The hardware and infrastructure behind AI: chips and memory, data centers and power, cloud compute, and the inference engines that serve models.
A Tsinghua paper lets agents drop and reposition KV cache instead of summarizing history. Its AKV engine, built on SGLang, cut API-equivalent cost 27.7% and added 1.7 points on average across three open models and four agent benchmarks, with 1.51x faster P95 latency.
Meta's Q3 revenue rose 26% to $64.6 billion, but $34.9 billion in capex pushed free cash flow to negative $2.1 billion. It also announced nonvoting Class C stock that it says can help finance growth; Reuters says it may avoid diluting Zuckerberg's control.
Recursive's Flint is a 40B MoE coding model with 2.5B active parameters, released under Apache 2.0. Here is what it scores in Recursive's own tests, how fast it runs on a Mac or RTX Spark, the hardware you need, and what the Eureka Machine built.
NVIDIA's DSX POD is a reference design for containerized Vera Rubin AI factories, built and tested by partners before delivery. Here is what is inside one container, how deployment changes, what it does not fix, and who ships it in early 2027.
DeepSeek open-sourced DeepInfer, the Apache 2.0 inference engine behind DeepSeek-V4.1-Flash, for NVIDIA SM100 and Huawei Ascend 950. What it includes, what it leaves out, and what DeepSeek's claim of 155 trillion tokens a day and $38 per GPU-hour does and does not show.
A new paper replaces Gemma 3 270M's 168M-parameter embedding table with frozen 512-bit codes. Trained on the same 2.05B tokens, it ties or beats the learned table on LAMBADA, HellaSwag and PIQA, and a ternary version is a 46 MB file running 615 to 740 tokens per second on a laptop CPU.
RTX Spark puts an NVIDIA Blackwell RTX GPU, a Grace CPU and up to 128GB of unified memory in one chip for Windows laptops. Here are the specs, the laptops, what NVIDIA says each memory size runs locally, Nemotron 3.5 Bolt, and retailer prices from $2,599.99.
Nvidia's Nemotron 3.5 Bolt 74B-A9B packs a model compressed from Super VL into 24GB with 2-bit experts for RTX Spark and DGX Spark. Here is what it needs to run, how its speed compares in Nvidia's own tests, and when Bolt E4B arrives.
Meta's engineering post on Prometheus, its Ohio supercluster that trained Muse, covers three Nvidia rack generations, a network spanning over 100 km, 8 spare GPUs per 72-GPU rack, a 99% cut in interruptions, and power swings of tens of megawatts.
A beginner's guide to running an AI model on your own computer with Ollama, LM Studio or llama.cpp: hardware needs, install steps, small models to start with, and our real test results on a CPU-only Linux machine, as of October 2026.
TorchNeuron makes Trainium2 and Trainium3 a standard PyTorch device called neuron. Here is what it supports, what the beta leaves out, how AWS documents setup today, and how it differs from the torch-neuronx XLA path and from CUDA.
Modal, which raised $750M at a $15.75B valuation last week, says Nvidia's Arm-compatible Vera CPU gave up to double the task throughput of an x86 Intel GNR machine in its sandbox benchmark. It will run Vera alongside Intel and AMD in a mixed fleet.
TorchNeuron runs PyTorch on Trainium with a one-line device change and is tested nightly on 100,000+ Hugging Face models. It removes the cost of trying another chip. Running fast still takes kernel work, as Deepgram's multi-month port shows.
Apple's M5 Ultra Mac Studio starts at $5,499, with up to 512GB of unified memory coming in late October and Thunderbolt 5 clustering for running large AI models at home. What ships now, what the memory and clustering claims mean, and who it is for.
OpenBMB's MiniCPM-V 4.7 is an Apache 2.0 vision model in 35B-A3B and 1B sizes. A new expert-scheduling method streams MoE weights from flash, and OpenBMB reports up to 20.51 tokens per second on an iPhone. Here is how it works, how it scores, and how to run it.
SemiAnalysis says more than 24GW across 40 campuses is in advanced construction, most of it due in 2027, and calls for +50GW in a single year. The same week, the FT counted 47+ stalled US projects. The two numbers measure different stages of the pipeline.
doubleAI's Maxwell sits inside running Python and PyTorch code, swaps in faster machine code, and attaches a formal proof that it behaves the same. Here are its published speedups, how it compares with Claude Code, Codex and Cursor, the $220M company behind it, and what is still unknown.
Jim Keller's RISC-V AI chip company closed a $1.5B Series E led by Trace Capital and Innovation X, more than doubling its value. It will deploy Galaxy servers at scale and expand the CPU and AI IP it licenses to Samsung and LG.
RegLayer counts at least 47 large US data center projects delayed, withdrawn or blocked in 2026, and Bernstein says only about 35% of the pipeline is credible. CoreWeave bond yields near 15% and paused IPOs show the money side. Here is how power, debt, chips and permits feed each other.
NVIDIA has released Nemotron 3.5 Super VL, an open-weights vision-language model designed for complex agentic workflows. The model features a 120-billion-parameter mixture-of-experts (MoE) architecture that activates 12 billion parameters per token, paired with a context window supporting up to 1 million tokens. NVIDIA announced it alongside Nemotron 3.5 Bolt, a 2-bit model for local agents on RTX Spark PCs. The release is available on Hugging Face in full-precision BF16 reference weights and q
