AIcapex

US Rates Strategy Joseph Abate Jim Martin Aug 5, 2026 joseph.abate@smbcnikko-si.com jim.martin@smbcnikko-si.com 1 212 893-1592 1 212 893-1580 AI capital expenditures For all the headlines about the AI buildout and the race between hyperscalers, the GDP data show relatively modest growth. We explore the accounting differences and look at the volume of financing they are raising in debt, loan, and equity markets. • Hyperscalers’ spending is large and growing. These companies are expected to double their spending on capital and software this year – to more than $770bn . Their recent earnings reports suggest they could spend as much as $1.3trn next year. • Despite double-digit growth, spending on IT equipment and software has accounted for 15% of real GDP growth since December 2024. However, last quarter, this share was over 50% . • About 75% of hyperscalers’ spending on computers and peripherals is imported – up from less than 35% in 2022 . Monthly imports of computer equipment from Taiwan have increased 4- fold since January 2025. • GDP data reliably track spending on AI infrastructure. But the use of AI as an intermediate input into other products is harder to measure. Business and consumer purchases of AI services appear in the corporate profit data in the national accounts . • The modest GDP effect, however, belies its impact on capital markets. Hyperscalers account for roughly 40% of the net supply in dollar-denominated corporate debt issuance this year – even though they make up just 5% of the IG debt market. Tech spending Hyperscalers recently updated their capex spending projections for 2026. Their spending is expected to nearly double from last year to reach $770bn this year . Next year, the hyperscalers are looking to spend over $1.3trn. And by 2030, these companies may have spent over $7trn on building out AI infrastructure. 1 Such rapid spending may already be straining the economy. Demand for memory chips has been so robust that companies are signing contracts for future production a year or more in advance and still, prices have risen more than 5-fold since last fall. Substitution with less state-of-the-art chips has pushed their prices up just as fast. S pending on data center construction has accelerated. The monthly “value put in place” for new data center projects has increased since the introduction of ChatGPT in November 2022 1 See, “The cost of compute: A $7trn race to scale data centers”, J. Noffsinger, M. Goodpaster, M. Patel, H. Chang, P. Sachdeva, and A. Bhan, McKinsey & Company, April 28, 2025. SMBC US Rates Strategy 2 (Figure 1). However, the most expensive component of data center construction – equipping them with computer servers and integrating networks – is not captured in this spending. Noffsinger et al, estimate that equipment spending related to buildout of datacenters is about 3x larger. 2 It is not difficult to see the effects of the technology spending in the GDP accounts. Real spending on IT equipment and software has grown 23 and 11%, respectively in the last year compared to a 2.1% gain in overall GDP . While this spending includes purchases made by other, non-hyperscalers, the rapid growth suggests these categories are fueled by the hyperscalers ’ infrastructure buildout. As spending on these items is growing much faster than GDP, their share of real output has risen from 5.5% in Q1 25 to 7.0% in Q2 26 (Figure 2). By comparison, spending in the initial phase of the dot-com boom was stronger . We compared the growth in IT equipment and software since the introduction of ChatGPT with a similar 14 quarter period beginning in March 1995. In that period spending on equipment and software rose nearly 80% and as a share of GDP, it doubled from 0.3% to 0.6%. Since November 2022, IT equipment spending has risen 52% and software expenditures are up 36%. Despite their overall small share of GDP, spending in IT equipment and software has added about 0.8% per quarter to real GDP growth since early 2025. Indeed, spending on software, computers, peripherals, chips, and servers accounts for nearly a quarter of GDP growth over the period . Related spending on research and development has also been robust in the last year, adding several tenths to quarterly GDP growth. But while the GDP figures are impressive, they are smaller than the spending figures reported by the hyperscalers. For example, last year, they reported spending nearly $400bn in capex but the nominal GDP figures for IT equipment and software were up just $240bn – and this amount includes non- hyperscaler spending. So, what accounts for the difference? 2 McKinsey (2025) Figure 1: Data center construction ($bn) Figure 2 : IT equipment and software (% GDP) Source: US Census Dept., SMBC Nikko Source: BEA, SMBC Nikko 0 1 2 3 4 5 6 Jan-22 Jan-23 Jan-24 Jan-25 Jan-26 2.0 2.5 3.0 3.5 4.0 4.5 Mar-22 Mar-23 Mar-24 Mar-25 Mar-26 Equip Software SMBC US Rates Strategy 3 Imports Much of hyperscalers’ spending i s on imported goods . 3 Since GDP measures spending on domestic production, the net amount of the AI buildout is smaller. Roughly 75% of the spending on computers and peripherals in the GDP statistics is imported – up sharply since the introduction of ChatGPT (Figure 3). Likewise monthly computer and semiconductor imports from Taiwan have risen sharply since January 2025 (Figure 4). After accounting for imports the net contribution to GDP growth since 2025 from spending on IT equipment and software is 15% . GDP measurement Tracing hyperscaler spending and earnings into the GDP statistics is complicated because the national income and product accounts are conceptually different than standard GAAP accounting . Not all the money spent by the hyperscalers shows up in the GDP data because some of this spending is for goods that are later used in the production of other (final) goods. Our sense is that much of the AI buildout is on final products that would show up in both their corporate reports and the GDP data . Data center construction appears in business fixed investment for non-office structures while spending on finished chips, servers and software is captured in the nonresidential investment figures. Hyperscalers include their finance leases in their projected capex . For national income accounting, the legal ownership of the equipment matters less than its “economic” ownership. Thus , if the lease effectively transfers the benefits of using the equipment to the lessee, then from a GDP perspective, the lessee has “purchased” the equipment. The “purchase” appears in the GDP accounts immediately and in full. Subsequent principal payments are not included in GDP although any interest paid to the supplier shows up in the gross domestic income data. 3 See, “Artificial Intelligence and the US Economy: An Accounting Perspective on Investment and Production”, L. Carpinelli, F. Natoli, and M. Taboga, Bank of Italy, January 19, 2026. Figure 3: Import share computers and periph. (%) Figure 4 : Imports from Taiwan ($bn) Source: BEA, SMBC Nikko Source: BEA, SMBC Nikko 30 35 40 45 50 55 60 65 70 75 80 Mar-22 Mar-23 Mar-24 Mar-25 Mar-26 0 5 10 15 20 25 Jan-22 Jan-23 Jan-24 Jan-25 Jan-26 Comp Semis SMBC US Rates Strategy 4 Hyperscaler money spent to design the chips used in the data centers may not appear in the final product figures. Instead, the BEA measures these expenditures by looking at company revenues and matching them up to purchases of intermediate products supplied by other companies. Because this is a revenue- based analysis, some of the outlays by the hyperscalers appear as income to other companies. Importantly, business and consumer spending on AI also appears in the income data rather than the output statistics . These revenue streams are blended with income from other services like web hosting and cloud computing. Moreover, although businesses are “renting” AI services, from the perspective of the national income accounts they are “purchasing” these services so the income derived by providing AI shows up as corporate profits rather than rents. Financing AI The modest GDP effect, however, belies its impact on capital markets. So far this year, hyperscalers have raised more than $200bn in bonds and loans and another $115bn in equity. 4 To put hyperscaler debt issuance in context we look at the overall new issuance of investment grade (IG) corporate debt. Through July, IG companies have raised $1.4trn and are on track to surpass the record $1.8trn raised in the 12 months through December 2020. While roughly 2/3 of this is refinancings, the remaining $350-400bn is “net supply” or additional corporate borrowing that increases leverage. Hyperscalers account for roughly 40% of the net supply in dollar-denominated corporate debt issuance this year – even though they make up just 5% of the IG debt market . Borrowing costs for hyperscalers have risen roughly 30bp compared to US Treasuries so far this year . By contrast, peer credit spreads have been unchanged – a meaningful move for issuers that skew higher quality and who are issuing 10-40y maturities. Concessions that IG bond issuers have to offer investors have risen. Even though credit valuations were similar in July and January the additional offering spread has risen to 5bp from 1bp. This may seem small, but it reflects a dramatic shift in market power from sellers to buyers of debt even as overall IG credit risk remains at historically low levels. While the credit risk of hyperscalers has not been questioned, their capex ambitions are raising investor concerns about the market ’s capacity to dige st this debt amid ever-rising capex spending . 4 See “The Unsettled Vibe That’s Creeping Through Markets”, S. Goldfarb and H. Lang, Wall Street Journal, July 26, 2026.