Merryn Talks Money: Higher Yields, Higher Gold (Podcast)

The weekly markets wrap highlights a surge in global bond yields alongside a move higher in gold, signaling shifting rate expectations. It also flags a study suggesting small language models (SLMs) could reduce the need for parts of the current AI data-centre buildout, creating risk for hyperscalers and AI-related investors. Overall, the update is more scenario-setting than a quantified catalyst.

Analysis

The market mechanism is less about the level of yields than the fact that the marginal buyer is demanding a higher term premium. That is structurally bearish for any asset whose valuation depends on cash flows pushed far into the future: long-duration tech, unprofitable software, REITs, utilities, and the capital-intensive AI buildout. If funding costs stay elevated for more than a few weeks, hyperscalers will get more disciplined on marginal data-center projects, and the knock-on winners are not the obvious GPU names but operators with scarce existing capacity, cheaper power, and better enterprise software distribution.

The small-language-model angle is a genuine second-order threat, but mostly to the narrative of infinite compute demand rather than to compute demand itself. SLM adoption would compress the growth rate of frontier-training capex first, then pressure cloud inference pricing as enterprises shift from large general-purpose models to narrower workflow tools. That creates a slower, more selective spending cycle: semis and colo names tied to incremental expansion are vulnerable over 1-3 months, while software vendors with embedded distribution can actually gain share if they can package smaller models into existing workflows without exposing customers to GPU economics.

The contrarian point is that higher gold alongside higher yields usually signals a policy-credibility problem, not a clean growth scare. If governments respond with fiscal restraint or coordination, yields can retrace fast; if they respond with financial repression, gold wins and duration loses. For the AI trade, the key falsifier is not a think-piece about SLMs but actual capex guidance: if the hyperscalers reaffirm or raise spend and power procurement, the market is likely over-discounting the near-term risk.

Key Decisions for Investors

• Tactically buy GLD / sell TLT on a 1-3 month horizon: this expresses the view that policy credibility risk is rising faster than growth risk. Risk/reward improves if 10Y yields hold above recent breakout levels; exit if real yields fall back materially or a policy backstop is announced.
• Fade the most duration-sensitive AI beneficiaries on strength: short SMH or NVDA into rallies, targeting a 1-3 month multiple reset if capex discipline emerges. Falsify if hyperscaler capex guidance re-accelerates or GPU lead times tighten again.
• Pair long software/platform names with recurring enterprise distribution vs short AI infra proxies: e.g., long MSFT/CRM versus short EQIX/DLR. The thesis is that value migrates from physical capacity to workflow capture if SLMs lower model-cost barriers. Stop if cloud consumption growth remains broad-based.
• Watch-list, not a trade yet: data-center power and utility capex names. Higher yields can delay project starts, but if governments respond with subsidized infrastructure, these could snap back quickly; wait for confirmation in capex guidance before acting.