Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Z.AI's API revenue jumps 28-fold to ¥825M
09/01/26, 07:35 PM

Z.AI, formerly Zhipu AI, generated ¥953.9 million in first-half 2026 revenue, up 399.7% year on year, while open-platform and API revenue reached ¥825.2 million, up 2,735.7%, in its official interim-results announcement. Cloud/API sales accounted for 86.5% of revenue, versus 15.2% a year earlier, a revenue-mix flip rather than simple top-line growth. Builders evaluating model suppliers should file that usage is scaling faster than Z.AI’s old licensing business; its GLM-5.3-Flash price-performance gap is the adjacent procurement fact.
Wire

Huawei spends ¥121.38B on H1 R&D
09/01/26, 07:32 PM

Huawei spent ¥121.38 billion on research and development in the first half of 2026, up 25.2% year on year, while revenue rose 9.55% to ¥467.82 billion, according to Xinhua’s report on the company’s semi-annual figures. The company named computing, cloud, connectivity, intelligent driving, and artificial intelligence among the strategic areas receiving that investment, making its domestic AI buildout a balance-sheet priority rather than a side project. Builders tracking chip and cloud supply should read it alongside the rising capital cost of regional AI capacity, because Huawei is funding the other side of the same compute race.
Wire

BigQuery Graph goes GA with 100x traversal gain
09/01/26, 06:36 PM

Google Cloud made BigQuery Graph generally available and says undirected traversal is now 100× faster than in preview, while GQL runs 2× faster. The GA announcement puts graph traversal, SQL, BigQuery ML, and AI functions in one petabyte-scale engine with no ETL, and adds an auditable context graph for agent actions. For builders grounding agents on connected enterprise data, this makes graph-RAG worth benchmarking against another vector index; OfficeQA’s harness results show why retrieval and orchestration—not model branding alone—decide production accuracy.
Wire

CrowdStrike ships SafeMind cyber models
09/01/26, 06:34 PM

CrowdStrike launched SafeMind with two cybersecurity models—Red Tempest for offensive work and Blue Solano for defense—and says its evaluation delivered 29% higher detection, 6× faster end-to-end remediation, and 99% lower cost than leading frontier and open-source baselines. The company’s launch release says the models are built with NVIDIA Nemotron open models and CrowdStrike telemetry, but the comparisons are vendor-reported rather than an independent benchmark. Security teams should put specialized cyber models on their own telemetry bake-off, alongside the paper’s analysis of shrinking exploit-disclosure windows, before changing a production control plane.
Wire

Tripo AI raises ¥3B for 3D foundation models
09/01/26, 05:49 PM

Tripo AI raised approximately ¥3 billion across Series B and B+ rounds while previewing Tripo P2.0, a 3D-native foundation model that adds quad topology, according to the company’s financing announcement. The capital will fund 3D data infrastructure, model training and inference, and commercialization across manufacturing, VR, games, and embodied AI. Teams evaluating generative 3D should test production topology and asset handoff—not only demo speed—as robot simulation turns 3D models into an engineering funnel.
Wire

Fervo signs Google for 396MW of geothermal power
09/01/26, 05:47 PM

Fervo Energy signed Google to a 15-year, 396 MW enhanced-geothermal power purchase agreement, delivered in four 99 MW tranches beginning in the third quarter of 2028, according to Fervo’s SEC filing. Google also gets an option for roughly 600 MW more, but that expansion is not committed; the filing puts potential contracted capacity near 950 MW only if Google accepts and the companies close a definitive agreement. For AI operators, the grid bottleneck already shadowing data-center growth is turning firm clean power into compute infrastructure.
Wire

Aranya raises $11M to run GPU clusters
09/01/26, 05:45 PM

Aranya raised $11 million after a customer cut GPU-cluster setup from six weeks to under 48 hours, while its distributed operating system reduced outages by 90%, according to the company’s funding announcement. The startup says it already manages more than $500 million of GPU hardware and turns bare metal into production-ready inference clusters in under 48 hours. For teams stitching together fragmented capacity, the data-center capex race behind GPU expansion makes cluster operations a capacity decision, not plumbing.
Wire

FaultSense cuts MoE fault tests 20x
09/01/26, 04:37 PM

Microsoft Research’s FaultSense cuts diagnostic tests for Mixture-of-Experts serving clusters by up to 20×, using an application-layer probe to localize faulty GPUs or communication paths without host instrumentation. The September 2026 paper frames gray failures as slow components that inflate serving latency without throwing explicit errors. For operators, this is a reminder to budget fail-slow detection alongside GPU capacity; the archive’s analysis of hyperscaler AI capex shows why a small diagnostic tax matters at rack scale.
Wire

Yext reports $440.8M ARR as AI search expands
09/01/26, 04:34 PM

Yext reported $111.1 million in fiscal Q2 2027 revenue and $440.8 million in annual recurring revenue, with $34.0 million in adjusted EBITDA at a 31% margin; it also completed its GoShine acquisition to extend brand visibility into AI search. The company’s SEC-filed earnings release says it released a Corvo AI prototype and is expanding its API and MCP offering. For builders, Yext’s quarter makes structured business data an operating input for agent discovery, not a search-marketing sidecar; the archive’s analysis of Shopify’s AI-referred catalog demand is the adjacent signal.
Wire

DeepInfra deploys 1,000 B300 GPUs in Oregon
09/01/26, 03:44 PM

DeepInfra and humans& announced a Hillsboro, Oregon cluster with more than 1,000 NVIDIA B300 GPUs and more than 2 MW of power capacity, in DeepInfra’s announcement. The dedicated DeepCluster deployment is aimed at training, inference, and research; DeepInfra says its managed platform supports 200+ open-source models and processes nearly 10 trillion tokens per week. The scale is a useful counterweight to token-price comparisons: teams planning frontier workloads should price dedicated power, deployment speed, and utilization beside the site’s analysis of Nvidia’s share of hyperscaler capex.