Temasek comfortable with current AI spending levels, executive says

Temasek remains positive on the AI buildout and plans to raise AI ecosystem exposure to as much as 15% of its portfolio within five years, from about 6% currently. The roughly $400 billion state investor, which holds stakes in OpenAI and Anthropic, argues that hyperscaler AI spending is supported by exceptionally strong balance sheets despite concerns over high interest rates and global bond-market volatility. Analysts expect hyperscaler capital expenditure to exceed $1 trillion by 2027, primarily for chips and data centers.

Analysis

This is not a new fundamental datapoint for NVDA; it is a marginal validation of institutional appetite for AI exposure. The more investable implication is that capital availability for private-model developers and data-center projects remains robust even as the discount rate rises, extending demand visibility for the compute stack. Near term, however, a rates-driven de-rating can overwhelm capex optimism: NVDA’s earnings sensitivity is positive, but its multiple remains materially more duration-sensitive than power, cooling, and networking beneficiaries.

The likely second-order winners are AI infrastructure bottlenecks with contracted revenue and less debate over application monetization: VRT, ETN, GEV and CEG should retain pricing power if data-center construction schedules hold. Private capital flowing into foundation-model companies also increases competitive pressure on listed software vendors whose AI revenue is still largely bundled into existing seats; that favors a relative long infrastructure / short high-multiple software basket rather than a blanket long technology position over the next 1-3 months.

Consensus may be underestimating the financing constraint rather than the chip-demand constraint. Hyperscalers can self-fund, but colocation developers, utilities and smaller cloud providers face higher borrowing costs and grid-interconnection delays; a rise in long-end yields can defer physical deployment even if GPU orders remain intact. The thesis is falsified by hyperscaler capex guidance falling below current run-rates, material increases in GPU lead times easing, or data-center power-project cancellations; conversely, backlog and booking acceleration at VRT/ETN would confirm that spending is translating into deployable capacity over 6-18 months.

Key Decisions for Investors

• Maintain NVDA as a core long only on rate-driven weakness rather than chase incremental headline support; add if the stock underperforms SOX by 10%+ without a reduction in hyperscaler capex guidance. Target a 6-12 month holding period; exit/reduce on evidence that aggregate hyperscaler capex growth rolls over meaningfully in the next two earnings cycles.
• Initiate a 1-3 month pair: long VRT and ETN, short an equal-dollar basket of AI-exposed software duration risk (IGV ETF). The trade monetizes physical-infrastructure scarcity versus uncertain software AI monetization; risk is a sharp decline in Treasury yields that re-expands software multiples faster than infrastructure earnings revisions.
• Watch CEG and GEV for grid-capacity contract announcements and utility rate-base visibility. Do not initiate solely on this item; enter after independently confirmed data-center power awards, with a 6-18 month horizon and a stop if power-project timelines slip or regulatory cost recovery is challenged.
• Avoid treating BLK or MA as direct AI expressions. Their relevant exposure is broader risk-asset and private-market fundraising conditions; rising real yields are a near-term valuation headwind, making them better held as diversified financial exposures than as a response to this catalyst.