JPMorgan: AI Infrastructure Investment at a Crossroads, Asian Tech Hardware Remains Attractive
nashnova research
J.P. Morgan's latest report frames AI infrastructure around three questions — compute economics, safety regulation, and capex funding — and concludes that AI investment remains healthy; Asian tech hardware has already priced in some downside, but no capex inflection signal has appeared.
Is buying compute power still a good business?
J.P. Morgan estimates that a model provider deploying 1 GW of compute capacity can generate $20–40 billion in annual revenue from selling AI tokens — well above the roughly $10 billion per GW level in 2025.
Some companies already report inference gross margins of 60–80%; separately, over 80 vertical AI firms have crossed $100 million in annual recurring revenue in a short span.
This means → compute investment is still converting into tangible profit; AI commercialization is past the pitch-deck stage and delivering real returns.
Will open-source models undercut the whole value chain?
J.P. Morgan says no. Historically, per-token costs fall 80–90%, yet lower costs drive AI adoption into more industries, expanding total demand.
In plain terms = prices drop, usage surges, and the pie grows larger — the same pattern that played out when smartphones became affordable.
Open-source models will also spawn more vertical applications; both closed- and open-source token consumption are expected to grow strongly.
Where will hyperscalers find the money?
Major hyperscalers still carry low leverage — net debt to equity sits at roughly 13%.
Strong compute demand and higher pricing are accelerating public-cloud revenue growth; operating cash flow is poised to become a key funding source, supplemented by equity issuance.
This means → even as some hyperscalers turn free-cash-flow negative, their balance sheets retain ample room to sustain capex expansion — not a binding constraint in the near term.
Which segments have the most pricing power into 2027?
Foundry — especially 8-inch mature nodes — is expected to raise prices 10–15%, up from 8–10% in 2026; OSAT, substrates, semiconductor equipment, and high-end CCL also have room to reprice.
IC substrates — the boards connecting a chip die to the circuit board — and advanced packaging remain key bottlenecks; the gap may widen further once Intel's EMIB-T packaging ramps in H2 2027.
By contrast, memory and PCB materials may see slower price increases — this reflects the fact that supply-demand cycles across segments are not synchronized.
Which optical-interconnect route is winning?
J.P. Morgan currently favors near-package optics — placing optical modules close to, but outside, the chip package.
Google, AWS, and Nvidia are all expected to adopt near-package optical solutions over the next two years.
Co-packaged optics — integrating optical modules directly inside the chip package — represents the long-term direction, but supply-chain readiness still needs more time.
Why are memory stocks flat despite healthy fundamentals?
J.P. Morgan sees memory supply-demand balance arriving no earlier than 2028; prices are expected to keep rising through 2027 — fundamentals remain sound.
Yet investor enthusiasm is muted: concerns about HBM — high-bandwidth memory, the critical companion to AI chips — spec downgrades and algorithm-efficiency gains reducing memory consumption have not faded.
In plain terms = the data is improving, but sentiment is stuck on "what if AI ends up needing less memory?" — a shift may require the next wave of AI models to clearly raise per-function memory demand.
市场有风险，内容仅供研究参考，不构成投资建议。