GPU prices keep climbing as AI data centres soak up memory
DRAM prices are forecast to rise 400 percent from 2024 levels by the end of 2026, and chipmakers say supply will stay tight well beyond that.
CRUNCHY TECH
EVERY PUBLISHED BRIEF
Browse every Crunchy Tech brief, grouped by date.
PUBLISHED
DRAM prices are forecast to rise 400 percent from 2024 levels by the end of 2026, and chipmakers say supply will stay tight well beyond that.
Arete-27B, built on Alibaba's Qwen3.8-27B, cut propaganda answers from 22 percent to under one percent in the company's own tests while keeping general performance largely intact.
Carnegie Mellon and US national lab researchers found most AI agent papers failed integrity checks, claiming experiments that never happened.
The SSL VPN bug, tracked as CVE-2026-19868, may already have been used in attacks, and SonicWall says there are no workarounds.
The first tranche of a round that could reach $20 billion has landed, but OpenAI scaled back its ambitions and is now courting wealthy individuals for the rest.
PUBLISHED
Censys found more than 800 password-free deployments of the open source AI penetration testing tool, which has been linked to attacks on Korean banks and ConnectWise.
A 10 billion yen plan adds AI-equipped experts, a cyber rescue team and active cyber defence after breaches hit dozens of companies and government agencies.
Token spend grew just 1.5% across August and September as price cuts and thriftier customers squeeze AI lab revenue, with Anthropic's IPO looming.
A hands-on comparison with this year's Android foldables finds Apple's first foldable polished but heavy, with a cramped cover screen and software that lags the competition.
Mira Murati's startup says its 17-billion-active-parameter model leads western open-weight rivals, with strong agentic and visual scores and a planned Apache 2.0 release.
A planning notice values the Home Office's public cloud hosting renewal at £881.6 million before VAT, up from the roughly £450 million contract AWS won in 2023.
A preliminary DSA ruling says TikTok’s For You feed traps vulnerable users in harmful rabbit holes, and the company must change it or face fines of up to 6% of global turnover.
Elon Musk says the Texas factory will switch to building the Cybercab self-driving taxi, ending a pickup that never neared its sales target.
An unauthorised account read archived customer contact details and some passports, driver's licences and utility bills; certificate systems were untouched.
An amendment to the Cyber Security and Resilience Bill would let ministers regulate how organisations deploy agentic AI and investigate incidents when safeguards fail.
A UN conference in Geneva this month could decide whether talks begin on a legally binding treaty for autonomous weapons, with Russia, the US and Israel pushing back.
During a coding test, an internal model crafted raw packets and probed AWS after its sandbox proxy failed. It never reached the target site, and OpenAI says no sensitive data was moved.
The company says its AI-native agent will build travel interfaces on the fly rather than chat, and will also handle in-trip support.
India's CERT warns of escalating AI-assisted attacks timed for weekends and holidays, with victims hit again after incomplete remediation.
PUBLISHED
A researcher's *sshole flaw exploits musl's handling of passwordless accounts, but only under narrow conditions; OpenSSH 10.6p2 fixes it.
CISA, the FBI and other departments say internet-connected refrigeration systems in supermarkets, military freezers and even a French blood bank are being targeted.
Anthropic's Personal Finance feature rolls out in beta to US Claude Pro, Max and Ultra subscribers, with spending breakdowns and affordability answers.
Sekoia ties a wide espionage campaign to APT29, using poisoned websites and a compromised analytics service to steal data from unpatched phones.
The cybersecurity-tuned model scores 48.5% on ExploitGym and stays in scope over 99% of the time, but only vetted Daybreak Red customers can use it for now.