Nvidia invests $3.5B in MediaTek and opens NVLink Fusion to rival custom chips
Nvidia is putting $3.5 billion into MediaTek via convertible bonds and giving it NVLink Fusion access, so the custom accelerators MediaTek builds for hyperscalers plug straight into Nvidia rack-scale systems — a week after AWS committed to 2 million more Nvidia GPUs for 2027-2028. The move concedes the custom-silicon chip itself while making Nvidia's interconnect the standard every AI factory builds around, including those of Amazon, Google, Microsoft, OpenAI and Anthropic as they design away from its GPUs. MediaTek expects $2 billion in custom data center ASIC revenue in 2026, and the deal extends to DGX Spark, RTX Spark consumer PCs, and software-defined vehicles.
techcrunch.com →- 02EU designates ChatGPT a very large search engine under the DSA
The European Commission classified ChatGPT as a very large online search engine — the first generative AI chatbot to carry the label — after OpenAI reported roughly 159 million average monthly EU users, with Reddit and Roblox designated very large online platforms. All three have until December 31, 2026 to run systemic risk assessments and open ad archives, researcher data access, and semiannual transparency reports. The unresolved question is scope: law professor Natali Helberger notes Article 40 access rights may or may not extend to training data and model weights, and the designation does not let the Commission test the models directly.
the-decoder.com → - 03China's CXMT starts small-batch HBM3E, one generation behind Samsung and SK Hynix
CXMT is producing HBM3E in small quantities, per The Information, while Samsung, SK Hynix and Micron mass-produce HBM4 — a real first for China's memory industry, but at roughly 25 percent yield per SemiAnalysis and three to five years behind technically. Alibaba's T-Head and Cambricon plan to ship products on it from 2027, loosening one export-control chokepoint on domestic AI chips, though CXMT's $8.6B Shanghai IPO earmarked none of the raise for HBM.
the-decoder.com → - 04Infostealers hijacked Claude session cookies, and Anthropic force-signed users out
Anthropic signed affected Claude users out, stripped saved payment methods, and refunded unauthorized charges after Vidar, LummaC2, StealC, RedLine, Acreed on Windows and Atomic Stealer on Macs harvested live session cookies from users' own machines. The compromise was local, not in Claude, and cookie replay walks past both passwords and 2FA — so session tokens are now the credential worth protecting, and signing out one service leaves every other active session on an infected machine still replayable.
helpnetsecurity.com → - 05OpenAI's ChatGPT ads reach a $1B annual run rate 200 days after launch
OpenAI says ChatGPT ads reached a $1 billion annualized run rate about 200 days after US testing began, with ads now live in more than 40 countries for free users and Go subscribers and a self-serve platform opening in India, Europe, the Middle East, and North Africa. Ad revenue is now a third pillar alongside subscriptions and API usage, and the timing matters — OpenAI is preparing an IPO and defending an $852 billion valuation that recurring consumer ad revenue helps underwrite.
the-decoder.com → - 06Claude Code and Codex misjudge elapsed time by 3x to 10x on coding tasks
Two MATS researchers tested Claude Code and Codex across 218 coding tasks and found both guess around 90 minutes regardless of difficulty, overshooting by 3x for Claude and 6-10x for Codex, with short tasks worst; Opus 4.8 and GPT-5.5 also rated their own output about 20 points above actual scores, once claiming 70 percent success against real marks of 7 and 14.5 percent. Giving agents a tool that reports elapsed time fixed the estimates almost every time, so any instruction like 'iterate for two hours' needs an explicit clock in the harness rather than the model's own sense of duration.
the-decoder.com → - 07Willison maps ChatGPT Work: Work Cloud's code sandbox reaches the open internet
Simon Willison splits ChatGPT Work into Work Cloud, on web and mobile, and Work Local, the desktop app he reads as Codex re-skinned for non-developers. Work Cloud is the substantive one: a code sandbox with internet access open by default, headless Chrome, a filesystem persisted across sessions, sub-agents, and site deploys via Cloudflare Workers. A $20/month subscriber can clone a repo, install its dependencies, and drive live sites from a phone. The same open egress plus 223 registered tools and 44 skills lands it inside Willison's lethal trifecta — private data, untrusted content, and a channel back to an attacker.
simonwillison.net → - 08Nissan and Honda to co-develop core ECUs and software for fiscal 2029 vehicles
Nissan and Honda will jointly develop standardized core electronic control units and software for software-defined vehicles arriving in fiscal 2029, the first concrete output of talks that started in 2024. Japan's second- and third-largest automakers are now pooling software scale against Toyota, which is pursuing the same shift separately with Subaru and Waymo, mirroring the shared-OS bloc Volkswagen, BMW, Mercedes-Benz and Stellantis formed in Europe.
bnnbloomberg.ca → - 09US restricts Chinese drones and humanoids as five Chinese makers take 86% of shipments
Washington tightened rules on foreign-made advanced robots and set drone tariffs effective in September, with component tariffs in 2027, while the FCC's Covered List expanded to foreign drones and advanced robots. The scale gap the restrictions do not touch: of roughly 22,000 humanoid robots shipped globally in the first half of 2026, five Chinese makers accounted for 86 percent. The likely outcome is a fragmented market rather than a clean split — Chinese firms competing on cost across Europe, Southeast Asia, and Latin America, US and allied makers holding security-sensitive segments, and Japanese, Korean, and Taiwanese manufacturers positioned in between.
techcrunch.com →