/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Sources: Moonshot is seeking access to more Nvidia Blackwell chips to prepare for Kimi K4's development, after training K3 on Nvidia chips, including Blackwell

The Information

Context & Ripple Effects

Moonshot's compute request follows reports that Kimi K3 was being prepared as a 2T–3T-parameter model and had already been trained on Nvidia hardware, including Blackwell. The move makes accelerator availability a concrete input to the lab's next model cycle.

Nvidia has positioned its GB200 Blackwell systems as particularly suited to mixture-of-experts workloads such as Kimi, in its performance claims for MoE deployments. That makes continued Blackwell access strategically relevant beyond a one-off hardware purchase.

First-order effects

  • Moonshot's Kimi K4 development schedule and training options become more dependent on securing additional Blackwell capacity.
  • Nvidia gains another reported demand source for Blackwell hardware from a lab already using its platform.

Second-order effects

  • Other model developers seeking comparable training capability may face sharper competition for available high-end accelerator capacity.
  • Moonshot has an incentive to keep its software and training stack aligned with Nvidia's Blackwell platform, raising the switching cost of moving a future model cycle elsewhere.

Third-order effects

  • If leading labs repeatedly need the newest accelerator generation for successive frontier-model releases, allocation of compute becomes a durable competitive gate alongside model research.
  • The pattern reinforces Nvidia's annual platform-upgrade cadence: labs may need to refresh infrastructure to remain competitive, though the extent depends on whether alternative compute platforms become viable for training.

The trend: Frontier AI development is becoming increasingly governed by access to current-generation accelerator capacity, not just model architecture and talent.

Discussion

  • @kyleichan Kyle Chan on x
    Kimi K3 partly trained on Blackwells along with Qwen3.8 and DeepSeek, according to The Information. https://www.theinformation.com/ ...
  • @sayreevan Evan Swarztrauber on x
    According to this report, Kimi K3 trained on export controlled NVIDIA Blackwell chips. The open vs. closed debate should not distract us from chip smuggling and lax enforcement. One can be pro-open-source AI while wanting to prevent advanced chips from going to China.
  • @leomschwartz Leo Schwartz on x
    Important reporting from @theinformation today on how Moonshot trained Kimi K3 using Nvidia chips and its plans for K4 https://www.theinformation.com/ ... [image]