Hardware News.
The chips, GPUs and data-center hardware underneath the AI boom, explained without the spec-sheet jargon. We cover new silicon from Nvidia, AMD, Intel, Apple and the custom-ASIC challengers, and why each launch matters for cost, performance and who wins.
Hardware
NAVER, NVIDIA Build Gigawatt Sovereign AI in Korea
NAVER will expand its sovereign AI infrastructure on NVIDIA DSX, starting at 55 megawatts and scaling toward gigawatt capacity to power next-gen HyperCLOVA X models and a Korean AI agent platform.
Hardware
New York Orders First US Data Center Moratorium
New York became the first state to freeze new AI data centers: Gov. Kathy Hochul signed an executive order on July 14, 2026 pausing environmental permits for facilities of 50 MW or more for up to a year while the state writes new energy, water, and cost rules.
Hardware
SK Hynix HBM4 Warning Triggers Record Memory-Stock Selloff
A brokerage note pegging SK Hynix's Q2 profit 8% below consensus, citing slower HBM4 shipments, sent SK Hynix down 15% in its worst single-day drop ever and dragged Micron, SanDisk, and Western Digital down about 6% in sympathy on July 13, 2026.
Hardware
SK hynix and NVIDIA Sign Multi-Year Deal for Vera Rubin Memory
SK hynix and NVIDIA have signed a multi-year partnership to co-develop memory for the Vera Rubin platform, Vera CPUs, RTX Spark, and Jetson Thor, locking in HBM supply for NVIDIA's next three GPU generations.
Hardware
Nvidia's LPU Bets Inference Needs Its Own Silicon
Nvidia unveiled a Language Processing Unit, a chip class purpose-built for LLM inference rather than training, signaling that the company sees the money moving from building models to serving them and wants to own that layer before rivals like Groq and Cerebras do.
Hardware
Intel Fixes Its 18A Yield Problem, Clearing Panther Lake Ramp
Research firm BlueFin says Intel has resolved the yield issues plaguing its 18A process and is ramping production toward roughly 30,000 wafers a month, the single biggest step toward making Panther Lake widely available and proving Intel Foundry can compete.
Hardware
Nvidia's Feynman Bets AI's Future on Light, Not Copper
Nvidia's post-Rubin Feynman architecture will be its first to use NVLink switches with co-packaged optics, moving photonics next to the switch chip because copper can no longer move enough bits per watt in AI racks.
Hardware
Google's Ironwood TPU Goes Live to Challenge Nvidia
Google's seventh-generation Ironwood TPU is now generally available, packing 9,216 chips into a single 42.5-exaflop pod and claiming a 44% lower cost of ownership than an Nvidia GB200 server.
Hardware
RAMageddon: why RAM prices just spiked 60% in one quarter
AI's appetite for HBM is cannibalizing the wafers that make ordinary DRAM, driving the worst memory shortage in 15 years: DRAM contract prices rose up to 63% and NAND up to 75% in a single quarter.
Hardware
Rapidus undercuts TSMC with $20K 2nm wafers for 2027
Japan’s state-backed Rapidus plans to price its 2nm-class wafers around $20,000 when it launches in 2027, a deliberate discount to TSMC’s leading-edge rates aimed at breaking the foundry’s pricing power.
Hardware
Qualcomm bets $4B on Modular to own the AI software stack
Qualcomm is in talks to buy Modular Inc. for about $4 billion, a move to bolt a CUDA-style software layer onto its silicon and turn its Dragonfly data-center chips into a real Nvidia alternative.
Hardware
TSMC Hikes Wafer Prices as Its Foundry Grip Tightens
TSMC is raising 3nm wafer prices by roughly 15% before year-end and lifting all leading-edge nodes 5 to 10%, using a 72% share of the foundry market to pass AI-era costs straight to Nvidia, AMD, and Apple. The price of every advanced chip is about to move.
Hardware
2026 is a GPU drought: RTX 50 Super slips to 2027
The RTX 50 Super refresh has reportedly slipped to CES 2027 and AMD's next-gen RDNA 5 is not expected until late 2027 or 2028, leaving 2026 as one of the thinnest years ever for genuinely new desktop graphics cards.
Hardware
Nvidia's Groq 3 LPU Bets Inference Isn't a GPU Problem
Nvidia is shipping a Language Processing Unit, the Groq 3 LPU, in liquid-cooled racks of 256 chips with 128GB of on-chip SRAM, a purpose-built inference part that concedes GPUs are not the ideal way to serve large models fast.
Hardware
Nvidia's RTX Spark Chip Storms the Windows PC Market
Nvidia unveiled its RTX Spark Superchip, a laptop processor for Windows PCs, pushing beyond graphics into the CPU that has long belonged to Intel, AMD and Qualcomm, and sending their shares lower.
Hardware
John Deere must open repairs under FTC right-to-repair deal
The FTC and five states secured a settlement on July 8, 2026 forcing John Deere to give farmers and independent shops the same repair software it reserves for dealers, for 10 years under federal supervision.
Hardware
Nvidia's 16-Hi HBM demand triggers a memory sprint
Nvidia asked SK Hynix, Samsung and Micron to deliver 16-high HBM4 stacks by Q4 2026, a density never shipped commercially. To fit 16 layers the suppliers must thin each DRAM wafer to about 30 microns without warping, and the winner locks in the richest slice of the AI-memory boom.
Hardware
RAMpocalypse: DRAM Prices Jump 90% as AI Eats Supply
DRAM prices surged about 90% in Q1 2026 as Samsung, SK Hynix and Micron shifted most output to AI memory, draining consumer supply and triggering a US price-fixing lawsuit.
Hardware
NVIDIA Halos Brings a Safety Stack to Physical AI
NVIDIA unveiled Halos, described as the first full-stack safety system for robotics and physical AI, unifying AI compute with functional safety across the robot stack.
Hardware
TSMC Starts 2nm Volume Production, Already Sold Out
TSMC began volume production of its 2nm-class N2 process, its first with gate-all-around transistors, and its 2026 capacity is already fully booked, with Apple taking more than half of the initial output ahead of Qualcomm, AMD and Nvidia.
Hardware
Nvidia's RTX Spark Superchip Puts an AI PC in a Box
Nvidia's RTX Spark is a single superchip pairing a 20-core Arm CPU with a Blackwell GPU and 128GB of unified memory, aiming to run local AI agents on Windows PCs shipping this fall from Dell, HP, Lenovo, Asus, and Microsoft Surface.
Hardware
Qualcomm's AI200 and AI250 Chase Nvidia in Inference
Qualcomm, best known for phone chips, is entering the data center with its AI200 and AI250 accelerators aimed at AI inference, the cheaper, higher-volume half of AI compute, with Saudi firm Humain signed as its first 200-megawatt customer.
Hardware
Samsung Ships First 12-Layer HBM4E Memory Samples
Samsung has sampled the industry's first 12-layer HBM4E memory, hitting up to 16 Gbps per pin and about 3.6 TB/s of bandwidth per stack at 48GB, aimed squarely at next-generation AI accelerators.
Hardware
Nvidia's Vera Rubin Platform Targets the Supercomputer
At ISC High Performance 2026, Nvidia unveiled the Vera Rubin platform, a full-stack successor to Blackwell that restores native double-precision (FP64) compute for scientific supercomputing. With 35 AI-HPC systems already in development across Europe, Nvidia is fusing the AI and traditional-HPC roadmaps into one architecture.
Hardware
Jim Keller's Tenstorrent Guns for Cerebras and Nvidia
Tenstorrent CEO Jim Keller says his startup's Blackhole Galaxy servers deliver up to 5x better cost than an Nvidia GB300 and will beat Cerebras on everything, betting the AI-chip war is now won on total cost of ownership and open RISC-V, not peak compute.
Hardware
AMD's MI455X and Helios Take Direct Aim at Nvidia
AMD's Instinct MI455X accelerator and its 72-GPU Helios rack, detailed at the Advancing AI 2026 event in July, are engineered to match Nvidia's Vera Rubin on training and inference, even as analysts dispute whether volume shipments slip toward 2027.
Hardware
Nvidia's RTX Spark superchip turns Windows into an AI OS
Nvidia's RTX Spark is a superchip that fuses a Blackwell RTX GPU (6,144 CUDA cores) with a 20-core Grace CPU and 128GB of unified memory over NVLink-C2C, delivering up to 1 petaflop of AI compute so agentic models run on-device inside Windows PCs shipping this fall.
Hardware
Nvidia's Vera Rubin Brings Native FP64 Back to Science
At ISC 2026 in Hamburg, Nvidia positioned Vera Rubin as a scientific supercomputer: each Rubin GPU delivers 200 TFLOPS of FP64, 288GB of HBM4, and 22 TB/s of bandwidth, with rack-scale systems hitting 5 petaflops of native double precision.
Hardware
SK hynix and Samsung Race to Ship 12-Layer HBM4E
Samsung shipped the industry-first 12-layer HBM4E samples on May 29, 2026 and SK hynix followed on June 18, both packing 48GB per stack at roughly 4TB/s to feed next year AI accelerators like Nvidia Rubin Ultra, but neither has cleared qualification or volume production yet.
Hardware
Intel's Crescent Island GPU Skips HBM for 160GB of LPDDR5X
Intel's Crescent Island is an inference-only data-center GPU on the new Xe3P architecture that packs 160GB of LPDDR5X, scalable to 480GB, deliberately trading HBM bandwidth for capacity and cost to dodge the memory supply crunch. Sampling begins in the second half of 2026.