Pandorex dossier

AI chips

Compute alone does not explain an AI chip. This dossier connects accelerators, memory, software and supply chains, from mobile devices to server racks.

18 articles

The questions that matter

Is the claim about training or inference, and which model was measured?
Do memory, energy use and software support the promised performance?
Is the chip shipping, available as a sample or only announced?

Start here: reporting in context

Selected reporting from our archive. Each article links to its sources.

Cloud & Infra

d-Matrix Connects Raptor to NVIDIA NVLink — Open Silicon, Controlled Rack

Raptor is meant to handle decode beside Vera Rubin. The chip has not taped out yet; initial MGX systems are due in late 2027.

September 10, 2026·2 min read

—

AI & Chips

Arm Spreads Mobile AI Across CPU and GPU as CSS for Mobile 2 Skips a Standalone NPU

Arm CSS for Mobile 2 combines C2 CPUs, SME2 and Mali neural accelerators; Arm does not list a separate NPU in the platform.

September 8, 2026·4 min read

—

AI & Chips

AI Is Driving Memory Prices Higher: Why 'RAMageddon' Goes Beyond HBM

DRAM and NAND are becoming a much larger cost block for cloud providers. The key point is not that AI simply consumes ordinary PC RAM, but that it shifts capacity and supplier priorities across the memory industry — all the way down to notebooks, smartphones and SSDs.

September 7, 2026·6 min read

—

Latest developments & archive

AI & Chips

TSMC Hits Record Third-Quarter Revenue, but the Margin Picture Is Still Missing

TSMC generated NT$1.49 trillion in third-quarter revenue. Monthly data show record momentum, but not yet the profit trend.

October 8, 2026·4 min read

—

Cloud & Infra

Vultr Orders HPE’s First AMD Helios Racks in $1.2 Billion Deal

Vultr has ordered HPE’s first AMD Helios racks. The open Ethernet design connects 72 MI455X GPUs per rack.

September 30, 2026·4 min read

—

AI & Chips

Alibaba Pulls V900 Roadmap Forward From Q3 to Q1 2027

Alibaba says the V900 will ship with 216 GB in Q1 2027. Its 3× performance figure remains a vendor claim, not an independent result.

September 22, 2026·2 min read

—

Patents

Applied Materials outlines coaxial bridges for chiplets and HBM

A prefabricated module with individually shielded lines could connect chips at high density. The filing includes performance figures, but no independent product validation.

September 19, 2026·2 min read

—

Patents

Nvidia filing: AI agents could signal their compute needs ahead of time

A newly published patent application describes forecasts of upcoming model calls, tokens and GPU memory to help servers plan resources.

September 19, 2026·2 min read

—

AI & Chips

Dimensity 9600 Pro: MediaTek Launches 2nm Chip With Dual NPU

MediaTek's 2nm flagship supports LPDDR6, UFS 5.0 and models up to 30B. Performance figures still come from in-house tests.

September 15, 2026·2 min read

—

Patents

Intel proposes moving the control chip inside the HBM stack

A newly published application places interface logic between memory layers to balance signal paths. Moving it higher instead favors cooling.

September 12, 2026·1 min read

—

Cloud & Infra

d-Matrix Connects Raptor to NVIDIA NVLink — Open Silicon, Controlled Rack

Raptor is meant to handle decode beside Vera Rubin. The chip has not taped out yet; initial MGX systems are due in late 2027.

September 10, 2026·2 min read

—

Cloud & Infra

NVIDIA Bundles Australia's AI Buildout: 2 GW Is a Target, Not Live GPU Capacity

Eight partners target up to 2 GW of AI infrastructure by 2027. The figure spans land, power and data-centre capacity.

September 10, 2026·2 min read

—

AI & Chips

Amazon Ties Up to $60 Billion of Business to Qualcomm as AI Chips Become a Hyperscaler Bet

Amazon received warrants for 25m Qualcomm shares, with vesting tied to as much as $60bn of chip-related business.

September 8, 2026·4 min read

—

AI & Chips

Arm Spreads Mobile AI Across CPU and GPU as CSS for Mobile 2 Skips a Standalone NPU

Arm CSS for Mobile 2 combines C2 CPUs, SME2 and Mali neural accelerators; Arm does not list a separate NPU in the platform.

September 8, 2026·4 min read

—

AI & Chips

Enflame's Online Tranche Was 6,109x Oversubscribed: Money Still Does Not Replace CUDA

Chinese AI-chip contender Enflame is raising about 6.1 billion yuan and attracted extreme IPO demand. That shows investor appetite and strategic momentum — but it does not yet prove that Nvidia has been matched technically or at the software-ecosystem layer.

September 7, 2026·6 min read

—

AI & Chips

AI Is Driving Memory Prices Higher: Why 'RAMageddon' Goes Beyond HBM

September 7, 2026·6 min read

—

AI & Chips

AMD Halo Station: 96 Cores and up to 576 GB of GPU Memory

AMD shows a workstation for large local AI models. The prototype combines Threadripper with Instinct accelerators; launch is planned for 2027.

September 5, 2026·1 min read

—

AI & Chips

Google Eyes Two New AI Chips with Marvell — The Next Power Struggle Moves to Custom Silicon

According to Reuters and The Information, Google is in talks with Marvell to develop two new AI chips. This is more than a chip rumor: it shows how hyperscalers in 2026 are trying to reduce dependence on standard GPUs and bring inference costs under control.

April 20, 2026·4 min read

—

AI & Chips

AI Chips 2026: Inference Hardware Becomes the Battleground NVIDIA, AMD, Intel and the Startup Wave

Training was yesterday 2026 is all about inference. NVIDIA, AMD, Intel and a wave of startups are fighting for the mass market in AI hardware. An analysis of the balance of power.

March 31, 2026·7 min read

—

AI & Chips

AMD MI450: Billion-Dollar Deals with OpenAI, Oracle, and Cisco

AMD secures massive AI chip contracts: 6GW OpenAI commitment, 50,000 MI450 chips for Oracle, 1GW Cisco deal through 2030.

March 28, 2026·6 min read

—

New Technology

Hardware Prices Q1 2026: RAM Explodes, GPUs Scarce, Hard Drives Sold Out

DRAM prices rise up to 130 percent, hard drives are completely sold out for 2026, budget PCs are disappearing from the market. A current overview with real numbers.

March 28, 2026·6 min read

—