JPM International Market Intelligence | Morning Briefing Sept 16

20 SECONDS: Short positioning into FOMC. After biblical de-rating 3x questions that will determine the next leg for AI infra. Security Software and Behind the Meter. Trading XA Vol Primer. K-shaped China Macro.

• Risk tone is modestly constructive this morning. AI infra is outperforming again, but rates and macro event risk are capping conviction.

• Zuckerberg called for independent evaluators to assess AI models, while US officials met Anthropic to discuss AI safety risks.

• The FOMC is today’s main event. A hike is seen as a done deal; uncertainty centres on Warsh and the press conference. Investor positioning has turned sharply short, and Jay sees the front end firmer in most scenarios. Equities could see some relief.

After biblical de-rating, JPM’s Gokul frames three questions that will determine the next leg for AI infra ecosystem:

• Selling compute is a great business: semi/hardware margins are at record highs and revenue per GPU hour is well above last year. Can buying compute also be sustainably attractive? Yes. Frontier-model vendors generate 60–80% inference gross margins and could earn $20–40bn of annual revenue per GW from AI tokens (vs $10bn in 2025). Open source should continue to support the rapid rise of vertical AI companies reaching $100m+ ARR. Open and closed models can coexist, like Android and Apple.

• AI model training is unlikely to slow, even if new model releases face alignment or safety restrictions. Open-source models continue to advance, while rapid progress in AI and RSI should create revenue pools beyond coding.

• There is ample room for strong capex growth over the next two years, supported by accelerating operating cash flow, low hyperscaler leverage (13% net debt to equity) and equity issuance. We see no imminent sign of the capex downturn being priced into Tech hardware.

• We still expect Tech leadership to be broader and more dispersed than in 1H26. Within AI infra one should look at optics, substrates, packaging and SPE— more broadly, inference over training. Security software should remain a winner as agentic AI expands the attack surface. AI-driven security spending is a necessity and could exceed $430bn across a three-year TAM. Behind-the-meter power should become more valuable given timing, deliverability, permitting and affordability constraints. A projected 50–80 GW grid shortfall by 2030 creates a structural opening for onsite power.