Global AI investment trade and infrastructure cycle Report Interpretation

Overview

HSBC addresses seven major investor debates around the global AI trade: monetization, model-price deflation, hyperscaler capex and financing, Korean memory, mainland China localization, and portfolio positioning. Its central view remains constructive, although the next phase depends on proof that AI revenue and unit economics can support the scale of infrastructure spending.

Core views

HSBC argues that AI monetization still has substantial runway. Reported annualized revenue run-rates are USD65bn for Anthropic and USD40bn for OpenAI, while enterprise adoption remains early: median AI spending is USD12 per employee per month, compared with USD650 for the top spending decile. AI spending in Financials and Manufacturing is up 3.5x and 2.5x year-to-date, respectively. Microsoft reported more than 30m paid M365 Copilot seats in Q2-26, with net additions more than doubling sequentially, while Alphabet said nearly 90% of the Fortune 100 use Gemini Enterprise. HSBC sees corporate budgets as capable of absorbing higher AI outlays: USD650 per employee per month across the average S&P 500 company would equal USD250bn annually, or 5% of EBITDA, 8% of wages and 27% of R&D. Better model capability is central to this thesis, because longer and more reliable task completion can move AI from individual prompts to end-to-end workflows. METR data show Claude Mythos completing tasks lasting more than three hours with an 80% success rate, versus roughly 20-minute tasks for a comparable success rate from Claude Opus in May 2025. HSBC identifies Anthropic's prospective S-1 as a test of the monetization thesis, with customer concentration, retention, gross margin after inference costs, compute commitments and cash conversion more important than headline revenue growth. On model pricing and competition, HSBC accepts that token prices are falling—average token prices were down 50% from their May 2026 peak—and that Chinese models are nearing 50% of token volume. However, it argues these changes need not reduce aggregate AI expenditure. It cites a Jevons-paradox effect: token volume has risen nearly threefold since June, and OpenRouter discounts drove daily usage of OpenAI Terra and Luna up 5.6x and 13.8x, respectively, versus 1.1x for an undiscounted model. The report distinguishes token volume from token spending: Anthropic and OpenAI have retained around 80% of spend since 2025. It also argues that cost per task is more relevant than cost per token for complex workloads; OpenAI Astra costs USD7.7 per million tokens, about 6.5x Kimi3's price, but its cost per task is only twice as high because of greater capability. HSBC expects US frontier models to retain an advantage in complex reasoning and agentic use cases, while Chinese and open-weight models gain share in high-volume, price-sensitive and specialized workloads. Hyperscalers can still monetize open-model adoption through cloud access, accelerator rental, storage and databases. The report recognizes the enormous scale of hyperscaler investment but sees limited evidence of aggregate overcapacity today. Consensus capex estimates for US hyperscalers are USD770bn for 2026 and USD1.1trn for 2027, revised upward by USD170bn and USD260bn since the Q2 earnings season. Outstanding purchase and uncommenced lease commitments total around USD2.5trn, while S&P Global Ratings estimates more than USD7trn of data-center and AI-related spending from 2025 to 2030. Against this, remaining performance obligations across Microsoft, Amazon, Alphabet and Oracle reached about USD2.35trn in Q2-26, up from USD815bn one year earlier. Microsoft commercial RPO rose 25% excluding OpenAI, and Alphabet's USD514bn Cloud backlog is mainly standard GCP contracts, with just over half expected to become revenue within 24 months. Firm rental prices for NVIDIA B200 and H100 GPUs also indicate that usable compute remains scarce, partly because power, construction and commissioning delay capacity additions. HSBC estimates AI-ecosystem annual revenue of about USD920bn by 2030, with enterprises contributing 75%, and expects hyperscaler ROIC to decline but remain at 17% by 2028. The overcapacity risk would become more material if capex persists after compute prices soften, backlog growth slows or utilization falls. HSBC nevertheless highlights a timing mismatch between investment and monetization. Hyperscaler free cash flow has turned negative and is expected to remain pressured through end-2027, as spending precedes the revenue opportunity expected by 2030. Debt issuance was nearly USD250bn year-to-date in 2026, and the five hyperscalers plus CoreWeave have over USD1trn in uncommenced lease commitments that could support roughly USD400bn-500bn of off-balance-sheet borrowing, against just over USD500bn in bonds outstanding. Financing flexibility can match funding to individual data-center assets, but credit-market sensitivity is rising: bond yields and CDS spreads have increased, and Oracle's downgrade to BBB- illustrates the potential pressure from negative free cash flow, leverage and customer concentration. GPU and AI servers are expected to account for about 70% of hyperscaler capex in 2026. HSBC expects depreciation and amortization to rise from USD110bn in 2024 to USD730bn-930bn by 2030, while consensus expects ROE to decline from a 38% Q2-26 peak to 23% by Q4-28. Thus, the cycle may be economically rational long term while still producing several years of weaker shareholder returns. HSBC remains bullish on Korean memory after the correction. The stocks are about 35% below their peaks, and its market-implied earnings model suggests investors shortened the expected earnings-cycle duration for Samsung Electronics and SK Hynix from about six years to 2.5 years. Foreign selling across emerging-market technology reached roughly USD150bn year-to-date, including USD60bn since June, while leveraged single-stock ETF assets fell from USD37bn at their peak to USD10bn. HSBC believes selling pressure has largely subsided and that lower volatility could support valuations. Fundamentals are supported by rising hyperscaler capex: it estimates about 70% of US hyperscaler capex goes to GPUs and servers, with nearly 30% of that spending on memory. HBM4, server DRAM, eSSD demand and supply tightness are identified as catalysts. The major downside is expanding conventional DRAM supply from CXMT, whose capacity may rise by about 200,000 wafers a month to 300,000 by early next year and whose DRAM share could reach 18% by 2028 from below 8%. HSBC sees long-term agreements, deposits and price floors as partial protection. It also highlights rising shareholder returns, including SK Hynix's KRW40trn buyback and cancellation and Samsung's expected KRW90trn-110trn of 2026 shareholder returns. For mainland China, HSBC prefers the domestically focused AI semiconductor and hardware stack, where inference growth, localization and policy support are driving demand. Domestic fabs supply less than 15% of the country's memory and compute demand, so added local capacity can replace overseas supply. Less efficient domestic chips require more chips, packaging, networking and power to produce equivalent usable compute, broadening potential beneficiaries beyond accelerators to equipment, servers, cooling, data centers and power infrastructure. Lithography remains a constraint because China has lacked access to state-of-the-art EUV equipment since 2019 and is unlikely to produce a domestic EUV machine before 2028. HSBC points to SMIC's Q2 results—AI-peripheral-chip revenue up about 40% quarter-on-quarter, utilization near 94%, and wafer ASP up 5.7% quarter-on-quarter—as evidence that localization is translating into fundamentals. It sees equipment suppliers such as Naura, AMEC and ACM Research benefiting as production shifts onshore. Conversely, it is more selective on optical-module companies with greater US exposure because of FCC restrictions, and on internet applications: Alibaba's external cloud revenue grew 45% year-on-year in Q2-26, but Tencent's AI initiatives reduced operating profit by about RMB10.5bn and its RMB52.8bn capex, up 176% year-on-year, pushed free cash flow negative. For portfolio positioning, HSBC favors areas with strong demand and monetization prospects where valuations are less stretched. It is increasingly tilted to emerging markets—Korean memory, Taiwan semiconductors and advanced packaging, and mainland China semiconductor equipment and power infrastructure. In Taiwan, growing use of custom ASICs is broadening demand beyond GPUs into advanced packaging; Asian OSAT companies have raised 2026 capex by more than 80%, yet CoWoS equipment lead times can reach 12 months and capacity is expected to remain tight through at least 2027-28. In the US, HSBC ranks semiconductors ahead of hyperscalers and software because semiconductor and cloud-infrastructure revenues are less dependent on which model or application succeeds. It views the recent software rotation as more short-covering-driven than a fundamental improvement. In Europe, it also focuses on semiconductors, citing strengthening analog-cycle data and potential price increases that could improve utilization and profitability.

Analysis framework

HSBC structures the report around seven investor debates. It compares adoption, spending, token usage, backlog, compute pricing, financing, earnings expectations, supply constraints and valuation signals to test whether AI revenue can justify the infrastructure buildout and to identify relative regional and sector exposures.

Methodology notes

• Industry AnalysisSupply-demand framework

  Compute, memory and advanced-packaging supply-demand analysis

  The report uses backlog, GPU rental prices, utilization, equipment lead times and supply constraints to judge whether AI infrastructure demand is exceeding available capacity.

• Corporate Fundamentals and FinanceROIC–WACC spread

  Hyperscaler returns on invested capital

  HSBC evaluates whether large AI capex can still earn adequate returns, forecasting ROIC at 17% by 2028 despite a decline.

• Event-Driven and Behavioral FinanceExpectation Gap and Expectation Management

  Market-implied earnings-cycle expectations for Korean memory

  The report compares current implied earnings assumptions with those at the market peak to argue that investor expectations have become more pessimistic.

• Other

  Jevons paradox

  HSBC uses the idea that lower token prices can stimulate sufficiently greater usage to increase total AI spending rather than reduce it.

Asset mapping & comparison

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

• Samsung Electronics (005930 KP)

  Korean memory beneficiary of AI-driven GPU, server DRAM and HBM demand

  Strengths

  Memory demand, long-term contracts and expected shareholder returns

  Weaknesses

  Conventional DRAM exposure

  Comparison

  Market-implied earnings-cycle duration has compressed materially from the peak

  Risks

  Greater conventional DRAM supply from Chinese producers

• SK Hynix (000660 KP)

  Korean memory beneficiary of HBM4, server DRAM and eSSD demand

  Strengths

  Multi-year customer agreements and KRW40trn buyback and cancellation

  Weaknesses

  High realized volatility

  Comparison

  The report groups it with Samsung as a sector re-entry opportunity after the correction

  Risks

  Chinese conventional DRAM capacity expansion

• Naura, AMEC, ACM Research

  Potential beneficiaries of mainland China's domestic semiconductor-equipment orders

  Strengths

  Exposure to onshore production and localization

  Comparison

  Beneficiaries of the domestic hardware buildout rather than AI application monetization

  Risks

  Technology constraints and localization execution risk

• US semiconductors and cloud infrastructure

  Preferred US AI exposure

  Strengths

  Direct exposure to training and inference growth regardless of the winning model or application

  Weaknesses

  Dependent on sustained capex and monetization

  Comparison

  Preferred over hyperscalers, which are preferred over software

  Risks

  Compute-price softening, weaker backlog growth and falling utilization

Key data

• Anthropic annualized revenue run-rateUSD65bnReported as of July 2026
• OpenAI annualized revenue run-rateUSD40bnReported in August 2026
• US hyperscaler capexUSD770bn in 2026e; USD1.1trn in 2027eConsensus estimates, revised up by USD170bn and USD260bn
• Hyperscaler RPO backlogApproximately USD2.35trnQ2-26, versus USD815bn one year earlier
• Hyperscaler financingNearly USD250bn debt issuance in 2026; more than USD1trn uncommenced lease commitmentsThe latter could support approximately USD400bn-500bn of off-balance-sheet borrowing
• Korean memory correctionApproximately 35% below peakHSBC views the correction as improving valuations, positioning and earnings risk
• SMIC AI-peripheral-chip revenue growthApproximately 40% q/qLatest Q2 earnings; utilization approached 94% and wafer ASP rose 5.7% q/q

Impact & implications

HSBC's constructive AI view rests on continued enterprise adoption, durable infrastructure demand and constrained supply, but it distinguishes long-term economic potential from nearer-term pressure on cash flow, leverage and returns. It favors regional and supply-chain exposures that benefit directly from AI infrastructure demand and sees greater selectivity as necessary in software and Chinese internet monetization.

Risks

• AI infrastructure supply could eventually grow faster than revenue available to support the capex.
• Free-cash-flow pressure, rising leverage, lease commitments and higher financing costs may weigh on hyperscaler shareholder returns.
• A softening in compute prices, slowing backlog growth or declining utilization would increase overcapacity concerns.
• Rapid CXMT capacity expansion could pressure conventional DRAM pricing and Korean memory earnings.
• Mainland China technology constraints, including restricted access to advanced EUV equipment, may limit localization progress.
• US-facing Chinese optical-module companies face structural pressure from FCC restrictions.

What to watch

• Anthropic's prospective S-1 disclosures on customer concentration, retention, post-inference gross margins, compute commitments and cash conversion.
• Whether enterprise AI spending and model capability continue to translate into sustainable monetization.
• Hyperscaler backlog growth, GPU rental prices, compute utilization and capex revisions.
• The pace of hyperscaler revenue growth relative to depreciation, financing costs and the expanding asset base.
• CXMT capacity additions, DRAM market-share gains and the protection provided by Korean memory long-term agreements.
• China localization indicators, including domestic-fab utilization, equipment orders and AI-hardware demand.