AI Chips and Infrastructure

Oct 29

The Verge reports that Arizonans embraced TSMC, which uses nearly 5 million gallons of water a day, while a bipartisan coalition now fights hyperscale data centers. This post explains why, what is disputed, and what it means for AI buildouts.

Oct 29

A New York Times report says US intelligence believes a Singapore data-center executive with access to Nvidia CEO Jensen Huang was an intelligence asset for China, and that President Trump was briefed and did not act. Here is what the report claims, what Nvidia and Megaspeed said, and what nobody has confirmed.

Oct 29

Samsung Electronics reported 195.3 trillion won revenue and 107.4 trillion won operating profit for Q3 2026, with HBM4 sales more than tripling and 2027 HBM sold out. A viral line says long-term deals may cover 70% of memory output through 2030. We checked what the call actually says.

Oct 28

Reuters reports, citing sources, that China's CXMT will begin limited HBM3E production despite yields of only 20% to 30%. This post lays out the claim, the earlier estimates it builds on, what no one has confirmed, and why AI builders should care.

Oct 28

Reuters reported on October 28, 2026 that Justice Department officials are asking whether Nvidia's $12.93 billion purchase of Hugging Face would hand the chipmaker control of a key channel for distributing AI. Here is what is sourced, what is not, and what it means for anyone who pulls open weights.

Oct 28

Customs at Kuala Lumpur International Airport seized 30 servers with advanced AI chips worth nearly RM49.5 million (about US$12.1 million), the fifth case linked to the same suspected syndicate this year. Here is what officials said, what they withheld, and how the permit rule works.

Oct 28

Chinese GPU maker Moore Threads reported third-quarter revenue of 1.552 billion yuan, up 1,773.9%, and a net profit of 49 million yuan. We lay out the figures, the JD Cloud 100,000-GPU plan, and what is and is not confirmed about China's domestic AI chip push.

Oct 28

NVIDIA released Nemotron 3.5 Transcribe, an open speech model that transcribes 40 languages, translates 31, labels speakers in 13 and answers questions about recordings, plus Granary v2, an open dataset of about 844,000 hours of speech. Here is what it does, how the numbers compare, and where it stops.

Oct 28

Planning their 2027 supply, PC makers including HP and Dell reportedly found an unglamorous worry: printed circuit boards and chip substrates. We pull together the reporting on lead times, materials and prices, and mark what is a single source.

Oct 28

SK hynix reported 104.6 trillion won in revenue and 83.33 trillion won in operating profit for Q3 2026, with HBM4 set to pass HBM3E in Q4 and long-term deals covering over half of 2027-2028 capacity. Here is what is confirmed, what missed, and what it means for the price and availability of AI compute.

Oct 27

China's AI and chip stock index has fallen about 40% in under four months, according to Reuters. We separate the verified numbers from the explanations, and spell out what it does and does not change for people who build with AI.

Oct 27

Cloudflare disclosed on October 26 that its GLM-5.3 model on Workers AI sometimes answered with fragments of other customers' requests. The cause was a context-window mismatch plus a bug in the open-source SGLang engine. Here is what happened and what operators should check.

Oct 27

Corning's Q3 2026 results, filed on October 27, put numbers on the fiber and connectivity side of the AI build-out: Optical Communications up 44% to $2.38B and Gen AI product sales growing more than 100%. We read the filing and separate what it states from what headlines add.

Oct 27

The Verge reports that data centers have become a wedge issue in close 2026 races from Texas to Pennsylvania. Candidates from both parties are walking back earlier support, and AI-linked super PACs are spending heavily. Here is the state-by-state picture and what it could mean for AI buildout.

Oct 27

On October 27, 2026, NVIDIA published how CUDA 13.4 enforces per-workload GPU memory limits and how a new open-source operator plus KAI Scheduler v0.18.0 expose it in Kubernetes. explainx.ai explains the modes, the annotations and when to use which.

Oct 27

An NVIDIA developer post shows three changes that speed up Ling-3.0-flash on a single DGX Spark: a runtime-compiled MXFP4 MoE kernel, an FP8 LM head and built-in multi-token prediction. The synthetic benchmark says 185% faster; the full GSM8K run says 42.5%. Here is the difference and the silent-fallback trap.

Oct 27

NVIDIA published a worked recipe for adapting an open Nemotron model to local rules: sample synthetic residents from Nemotron-Personas-France, generate grounded conversations, LoRA fine-tune, merge, and compare against the base model.

Oct 27

NVIDIA published a developer post on October 27, 2026 about building games for RTX Spark: a faster DLSS 5 model, the Nemotron 3.5 Bolt E4B model for on-device AI characters, eight RTX Kit SDKs, Arm64 Nsight tools and updated MotionBricks motion models.

Oct 27

At Oracle AI World on October 27, 2026, Oracle and Google announced that Google TPUs will be available through Oracle Cloud Infrastructure starting in calendar 2027. We separate what both companies actually said from what they left out, including pricing, regions, capacity and which TPU generation.

Oct 27

At Oracle AI World, Oracle and NVIDIA said Vera Rubin NVL72 and Vera CPUs are generally available on OCI, with an initial 150,000 CPUs, and OpenAI said it is bringing Vera into production for agent sandboxes. We separate the shipped parts from the planned ones.

Oct 27

Supply-chain sources say TSMC is preparing a CoWoS advanced packaging price rise of up to 20%, as equipment lead times stretch. TSMC has not confirmed any number. We separate what is traced to a source from what is repeated on social media, and explain why packaging prices matter for AI.

Oct 26

At GTC Berlin on October 22, Jensen Huang said Nvidia licenses its agent software at about $2,000 per GPU per year. A $50 billion headline followed, but that number is an analyst scenario, and Nvidia has not said which software is covered, when it starts, or how it relates to the $4,500 AI Enterprise list price.

Oct 26

NVIDIA Mission Control 2.5 takes the autonomous recovery engine that used to cover only Slurm and applies it to Kubernetes workloads scheduled by Run:ai. NVIDIA publishes no recovery-time numbers, so this guide explains the mechanism and what to ask before adopting it.

Oct 26

A chart showing Vera Rubin training a 10T-parameter model on 100T tokens in a month with about 16,000 GPUs instead of 64,000 is spreading on X. We read Nvidia’s own page and footnote: what the claim is, what it assumes, and what independent data would be needed.

Oct 26

Generator step-up and substation transformers now take more than 160 weeks to arrive, according to Wood Mackenzie figures reported by Reuters and Data Center Knowledge. This post sorts the numbers, the sources that disagree, and what the wait means for AI capacity.

Oct 26

Unconventional AI, the $4.5 billion startup led by Naveen Rao, announced on October 26 that it is acquiring Israeli AI inference chip company NeuReality for an undisclosed cash-and-stock price. It is the first acquisition since the startup launched, and it is about people more than products.

Oct 25

Oct 25

On October 25, Elon Musk said SpaceX closed $40 billion of debt for GPUs, that debt markets have almost entirely shut off AI data center borrowers, and that 200+ neoclouds face "death and consolidation". Earlier reporting described the $40 billion as early-stage talks expected to close in 2027. explainx.ai sorts the claims from the record.

Oct 25

Two open-source Linux kernel modules remove the half-rate cap Nvidia applies to FP32-accumulate tensor math on GeForce cards. The RTX 4090 goes from 165.6 to 331.6 TFLOPS in synthetic tests; real matrix multiplies gained 12 to 34 percent in one owner run. Here is what is verified, what is not, and who should care.

Oct 24

In a Wall Street Journal interview, Ciena CEO Gary Smith, who lived through the 2001 telecom crash, describes AI demand outpacing supply of optical networking parts. explainx.ai checks that against Ciena's September earnings release and earlier interviews.

Oct 24

On October 24, 2026, a post citing Silicon Data rental indices said Nvidia H100 GPU rental costs are up 45% since December, and investor Chamath Palihapitiya said power is the binding constraint. The public Silicon Data index shows about 44% from its December base.

Oct 23

After buying the 122-acre former George Washington University campus in Loudoun County for $427 million, Amazon has withdrawn its plan for four data centers there. The deal, called Project Potomac Promise, trades the project for land, housing money and a ban on data centers at the site. Here is what each side said, what is still undecided, and why it matters for AI infrastructure.

Oct 23

On October 21, 2026 Swedish inference provider Berget AI published an open-source, per-request carbon method built with SEI, WattTime and Klimatkollen. This guide explains the numbers, what the method includes and leaves out, and what developers on Hacker News pushed back on.

Oct 23

Israeli security firm Lava says it identified 32 Nvidia Blackwell Ultra B300 GPUs on the network of state-owned China Telecom, found through publicly reachable GPU monitoring endpoints. Nvidia did not deny it. Here is the evidence trail, what it cannot show, and why the detection method matters.

Oct 23

Modular released MAX 26.7 on October 22, 2026 and opened the graph compiler, runtime and hardware backends. Social posts called it fully open for development, extension and commercial use. Modular's own wording is "source available", and the repo carries a MAX License with usage-data, branding and AI-training clauses. Here is what is confirmed.

Oct 23

Nvidia put Nemotron 3.5 Bolt E4B on Hugging Face on October 23, 2026. It is an 8B-total, 4B-effective post-train of Gemma 4 E4B tuned for multi-turn tool use on phones, Jetson boards and robots, with a Q4_K_M GGUF and a speculative-decoding drafter.

Oct 23

NVIDIA published a Kazakhstan compute update on October 22, 2026 during its GTC Berlin run. It lists GPUs already deployed by Kazakhtelecom, planned Vera Rubin capacity in Ekibastuz, and a robotaxi study with inDrive and Pony.ai. This post separates deployed hardware from plans and ties it to the June 2026 Data Center Valley deal.

Oct 23

NVIDIA OpenShell, the open-source runtime for sandboxing autonomous AI agents, hit 409 points on Hacker News. Here is what it does, how to try it, where it sits next to Docker and other sandboxes, and what developers complained about.

Oct 23

NVIDIA published Part 1 of a series on the TensorRT Global Performance Tuner, a TensorRT 11.1 feature that tries several build routes for your model, checks accuracy, and picks the fastest engine. Here is the idea in plain language, what the docs confirm, and what NVIDIA has not yet published.

Oct 22

Bolt announced that its planned fleet of at least 25,000 autonomous Lucid vehicles will use Nvidia's Alpamayo Stack self-driving software on Hyperion hardware, starting with data collection in Madrid in 2027. We separate the confirmed from the undated and explain where Nvidia now competes.

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 22

Oct 21

Oct 21

Oct 21

Oct 21

Oct 21

Oct 21

Oct 21

Oct 21

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 20

Oct 19

Oct 19

Oct 19

Oct 19

Oct 19

Oct 18

Oct 18

Oct 18

Oct 17

Oct 17

Oct 16

Oct 16

Oct 16

Oct 16

Oct 16

Oct 16

Oct 16

Oct 16

Oct 15

Oct 15

Oct 15

Oct 15

Oct 15

Oct 15

Oct 15

Oct 15

Oct 15

Oct 15

Oct 14

Oct 14

Oct 14

Oct 14

Oct 14

Oct 14

Oct 14

Oct 14

Oct 14

Oct 13

Oct 13

Oct 13

Oct 13

Oct 13

Oct 13

Oct 12

Oct 12

Oct 12

Oct 12

Oct 12

Oct 11

Oct 11

Oct 10

Oct 10

Oct 9

Oct 9

Oct 9

Oct 9

Oct 9

Oct 9

Oct 9

Oct 9

Oct 9

Oct 9

Oct 7

Oct 7

Oct 5

Oct 3

Oct 3

Oct 3

Oct 1

Oct 1

Sep 29

Sep 28

Sep 28

Sep 28

Sep 24

Sep 23

Sep 22

Sep 21

Sep 17

Sep 17

Sep 17

Sep 17

Sep 17

Sep 17

Sep 17

Sep 16

Sep 16

Sep 14

Sep 13

Sep 11

Sep 10

Sep 10

Sep 8

Sep 7

Sep 7

Sep 5

Sep 4

Sep 4

Sep 4

Sep 3

Sep 3

Sep 3

Sep 3

Sep 2

Sep 1

Aug 30

Aug 27

Aug 27

Aug 26

Aug 25

Aug 24

Aug 24

Aug 24

Aug 21

Aug 20

Aug 19

Aug 18

Aug 17

Aug 16

Aug 13

Aug 12

Aug 11

Aug 10

Aug 10

Aug 9

Aug 7

Aug 7

Aug 6

Aug 5

Aug 5

Aug 5

Aug 3

Jul 31

Jul 30

Jul 29

Jul 27

Jul 27

Jul 26

Jul 26

Jul 26

Jul 26

Jul 26

Jul 26

Jul 23

Jul 23

Jul 21

Jul 21

Jul 21

Jul 21

Jul 16

Jul 14

Jul 9

Jul 7

Jul 7

Jul 6

Jul 2

Jun 29

Jun 27

Jun 24

Jun 21

Jun 20

Jun 16

Jun 14

Jun 6

Jun 4

Jun 1

May 30

May 17

May 15

May 8