Every eighteen months or so, the market rediscovers the same fear: that the enormous capital being poured into frontier AI labs is a bubble waiting for a Chinese lab to pop it. In January 2025, it was DeepSeek R1. This week, it was Moonshot AI’s Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window that, on Moonshot’s own published benchmarks, beats Claude Opus 4.8 on a majority of evaluations, while still trailing Claude Fable 5 and GPT-5.6 Sol. The market’s reaction was immediate and familiar, with a broad sell-off in AI-related equities and a wave of commentary invoking “DeepSeek 2.0”.

It is worth remembering what actually happened after DeepSeek 1.0. The apocalyptic reading in January 2025 was that a cheap, open, locally runnable model had rendered the entire Western AI capital stack, the GPUs, the data centres, the power contracts and the labs themselves, unnecessary. Seventeen or eighteen months later, capital expenditure on AI infrastructure is significantly higher, not lower. The frontier labs are generating more revenue, not less: by our estimate, the industry as a whole is now doing something on the order of USD 175 billion in annualised revenue, and OpenAI and Anthropic alone have gone from a combined USD 8 billion to something like USD 110 billion over that stretch. DeepSeek did not doom anyone’s expectations. It was a genuine achievement that, in retrospect, changed the efficiency frontier of training without changing who sits at the top of it.

Are we finally gaining more visibility on the long-awaited SFDR 2.0?

Emerging market (EM) debt remains one of the most diverse and opportunity-rich segments within fixed income. The asset class encompasses both hard currency debt and local currency debt. Both segments offer exposure to emerging economies, but their drivers of return, valuation dynamics and attractiveness differ meaningfully.