Hardware
Nvidia launches 64GB DGX Spark at $4,999 as 128GB prices climb
A cheaper 64GB configuration of the desktop AI system arrives on 23 October as memory prices push the 128GB model to around $6,950.
Still
Crunchy.
YESTERDAY, STILL WORTH KNOWING
Rising memory costs push up hardware prices, Steam sets sales records, and AI-generated content and game ports bring different kinds of trouble.
Memory costs squeeze Nvidia from both ends
Nvidia's Saturday news shows AI data centre demand for memory rippling outwards. The company has added a 64GB version of its DGX Spark desktop AI system, on sale from 23 October at $4,999, while the 128GB model that launched a year ago at $3,999 now sits at roughly $6,950 to $9,000. The new configuration keeps the same GB10 chip, software stack and 200GbE networking, and two units can be clustered to pool 128GB and run models up to 200 billion parameters, with a Sync Cluster Assistant handling setup automatically.
The same pressure hit a seven-year-old streaming box. The Shield TV Pro has jumped 50% to $299.99 despite unchanged hardware, with Nvidia citing component costs, and the cheaper standard Shield TV has reportedly been discontinued. Buyers are left with one pricier option that is already out of stock at most retailers.
AI floods the pipes, from bug bounties to token bills
Google has suspended product vulnerability submissions to its Open Source Software Vulnerability Reward Program, citing a flood of invalid AI-generated bug reports that swamped engineers with hallucinated findings. It follows Linux maintainers reporting they were overwhelmed by a record 2,000 kernel vulnerabilities per release, and Intel suspending its own bounty. The freeze does not affect supply chain reports or older submissions, with an update promised by early 2027.
The volume problem extends to usage itself. OpenRouter data charted by Andreessen Horowitz shows AI agents consumed 7.3 trillion tokens in August against 1.4 trillion for humans, a fivefold gap, though over 85% of agent tokens were cached prompts, meaning the machines were mostly rereading their own context. Cached tokens are cheap but still occupy memory, and a16z links the trend to demand for high-bandwidth memory that Micron expects to stay scarce into 2028.
AI's messy human fallout, from smuggling to safety exits
Greg Lui, chief executive of California's Earthmade Computer, has been charged with smuggling over $300m of export-controlled Nvidia A100 and H100 chips to China, allegedly routing servers through Malaysia and Singapore with false paperwork before delivery to Chinese buyers, including the government. He faces up to 20 years if convicted, in a string of similar prosecutions showing that demand for restricted chips keeps a smuggling market alive.
OpenAI safety researcher David Robinson has left with a public essay in The Atlantic arguing the industry's trial-and-error method means bigger mistakes as systems grow, and that AI firms should run like nuclear plants with layered redundancy. His departure continues a pattern dating back to Jan Leike in 2024. Meanwhile Meta has open sourced the code for building DIY Muse AI gadgets on hobbyist hardware such as ESP32 boards and Raspberry Pis, weeks after the chatbot's filesystem exposure was declared intended behaviour.
Gaming thrives while credit and cables fray
Steam is estimated to have taken a record $1.7 billion in September, up 13% on 2025, despite costly PC hardware, with established intellectual property generating nearly 80% of top-game revenue. Yet the fan porting scene is feuding: DarioSamo, who led the hand-built Banjo: Recompiled port, has disowned an AI-generated Banjo-Tooie port built on his team's research, calling it a race to the bottom. Capcom, by contrast, plans AI-assisted development tools for its RE Engine while insisting it does not generate art.
Hardware troubles persist offline too. A $5,245 prebuilt RTX 5090 PC melted its power cables after a year boxed up, leaving the buyer caught between Digital Storm's expired parts warranty and PNY's refusal to cover non-PNY cabling, though PNY is now reviewing the case. On a brighter note, extreme overclocker Kingpin is returning with ASUS after EVGA's exit, and PewDiePie has released Ajax, an uncensored fine-tune of Qwen3.5-9B that runs at home, saying OpenAI banned him twice for distillation, the practice of training one model on another's outputs.