DePIN Compute Watch

NVIDIA 擴展 Jetson Thor 產品線，新增 T3000 與 T2000 模組，針對邊緣 AI 與機器人提供更低成本、功耗選擇。

• T3000 定位：865 TFLOPS FP4 性能、32GB 記憶體，強調記憶體頻寬優勢與多模態推論效率。
• T2000 定位：400...

Japan's Noetra Corp ordered 27,500 Nvidia Rubin GPUs for a sovereign AI data center, backed by ¥387.3 billion initial subsidy and up to ¥1 trillion...

Grayscale's Bitcoin Miners ETF is rebranding as the Grayscale AI Compute ETF in September 2026, expanding beyond miners to specialized AI hardware and...

• Kubernetes為無狀態HTTP請求設計，無法處理AI代理長時間狀態化、依賴先前步驟的流程
• 新需求包括毫秒級沙箱啟動、耐久狀態管理、多代理協調及安全憑證傳遞
• Kubernetes社群已推出Agent Sandbox CRD，顯示傳統抽象不再適用
• 此基礎設施落差可能推升對DePIN網路中靈活、低延遲算力資源的需求

VivoPower 指出，具電力保障且再生能源支援的 AI 算力站點日益稀缺，已成為基礎設施成長的主要限制。 這對依賴分散式能源的 DePIN 算力網路而言，凸顯電力瓶頸將是未來部署的關鍵挑戰。

實務GPU指數顯示B200/B300等新卡一年內價格翻倍，H100中位數約$3.15/hr；學術論文指出算力無法儲存，期貨價格將反映預期現貨減風險溢酬。

DePIN提供者可參考指數設定現貨報價，租用者則用學術模型評估RWA長期價值與避險壓力。

兩者互補：市場數據提供即時行情，理論框架解釋為何合約價與現貨價走勢分歧。

Mesh LLM 透過 iroh 協議，讓 MacBook Pro 用戶只需一鍵「mesh-llm --auto」即可貢獻 VRAM 參與分散式模型推理，證明消費者硬體能直接加入 AI 工作負載，擴展 DePIN 算力敘事超越專業 GPU 租賃。

A solo animator created an 11-minute 4K film with 150+ shots and over 27,000 frames on Render Network, completing what would have taken months on a...

No significant updates today.

Anthropic's CFO reveals the company deliberately gives research the largest share of compute, ahead of customer inference workloads. The strategy...

• Sunrun pilots Nvidia GPU nodes in solar-equipped homes, creating distributed edge data centers focused on AI inference to bypass grid bottlenecks
  -...

• DeepInfra launched its first non-US site in Toronto with 1,000+ B300 GPUs and 1.7 MW capacity, fully owned and operated for low-latency inference.
  -...

AMD is pursuing Nvidia-style integrated AI solutions through its 5C partnership for gigascale campuses and rack-scale designs, aiming to offer full...

Spheron's marketplace model delivers lower bare-metal GPU costs than Together AI's centralized options for most workloads.

• H100 cluster: Spheron at...

CNCF's framework stresses workload-specific placement over rigid cloud or on-prem choices, factoring sovereignty regulations like GDPR alongside...

India's tier-2 cities are rapidly emerging as strategic locations for data center growth amid metro limitations.

• AI workloads drive demand for...

96GB VRAM enables FP8 inference for 70B models with ~26GB KV cache headroom, roughly doubling single-GPU concurrency versus 80GB cards and avoiding...

Prime Intellect's $130M Series A at $1B valuation shows crypto-native capital actively backing decentralized compute infrastructure. The startup's...

Grok 4.5's token efficiency claims could reduce compute per task, even as it runs on tens of thousands of Nvidia GB300 GPUs and targets developer...

Akash's reverse auction lets tenants post GPU requests while providers compete on price, settling leases in AKT and delivering H100 rates around...