“Overseas # Overinvestment Amid Widening Quality Divide Technology Henry Huang Jeff Pu, CFA SFC CE No. BYA749 SFC CE No. BNO719 henryhuang@gfgroup.com.hk jeffpu@gfgroup.com.hk ## What's New? On Oct 22, Nvidia's CEO flagged capital as potential bottleneck for AI boom. We provide ou…”
原始全文
Overseas # Overinvestment Amid Widening Quality Divide Technology Henry Huang Jeff Pu, CFA SFC CE No. BYA749 SFC CE No. BNO719 henryhuang@gfgroup.com.hk jeffpu@gfgroup.com.hk ## What's New? On Oct 22, Nvidia's CEO flagged capital as potential bottleneck for AI boom. We provide our view on hyperscalers' investment incentive, capital availability and financing dynamics. In sum, we believe AI Red Queen Effect will sustain CSPs overinvestment, yet funding will increasingly favor larger players. We expect IG providers including Top-5 CSPs and SPCX compute addition to account for $84\% / 83\%$ of U.S. capacity build 21GW/38GW in 2026E/2027E. ## Comments Compute remains tight amid robust AI demand: Recent commentary from hyperscalers continues to point to tight compute availability and robust AI demand. Microsoft sees Azure demand continues to exceed available capacity; Amazon and Alphabet both highlighted supply constraints; Nvidia views unconstrained demand potentially doubling in 2027 vs. \~70% supply coverage; TSMC also noted demand well above their silicon supply. Our U.S. Tour (report of U.S. Tour Oct 23) echoes the tightness amid thermal and interconnect bottlenecks, with GB300 rack bids rising from \$5.5m to \$6.5m and fully configured VR200 racks reaching \~\$8m, suggesting incremental compute supply continues to be absorbed. AI Red Queen Effect to drive CSPs overinvestment, with robust B/S: We believe the AI race exhibits a Red Queen Effect, where hyperscalers are more willing to risk overinvestment than underinvestment, as slowing capex could cede AI leadership and weaken their competitive positions. Despite our upward-revised capex forecasts, with U.S. Top-5 CSP capex to grow +85%/+53%/+20% in 2026–2028, and widening credit spreads, with 20y+ hyperscaler spreads rising to \~118bps, we expect their balance sheets to remain healthy into 2028, where net leverage (EBITDA basis) expected to be at 0.6x with interest coverage at 14x, leaving ample headroom for further investment. Capital remains sufficient across financing markets: We see sufficient capital to support AI buildout across both public and private markets. Based on our forecast, \~\$8tn USD IG market alone could offer \$550bn of incremental hyperscaler debt capacity into 2028 from issuer-concentration perspective, with additional headroom across other global IG markets. More importantly, \~\$4.5tn private market dry powder provides a sizable alternative funding pool as public markets face rising issuer concentration and market saturation, while project finance and ABS could also further broaden funding channels. AI financing to become more diversified: AI vendors are increasingly tapping different markets and instruments to broaden their funding base. Alphabet raised C\$8.5bn in Canada, €9bn in Europe, and also \$49.6bn through common equity and mandatory convertibles; CoreWeave secured \$8.5bn private-credit facility anchored by Blackstone; Apollo provided \$3.5bn for the Valor/xAI compute infrastructure transaction. Meanwhile, Meta suspended share buybacks this year (vs. \$26.2bn in FY25), and launched a \$35bn equity offering alongside a \$60bn Class C ATM program to preserve capital for AI investment. We expect such diversification across global bonds, private credit, equity and hybrids to accelerate as funding needs rise. Investors turn more selective as quality divide widens, favoring larger AI vendors: Despite capital availability, investors are becoming more selective on borrower quality, collateral and financing structure, with new issue concessions on some deals rising from 2–3bps historically to \~20bps. Recent price declines concentrated in unsecured debt from lower-quality AI issuers, while secured loans backed by chips or lease payments from high-quality vendors retain resilient. The divide is evident between hyperscalers and Neoclouds, with hyperscaler 20y+ spreads at \~118bps vs. CoreWeave yields rising to \~15% by mid-Oct; within Neoclouds, with larger players retaining funding access while weaker ones face tighter capital. SpaceX's CEO Elon Musk also highlighted this divergence, noting SpaceX secured \~\$40bn of GPU-related debt given its IG rating and diversified revenue, while expecting only few Neoclouds to survive. In sum, we don't see capital availability as a major concern for now, but more selective funding should favor larger, better-capitalized players. Risks: 1) AI demand slowdown, 2) infrastructure deployment, 3) supply constraints, 4) competition. ## AI Red Queen Effect Amid Compute Undersupply ## Compute remains tight amid robust AI demand Hyperscalers have continued to highlight tight compute availability and robust AI demand in recent earnings calls and management commentary. Microsoft's CFO stated in recent F4Q26 earnings call that "Azure demand continues to exceed available capacity" and noted capacity was still constrained; while CEO highlighted earlier this year that AI demand growth is not just linear, requiring significantly more upstream capacity before supply constraints ease; Amazon's CEO stated this month that "Even with 2026 capex raised to \$220bn, still cannot meet all 2026-27 demand." with significant customer commitments already extending into 2028; Alphabet mentioned they were "operating in a supply-constrained environment" as strong demand from both external Cloud customers and internal workloads. Meanwhile, Nvidia's CEO indicated on the recent earnings call that unconstrained customer demand could roughly double in 2027, while available supply at the time could cover only \~70%. At GTC Berlin in Oct 26, Huang reiterated that "demand is extraordinary, while capacity is the constraint". From foundry perspective, TSMC also highlighted on 3Q26 earnings that "The demand actually is still far above our capability today to supply, and we are working very hard to narrow the gap." Per our OCP 2026 Themes and U.S. Tour Readouts report on Oct 16, our recent U.S. supply chain checks also point to compute tightness. High-performance compute remains constrained by thermal and interconnect bottlenecks, while fast-growing workloads such as AI coding remain compute-constrained. Meanwhile, AI rack pricing continues to rise, with GB300 bids increasing from \$5.5m to \$6.5m for small volume purchases and fully configured VR200 racks reaching \$8m, further suggesting that incremental compute supply continues to be readily absorbed by strong demand. CoWoS capacity also reflects hyperscalers' demand. Following our earlier raise in May Edition on May 12 in which we revised up TSMC's CoWoS capacity from 175KPM to 195KPM by end-2027, the Street has moved up to our numbers, and we have fine tuned our estimate to 200KPM by end-2027 on our Sept Edition on Sept 14. Figure 1: TSMC, ASE, and Amkor quarterly CoWoS capacity (kwpm) trend  Sources: Company data, GF Securities (Hong Kong) Brokerage ## Al Red Queen Effect to keep overinvestment race going The current AI race exhibits a Red Queen Effect, where hyperscalers must keep investing to maintain competitive positions as rivals simultaneously scale models and infrastructure, while slowing investment unilaterally risks leaving a player structurally disadvantaged. This creates an asymmetric payoff between overinvestment and underinvestment, where overinvestment may weigh on near-term FCF and returns, while underinvestment risks ceding AI leadership and potentially weakening existing core businesses. This dynamic is already reflected in hyperscalers' compute allocation strategies, where Google prioritizes frontier model training and Gemini R&D over incremental Cloud monetization; Microsoft allocates scarce capacity to first-party products such as Copilot rather than maximizing near-term Azure growth; Meta takes "build ahead" approach, viewing "compute per researcher" as a strategic advantage and retaining the flexibility to redeploy excess capacity to ads and recommendations should superintelligence take longer to materialize. Meanwhile, OpenAI continues to pursue aggressive compute scaling, with management believing "twice the compute, twice the revenue." That said, hyperscalers appear more willing to risk overinvestment than underinvestment, with Red Queen Effect reinforcing the incentive for all players to stay in the game. Meanwhile, this also creates a similar prisoner's dilemma for shareholders and boards that pushing for lower investment to protect near-term FCF but could come at the cost of allowing competitors to pull ahead. Meta has already showcased the board chooses to stay in the game as well, with Zuckerberg arguing that “underinvesting is riskier than overinvesting” and the board approving a new non-voting Class C share structure to support AI investment and preserve founder leadership. That said, unless competitors collectively slow investment, there is limited incentive for any single player to move first. While shareholder pressure may moderate the pace of capex growth, we therefore see it is unlikely to end the AI infrastructure race as long as demand remains strong and external funding remains available. ## U.S. CSP capex to be +85%/+53%/+20% in 2026/2027/2028 Following our Addressing 2028 Capex and AI Platform Value Shift report on Aug 6, we have revised upward our CSP capex forecast in our Oct Edition on Oct 15, raising our 2027E/2028E U.S. Top-5 CSP capex estimates from \~\$1.2tn/1.4tn to \~\$1.3tn/1.5tn. This implies 2027E capex growth of +53% YoY, up from our previous +45% forecast, while 2028E growth forecast remains unchanged at +20% YoY. Figure 2: U.S. Top-5 CSP capex (USD bn) 2023 2024 2025 2026E 2027E 2028E MSFT 41 76 118 190 268 322 AMZN 53 83 132 220 314 377 Google 32 53 91 205 342 410 Meta 28 39 69 141 225 270 Oracle 7 11 35 67 111 133 Top-5 161 261 446 823 1,260 1,512 YoY 62% 71% 85% 53% 20% Sources: Company data, Bloomberg, GF Securities (Hong Kong) Brokerage ## Hyperscalers' B/S to Remain Healthy into 2028 Despite Rising Rates ## Credit spreads widen amid rising debt supply Credit spreads across AI-related issuers have widened recently amid growing investor concerns over AI investment and financing. According to Reuters' report on July 29, median spreads for Amazon, Alphabet, Meta and Oracle had widened to 40bps/60bps/118bps across 2–4y/5–7y/20y+ maturities, versus 30bps/50bps/109bps in 2025. More recently, Reuters reported on Oct 15 that CDS spreads across technology companies reached record highs, with CoreWeave's 5y CDS exceeding 1,100bps and Oracle's approaching 300bps, while protection costs for Amazon, Alphabet and Meta also increased. Bloomberg also reported on Oct 23 that global corporate bond spreads widened to 129bps, the highest since May 2025, while Meta's 5y CDS reached a record 115–119bps following its latest capex update, while Amazon's remained near the highest level since March 2020. We believe the widening spreads reflect the surging bond supply. In 2025, U.S. Top-5 CSPs' total debt issuance reached \$108bn, while 2026 YTD has already exceeded \$235bn. Based on our forecast, given capex expands further by +85%/+53%/+20% YoY over 2026-2028 and FCF turns negative for CSPs, we estimate U.S. Top-5 CSPs' debt issuance to reach \$257bn/446bn/450bn in 2026-2028, assuming the growing new profit on cloud revenue. Figure 3: Hyperscalers debt issuance (USD bn)  Sources: Company data, GF Securities (Hong Kong) Brokerage ## Hyperscalers' balance sheet to remain healthy into 2028 despite widening spreads As said, we revised upward the CSP capex forecast in 2027/2028, and therefore accordingly fine tuned the debt issuance estimate, following our Addressing 2028 Capex and AI Platform Value Shift report on Aug 6. Based on our updated forecast into 2028, despite higher debt issuance and rising spreads, we believe hyperscalers' balance sheets should remain robust. Top-5 CSPs' gross leverage (EBITDA basis) is expected to be 1.1, net leverage (EBITDA basis) to be 0.6, and the net debt to-equity ratio of only 23%, while CFO-to-debt to be 1.0 and interest expense coverage still up to 14. In theory, hyperscalers' intact balance sheet metrics and robust 14x interest coverage should afford them ample runway to issue additional debt for financing capex. Figure 4: U.S. Top-5 CSP leverage and solvency metrics 2025 2026E 2027E 2028E Key assumptions Revenue YoY 15% 18% 17% 16% EBITDA YoY 22% 21% 24% 22% Gross Debt (bn) 297 537 910 1,273 Net Debt (bn) 100 220 512 785 Key ratios Gross leverage (EBIT basis) 0.6 1.0 1.3 1.5 Gross leverage (EBITDA basis) 0.5 0.7 1.0 1.1 Net leverage (EBIT basis) 0.2 0.4 0.7 0.9 Net leverage (EBITDA basis) 0.2 0.3 0.5 0.6 Net debt / equity 7% 11% 19% 23% CFO / debt 2.0 1.4 1.1 1.0 Interest expense coverage 43 25 18 14 Sources: Company data, GF Securities (Hong Kong) Brokerage Based on our forecast, we see hyperscalers still have substantial headroom to raise additional debt while maintaining investment-grade credit profiles, supported by their strong balance sheets. Assuming net leverage increases to 2.0x in 2028, we estimate over \$1.7tn of incremental debt capacity across the Top-5 U.S. CSPs, with these leverage levels still broadly consistent with solid IG credit profiles. ## Capital Remains Ample with Shifting Financing Dynamics ## Capital remains ample across financing markets The public credit market still appears to offer some headroom to absorb incremental hyperscaler debt supply. Within \~\$8.0tn USD IG Corporate Index, the three largest individual issuers currently have no more than \$175bn of index-eligible debt outstanding each and none represents more than 2.3% of the index. According to Bloomberg, U.S. Top-5 CSPs' combined weight in the USD IG Corporate Index increased from \~2.25% in Aug 2025 to \~5% in Jul 2026. Simply assuming the size of the USD IG market grows more slowly than before and each CSP's index weight could rise further toward the current ceiling level, it would imply more than \$550bn of incremental debt capacity in the USD IG market alone from an issuer-concentration perspective. Based on our debt issuance forecasts and assuming CSPs' debt issuance maintains 60-70% absorbed by the USD market, this headroom could support hyperscalers' debt issuance through at least 1H28. Beyond USD IG, additional public markets are also available across the EUR IG market with current size of \$3tn, and the combined \~\$1.3tn GBP/JPY/CAD/AUD/CHF IG markets, providing further opportunities to diversify funding across currencies and investor bases. More importantly, private markets provide a much more substantial pool of capital as traditional public markets become increasingly constrained by issuer concentration and market saturation. According to Preqin, \~\$4.5tn of dry powder is available across private equity, infrastructure, real estate and credit strategies, providing a significant alternative source of funding for the multi-year AI infrastructure buildout. Meanwhile, project finance is expected to play a growing role in data center buildouts, while completed assets might be able to tap the ABS market for refinancing. Net, we see ample capital available across public and private financing markets, with funding likely to increasingly migrate across different pools of capital. Figure 5: Global private capital dry powder (USD tn)  Sources: Preqin, GF Securities (Hong Kong) Brokerage ## AI financing to become more diversified While USD IG remains the primary funding market, AI vendors are increasingly tapping other regional bond markets to broaden their investor bases. Alphabet raised C\$8.5bn in Canada earlier this year, before Amazon followed with a record C\$14bn offering, while hyperscalers have also become major issuers in the Swiss market and Alphabet tapped sterling with a 100-year bond. In Europe, Alphabet also launched its largest euro bond offering of at least €9bn in May, demonstrating that sizable AI-related financing is increasingly being distributed across global credit markets. Meanwhile, private markets are also taking a larger role in the AI buildout. Private credit has already funded several large-scale AI infrastructure transactions in 2026. CoreWeave raised \$8.5bn facility in March anchored by Blackstone Credit & Insurance, while Apollo-managed funds provided a \$3.5bn capital solution for a \$5.4bn Valor/xAI compute infrastructure transaction in January. We expect this shift to accelerate from 2028 as issuer concentration and market saturation become increasingly binding in traditional IG markets. Funding structures are also becoming more diversified. As FCF turns negative, hyperscalers are increasingly combining debt with equity, hybrid securities and changes in capital allocation. Alphabet raised \$49.6bn through common equity and mandatory convertible preferreds in June; Meta suspended share buybacks this year (vs. \$26.2bn in FY25), and launched a \$35bn equity offering alongside a \$60bn Class C ATM program to preserve capital for AI investment. In sum, we expect AI vendors across hyperscalers, semiconductor companies, data center developers and Neoclouds to increasingly diversify their funding sources, tapping a broad range of public and private credit markets alongside equity and hybrid instruments. ## Investors turn more selective as the quality divide widens Despite ample capital availability, investors are becoming increasingly selective — not every AI borrower, duration or price is equally attractive. Investors' appetite for long-duration AI debt has softened, with the number of insurance investors placing over \$50m orders in 30-year tranches of large AI financings falling from \~15 in 1Q26 to roughly half that level by end-2Q26. Meanwhile, new issue concessions on some deals have also increased from a typical 2–3bps historically to up to \~20bps in a recent large hyperscaler transaction. That said, credit markets are increasingly differentiating by borrower quality, collateral and financing structure. Recent price declines have been concentrated in unsecured debt from lower-quality AI issuers, while secured loans backed by chips or lease payments from high-quality hyperscalers have been more resilient. The divergence is particularly evident between hyperscalers and Neoclouds. Major hyperscalers' median spreads on over 20-year bonds have widened to \~118bps in 2026, while CoreWeave's bonds issued in Jun 2026 at a 9.675% yield have since fallen sharply, pushing yields to \~15% in mid-Oct, according to Reuters, and its 2032 unsecured bonds traded below 80 cents on the dollar. The divide is also widening within Neoclouds themselves — larger platforms with contracted revenues, valuable GPU collateral and stronger counterparties can still access secured financing, while smaller and less-capitalized players seem to face increasingly constrained funding. On Oct 26, SpaceX's CEO Elon Musk also highlighted the widening quality divide, saying that debt markets had “almost entirely shut off” access to capital for companies without substantial revenue outside AI data centers, while SpaceX was able to secure US\$40bn of GPU-related debt given its investment-grade rating and diversified revenue from Starlink and launch services. Musk expects this funding divergence to drive significant consolidation among the 200+ Neoclouds, with only a limited number ultimately surviving. In sum, we do not view the availability of capital as a major concern for now, with the key question shifting toward who accesses capital and on what terms. As the AI investment cycle matures, credit quality, collateral and financing structure should become increasingly important, favoring larger and better-capitalized AI infrastructure players, while raising the cost of capital and potentially cons…