Tech News Roundup: Google's TPUs Land in Meta's Data Centers, Barclays Cuts 5,000 Jobs for AI, DeepSeek's $15B Raise & More (Oct 22, 2026)
Google TPU hardware (image via Tom's Hardware)
Thursday's tech news starts inside Meta's data centers, where Google's TPUs are now serving AI in production, a real crack in Nvidia's software lock. Barclays is cutting 5,000 jobs and crediting AI for part of it, DeepSeek is reportedly raising about $15 billion, a New Mexico court set a $25 billion penalty for Meta, and Jensen Huang says there's no memory glut coming. We've also got updates on iPhone Duo demand, PHP2Shell attacks that started before the patches, and Microsoft's findings on the ConnectWise breach. Plus: Railway and Quest Diagnostics breaches, a critical GitLab fix, App Store review delays, Walmart welcoming AI shopping agents, Hugging Face's DNA chatbot, and TSMC's search for rare-earth-free gear. Here's the Lanbrew tech roundup for October 22, 2026.
Google's TPUs are now running AI in Meta's data centers
Meta, the company behind the PyTorch AI framework, is now serving AI models in production on Google's tensor processing units. Google announced TorchTPU, a PyTorch-native way to run on its custom chips that it built together with Meta's PyTorch team, along with new TPU versions of the popular SGLang and vLLM inference engines. The big deal here is that most PyTorch models no longer need to be rewritten for Google's JAX framework to run on TPUs.
Meta's PyTorch engineering director said the company can't be tied to a single hardware architecture, which is a pretty clear signal aimed at Nvidia. Meta tried TPUs for research years ago before dropping them, and it reportedly signed a multi-year rental deal more recently. Early numbers look strong too: one test serving Moonshot's 2.8-trillion-parameter Kimi K3 on 16 Ironwood TPUs hit about 700 tokens per second for a single user, versus 452 on a comparable setup of 16 Nvidia GB200s. Amazon open-sourced its own PyTorch backend for Trainium a day earlier, so the push to loosen Nvidia's software grip is picking up.
Source: Tom's Hardware
Barclays (image via The Guardian)
Barclays will cut 5,000 jobs as AI makes staff "more productive"
Barclays is cutting about 5,000 jobs, roughly 5% of its 93,000 employees, as part of a plan to save more than £1.1 billion in 2027. The bank says about 60% of the cuts will hit support functions, including fewer software developers, because AI is making the people who remain more productive. About 1,000 jobs have already gone since January.
CEO C.S. Venkatakrishnan said most of the savings so far come from bigger technology investments and simpler processes, calling AI the "icing on the cake." The Unite union pushed back, saying the bank should be retraining workers instead. The news landed alongside strong results: third-quarter pre-tax profit rose 12% to £2.3 billion, and Barclays is returning another £500 million to shareholders. It follows Standard Chartered's plan to cut more than 7,000 jobs as it leans on AI.
Source: The Guardian
DeepSeek (image via TechCrunch)
DeepSeek is reportedly raising about $15 billion
Chinese AI lab DeepSeek has raised around $15 billion (about 100 billion yuan), one of the biggest funding rounds ever in China, according to a Bloomberg report cited by TechCrunch. Battery giant CATL and Tencent are among the potential investors. That's a big jump from June, when DeepSeek raised more than $7 billion at a $50 billion valuation, and the company has been linked to a possible IPO in China as early as next year.
DeepSeek is spending heavily on AI infrastructure and open models for markets at home and abroad. The report comes two days after US federal civilian agencies were told to remove DeepSeek's technology from their systems over security concerns. Rivals are raising big money too: Moonshot AI landed $2 billion in May, and Zhipu and MiniMax have been public in Hong Kong since January.
Source: TechCrunch
Meta (image via CNBC)
A New Mexico court orders Meta to pay $25 billion in its privacy case
A New Mexico judge has ordered Meta to pay $25 billion in penalties in a data privacy case that traces back to the Cambridge Analytica scandal. In September, a Santa Fe jury found Meta liable under the state's Unfair Practices Act for misleading residents about how it collects and protects user data, and counted more than 43 million violations. Each one could carry up to $5,000, which put the theoretical total at $219.5 billion before the court set the final figure at $25 billion.
The judge's order also adds 15% annual interest until the penalty is paid. Meta says it is confident it will win on appeal, and its chief legal officer called the ruling a barrier to doing business in the state.
Source: CNBC
Nvidia CEO Jensen Huang (image via CNBC)
Jensen Huang says there's no memory glut coming
Memory chip stocks have taken a beating this past week over fears of a supply flood by 2029, but Nvidia CEO Jensen Huang isn't buying it. "You can always use more" memory, he told CNBC at GTC Berlin, pointing out that Nvidia already puts phone-style LPDDR memory into data centers on top of the HBM stacked next to its GPUs.
The selloff was sparked by a South Korean column warning that the memory "supercycle" could end in 2029, plus a new chip design from China's CXMT that could let it grab share from SK Hynix, Samsung and Micron. Asked whether Nvidia might buy CXMT memory, Huang said he's "open-minded," noting that nearly every company relies on Chinese suppliers somewhere in its supply chain.
Source: CNBC
iPhone Duo (image via The Verge)
Update: Apple reportedly boosts iPhone Duo orders by 40% and adds a "Made for iPhone Duo" badge
What changed: last week, slipping iPhone Duo delivery dates left it unclear whether Apple was dealing with heavy demand or tight supply. A DigiTimes supply-chain report, covered by 9to5Mac, now says Apple has raised its first-year shipment target from 9.7 million to 14 million units, a jump of more than 40%. That lines up with some buyers seeing their delivery estimates move earlier.
Apple is also getting the App Store ready for Friday's launch. A new "Made for iPhone Duo" badge will mark apps built for the foldable's two screens. Netflix can fill the inner display with a movie while showing clips outside, Prime Video works half-folded, and Word, Outlook and Zoom get side-by-side layouts. The iPhone Duo starts at $1,999 and arrives Friday, with limited walk-in stock.
PHP2Shell (image via BleepingComputer)
Update: PHP2Shell attacks started four days before the patches
What changed: Lanbrew covered the PHP2Shell patches and the site outages they caused, and now we know attackers were ahead of everyone. Amazon Threat Intelligence says its honeypots caught the first attempts to exploit the two critical PHP flaws on October 17, four days before fixes came out on October 21. The PHP team says it had to move up its release after details, including the CVE numbers, leaked early.
Since then the attacks have exploded. Cloudflare CEO Matthew Prince says exploit attempts went from thousands in the hours before disclosure to hundreds of thousands within two hours and are now in the millions, with attackers tweaking payloads to slip past firewall rules. If you run PHP, patching is the only real fix, and the UK's NCSC is telling organizations to update or take servers offline.
Source: BleepingComputer; The Register
Storm-0940 (image via The Register)
Update: Microsoft says the ConnectWise breach opened the door to 2,250 Microsoft 365 tenants
What changed: when ConnectWise pulled its Cloud Backup service last week, nobody had said who was behind the breach. Microsoft now attributes it to Storm-0940, a China-linked espionage group active since at least 2021. Using stolen ConnectWise application credentials, the attackers tried to get into more than 2,250 Microsoft 365 tenants, at times hitting 2,900 attempts a minute, and didn't need anyone to log in interactively.
The new twist is AI. Microsoft says the group used ARTEX, an open-source penetration-testing framework that uses AI agents to automate everything from scouting to credential attacks, and in some cases went from choosing a target to privileged cloud access in under three hours. An earlier password-spraying campaign of more than 151,000 attempts failed before the group switched to the ConnectWise identities.
Source: The Register
Railway breach (image via TechCrunch / Getty)
Railway tells every user to rotate their secret keys after a breach
App hosting platform Railway says hackers may have accessed encrypted environment variables, and in some cases got the decrypted contents. Those variables usually hold passwords and API keys that apps use to reach databases and cloud services, so anyone holding them could log in as the owner. Railway thinks only a subset of users was affected but is telling everyone to revoke and rotate their keys, and it paused customer deployments while it investigates.
Railway hasn't said how many people were hit or how the attackers got in, but it believes the breach was part of a wider campaign aimed at compromising AI inference providers. AI coding tool OpenCode reported a similar theft of customer secrets earlier this week.
Source: TechCrunch
GitLab (image via BleepingComputer)
GitLab patches a max-severity flaw that lets attackers run code
GitLab has released fixes for CVE-2026-110264, a maximum-severity bug that could let an attacker with no login run code on a GitLab server. The problem is an integer overflow in GitLab's regular expression compiler, which can be triggered with a specially crafted regex. It was reported through GitLab's HackerOne bug bounty program.
Patched versions include 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10 and 18.11.13 for both Community and Enterprise editions. GitLab.com is already fixed, but self-hosted admins should upgrade now. Shadowserver counts more than 2,700 GitLab Community Edition instances exposed online.
Source: BleepingComputer
Quest Diagnostics says hackers accessed patient records
Quest Diagnostics, one of the biggest medical lab companies in the US, says someone got unauthorized access to Quanum LSM, the system some clients use to order tests and look up results. Quest found out on Wednesday and believes certain patient records were accessed, though it's still working out how far the breach went. It has called in outside security experts and notified federal law enforcement.
The company says the incident hasn't seriously disrupted operations or patient services. It's the latest in a string of attacks on healthcare companies, which have recently hit device makers such as Abbott, Stryker, Medtronic and Boston Scientific. Quest shares slipped in premarket trading even as its quarterly revenue of $3.10 billion beat estimates.
Source: Reuters (via Yahoo Finance)
Apple (image via The Register)
Developers say Apple's App Store reviews are stuck for weeks
More and more developers are complaining on Apple's forums and Reddit that their apps have been sitting in "Waiting for Review" for weeks, with some waiting since mid-September. Apple's own guidelines say it usually reviews half of submissions within a day and 90% within two days.
This has happened before. Developers reported multi-week delays back in February, sometimes for simple bug fixes, and earlier reports blamed a flood of AI-coded apps, with submissions said to be up 84% this spring. The Register says Apple hasn't responded to its questions yet. The timing is awkward with the iPhone Duo launching Friday and Apple pushing developers to optimize for it.
Source: The Register
Walmart (image via The Verge / Getty)
Walmart is opening its store to AI shopping agents
While Amazon has been blocking AI shopping agents, Walmart is inviting them in. Starting next month, you'll be able to point an AI agent at Walmart's site with guardrails: by default every purchase needs your approval, though you can set a budget, like up to $100. Each agent gets its own "Agent Profile," so you can review what it bought separately and shut it off at any time.
Walmart has written rules for agents too. They can't go beyond what the shopper asked for, can't change payment methods, and only see the information they need. Its new Agent Gateway, now in beta, connects agents through the Model Context Protocol, so they can find your usual items without seeing your full shopping history. Sam's Club is getting the same support.
Source: The Verge
Hugging Face Carbon-X (image via Wired)
Hugging Face's new open model lets you "chat with DNA"
Hugging Face has released Carbon-X, an open model that understands both genetic sequences and plain English and is small enough to run on a laptop. You can ask it how a DNA change might affect gene activity or health and get a conversational answer back, which Hugging Face hopes will help researchers working on drugs and gene therapies and maybe one day help patients make sense of their own genomes.
It was built with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy, starting from an existing language model taught to read DNA six letters at a time. Wired compares it to a ChatGPT for DNA: DeepMind's AlphaGenome gives numerical predictions and needs more compute, while Carbon-X answers in plain language. Nvidia agreed to buy Hugging Face for nearly $13 billion in September.
Source: Wired
TSMC (image via Tom's Hardware)
TSMC wants chipmaking gear without Chinese rare-earth magnets
TSMC has started asking its equipment suppliers whether they can provide factory automation machines with motors that don't need rare-earth magnets, according to a Reuters report covered by Tom's Hardware. The goal is to reduce its reliance on materials that China dominates and avoid disruptions to its fabs from Chinese export controls.
TSMC's highly automated fabs, which make chips for Nvidia, Apple, Broadcom, Qualcomm and many others, are full of machines with powerful permanent magnets made from elements like neodymium, dysprosium and terbium. Swapping in alternative motors would be a long process, but it shows how seriously chipmakers are taking supply-chain risk.
Source: Tom's Hardware
Comments
Post a Comment