Asian Tech Stocks Plunge 25-30% as Market Prices in 'EPS Cuts or Capex Reductions'! JPMorgan Predicts the Opposite

Asian tech stocks have experienced their third deep correction of over 20% in this AI-driven upcycle, with market pricing already reflecting pessimistic expectations of imminent EPS downgrades or capital expenditure cuts by hyperscalers. JPMorgan believes the reality is moving in the exact opposite direction.

According to Zhuifeng Trading Desk, JPMorgan stated in its Asian Tech Strategy report released on August 5 that although Asian tech stocks and the PHLX Semiconductor Index (SOX) have cumulatively dropped 25% to 30% recently, fundamentals show no signals indicating substantial weakness in the next 6 to 12 months. The bank explicitly stated that EPS estimates will continue to be revised upward in the coming quarters, the breadth of revisions will further expand, and the 2027 capital expenditure guidance for hyperscalers is also trending upward.

In the report, JPMorgan announced that now is the time to buy Asian tech stocks. The bank believes that valuations have become reasonable after this round of correction, with Asian tech stocks (excluding memory) currently trading at approximately one standard deviation above their 10-year average P/E ratio, which is not excessively expensive.

AI Scaling Laws Remain Intact, Compute Demand Accelerates

JPMorgan emphasized that the core logic of this upcycle—the AI Scaling Laws—remains valid. The rule of thumb that "increasing compute power by 10x for model training yields approximately a 2x improvement in intelligence" largely persists. Multiple frontier AI labs are competing to breakthrough model capabilities, with some vendors making progress in Recursive Self-Improvement (RSI), which is expected to accelerate the evolution pace of frontier models.

Public cloud revenue data confirms the strong momentum in AI compute consumption. The combined year-over-year growth rate of public cloud revenue for the four major cloud providers—Google, Microsoft, Amazon, and Oracle—has accelerated for several consecutive quarters. In Q2 2026 alone, new revenue increased by approximately $15 billion, nearly doubling compared to Q1. Meanwhile, contract backlogs for the three major public cloud providers continue to expand: Google Cloud's backlog has exceeded $514 billion, Amazon AWS reached $496 billion, and Microsoft's commercial remaining performance obligations stood at $678 billion. This provides solid support for hyperscalers to maintain high-intensity capital expenditure.

The rise of open-source large language models is not viewed as a negative. The bank believes that the increasingly fierce competition in frontier models and the frequent shifts in leadership are actually beneficial for the continued expansion of tech hardware demand and AI compute capital expenditure.

Hyperscaler Capital Expenditure to Remain Strong in 2027; Financing May Cause Short-Term Volatility

Capital expenditure for hyperscalers will maintain strong growth in 2027. According to forecasts by analysts Doug Anmuth and Samik Chatterjee, the combined capital expenditure of seven major hyperscalers—Amazon, Microsoft, Google, Meta, Oracle, Coreweave, and SpaceX—is expected to increase by approximately 103% year-over-year to about $901 billion in 2026, and further grow by 65% to approximately $1.49 trillion in 2027. SpaceX is also regarded as an emerging heavyweight in capital expenditure, poised to rival traditional cloud providers.

Regarding market concerns about the sustainability of capital expenditure, hyperscaler free cash flow will turn negative in the second half of 2026 and in 2027. However, the report argues that their balance sheets remain overall healthy—as of Q2 2026, the combined net debt-to-equity ratio for hyperscalers was approximately 12%. The bank expects these companies to replenish funds through equity and debt financing channels and will not shrink AI compute investments in 2027.

At the inventory level, there are no signs of inventory buildup for any key AI components (GPUs, ASICs, memory, etc.), which is distinctly different from the cyclical peak characteristics seen in automotive parts in 2022 or DRAM chips for cloud providers in 2017. The bank expects that AI chip demand will continue to exceed supply chain capacity and data center power budget limits over the next 12 months.

Semiconductor Equipment and IC Substrates Offer Best Allocation Value; Memory Narrative Carries Risks

In terms of sub-sector allocation, JPMorgan believes Semiconductor Processing Equipment (SPE) is the most advantageous segment over the next 12 months. The bank expects significant upward revisions to capital expenditure forecasts for TSMC and major memory manufacturers, with equipment and cleanroom space likely becoming the next bottleneck in 2027-2028.

IC substrates are listed as the segment with the most solid fundamentals within the components sector. The bullish logic includes: continuously increasing package sizes for AI accelerators, strong demand for server CPUs, accelerated penetration of EMIB-T packaging technology starting late 2027, and incremental demand from CPO-related solutions. Supply concentration is high, making large-scale capacity expansion difficult in the next two years, while profit margins remain far below previous peaks, leaving ample room for EPS upward revisions.

The memory sector is a notable exception where the bank holds a relatively cautious view. Although supply and demand fundamentals are healthy and the supply gap is expected to persist for 2-3 years, NVIDIA and AMD plan to launch low-HBM-density accelerators (Vera Rubin and MI455 with 8-HBM4 stacking) and reduce SoCAMM memory configurations for the 2027 Vera CPU to cope with DRAM supply tightness and cost pressures. This is highly similar to specification-downgrading behaviors in historical cycles, which may create a valuation ceiling for memory stocks. After a deep correction of over 40%, memory stocks may see a strong rebound in the next 6 months, but it will be difficult to recover the previous highs of May 2026 in the short term.

Interconnect Efficiency and Power Supply May Become Key Variables in the Next Phase

The report also offers forward-looking judgments on the potential shift of mid-term bottlenecks. Currently, the Model FLOPS Utilization (MFU) for GPU and ASIC clusters is generally low, ranging from only 20% to 40%, with some large clusters even below 20%. As capital expenditure rises, power budgets tighten, and chip supplies remain constrained, AI infrastructure providers and model labs will place greater emphasis on improving cluster efficiency. Interconnect technologies—including CPO and optical connections, network overhead optimization, etc.—are expected to become the next key bottleneck, driving the accelerated application of 3DSoIC advanced packaging, SRAM memory hierarchy expansion, and CPO on interposers.

Over a longer timeframe, the report believes that AI chips will remain the primary bottleneck for compute infrastructure for most of 2026-2027. However, in the next 18 to 24 months, as semiconductor supply continues to ramp up, data center deployment delays and power supply issues (including grid access and behind-the-meter power solutions) may replace chips as the core variable constraining AI compute expansion. This risk factor is expected to come into clearer view from the second half of 2027 to 2028.