Amazon's reported Decart talks are about Trainium software, not another Whole Foods
WSJ and Reuters put Amazon in talks near $7 billion. Decart already demos on Trainium3. No deal is signed, and CTech's later figure is lower.
Read the pieceCompute & power
Chips, clusters, energy, inference economics — who can run what, at what price.
19 pieces
New in Compute & power
5 recent pieces
More in this topic
14 stories
CUDA 13.4 upgrade watchlist: CDMM default, driver and toolkit split, MPS V3
Checklist for fleet owners moving to CUDA Toolkit 13.4 after NVIDIA's 9 September 2026 release. Derived from the primary; not in-house lab results.
MLPerf Inference v6.1 preview table for Vera Rubin NVL72 vs GB300 NVL72
This article covers NVIDIA's Vera Rubin NVL72 preview claims in the 16 September 2026 MLPerf Inference v6.1 round and the supporting MLCommons artifacts.
CUDA 13.4: Windows on Arm, Rubin sm_107 preview, MPS V3, and a CDMM default flip
NVIDIA's 9 September 2026 toolkit release is an operator news item: new platform support, shared-GPU controls, and a coherent-platform memory default change.
Jalapeño first silicon: OpenAI's inference ASIC claims 1.5-1.9x work/watt and a nine-month tapeout
OpenAI's first-results post (25 August 2026) and Broadcom unveil put measured first-party silicon on the record. This Stack piece inventories the architecture claims, maturity, and pick-vs against merchant GPUs.
Vera Rubin NVL72's first MLPerf Inference v6.1 numbers: up to 3.7x vs GB300 on Qwen3-VL
NVIDIA's 16 September 2026 preview submission puts Vera Rubin NVL72 on DeepSeek-R1 and Qwen3-VL. This article covers the benchmark results and software stack, separate from earlier Vera Rubin production news.
CUDA 13.4 operator stack: Windows on Arm, Rubin preview, MPS V3, and a CDMM default flip
NVIDIA's 9 September 2026 toolkit release extends CUDA to Windows on Arm, previews Rubin (sm_107), modernizes shared-GPU control, and changes memory-management defaults on coherent platforms.
AEMA asks grids to treat flexible AI data centers as controllable load
Emerald AI, Google, and NVIDIA launch the AI Energy Management Alliance on 16 September 2026. The pitch is faster interconnection for facilities that can cut or shift demand when the grid is stressed.
CUDA Rust: two tracks, neither production-ready
NVIDIA opened native GPU kernels in Rust on 8 September 2026. cuda-oxide is SIMT on pinned nightly. cutile-rs is Tile on stable Rust. Pick by control need, not by logo.
Amazon's warrant is the custom-silicon hedge on paper
Qualcomm's 8-K prices a purchase path up to $60B in server chips against a 25M-share warrant. Only 3.75M shares vested on day one. Architecture, node, and foundry remain not disclosed in the filing.
NVIDIA is buying the open-model town square. The stays-open pledge is the product.
NVIDIA announced a $12,930,300,000 acquisition of Hugging Face. The 8-K adds about $11.9 billion payable to stockholders, up to $1 billion in retention equity, close expected in the first half of 2027, and required regulatory approvals. Hub neutrality is NVIDIA's public promise, not a close condition printed in the filing.
SGLang ships day-zero MoE kernels faster than most teams pick a runtime
Apache-2.0 serving framework. v0.5.14 landed 26 June 2026 with DeepSeek-V4 and Waterfill/LPLB load balancing.
NVIDIA says Vera Rubin is in full production
10x vs Grace Blackwell, SPDX of the MGX design, export licenses, and a firm shipment week are not disclosed.
DeepSeek-V4 paper reports 27% FLOPs and 10% KV cache at 1M tokens versus V3.2
arXiv:2606.19348 (April 26, 2026); MIT checkpoints on HuggingFace; CSA/HCA hybrid attention.
H200 and MI325X China licenses go case-by-case, still not a general license
BIS final rule 2026-00789, effective 15 January 2026, reviews U.S. exports of those SKUs case by case if exporters certify U.S. supply, a 50 percent TPP cap, KYC, and U.S. third-party testing. Reexports stay denied.
One brief a week.
New releases, then the constraint.