Explosive growth in computing power orders: why is the "AI trade" receding? JPMorgan reveals the time lag between AI investment and productivity realization.

Title context: Explosive growth in computing power orders: why is the "AI trade" receding? JPMorgan reveals the time lag between AI investment and productivity realization. Text: Amid the sharp selloff and violent swings in global technology stocks, especially in the AI computing power and semiconductor sectors, a newly released research report from Wall Street financial giant JPMorgan shows that AI capital expenditure has already translated into demand for chips, servers, and power equipment, but the productivity dividend sufficient to drive sustained cost reduction and efficiency gains across the entire economy has not yet been realized in tandem. The core message of the research report means that upstream core chip and key component suppliers tied to the booming global AI infrastructure buildout can continue to secure orders and profit growth, while a broad swath of technology stocks, including neocloud players and large cloud computing platforms, simultaneously face valuation pressure from higher interest rates and lengthening payback cycles. The JPMorgan global economics research team led by Joseph Lupton published a report titled "The Productivity Boom: AI's Contribution Is Not as Large as It Looks" on November 10, extending the team's prior macro judgment of "strong economic growth, sticky inflation, persistently high interest rates, and future AI-driven productivity transformation," and revealing that there is a waiting period between AI productivity and the realization of the AI dividend and trillion-dollar-scale AI capital expenditure, with the broad productivity dividend potentially arriving later than market expectations. After AI equipment is deployed, application development, data integration, process restructuring, and employee training are still needed to fully unlock productivity gains, and this process may take longer than market consensus expects. The JPMorgan research report cites the historical phenomenon from 1987 that "computers are everywhere except in the productivity statistics," as well as the subsequent productivity improvement in the 1990s, to illustrate that there is a lag in technology diffusion. Suppliers receiving orders, computing power operators achieving profitability, and enterprises forming broad productivity gains after applying AI are three different realization processes. When financing costs rise and the market shortens the time it is willing to wait for returns, strong orders and pressure on stocks and bonds can absolutely coexist. JPMorgan's research provides a macroeconomic-level explanation for this "AI investment return time lag." The JPMorgan research report shows that more conservative academic research estimates that AI will bring a cumulative increase in total factor productivity (TFP) of less than 0.5% over the next decade; more optimistic academic research estimates that AI could raise annual productivity growth by 0.81.3 percentage points. After CoreWeave sounds the credit alarm, JPMorgan reveals the "return time lag" Computing power orders continue to expand, but the credit market has begun to ask: how much capital do companies need to invest, and how long must they wait, to obtain cash returns sufficient to cover financing costs? CoreWeave's latest disclosure shows revenue backlog reaching US$132.5 billion, but third-quarter net interest expense was as high as US$846 million, clearly exceeding operating profit of US$207 million, vividly illustrating the tension between demand growth and financing burden. Before the AI neocloud leader CoreWeave released its latest earnings report, the credit market had already sounded the alarm that AI investment returns might fall far short of expectations and that even debt pressure could cause CoreWeave to go bankrupt before it receives strong cash flow. From September 25 to October 26, the five-year credit default swap (CDS) spread of CoreWeave (CRWV.US), the stock at the very center of the "AI computing power selloff storm" and "AI bubble burst narrative," rose from 835 to 1,148 basis points; under a 40% recovery rate assumption, the model estimates a corresponding five-year cumulative implied default probability of about 62%this indicator also heavily includes credit risk premiums and cannot be directly equated with the company's actual default probability over the next five years; in addition, Meta's CDS spread for the same maturity rose from 98 to 121. The CoreWeave CDS spread subsequently narrowed to 1,041.4 basis points on the 27th, indicating that pressure began to ease; on November 5, its five-year CDS further rose above 1,289 basis points, then pulled back somewhat. These data also continuously highlight that the 10-year U.S. Treasury yield, sitting at historically high levels, raises the overall financing benchmark, while credit spreads are further tiered according to cash flow, leverage, and dependence on refinancing. Joseph Lupton of JPMorgan's global economics research team pointed out in his November 10 research that recent labor productivity improvement comes more from capital deepening and higher intensity of factor utilization, and that the broad efficiency dividend AI brings through process restructuring, data integration, and organizational adjustment still needs time to materialize. This also largely means that orders for chips, servers, and power equipment can grow first, while sustained investment returns for computing power operators and end users may not arrive in tandem; in an environment of rising interest rates and widening credit spreads, this waiting period itself will increase financing burdens and depress the present value of forward returns. The production and revenue increments during the AI buildout phase are first transmitted to chip, server, and related equipment manufacturers, which is why the share prices of Asia's AI infrastructure-related manufacturing leaders have been extraordinarily strong in recent years. If stock correlation rises further and triggers unwinding of crowded dispersion trades, the adjustment could be amplified. What will determine whether AI assets can regain a premium will be the speed at which orders convert into cash, whether financing capacity can support deliveries, and whether cost reduction and efficiency gains on the application side can be gradually realized. Chip rebound collides with rate-hike repricing, market reweighs AI investment returns At the U.S. stock market close on November 5 Eastern Time, the Nasdaq 100 index plunged 6.36%, the largest single-day drop since March 2020, with constituent stocks collectively losing about US$2.3 trillion in market value; the Philadelphia Semiconductor Index plummeted 12.25%. After Democrats won control of both chambers of Congress, the market worried that AI regulation and data center construction constraints would tighten, potentially delaying project delivery, equipment procurement, and revenue realization. Policy concerns triggered valuation adjustments, and systematic strategies such as trend-following subsequently cut positions, further amplifying the selloff. On November 9, pessimistic expectations were somewhat revised, with the Nasdaq 100 and the Philadelphia Semiconductor Index rebounding 3.40% and 6.85%, respectively; Micron and SanDisk rose 6.25% and 6.16%, respectively. However, NVIDIA Corporation bucked the trend and fell about 4.3% due to company-specific news. By Tuesday, November 10, Micron and SanDisk again pulled back about 1.1% and 3.0%, respectively, showing that funds are still repeatedly weighing strong industry demand against macro valuation pressure, and notably, the magnitude of this technology stock pullback was clearly milder, with the VIX also falling from 19.92 to 19.46, largely indicating that the market is reassessing pressure from the Federal Reserve's rate path and has not repeated last Thursday's panic selling. Data released on November 10 showed that U.S. October headline and core CPI rose 0.4% and 0.3% month over month, respectively, above expectations of 0.3% and 0.2%; the actual core reading hit the condition previously set by Fed Governor Waller, a permanent FOMC voter, for "considering a 50 basis point rate hike in December." The Dow, S&P 500, and Nasdaq Composite subsequently closed down 0.30%, 0.38%, and 0.74%, respectively. Market pricing showed that the probability of a 50 basis point hike in December rose to 11.8%, with 25 basis points still the main expectation; traders also increased bets on three more 25 basis point hikes from current levels through mid-2027, including the expected December hike. What has changed in market concerns is that this Fed tightening cycle may be faster and longer. Higher output per hour does not mean AI has already brought a full efficiency revolution JPMorgan first disentangles three concepts that are easily conflated: capital deepening, factor utilization, and total factor productivity (TFP). More equipment and software per worker can raise output per hour; more intensive equipment operation and reduced idle time in workflows can also raise output; and only after stripping out these changes, being able to produce more goods and services with given inputs is closer to broad efficiency improvement. Capital deepening itself can also include technological progress embedded in new equipment, so this decomposition cannot be understood as AI making no contribution at all. Capital deepening also includes technological progress embedded in new equipment. For example, more advanced AI chips improve computing efficiency, and part of their contribution is counted in capital deepening. Therefore, JPMorgan believes that recent productivity improvement mainly coming from capital deepening does not mean AI has made no contribution, but rather that AI has not yet formed sufficiently visible TFP gains through broad application and business process transformation. Key figures show that U.S. labor productivity grew at an annualized 2.6% over the past three years, clearly above the 1.4% of the previous economic expansion. But the decomposition cited in the report shows that capital deepening explains about half, factor utilization explains about 20%, and TFP contributes about one-third, with the latter growing at an annualized rate of about 0.9% over the past three years. Technology and non-technology capital deepening together contributed about 11.5 percentage points, of which data center-related investment may have contributed up to about 0.5 percentage points on its own. These approximate shares indicate that recent productivity improvement has relied heavily on increasing capital and using existing inputs more fully. More noteworthy is the incremental trend: the report argues that utilization-adjusted TFP growth has tended to stagnate over the past year. Data updated by the San Francisco Fed on November 9 provides similar evidenceover the four quarters through the third quarter of 2026, unadjusted TFP growth was 1.27%, while adjusted TFP fell 0.22%. This supports the judgment that "the broad efficiency dividend has not yet fully emerged," and it cannot be used to conclude that individual companies' AI applications have no benefits. From the perspective of enterprise deployment mechanisms, buying GPUs and connecting to large models completes only part of the investment. Productivity returns also require usable data, application development, permission governance, workflow restructuring, and employee training. In the early stage of deployment, these supporting investments occur first, and process benefits are released later, forming the productivity J-curve emphasized in the research report. Therefore, growth in computing power procurement, growth in cloud service revenue, and efficiency improvement at end enterprises are interrelated but not synchronized realization processes. Suppliers reap orders first, the economy still must pass through costs and interest rates JPMorgan divides the technology boom into two phases: investment and return. In the first phase, AI capital expenditure and consumption demand brought by rising asset prices come first, and capital goods suppliers are the first to receive revenue; in the second phase, after technology is widely applied, supply expands and unit costs fall, and only then do real income and corporate profits gain more durable support. That is, the first phase is investment and consumption expanding first, with demand growth outpacing supply improvement; only the second phase is technology application diffusion, productivity improvement, and unit cost decline." Report data show that global capital expenditure has grown at an annualized rate of about 7% since early 2025, and U.S. technology capital expenditure has grown about 30% over the past year. Heavy reliance on imports for equipment means the contribution of investment spending to U.S. GDP is partly offset by imports, while manufacturing-stage gains flow more to Asia. On the report's basis, global manufacturing has grown 5.6% since 2024, with Asia contributing 4.7 percentage points; the Asia scope in the related charts excludes China and India. Recent supply chain disclosures are consistent with this sequencing. Microsoft Corporation's October 28 conference call showed quarterly capital expenditure of US$45.2 billion, about three-quarters of which was for shorter-lived assets such as CPUs and GPUs, and explicitly included higher memory and storage costs. TrendForce on November 4 still observed rising server DRAM contract prices, HBM4 volume ramp-up, and a higher share of long-term agreements, while noting weak momentum in the spot market. Suppliers' price increases and revenue are precisely cloud providers' procurement costs; upstream prosperity can lead downstream full realization of investment returns. This also explains why improving productivity figures have not immediately relieved inflation pressure. More intensive use of equipment and personnel may increase overtime, maintenance, and wear-and-tear costs; although new capital expands supply, demand growth during the construction period may be faster. JPMorgan further points out that investment financing demand is rising, while the wealth effect encourages household consumption, and the economy may need a higher neutral real interest rate to attract savings and support capital formation. Notably, JPMorgan is discussing changes in the equilibrium interest rate, meaning that even if inflation falls in the future, the financing environment may not quickly return to the previous low-rate state. The JPMorgan research report also raises a condition easily obscured by the AI narrative: consumption cannot be sustained long term by rising stock prices and declining savings, and expansion ultimately needs labor income to take over. The institution expects hiring to recover cyclically, but weak labor supply may make inflation harder to bring down. Extrapolating from this framework, the JPMorgan research report shows that the investment focus across the entire AI industry chain in the future should fall on whether orders can be converted into cash, whether customers can continue to obtain financing, and whether end-user efficiency improvement can support the next round of procurement. Earnings growth at memory, power, and interconnect suppliers can offset some valuation pressure; platform and application companies that can convert AI investment into verifiable revenue or cost savings are better positioned to support long-term valuations. Growth brought simply by increasing capital investment will see its contribution to economic growth weaken once expansion slows.