\n\n
Key developments:
- China’s operational AI compute capacity has reached 24GW and another 50GW is under construction, narrowing the gap with the United States’ 56GW total, according to a Financial Times analysis of SemiAnalysis data.
- Jeff Holden’s startup Atomic Machines, backed by $250 million, unveiled the PrimeSwitch PS‑150 relay that switches in about 50 µs—roughly 1,000× faster than traditional contactors—targeting 800 V DC power distribution for AI data‑centers.
- Microsoft began pre‑orders for the Surface RTX Spark Dev Box, priced at $5,999, featuring Nvidia’s Blackwell‑based RTX Spark (6,144 CUDA cores, 20‑core Grace CPU, 128 GB unified memory) and delivering ~1 petaflop of AI performance, sufficient for running 120‑billion‑parameter models locally.
- Nvidia announced its NeMo‑DCR framework, achieving a 35× speedup for refitting trillion‑parameter models, streamlining large‑scale training workflows.
- Mistral released a preview of Large 4, a 1‑trillion‑parameter multimodal model with 49 B active parameters, trained on 4,000 Nvidia Grace Blackwell GPUs, showcasing Europe’s push into frontier AI.
- South Korea disclosed a $3.5 billion investment plan for domestic frontier AI, earmarking roughly 29,000 GPUs to achieve parity with China’s open‑source models by the late 2020s.
Why they matter:
- The rapid expansion of China’s AI compute infrastructure signals a power shift in global AI capacity, potentially reshaping competitive dynamics in AI research and commercial services.
- Atomic Machines’ ultra‑fast relay could dramatically reduce latency and power loss in AI data‑center power grids, enabling higher density deployments and lower operational costs.
- Microsoft’s RTX Spark Dev Box brings petaflop‑class AI compute to individual developers, lowering the barrier to entry for building and testing large models without cloud dependence.
- Nvidia’s NeMo‑DCR speedup cuts the time needed to adapt trillion‑parameter models, accelerating AI innovation cycles and reducing cloud‑compute expenses.
- Mistral’s Large 4 demonstrates that non‑US players can produce trillion‑scale models, diversifying the ecosystem and fostering competition in multimodal AI.
- South Korea’s sizable GPU investment underscores the geopolitical race for AI leadership and suggests future regional hubs for AI‑driven industries.
What to watch next:
- Follow how Chinese utilities and local governments manage the power demands of the expanding 24GW+ AI compute base and whether regulatory bottlenecks emerge.
- Monitor adoption rates of Atomic Machines’ PrimeSwitch across major data‑center operators and any subsequent hardware iterations.
- Track early developer feedback on the RTX Spark Dev Box and whether it spurs a wave of on‑premise AI applications.
- Watch Nvidia’s rollout of NeMo‑DCR tooling and its impact on the cost structure of trillion‑parameter model training.
- Observe Mistral’s full model release schedule and any benchmark results against competing large‑scale models.
- Keep an eye on South Korea’s GPU deployment timeline and the emergence of homegrown AI products that leverage the funded hardware. [[1]]