Semiconductors & AI Hardware
TensorRT’s New Tuner Tests Engine Routes. Do the Speedups Transfer?
TensorRT 11.1 searches alternative engine build routes. NVIDIA’s reported gains warrant evaluation, while accuracy coverage and deployment conditions limit what teams can infer.
NVIDIA Reports Faster Mamba-2 Forward-Pass Kernels on B300
NVIDIA’s Mamba-2 kernel report raises a practical adoption question: what evidence would justify integrating the implementation into a real workload?
NVIDIA’s $4,999 DGX Spark 64GB: What Can It Run Locally?
NVIDIA’s DGX Spark 64GB starts at $4,999. Examine its model-capacity claim, shared memory, two-system option and questions to ask before ordering.
Can Volato Deliver Its AMD GPU Orders in Ohio? The Evidence to Watch
Volato has signed two AMD GPU cluster orders, but Ohio power, current financing and accepted service remain separate proof points. Here is what public filings establish.
Volato’s $1.17 Billion AMD GPU Orders: What the Customer Agreed to Pay, and When
Volato disclosed two four-year AMD GPU orders worth about $1.17 billion, with roughly $133 million in contractual prepayments. Here is what the filings say about payment timing and contract conditions.
Akamai’s Anthropic Deal Targets CPUs. The Hardware Mix Is Still Unknown.
Akamai has disclosed a large CPU cloud commitment and a staged hardware buildout. Its processor choice, server design, and Anthropic-specific power needs remain undisclosed.
Alibaba Says Its V900 Chip Is 3× Faster. Commercial Proof Is Slated for 2027.
Alibaba has unveiled the Zhenwu V900 AI processor and targets commercial release in Q1 2027. Its 3× performance claim and 500,000-card scale have not been independently demonstrated.
CXMT’s G5 DRAM Is a Mobile-Memory Milestone, Not an HBM Breakthrough
CXMT’s fifth-generation DRAM platform could expand China’s LPDDR5X supply and add pressure to mobile-memory pricing. But its headline efficiency claim says nothing yet about good-die yield, global availability, or the stacked HBM used by AI accelerators.
Qualcomm’s Amazon Deal Is a $60 Billion Ceiling, Not an AI Chip Victory Yet
Qualcomm’s AWS collaboration pairs custom inference silicon with 1.6T optics, but its $60 billion warrant threshold is conditional—not guaranteed chip revenue.