AI capital expenditure boom continues, with growth downshifting in 2028 and driving value migration toward optical interconnect

Overview

The report develops around two main themes: first, estimating AI capital expenditure, free cash flow, debt issuance and debt-servicing capacity of the top five U.S. cloud service providers from 2026 to 2029; second, analyzing memory configuration reductions, NVL576 scale expansion and NPO optical interconnect adoption for Nvidia's Vera Rubin and Rubin Ultra platforms. The conclusion is that AI infrastructure demand remains strong, but capital expenditure growth will slow markedly in 2028, while hardware value may migrate from memory to GPU, CPU, optical engines, TIA and laser drivers.

Core views

Recently, Amazon, Google, Meta and Oracle raised capital expenditure guidance, while cloud business backlog and CoWoS capacity expansion jointly validate strong demand in 2026-2027. As free cash flow comes under pressure, cloud service providers will rely more on debt financing, but overall leverage and interest coverage capacity remain manageable except for Oracle. After 2028, wider financing spreads, rising capital return requirements and price competition from Chinese large models will prompt capital expenditure growth to downshift. At the platform level, Nvidia controls total system cost by reducing LPDDR5x and HBM capacity, and offsets lower per-card memory configuration with NVL576 and NPO optical interconnect; therefore, the allocation value of GPU, CPU and the optical chain is superior to memory.

Analysis framework

The report combines cloud service providers' latest earnings-call capital expenditure guidance, cloud business backlog, TSMC and OSAT CoWoS capacity, free cash flow forecasts, debt issuance and debt-servicing metrics to build capital expenditure scenarios for 2026-2029; it also uses the Vera Rubin roadmap, VR200 NVL72 bill-of-material cost breakdown, HBM layer count and capacity plans, and NPO adoption rate to estimate platform content value and optical engine market size.

Methodology notes

• capital expenditure forecastguidance and supply chain cross-validation framework

  Cross-validate cloud service provider management guidance with changes in advanced packaging capacity.

  Based on capital expenditure guidance from Amazon, Google, Meta, Microsoft and Oracle, and using CoWoS capacity expansion plans from TSMC, Amkor and ASE to test the feasibility of AI compute demand, the report forecasts capital expenditure by the top five cloud service providers.

• credit and debt-servicing analysiscapital expenditure—free cash flow—debt financing linkage framework

  Assess the continuous impact of capital expenditure expansion on cash flow, debt issuance needs and balance sheets.

  Through total debt, net debt, EBITDA leverage, net debt-to-equity ratio, operating cash flow-to-debt ratio and interest coverage ratio, the report judges cloud service providers' debt financing capacity and potential constraints.

• competitive scenario analysislarge-model pricing and capital return pressure framework

  Incorporate model price competition and market share changes into the assessment of returns on AI capital expenditure.

  By comparing per-million-token prices and token share changes of frontier large models in China and the United States, the report evaluates the impact of low-price competition on U.S. large-model margins, cloud service provider monetization capability and subsequent capital expenditure growth.

• platform value analysissystem bill-of-material cost and content migration framework

  Identify increases and decreases in supply-chain value from rack bill-of-material costs and architecture changes.

  The report estimates the cost share of HBM, LPDDR5x and SSD in the VR200 NVL72 rack, and analyzes the path of value migration from memory to optical engines, TIA and laser drivers after Rubin Ultra configuration reduction, NVL576 expansion and NPO adoption.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• NVDA

  Core beneficiary of AI accelerators and the Vera Rubin platform; the report assigns a Buy view.

  Strengths

  Vera Rubin mass production and shipment plans remain in the second half of 2026; NVL576 and NPO can expand the scale domain and enhance platform performance.

  Weaknesses

  Rubin Ultra reduces HBM and LPDDR configuration, indicating customers are more sensitive to total system cost and returns on capital.

  Comparison

  Compared with relying purely on higher memory capacity, Nvidia maintains platform performance and system value through larger-scale optical interconnect.

  Risks

  Slowing AI demand, platform deployment delays, supply constraints, intensified competition and execution of the technology roadmap falling short of expectations.

• INTC、AMD、ASE

  CPU and advanced packaging-related names preferred by the report, all assigned Buy views.

  Strengths

  CPU demand is driven by expansion of AI server platforms; AMD also participates in AI accelerator competition, while ASE benefits from CoWoS capacity expansion.

  Weaknesses

  The report does not provide independent earnings forecasts or target prices for each company, and earnings realization depends on platform share and capacity utilization.

  Comparison

  Compared with commoditized memory, CPU and advanced packaging have higher allocation priority in system expansion.

  Risks

  Customer concentration, competition, capacity ramp-up, advanced packaging yield and slowing capital expenditure growth.

• LITE、MRVL、SMTC、TSEM、Bizlink、Browave

  Beneficiaries in the NPO optical interconnect supply chain, among which LITE, MRVL, SMTC and TSEM receive Buy views.

  Strengths

  NVL576 expands the optical scale domain, and NPO increases the unit content value of optical engines, TIA and laser drivers; Marvell is viewed as a major beneficiary, and Semtech as a secondary beneficiary.

  Weaknesses

  NPO is still in the adoption stage, and shipment forecasts are sensitive to assumptions about adoption rate, platform timing and unit value.

  Comparison

  NPO manufacturing and deployment flexibility are superior to CPO, making it suitable as a reliable near-term path; CPO is better in power consumption and latency, but advanced packaging complexity and yield risks are higher.

  Risks

  Delayed NPO adoption, earlier-than-expected CPO substitution, declines in optical engine prices, platform shipments falling short of expectations and supply-chain execution risks.

• Hon Hai、SMCI

  Beneficiaries in the Nvidia GPU server and system assembly chain; SMCI receives a Buy view.

  Strengths

  Rapid growth in cloud service provider capital expenditure and Vera Rubin platform upgrades support AI server deployment demand.

  Weaknesses

  Memory configuration reduction will require server vendors to rebalance system bills of materials, and total system value and margins may undergo structural changes.

  Comparison

  System vendors benefit from growth in overall rack volumes, but incremental value may flow more to GPU and optical interconnect components.

  Risks

  Infrastructure deployment delays, supply constraints, customer capital expenditure slowdown and system costs exceeding expectations.

• top five U.S. cloud service providers

  Main drivers of AI infrastructure demand and also the parties bearing debt financing and capital return pressure.

  Strengths

  Cloud business revenue and backlog are strong, and most companies have solid balance sheets with substantial financing capacity through 2029.

  Weaknesses

  Free cash flow is under clear pressure, and capital expenditure is increasingly dependent on debt financing; Oracle faces greater financial pressure than other cloud service providers.

  Comparison

  Amazon, Google, Meta and Oracle have recently raised capital expenditure guidance, with Microsoft being the main exception.

  Risks

  Insufficient bond market capacity, widening financing spreads, longer maturities, AI monetization falling short of expectations and price competition from Chinese large models.

• HBM and LPDDR memory supply chain

  A sector on which the report is relatively cautious, affected by AI platform configuration reductions and cost-control pressure.

  Strengths

  Overall AI server shipments and HBM demand are still supported by capital expenditure expansion.

  Weaknesses

  Rubin Ultra may reduce from 12-layer to 8-layer HBM, and Vera CPU's LPDDR5x capacity may also be halved or even reduced to one-quarter of the original specification.

  Comparison

  Compared with GPU, CPU and optical interconnect, memory is more commoditized, and unit content value faces downside risk.

  Risks

  Further customer configuration reductions, price competition, decline in unit capacity value and continued optimization of platform bill-of-material costs.

Key data

• capital expenditure of the top five cloud service providersUSD823 billion in 2026; USD1.19 trillion in 2027Corresponding to year-on-year growth of 85% and 45%, reflecting recent upward revisions to capital expenditure guidance and strong AI infrastructure demand.
• capital expenditure growth downshiftUSD1.428 trillion in 2028; USD1.642 trillion in 2029Corresponding to year-on-year growth of 20% and 15%, a clear slowdown from 45% in 2027.
• debt issuance scaleUSD257 billion, USD419 billion, USD421 billion and USD432 billion from 2026 to 2029, respectivelyDebt issuance in 2026 and 2027 is expected to equal 31% and 35% of capital expenditure, respectively.
• 2029 debt-servicing metricsnet leverage of 0.5x; net debt-to-equity ratio of 18%The table estimates interest coverage of around 16x, while the text judges it at around 14x; overall, debt-servicing capacity outside Oracle remains manageable.
• cloud service provider free cash flow pressureGoogle was negative USD5.9 billion in the second quarter; Amazon was negative USD7.6 billion over the past twelve monthsMeta's second-quarter free cash flow fell 91% year-on-year to USD784 million, while Oracle's fiscal 2026 free cash flow was negative USD23 billion.
• CoWoS capacityTSMC is expected to reach 195KPM by the end of 2027Amkor is expected to reach 15, 25 and 40KPM by the end of 2026 to 2028; ASE is expected to reach 15, 45 and 55KPM.
• Rubin Ultra HBM plan8-layer HBM4 at 192GB; 8-layer HBM4E at 256GBBelow 288GB for 12-layer HBM4 and 384GB for 12-layer HBM4E; the report believes the configuration reduction is mainly driven by cloud service provider customers.
• effect of LPDDR5x configuration reductionrack cost can fall from USD1.172 million to USD586,000 or USD293,000By reducing per-CPU capacity from 1536GB to 768GB or 384GB, the share of storage and memory costs in the total system bill of materials can be controlled at around 20%.
• Nvidia platform NPO optical engines5.8 million units in 2027; 8.1 million units in 2028Estimate based on an NPO adoption rate of around 20% in 2027, with Rubin Ultra and Feynman platforms expected to adopt NPO.
• overall NPO optical engine marketapproximately 11.3 million units in 2027; approximately 40 million units in 2028Corresponding to TIA and laser driver market sizes of approximately USD819 million and USD3.087 billion.
• Semtech's NPO revenueUSD205 million in 2027; USD926 million in 2028Based on an assumed market share of around 30%, benefiting from growth in demand for high-speed TIA and laser drivers.
• changes in Chinese large-model shareDeepSeek rose from 9.1% to 18.1%; MiniMax rose from 2.3% to 9.7%From January to June 2026, Chinese models expanded token share with a higher performance-price ratio, which may depress U.S. large-model margins.

Impact & implications

In the short term, upward revisions to cloud service provider guidance and advanced packaging capacity expansion support continued rapid growth in AI hardware demand, benefiting GPU, CPU, AI servers and optical interconnect supply chains. In the medium term, the financing structure for capital expenditure will gradually shift from operating cash flow to debt, increasing the importance of credit spreads, free cash flow and return on investment. Rubin Ultra memory configuration reduction means unit content value for HBM and LPDDR faces pressure, while NVL576 and NPO adoption will increase the system value of optical engines, TIA and laser drivers. Although capital expenditure will still grow in 2028, growth will fall significantly, and industry valuation should place greater emphasis on share, earnings realization and platform content value rather than relying solely on total capital expenditure growth.

Risks

• AI demand growth slows, causing cloud service providers to cut capital expenditure or delay equipment procurement.
• Data center power, networking, facilities, system integration and other infrastructure deployment progress falls short of expectations.
• Supply constraints emerge in advanced packaging, GPU, optical devices or other key components.
• Competition intensifies among GPU, custom ASIC, cloud platforms and large models.
• Insufficient bond market capacity, rising issuer concentration and widening credit spreads increase cloud service providers' financing costs.
• Low-price competition from Chinese large models compresses U.S. large-model margins, weakening the monetization capability of massive AI capital expenditure.
• Technical validation, mass production and adoption progress for Vera Rubin, Rubin Ultra, NVL576 or NPO falls short of expectations.
• The extent of memory configuration reduction or changes in system bill-of-material costs exceeds current assumptions.

What to watch

• Subsequent capital expenditure guidance from Amazon, Google, Meta, Microsoft and Oracle, and compute capacity commitments for 2027-2028.
• Changes in cloud service providers' free cash flow, debt issuance scale, credit spreads and interest coverage ratios.
• CoWoS capacity expansion progress, utilization rates and yields at TSMC, Amkor and ASE.
• Vera Rubin mass production and shipment pace in the second half of 2026, and whether Rubin Ultra launches according to the established timetable.
• Whether Rubin Ultra ultimately adopts 8-layer or 12-layer HBM4 or HBM4E, and Vera CPU's LPDDR5x capacity configuration.
• NVL576 rollout progress, NPO adoption rate and Nvidia platform optical engine shipments.
• Market shares and unit value of vendors such as Marvell and Semtech in high-speed TIA and laser drivers.
• Chinese large-model token share, price changes and their impact on U.S. model margins and cloud business monetization.
• Whether 2028 capital expenditure growth falls to 20% as forecast, and whether investment focus shifts from expansion to returns.