DeepSeek's DSpark software undermines Nvidia's new hardware decode bet

What happened

DeepSeek released DSpark on June 27, 2026, a speculative decoding module that improves per-user generation speed by 60% to 85% on V4-Flash and 57% to 78% on V4-Pro. The module is open-sourced under the MIT license and already extends to Qwen and Gemma model families. Meanwhile, Nvidia is shipping its Groq 3 LPX specialized decode rack in the second half of 2026, designed to pair with Vera Rubin GPU systems and deliver up to 35 times higher inference throughput per megawatt for trillion-parameter models.

Why it matters

Nvidia's strategy assumes customers will make a separate purchase decision for LPX decode hardware on top of their existing GPU investment. DSpark and DeepSeek's MLA architecture—which reduces memory requirements to roughly 10% of what V3.2 needed for million-token conversations—are free, open-source, and already in production. If software solutions running on standard Rubin GPUs alone deliver sufficient efficiency, the business case for LPX becomes harder to justify. AWS and Cerebras have also announced a competing disaggregated prefill-decode architecture launching in the same timeframe, widening the competitive field.

What to watch

Nvidia's full-year fiscal 2027 Data Center revenue consensus is near $343 billion（約55兆円）. The critical question is whether LPX attach rates meet expectations as customers choose between a second rack purchase, general GPU-based software optimization, or hyperscaler-native decode alternatives. Geopolitical protections secure some hardware supply but not the open-source ideas behind DSpark and similar software approaches.