JPMorgan: "Open source disruption" and "AI safety" are not issues, there is still room for capital expenditure in the next two years, semiconductor equipment will become the "new bottleneck"

Recently, market confidence in AI infrastructure investment has wavered. The rise of open-source large models, increasing calls for AI safety regulation, and negative free cash flow among hyperscale cloud providers—these three combined pressures have kept sentiment in the tech hardware sector persistently subdued.

According to Odaily Planet Daily, JPMorgan analysts including Gokul Hariharan addressed these three major concerns one by one in an Asian tech strategy report released on September 16. Their conclusion: The fundamentals of computing power demand remain solid, AI safety regulation concerns are a short-term disturbance, the Capex cycle will continue, and semiconductor equipment will become the most critical bottleneck in the entire supply chain around 2027.

"Buying Computing Power" Is Also a Good Business

The first market concern is that while "selling computing power" is profitable, can "buying computing power" (i.e., AI model providers and vertical AI application companies) also remain profitable over time?

Analysts point out that multiple reports now show current inference business gross margins reaching 60% to 80%.

The report further estimates: Assuming a model provider sells AI tokens, each GW of computing power can generate $20 billion to $40 billion in annual revenue, far higher than the approximately $10 billion per GW expected in 2025.

Citing data from Sapphire Venture, the report notes that over 80 vertical AI companies have achieved annual recurring revenue (ARR) of over $100 million in an extremely short period, validating value creation on the demand side of computing power consumption.

Open Source Is Not a Threat, But a Demand Catalyst

The second market concern is that open-source large model proliferation will depress token prices and erode the profit margins of model providers.

JPMorgan holds the opposite view. Analysts believe that historically, open-source models have driven token costs down continuously (with long-term declines of 80% to 90%), and this trend will accelerate AI’s penetration in various industries rather than weaken overall computing power demand.

Analysts lay out three logical points:

1. With computing power supply still tight, both proprietary frontier models and open-source model token consumption will continue strong growth;

2. The popularity of open source will drive AI deeper into broader application layers;

3. More vertical AI application companies will utilize low-cost open-source tokens to solve industry-specific problems, creating new revenue streams.

Analysts especially point out that the token cost’s temporary rebound in H1 2026 is an "outlier" caused by severe computing power shortages combined with the rise of agent AIs, not a trend reversal.

AI Safety Regulation: Short-term Disturbance, Unlikely to Change Training Pace

The third concern comes from regulation. According to Bloomberg, the CEO of Anthropic recently publicly called for slowing the pace of AI model improvements, raising market concerns about the outlook for computing power demand.

Analysts believe this concern is similarly exaggerated.

They observe: "AI models are evolving even faster, and the proximity of RSI (Recursive Self-Improvement) only proves that the technology’s evolution and AI scaling laws have not slowed."

Analysts further note that even if new model launches may be constrained by alignment restrictions and other factors, the training pace of frontier models themselves is unlikely to slow. At the same time, open-source models will not stop, which will instead pressure frontier model labs to accelerate training in order to maintain their lead.

JPMorgan believes that the issue of AI safety is actually a sign of regulation catching up to rapid technological evolution, not a signal of demand weakening. The report expects generative AI to open new market spaces in the next 1 to 2 years, similar to the new programming and agent workflow markets formed over the past 18 months.

Capital Expenditure: Leverage Still Low, Cash Flow Accelerating

The financing ability of hyperscale cloud providers’ Capex is the fourth market concern. Many cloud providers have entered a phase of negative free cash flow, and external financing pressure is rising.

JPMorgan’s judgment is: there is still room for Capex growth in the coming years.

There are three reasons:

1. Large hyperscale cloud providers’ net debt-to-equity ratios are currently only 13%, meaning leverage remains low;

2. Public cloud business is accelerating and carries higher pricing, and will be a major source of operating cash flow growth;

3. Equity financing from hyperscale cloud providers and downstream buyers of computing power can provide additional financial support.

Analysts forecast that combined Capex for Amazon, Microsoft, Google, Meta, Oracle, Coreweave, and SpaceX will rise from $443 billion in 2025 to an estimated $933 billion in 2026 and further to $1.577 trillion in 2027, with YoY growth rates of 71%, 110%, and 69%, respectively.

Analysts also warn that variables such as rising interest rates and changes in risk appetite within some AI infrastructure funding areas require ongoing attention.

Semiconductor Equipment: The Next Bottleneck Is Forming

On the supply chain side, JPMorgan believes that looking ahead to 2027, semiconductor equipment (SPE) will shift from being a bottleneck in “cleanrooms and other areas” to becoming the single most critical constraint in the entire supply chain.

On the demand side, drivers include:

• TSMC will significantly ramp up equipment investments for N2, N3, and A14 process nodes;

• Intel, Samsung foundry, Terafab and other challengers will also increase Capex;

• Mature process fabs are starting their first new investment cycle in four years;

• DRAM and NAND Flash manufacturers will rapidly expand equipment purchases as cleanroom capacity comes online in 2027–2028;

• Chinese memory investment is also set to accelerate significantly over the next two years.

On the supply side, JPMorgan points out that semiconductor equipment vendors are now enjoying rare pricing power—due to capacity constraints, rising input costs, and urgent orders from multiple customers.

Packaging and Substrate: Persistent Bottlenecks, Optical Interconnect on the Rise

At the component level, JPMorgan believes IC substrates and advanced packaging remain key bottlenecks.

Substrate supply is highly concentrated among leading manufacturers such as Unimicron and Ibiden, with capacity expansion cycles as long as 2.5 years. As AI accelerators increase substrate demand per package and EMIB-T packaging technology is introduced at the end of 2027, the supply-demand gap is expected to widen further.

For optical interconnect, JPMorgan expects Google, AWS, and Nvidia to comprehensively adopt NPO (Near Package Optics) solutions over the next two years to address performance bottlenecks in GPU-to-GPU, GPU-to-CPU, and storage interconnect. CPO (Co-Packaged Optics) is the long-term direction, but the supply chain still needs time to mature.

Memory: Fundamentals Healthy, But Sentiment Needs Repair

The memory sub-sector is an area where JPMorgan holds a more cautious stance.

The report notes that concerns over HBM spec downgrades, improved algorithm efficiency (e.g., recurrent Transformers) reducing memory demand, and KV cache migration to lower-tier storage like DDR or NAND, have made the investment logic in this segment relatively noisy.

But the fundamentals are not pessimistic: supply-demand balance is not expected before 2028, and prices are set to continue rising until 2027. According to JPMorgan, if AI sentiment recovers in Q4 2026, memory stocks could rally along with it. But before that, investor participation may remain lower than in other tech sub-sectors.

In 2027, Which Segments Will See Sharper Price Increases?

JPMorgan has identified sub-sectors where price increases are expected to accelerate in 2027 compared to 2026:

• Foundry: TSMC is expected to raise prices across all process nodes in 2027 (whereas 2026 increases are limited to advanced nodes); mature 8-inch nodes may see hikes of 10%–15%, higher than the 8%–10% in 2026;

• OSAT: Price increases will expand to wire bonding and mature flip chip packaging;

• Substrate: The price hike cycle kicks off, and EMIB-T’s introduction will further tighten supply-demand;

• High-end CCL (M7 and above): Driven by both AI demand and server spec upgrades, supply growth continues to lag demand;

• Memory: The pace of price increases is expected to slow;

• PCB Materials: Pricing momentum is also weakening.

Analysts note that EPS revision trends in the Asian tech hardware supply chain remain healthy, and valuations are attractive.

Although current market sentiment is weak, there are no clear signs yet of an imminent downturn in Capex cycles. The key catalyst driving the next bull run will still come from further evidence of ongoing monetization capacity at the model layer and AI application layer.