Tech News Roundup: Anthropic's $80B Loss, OpenAI Lands on Google Cloud, eBay Cuts 9% of Staff & More (Oct 21, 2026)

Anthropic (image via Reuters / BNN Bloomberg)

Wednesday's tech news starts with a peek at Anthropic's books: its draft IPO filing shows an $80.7 billion first-half loss, though most of that is a paper charge tied to convertible bonds. Elsewhere in AI, OpenAI is bringing its models to Google Cloud, eBay is cutting about 9% of its staff, Nvidia unveiled AI data centers that ship on a truck, Super Micro blew past its own sales forecast, Samsung and SK Hynix are upgrading their China memory plants, and ChatGPT ads are expanding to 26 more countries. Plus: a nationwide Orange mobile outage in France, a deadly accident at TSMC's 2nm fab site, Microsoft trimming Copilot from File Explorer, AMD and Intel teaming up on PyTorch, Meta's army of spare GPUs, a SafePay ransomware warning, Zeabur shutting down after a second breach, and JustWatch launching its own streaming service. Here's the Lanbrew tech roundup for October 21, 2026.

Anthropic's IPO filing shows an $80.7 billion first-half loss

Anthropic's draft IPO prospectus, reviewed by Reuters, shows the Claude maker lost $80.7 billion in the first half of 2026, up from a $13 billion loss a year earlier. That number looks wild, but most of it isn't money going out the door: about $72 billion is an accounting charge tied to the rising estimated value of bonds that can convert into Anthropic shares. As the company's valuation has climbed, those notes have become worth a lot more on paper.

The real spending is still huge. The S-1 shows Anthropic spent nearly $9 billion on computing capacity and infrastructure in the first half, up from $2.3 billion a year earlier, and it has committed to at least $518 billion over a decade for cloud, compute and infrastructure. The company told investors last week that third-quarter revenue rose 45% to $16.8 billion, and it's expected to go public in November. Reuters says some investors are now weighing a valuation of around $1.5 trillion, and many see the listing as a test for the wider AI market ahead of OpenAI's own possible IPO next year.

Source: Reuters (via BNN Bloomberg)

OpenAI (image via CNBC)

OpenAI's models are coming to Google Cloud

OpenAI is bringing its AI models to Google Cloud, which means its GPT models will now be sold through all three of the biggest cloud providers. In the coming weeks, developers will be able to pick OpenAI models from Google's Vertex AI Model Garden and use them in Google's Antigravity coding tool. It's a notable team-up between two companies that compete head-to-head with ChatGPT and Gemini.

The deal is part of OpenAI's push to sell more to businesses. Microsoft's exclusive license on OpenAI's technology ended less than six months ago, and Amazon announced plans to host OpenAI models the very next day. Google's cloud business has been on a tear, with growth speeding up to about 82% in the second quarter, and Google, Amazon and Microsoft all report earnings next week.

Source: CNBC

eBay (image via Reuters / BNN Bloomberg)

eBay is laying off about 9% of its staff

eBay will cut about 9% of its workforce, or roughly 1,100 jobs, according to an internal memo from CEO Jamie Iannone that Reuters reviewed and eBay confirmed. Iannone told staff the goal is to simplify operations and move resources to the company's fastest-growing areas, and that affected employees will hear in the coming days. It's eBay's second round of cuts this year, after about 800 jobs went in February.

The cuts come while eBay is actually doing well. Its push into authenticated luxury goods, collectibles and refurbished items, plus its $1.4 billion purchase of resale app Depop, has helped its stock climb nearly 40% this year. "Our strategy is working," Iannone wrote, but he said capturing what's next means making "thoughtful choices" about where to invest. eBay reports earnings next week.

Source: Reuters (via BNN Bloomberg)

Nvidia at GTC in Germany (image via CNBC)

Nvidia will ship AI data centers in a box

At GTC in Berlin, Nvidia announced the DSX Pod, a shipping-container-sized module that packs seven Vera Rubin NVL72 racks (504 GPUs and 252 Vera CPUs) with all the cabling and liquid cooling already installed at the factory. Each pod weighs about 28 metric tons and is built to travel by truck on normal roads. Once it arrives, it just needs to be placed, hooked up to power, networking and cooling, and tested.

Nvidia says the approach can cut the time it takes for a new AI data center to start serving tokens from about six months to as little as two weeks, because everything is checked before it leaves the factory. Partners including Foxconn, HPE, Dell, Lenovo, Gigabyte, Quanta, Pegatron and Schneider Electric will build them, and deliveries are expected to start in early 2027. Nvidia didn't give a price, but analysts say a single Vera Rubin rack can run more than $7 million. The Register's take: they're basically AI trailer homes.

Source: CNBC; The Register

Super Micro (image via Reuters / CNA)

Super Micro's AI server sales blow past its own forecast

Super Micro now expects revenue of $18.5 billion to $19.5 billion for the quarter that ended September 30, way above its earlier guidance of $14.5 billion to $15.5 billion and analysts' estimate of about $15.2 billion. The server maker credited stronger demand for its AI platforms, and it also raised its expected gross margin to 12% to 12.5%, up from 10.4% to 10.8%.

The stock jumped more than 11% on Wednesday, and rivals Dell and HPE rose too. It's a welcome bit of good news for AI hardware after last week's selloff over data center debt. Super Micro will report full results on November 3.

Source: Reuters (via CNA); Supermicro

Memory chips (image via Tom's Hardware)

Samsung and SK Hynix are upgrading their China memory plants

Samsung and SK Hynix are refitting their factories in China to make more advanced NAND flash memory, Reuters reports, citing four people familiar with the plans. Samsung's plant in Xi'an, which makes about 40% of the company's NAND, is moving from roughly 200-layer chips to 286-layer V9 NAND under a project nicknamed "Monami," with output targeted at 5,000 wafers a month this December and 40,000 by December 2027. SK Hynix wants to move its Dalian plant, the former Intel NAND site that now makes chips for Solidigm, from 192 layers to 240 or more.

More layers mean more storage per chip and lower costs, which matters as AI data centers soak up huge amounts of SSD storage. Both companies confirmed the upgrades and said they're working closely with US officials, since they need yearly US licenses to install some US-made equipment at their China plants. Samsung said it isn't adding capacity or building new factories. Tom's Hardware notes the moves are partly about holding off fast-growing Chinese memory makers.

Source: Reuters (via CNBC); Tom's Hardware

ChatGPT ads (image via The Next Web)

ChatGPT ads are expanding to 26 more countries

OpenAI is rolling out ads in ChatGPT to 26 more countries, bringing the total to 89. The new markets include 14 more Latin American countries such as Argentina, Chile, Colombia and Peru, plus places like Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan. OpenAI says more than 100,000 advertisers have run campaigns since ads launched a little over eight months ago, up from about 20,000 in late August, according to The Information.

Ads only show up on the Free and Go plans, while Plus, Pro and Enterprise stay ad-free. OpenAI says ads are clearly labeled, kept separate from answers, and don't change what ChatGPT says, and that it doesn't share conversations with advertisers or sell user data. The ad business reportedly hit a $1 billion annual run rate in August, and OpenAI says more ad formats and measurement tools are coming next.

Source: The Next Web; OpenAI

Orange's headquarters near Paris (image via The Next Web)

A nationwide Orange outage knocked out mobile service in France

Orange, France's largest telecom operator, lost much of its mobile network on Wednesday. Starting around 11:20am Paris time (4:20am CT), customers across the country couldn't make calls, send texts or use mobile data, and its budget brand Sosh and some 5G home internet boxes went down too. Downdetector logged around 12,000 reports at midday, compared with about nine on a normal day, with problems in Paris, Lyon, Marseille, Toulouse, Lille, Bordeaux and Strasbourg.

Orange told Reuters its technical teams were working to find the cause and restore service, but it hadn't named one by early afternoon. Most fiber lines kept working, so many people could still call and message over Wi-Fi, and some phones started getting signal back after about an hour. It's at least the third major Orange mobile outage this year, and it came a day after the Google Cloud power failure in London, though nothing links the two.

Source: The Next Web; Reuters (via CNA)

TSMC (image via Tom's Hardware)

Three workers killed in an accident at TSMC's 2nm fab construction site

Three contract workers died and another was hospitalized in intensive care after a fall at the construction site for TSMC's 2nm Fab 22 in Kaohsiung, Taiwan, Tom's Hardware reports, citing Taiwan's Central News Agency. Investigators at the Southern Taiwan Science Park Bureau suspect a broken bolt on a steel connection plate sent the four workers and steel parts falling during hoisting work on the site's fourth phase.

Steel hoisting at the site has been halted while labor inspectors investigate. Local reports count at least half a dozen other fatal accidents at TSMC construction sites in Taiwan since September 2024.

Source: Tom's Hardware

Copilot in File Explorer (image via Wccftech)

Microsoft is pulling eight Copilot actions out of File Explorer

Microsoft is trimming Copilot from Windows 11's File Explorer. Eight AI actions in the right-click menu for OneDrive files are going away: Summarize, Create FAQ, Compare Files, Ask a question, Explain image, Explain images, Compare images and Extract text. The removal starts rolling out in mid-November and should be done worldwide by early December.

The features aren't fully gone. You can still reach Copilot from the "Ask Copilot" menu entry, the Copilot icon in File Explorer Home, or the OneDrive activity center. It's the latest step back from Microsoft's AI-everywhere approach in Windows, after it already removed Copilot buttons from Snipping Tool and Photos and renamed Copilot in Notepad to "Writing Tools."

Source: Wccftech

AMD and Intel (image via Tom's Hardware)

AMD and Intel team up on one PyTorch path for CPU AI

Longtime rivals AMD and Intel are merging their CPU tuning work for PyTorch. AMD is folding its ZenDNN library, which it maintained as its own version of Intel's oneDNN, into the open-source oneDNN project, so a single library will carry optimizations for both companies' chips. AMD is also joining the UXL Foundation, which publishes oneDNN, as a steering member.

The idea came from the x86 Ecosystem Advisory Group the two companies set up in 2024. For developers, it should mean one code path for running AI inference on x86 CPUs instead of separate setups for AMD and Intel hardware.

Source: Tom's Hardware

Meta's Prometheus cluster (image via The Register)

Meta keeps thousands of spare GPUs waiting for others to fail

Meta shared how it keeps AI training running on Prometheus, its giant supercluster in New Albany, Ohio, and one trick is keeping a lot of hardware on the bench. In its Grace Blackwell racks, which have 72 GPUs per NVLink domain, only 64 are ever in use, and the other eight sit ready as hot spares. The Register says more than 11% of the cluster's GPUs are either dead or idle at any time, waiting to step in.

At a cluster this size, something is always breaking. Meta says that early on it saw about 1,000 interruptions per day for every 8,000 GPUs, each forcing training to roll back to the last checkpoint. After a year of fixes, it's down 99%, to about 11 a day. Prometheus is planned to reach a gigawatt and about a million accelerators.

Source: The Register

SafePay ransomware (image via BleepingComputer)

CISA says SafePay ransomware has hit more than 600 critical infrastructure groups

CISA, the FBI and partner agencies in the US, Germany and the UK have put out a joint advisory on SafePay, one of the busiest ransomware gangs around. They say it has breached at least 600 organizations in critical infrastructure sectors across North America, South America, Europe and Australia since August 2024. The group steals data before encrypting systems and threatens to leak it if victims don't pay.

SafePay usually gets in through exposed credentials, brute-force password attacks, phished accounts and VPN or firewall flaws, and it tries to block recovery by deleting shadow copies and wiping backup storage. The agencies recommend multifactor authentication, tight limits on admin tools, and quick patching of edge devices, and they warn that paying a ransom doesn't guarantee files come back.

Source: BleepingComputer

Zeabur (image via BleepingComputer)

Zeabur is shutting down after its second breach in two months

Developer hosting platform Zeabur is shutting down after another serious security breach. On October 14 and 15, an attacker used leaked internal credentials, which had been stored in plain text, to break into Zeabur's shared clusters, some dedicated servers and internal systems, and then read its main database. The company says the attack touched clusters in nine regions, 321,810 accounts and about 3.2 million environment-variable records, which often hold API keys and passwords.

It follows a breach in late August, and Zeabur thinks the attacker probably got the credentials then. It admits its admin tokens never expired, hadn't been changed since January 2024, and were missed in the post-August cleanup. Zeabur users should rotate any secrets they stored there, revoke active sessions and rebuild affected servers.

Source: BleepingComputer

JustWatch TV (image via Ars Technica)

JustWatch, the streaming guide, launches its own streaming service

JustWatch, which has helped people figure out where to stream movies since 2015, now has its own service. JustWatch TV offers some movies free with ads that aren't free anywhere else, like Green Room, Locke and Sing Sing, and you can watch without making an account. If you'd rather skip ads, it costs $4 a month, or $50 for a lifetime pass for a limited time.

The library is small for now, a few hundred titles from about 30 partners including Paramount and New Regency, according to the company. JustWatch says the service will also feed its advertising business, which uses audience data to plan movie marketing campaigns for studios and streamers.

Source: Ars Technica