AI capex continues high growth, with value shifting from memory to GPUs, CPUs, and NPO optical interconnects

Overview

The report analyzes U.S. cloud service provider capex from 2027 to 2029, debt financing capacity, and value migration in Nvidia’s next-generation platforms. Recently, Amazon, Google, Meta, and Oracle raised capex guidance, while cloud business order backlogs and CoWoS capacity expansion jointly validate demand for AI infrastructure. The research judges that high growth will continue in 2027, but after 2028, cloud service providers will focus more on free cash flow, financing costs, and investment returns, causing capex growth to gradually cool. Meanwhile, Vera Rubin and Rubin Ultra control BOM costs by reducing LPDDR5x and HBM specifications and expand optical scale-up, driving value to migrate from memory toward NPO optical engines, TIAs, and laser drivers.

Core views

First, U.S. Top 5 CSP capex is expected to rise from US$446bn in 2025 to US$1,190bn in 2027, with guidance increases from Amazon, Google, Meta, and Oracle and CoWoS capacity expansion serving as the key validation. Second, capex reliance on debt financing has increased significantly, but by 2029, Top 5 CSP net leverage is expected to be only 0.5x and net debt-to-equity 18%, remaining broadly controllable except for Oracle. Third, widening financing spreads, pressure on free cash flow, and price and share competition from Chinese large models will prompt capex growth to fall to 20% in 2028. Fourth, the Nvidia platform controls system costs by lowering memory specifications while expanding the scale-up domain through NVL576 and NPO optical interconnects, with optical devices and analog chips set to capture higher value.

Analysis framework

The research combines the latest capex guidance from cloud service provider earnings calls, order backlogs, free cash flow, debt issuance, and debt-servicing metrics to establish forecasts for 2026 to 2029; uses CoWoS capacity from TSMC, Amkor, and ASE as cross-validation for demand; evaluates the impact of memory specification reductions through VR200 NVL72 system BOM scenario analysis; and then calculates the market opportunity for NPO, TIAs, and drivers based on the Nvidia platform roadmap, NPO penetration, optical engine quantities, and value per system.

Methodology notes

• Capex ForecastingBottom-up Top 5 CSP Forecast

  Combination of company guidance and capacity validation

  Forecasts capex for MSFT, AMZN, Google, Meta, and Oracle separately, and uses cloud business order backlogs and changes in CoWoS capacity to validate demand strength.

• Credit and Debt-Servicing AnalysisFree Cash Flow–Debt Financing Framework

  Assessing the sustainability of capex financing

  Evaluates cloud service providers’ financing capacity through the ratio of debt issuance to capex, gross leverage, net leverage, net debt-to-equity, CFO-to-debt ratio, and interest coverage ratio.

• Cost AnalysisSystem BOM Scenario Analysis

  Impact of memory specification adjustments on full-system cost

  Compares original specifications, the main specification-reduction scenario, an extreme specification-reduction scenario, and a no-adjustment scenario to calculate HBM, LPDDR5x, and SSD costs and their share of rack BOM.

• Market SizingNPO Optical Engine Volume-Price Model

  Platform shipments multiplied by penetration rate and value per system

  Calculates shipments and market size based on NVL72 and NVL144 deployment scale, NPO adoption rate, the number of optical engines required per platform, and the unit prices of TIAs and drivers.

• Competitive AnalysisLarge Model Share and Price Comparison

  Impact of model cost-performance on AI investment returns

  Compares monthly Token share and price per million Tokens for Chinese and U.S. frontier large models to assess the pressure from price competition on U.S. large model margins, commercialization, and cloud service provider capex returns.

Asset mapping & comparison

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

• NVDA, SMCI, and Hon Hai

  Core beneficiaries of GPU platforms and AI servers

  Strengths

  Vera Rubin mass production schedule remains on track, NVL576 expands the scale-up domain, and AI infrastructure capex growth supports platform demand.

  Weaknesses

  Memory specification reductions reflect rising customer sensitivity to system cost and investment returns.

  Comparison

  Compared with commoditized segments such as memory, GPU platforms have stronger system control and value capture capabilities.

  Risks

  Changes to Rubin Ultra schedule, slowing AI demand, supply constraints, catch-up by competing platforms, and delays in infrastructure deployment.

• INTC, AMD, and ASE

  Beneficiaries of CPUs, custom accelerators, and advanced packaging

  Strengths

  Cloud service providers continue to expand computing capacity, and CoWoS capacity expansion validates demand for advanced packaging; AMD’s Meta-customized MI450 has also shifted to an 8-Hi solution.

  Weaknesses

  Rapid changes in platform specifications may lead to volatility in product mix and unit value.

  Comparison

  CPU and packaging segments benefit from the expansion of overall computing platforms, but platform pricing power is usually weaker than that of leading GPU suppliers.

  Risks

  Slower customer capex, product competition, yield risk, and capacity ramp-up risk.

• LITE, MRVL, SMTC, TSEM, Bizlink, and Browave

  Beneficiaries of NPO optical engines, TIAs, drivers, and connectivity solutions

  Strengths

  NVL576 and larger scale-up domains increase the number of optical interconnects, while NPO’s removal of certain module-side DSP functions increases the value of TIAs and laser drivers.

  Weaknesses

  Market opportunity is highly dependent on assumptions regarding NPO adoption rate, platform deployment pace, and the number of optical engines per system.

  Comparison

  NPO has better manufacturability and deployment flexibility than near-term CPO solutions, while CPO may still be the longer-term end-state architecture; MRVL is viewed as a primary beneficiary and SMTC as a secondary beneficiary.

  Risks

  Earlier-than-expected maturity of CPO, lower-than-expected NPO penetration, price declines, customer concentration, and platform design adjustments.

• HBM, LPDDR5x, and SSD supply chain

  Suppliers of AI server memory and storage components

  Strengths

  Overall expansion in AI rack shipments can still provide demand support.

  Weaknesses

  CSPs are pushing Vera CPU and Rubin Ultra to reduce memory specifications to avoid a significant increase in memory cost as a share of BOM.

  Comparison

  Compared with GPUs, CPUs, and optical interconnects, memory has more commodity-like attributes, and unit content value faces pressure from specification reductions.

  Risks

  Further specification reductions, price competition, customer inventory adjustments, and oversupply.

• MSFT, AMZN, Google, Meta, and Oracle

  Main AI capex entities and sources of cloud infrastructure demand

  Strengths

  Cloud business order backlogs are strong, and most companies have sufficient balance sheet strength and interest coverage capacity to continue investment through debt issuance.

  Weaknesses

  Free cash flow is under significant pressure, and capex reliance on debt financing is rising; Oracle’s financial pressure is relatively prominent.

  Comparison

  Except for Oracle, most leading CSPs’ leverage remains controllable through 2029, but financing costs and return requirements will limit growth.

  Risks

  Widening credit spreads, insufficient bond market capacity, AI commercialization returns below expectations, and price competition from Chinese large models.

Key data

• Top 5 CSP capex2026E US$823bn; 2027E US$1,190bn; 2028E US$1,428bn; 2029E US$1,642bnCorresponding to YoY growth of 85%, 45%, 20%, and 15%.
• Top 5 CSP debt issuance2026E US$257bn; 2027E US$419bn; 2028E US$421bn; 2029E US$432bnEquivalent to 31% and 35% of capex in 2026 and 2027, respectively.
• 2029 debt-servicing metricsGross leverage 1.1x; net leverage 0.5x; net debt-to-equity 18%; interest coverage 16xMetrics are based on aggregate forecasts for the Top 5 CSPs and exclude off-balance-sheet obligations; Oracle’s balance sheet risk is relatively higher.
• CoWoS capacity forecastTSMC 195KPM by end-2027; Amkor 15/25/40KPM by end-2026 to end-2028; ASE 15/45/55KPMAdvanced packaging capacity expansion is used to cross-validate AI accelerator demand.
• VR200 NVL72 memory and storage costMain specification-reduction scenario US$1.515m; extreme specification-reduction scenario US$1.221m; no-specification-reduction scenario US$2.101mThe extreme specification-reduction scenario can keep memory and storage costs at about 19% of BOM, while no adjustment would raise the share to 29%.
• Rubin Ultra potential HBM specifications8-Hi solution is 192GB HBM4 or 256GB HBM4E12-Hi options still include 288GB HBM4 and 384GB HBM4E.
• Nvidia platform NPO optical engine shipments2027E 5.8m; 2028E 8.1mBased on assumptions including a 20% NPO adoption rate in 2027.
• Overall NPO optical engine market2027E 11.3m; 2028E 40mCorresponding TIA and driver market size is approximately US$819m and US$3.087bn.
• Semtech NPO revenue forecast2027E US$205m; 2028E US$926mBased on assumptions of approximately 30% market share and front-end chip value per optical engine.

Impact & implications

AI infrastructure investment remains in an expansion phase, but value distribution across the industry chain is changing. In the short term, high capex and advanced packaging capacity expansion continue to support demand for GPUs, CPUs, and related equipment; in the medium term, cloud service providers will rely more on debt financing, making interest rates, credit spreads, and free cash flow key constraints on valuation and order sustainability. At the platform end, lowering memory specifications to control BOM will suppress unit value for memory, but will drive larger-scale optical interconnects to maintain system performance, so the content value of NPO optical engines, TIAs, and laser drivers is expected to increase significantly. CPO has long-term advantages in power consumption and latency, but given near-term limitations from advanced packaging complexity and yield, NPO is more likely to become a reliable transition and scale-up path.

Risks

• AI demand growth is lower than expected.
• Delays in data center power, networking, and other infrastructure deployment.
• Supply constraints in advanced packaging, chips, optical devices, or other key components.
• Intensifying competition in GPUs, ASICs, cloud platforms, and large models.
• Deterioration in cloud service provider free cash flow, rising financing rates, or insufficient bond market capacity.
• Low-price competition from Chinese large models compresses U.S. large model margins and capex returns.
• Vera Rubin or Rubin Ultra technology validation, mass production, and shipment progress falls short of expectations.
• NPO penetration and optical engine shipments are below forecasts, or the CPO roadmap replaces NPO more quickly.

What to watch

• Subsequent capex guidance from Amazon, Google, Meta, Microsoft, and Oracle and compute capacity bookings for 2027 to 2028.
• Top 5 CSP quarterly free cash flow, debt issuance scale, credit spreads, and interest coverage ratios.
• Token share and pricing changes of Chinese large models and their impact on U.S. model commercialization.
• CoWoS capacity expansion, utilization, and yield at TSMC, Amkor, and ASE.
• Vera Rubin mass production and shipment progress in 2H26, and whether the Rubin Ultra schedule is maintained.
• The final HBM4/HBM4E stack count, capacity, and power specifications adopted by Rubin Ultra.
• NVL576 deployment pace, NPO adoption rate, and the number of optical engines per platform.
• TIA and laser driver orders, market share, and value per system for vendors such as MRVL and SMTC.