top five net sells were SpaceX (SPCX.US) $249 million, Marvell (i.e., Marvell Technology, Inc., MRVL.US) $95 million, Intel Corporation (INTC.US) $76 million, NuScale Power (SMR.US) $60 million, and AMD (AMD.US) $52 million. The report also clearly points out that retail investors continued to buy themes such as the AI data center electrification industry chain, the most core and primary beneficiaries of AI and data center infrastructure, the Magnificent Seven, growth stocks, and AI software and AI application commercialization. This fully highlights that retail investors have not abandoned AI, but have concentrated buying in the Magnificent Seven and in some of the most core companies in storage, equipment, and cloud platforms that are also the most preferred overall by capital; the non-Magnificent Seven technology sector saw overall net outflows, but that does not mean every company within it was sold. The JPMorgan analyst team upgraded Meta to "Overweight," bullish on its frontier models, Muse agent, and model API services, expanding AI commercialization from advertising further into new products and revenue sources; in cybersecurity, it is particularly bullish on Okta, Palo Alto Networks, CrowdStrike, Varonis, and Zscaler. The core logic is that AI applications expand the attack surface, making security companies infrastructure partners of foundation model companies, and making security spending more critical. It is worth noting that retail investors were still overall net sellers of this security software basket, indicating that institutional fundamental views and short-term retail trading are not completely aligned. Oracle Corporation, a leader in AI cloud computing power infrastructure, is an important positive case in fund flows: the JPMorgan report cites its 30% year-over-year revenue growth, triple-digit cloud infrastructure revenue growth, and a $26 billion quarter-over-quarter increase in remaining performance obligations (RPO), arguing that these results help address doubts about continued order growth, conversion of orders into revenue, and subsequent financing; on September 11, single-day net retail buying of Oracle Corporation reached as high as $87 million, representing dip buying after earnings. In addition, on September 15, Skyworks and Qorvo received net buying of $7.5 million and $400,000, respectively, against the backdrop of news that their merger transaction is expected to close within the year, and this should not be directly classified as AI computing power buying. Data on financing, applications, and delivery in the AI industry provide a fundamental backdrop for core assets continuing to attract attention, but they are supplementary evidence beyond fund flow observations. According to media reports, OpenAI is discussing new financing at a valuation exceeding $1.2 trillion; Anthropic, according to Reuters, is preparing for a potential IPO at a valuation of about $2 trillion with financing of up to $100 billion, which could become the largest initial public offering in historyboth are in the stage of financing discussions or preparations. On the specific AI application side, Astra strengthens capabilities in coding, browsing, computer operation, and complex workflows, broadening the range of tasks AI can perform; OpenAI also confirmed that starting September 10 it suspended new registrations and upgrades for the $200-per-month Pro 20X package, with existing subscriptions unaffected. On the procurement side, Anthropic has disclosed agreements with Amazon.com, Inc. for up to 5 gigawatts, agreements with Alphabet Inc. Class C and Broadcom Inc. for 5 gigawatts, $30 billion in Azure capacity, and a $50 billion U.S. AI infrastructure investment related to Fluidstack. On the delivery side of AI computing power infrastructure resources, NVIDIA Corporation's data center revenue for the second quarter of fiscal 2027 reached $89 billion, up 117% year over year; South Korea's August exports rose 68.7% year over year to $98.26 billion, of which semiconductor exports reached $46.65 billion, and cumulative exports for the year through early September had already reached $709.4 billion, exceeding the full-year record for 2025. From an investment logic perspective, these data translate "AI demand expansion" into application usage, capacity procurement, and hardware revenue, providing support for the long-term growth of NVIDIA Corporation, storage, and cloud infrastructure companies. This is also why short-term retail fund flows focus on shrinking buying while continuing to select some core AI computing power investment targets among them. At the technical level, the most noteworthy incremental development is that demand for AI computing power resources is expanding from centralized pre-training to post-training, continuous inference, and agent execution, which also helps explain why capital is re-screening within hardware and software. JPMorgan, citing SemiAnalysis, pointed out that pre-training's share in the computing power allocation structure it discusses has fallen below 15%, end-token demand is expanding, and post-training techniques have also become an important direction for expanding model capabilities; the decline in share describes a structural change and is not equivalent to a reduction in absolute computing power demand. Based on the "Jevons paradox," overall future AI demand will continue to expand as computing power costs decline. From a system architecture perspective, user scale, task frequency, number of invocation rounds, and context length jointly affect total token processing volume: prefill needs to process input, decode continuously generates output, and the key-value cache (KV Cache) places requirements on VRAM capacity and memory bandwidth; multi-tool agents also require CPUs to execute programs, storage chip systems to provide data and caching foundations, high-performance Ethernet network equipment to connect computing resources, AI data center optical interconnect/optical communication systems to provide high-speed transmission modes between data, and cloud platforms to complete scheduling and service delivery. This essentially forms a complete demand chain of "accelerated computingmemory and storageinfrastructure deliverysoftware commercialization," and the buying recorded in JPMorgan's fund flow report for NVIDIA Corporation, SanDisk, ASML Holding NV ADR, and Oracle Corporation corresponds to different links within it. Retail investors' fund flows still selectively favor AI computing power and technology themes, but they are no longer indiscriminately chasing every AI label. Instead, as overall buying cools, they continue to allocate to the core of the computing power theme and to some critically important bottleneck-type companies in the AI computing power industry chain.