If you’re of a certain age, you’ll remember the way the web used to be: sites made and developed by humans.

You could load up a directory like AltaVista or Yahoo and click through categories including Home & Family, Science, and Computers. In those listings you’d find pages put together by hobbyists, academics, and tech nerds.

People gathered around forums and bulletin boards, and web developers tended to avoid putting images on their sites because it would slow down page loading. The idea of watching a video over the internet seemed almost impossible.

In the decades since those formative years, the web has transformed almost beyond recognition. Almost everything is now monetized to the nth degree, from search results and forums to social media and news outlets. Powerful apps and tools now run inside browsers, while AI is hard at work removing the need for the open web altogether.

There are now a good number of companies that are keen to see the spirit of the old web return, and that includes Kagi. For several years Kagi has been running a search engine based on privacy and security. Now Kagi has expanded its efforts beyond its Google-alternative search tool to curate a “small web” of sites driven and written by humans.

The Kagi Search Engine

As Kagi’s main business is its search engine, it makes sense to start here. Searching with Kagi will cost you a small fee (starting at $5 a month, plus sales tax) for 300 searches. The standard-level AI assistant is included in that price; it’ll help you make sense of what you’re seeing. You get unlimited searches and more AI assistant tokens for $10 a month. If you want to see what Kagi is like first before committing, there’s a free trial available that allows you 100 searches.

The argument that Kagi puts forward for why you should pay for search is pretty persuasive. As Kagi explains it, searching the web is always going to cost you—it’s just a question of whether you pay directly with dollars, by giving up data about your online activity, or by sifting through an increasing number of ads and sponsored links.

Besides making search more private, Kagi also wants to make it better by serving up results that aren’t influenced by paid promotions or whatever Google’s favored business practices happen to be from month to month. Kagi has its own indexes for web and news, and is built to promote “relevant, high quality results” and noncommercial sites.

To get started with Kagi search, all you need is an email address and password to manage your account. You can then get started with your 100 free searches. Open the search options panel to limit results to images, videos, podcasts, maps or news, or to filter by time or language.

For me, it feels a little bit like stepping back in time. Searching for a band, for example, brings up its official website, Wikipedia page, social media accounts, and YouTube channel—all free of promoted results, related searches, and sponsored content (such as concert tickets). It’s more the way search results used to appear.

Kagi’s Small Web Project

There’s now more to Kagi than search. The company offers access to a multitude of AI models (of course), and is busy building out its Small Web project, which aims to highlight the best of the human-made and human-curated internet. The company recently launched apps for Android and iOS too.

The idea is to avoid a web “dominated by algorithms, ads, and AI-generated noise,” in Kagi’s words. It’s a little like the old StumbleUpon, if you remember that far back. There’s a Next Post button for jumping to something else at random, allowing you to favorite what you see or to search for something specific. There’s a Share button too, of course.

Tiny Vinyl now has a dedicated belt drive record player with Bluetooth, built in stereo speakers, and accessories for storing and displaying its growing collection of miniature records.

Last year, we covered Tiny Vinyl, the 4-inch record format that puts one song on each side and turns physical music into something closer to a collectible sonic snack. Apparently the snack now requires its own table setting.

The new Tiny Vinyl Player is a $49.99 record player designed exclusively for the company’s miniature vinyl format. It is available through Target in a black leatherette clamshell cabinet with built in stereo speakers, Bluetooth output, and a wired audio connection for external speakers or headphones.

Related Reviews:

Why Does Tiny Vinyl Need Its Own Player?

Strictly speaking, it does not.

Tiny Vinyl records spin at 33⅓ RPM and can play on many conventional manual turntables. The problem is that their grooves sit very close to the spindle. Automatic stop and return mechanisms can engage before the song is finished, while some tonearms simply cannot travel far enough toward the center of the platter.

The dedicated player removes that uncertainty and gives collectors a simple way to use the format without checking tonearm geometry or disabling automatic functions. It also makes Tiny Vinyl more approachable for younger listeners and casual collectors who do not already own a turntable.

Tiny Vinyl says customers began requesting a player almost as soon as the records reached the market. Target also encouraged the company to develop one after the retailer began carrying the format across its stores. According to the company’s founders, more than one million Tiny Vinyl records have now been pressed.

Perhaps the tiny ecosystem is becoming less tiny.

What Does the Tiny Vinyl Player Actually Play?

One format. No plot twist.

The Tiny Vinyl Player supports the company’s 4-inch records at 33⅓ RPM. It does not play conventional 7 inch singles, 10 inch records, 12 inch LPs, or standard 45 RPM releases.

Each Tiny Vinyl record holds up to four minutes of music per side and is issued as a sequentially numbered limited edition. The records use 15 grams of 100 percent bio attributed vinyl, compared with approximately 140 grams for a conventional LP.

That limited playing time means this is not a format for albums, live recordings, or the 17 minute version of anything. It is designed for two songs, collectible artwork, and fans who enjoy owning a physical object connected to a particular artist.

Tiny Vinyl’s official catalog now includes more than 90 released or identified records, most of which sell for $14.99 at Target. The lineup spans artists such as Rihanna, Ariana Grande, Bruno Mars, Chappell Roan, Britney Spears, The Rolling Stones, Black Sabbath, Blink 182, Jimi Hendrix, BTS, and John Williams.

Tiny Vinyl Player Features

The Tiny Vinyl Player is a compact belt drive design measuring approximately 10 inches wide, 5.5 inches tall, and 6.5 inches deep. It weighs 3.88 pounds and requires AC power because there is no internal battery.

Key features include:

• 33⅓ RPM playback
• Support for 4 inch Tiny Vinyl records
• Belt drive mechanism
• Built in stereo speakers
• Bluetooth output for compatible speakers and headphones
• Wired audio output
• Black leatherette clamshell cabinet
• Included power cord
• Compact 10.04 by 5.51 by 6.46 inch dimensions

The player is designed to work immediately out of the box, although records are sold separately.

Please Connect Better Powered Speakers

Nobody should expect room filling bass, meaningful stereo separation, or much dynamic impact from speakers squeezed into a $49.99 record player that is only slightly larger than a shoebox.

Tiny Vinyl acknowledges that the built in speakers are intended for convenient, casual listening rather than high quality playback. Two colleagues who have already purchased it described the sonic playback as “thin” and “lacking bass impact.”

Color me surprised. Not really.

We have not tested the Tiny Vinyl Player, so this is not a sonic verdict. It is, however, a very strong recommendation to connect the player to a better pair of powered speakers.

Use the wired audio output when possible, or connect through Bluetooth if convenience matters more. Even an affordable pair of powered bookshelf speakers should provide greater output, stronger bass, and a more convincing sense of scale than the internal speakers.

Will that transform a 4-inch record into an audiophile pressing? Of course not. But it should make the experience considerably more enjoyable, which is presumably why you bought a record that can hide behind a drink coaster in the first place.

Tiny Vinyl Crate

The $19.99 Tiny Vinyl Crate provides storage for up to 32 records. It uses a wood frame with black leatherette covering and a velveteen interior, and it is sized so the Tiny Vinyl Player can sit directly on top.

The crate can also accommodate compact discs and cassettes, should your miniature record collection fail to consume every available inch of storage.

Tiny Vinyl 2×2 Frame

For collectors more interested in displaying the artwork, the $14.99 Tiny Vinyl 2×2 Frame holds four record jackets behind a clear shatter resistant window.

The black wood frame measures approximately 11 by 11 inches, uses magnetic closures for easier access, and can be mounted on a wall or placed on a flat surface. The records themselves are not included.

The Bottom Line

The Tiny Vinyl Player is not pretending to replace a conventional turntable, and anyone expecting it to compete with a proper analog system has wandered into the wrong aisle at Target.

This is a compact, affordable player for collectors who already own Tiny Vinyl records, younger listeners discovering physical music, or anyone whose automatic turntable cannot reach the format’s unusually close inner grooves. It also makes a fun desk system or gift without requiring the space and expense of a full size turntable setup.

Owners of a compatible manual turntable do not need it. Listeners expecting full size LP performance should avoid it entirely.

For everyone else, the Tiny Vinyl Player completes an amusing and increasingly popular collectible ecosystem. Just connect it to better powered speakers. Tiny records are one thing. Tiny sound does not have to be part of the deal.

Price & Availability

The Tiny Vinyl Player and accessories are available exclusively through Target at the time of publication.

Related Reading:

Photo credit: NASA/JPL-Caltech
Radar signals from the joint US-India NISAR satellite have mapped a mountaintop in East Antarctica that forms a clear hummingbird shape once the data is processed into color. Scientists at the mission nicknamed the result “the hummingbird” almost as soon as the image appeared.

Nunatak Zaterjavshijsja is a granite mountain in East Antarctica’s Prince Charles Mountains that rises from the surrounding ice sheet approximately 40 kilometers southwest of Mount Newton. A glacier lumbers past this impediment on its journey northeast to the ocean. The force of the mountain causes the ice to strain and bend, eventually forming deep cracks in the surface known as crevasses.

Sale

LEGO Art The Milky Way Galaxy Building Set – Space Decor DIY Kits for Adults for Bedroom wall, Ages…

• Milky Way wall art décor – Create a large Milky Way Galaxy piece of art with this LEGO building kit for adults
• LEGO 3D-effect wall décor – A Milky Way Galaxy artwork created from 3,091 layered LEGO bricks and pieces to produce a 3D effect with depth and…
• Packed with features – Wall art includes some of the Milky Way’s most famous stars, clusters and other spectacles, including Trappist-1, The…

NISAR’s L-band radar collected observations in August 2025, when the satellite was still being tested. It operates by transmitting microwaves towards the surface, which have horizontal polarization, and recording what comes back. Surfaces that bounce the signal back with the same direction as it was sent appear in magenta, which is the smoothest type of ice. Anything that returns with a changed polarization of the signal or scatters after entering the ice registers as green. These greens distinguish the faces of crevasses and other irregular structures. Where both types of signals come back strongly, you get white.

When the finished image is viewed, the peak forms a bright, oval center with a darker section smack in the middle, similar to an eye. Fine green lines radiate from the top and bottom of the core in patterns resembling feathers or wings. A long thin green chunk extends away from the oval, giving it a pointed beak appearance. The ice stream that continues to go northeast becomes the tail of the whole thing. Magenta fields surround the smoother areas of the fractured zone, while the deepest cracks light the brightest green.

Seongsu Jeong, a signal analysis engineer at NASA’s Jet Propulsion Lab who produced the image, explained that the radar can look past the snow and into the ice, revealing features that optical sensors cannot detect. A Landsat 9 image of the same region obtained in November 2025 is almost entirely white, with just weak traces of a mountain existing.

NISAR is equipped with both L-band and S-band radars, as well as a 12 meter mesh reflector, the largest radar antenna ever flown by the US. Every 12 days, the satellite covers nearly the entire surface of the Earth, including land and ice. The mission’s processed L-band data was made available to the public on July 20, 2026, only a few days before the first anniversary of its launch from India’s Satish Dhawan Space Center on July 30, 2025.

The beak shape was a complete coincidence, but the detail behind it is extremely valuable. The green fracture lines indicate exactly where the glacier is under stress and how the ice moves around the nunatak. Scientists may use this knowledge to investigate how glaciers move, the dynamics of ice sheets, and how Antarctic ice responds to impediments well inland from the shore. The same radar approach that transformed a single peak into a bird shape is already demonstrating the ability to provide repeated, all-weather data required to track these changes over time.
[Source]

Tour de France 2026 continues today with Stage 19, the first of two successive summit finishes at the iconic Alpe d’Huez. Just the name conjors images of legendary Tour moments of years gone by, from Fausto Coppi’s incredible climb to victory back in 1952, to Tom Pidcock’s seat-of-your-pants descent 70 years later. So good that they included it twice in this year’s route, the riders will have to do it all again tomorrow (Via a different route) on the day before the finish.

Today’s stage is a 127.9km ride from Gap which includes 3,500m of elevation gain, starting with the Category Two Col Bayard, a 4.8km long climb with an average gradient of 7.2%. The riders get some respite with a short descent and a brief flat section, before taking on the Category One Col du Noyer (7.2km at 8.5%) which takes them to the summit. They will be far from finished, though, with 100km still to go which includes the Category Three Col d’Ornon (5.4km at 6.4%) before they tackle Alpe d’Huez with its familiar 21 hairpin bends.

At Samsung’s Galaxy Unpacked event, the phone-maker announced three new folding phones, including its first-ever Ultra foldable phone. The Galaxy Z Fold 8 Ultra is the true successor of last year’s Fold 7, while the Fold 8 enters with an all-new wide-fold design.

These devices are still priced high enough to put them out of reach for many, but they pack subtle design improvements that might make a big impact in daily life. I love the new anti-reflective coating, easier-to-unfold design and battery upgrades, but the Galaxy Z Fold 7 remains a powerful and capable device – especially if you can get it at a rare discount.

The two book-style foldables differ mainly in experiential upgrades rather than spec bumps. It might be hard to get an idea of what’s exactly new. That’s why I put them side by side during my hands-on to find out.

Design and display

The Samsung Galaxy Z Fold 8 Ultra is roughly the same size as the already thin-and-light Fold 7. It weighs the same but is a millimeter slimmer when unfolded. Both phones are comfortable to hold, but the new model is easier to open.

Samsung has made the Ultra’s sides slightly slimmer so its bezels don’t run all the way to the corners. This gives you a slight gap between the sides to make it easier to unfold. In comparison, the Fold 7 has flat sides with zero gap when the phone is folded shut, making it more difficult to unfold.

Both phones have a 6.5-inch AMOLED cover screen and an 8-inch AMOLED folding display with a 120Hz variable refresh rate. The Galaxy Z Fold 8 Ultra’s inner screen is brighter at 3,000 nits (versus 2,600 nits on the Fold 7). It also has a matte finish, along with my long-requested feature: an anti-reflective coating for better legibility in harsh lighting conditions and outdoors.

I held both phones indoors in direct light, and the difference was noticeable. My Fold 7 was more susceptible to glare and harder to see under direct light when compared to the new model. 

Samsung has also improved the crease situation. It is minimal and can’t be seen unless you’re specifically looking for it. The phone-maker has added a new display structure that combines titanium-alloy film with an enhanced titanium plate to strengthen the display and balance the “hinge mechanics, display tension and magnetic force to make unfolding feel smoother, lighter and more natural,” says Samsung.

Both foldables are IP48 rated for dust and water resistance. So, they aren’t truly dust-tight like the Google Pixel 10 Pro Fold and Honor Magic V6, but they can survive being dunked in water for at least 30 minutes (the “8” in IP48). But they won’t protect against dust particles smaller than 1mm (the “4” in that rating).

Performance and battery

The Galaxy Z Fold 8 Ultra is powered by the Qualcomm Snapdragon 8 Elite Gen 5 for Galaxy chip, as compared with the Fold 7’s Snapdragon 8 Elite processor. Both phones come in three storage variants: 256GB, 512GB and 1TB, with the top-end model getting 16GB of RAM (the 256GB and 512GB models are again paired with 12GB of RAM).

Samsung says it has created more space in the internals for better heat dissipation, so you can theoretically expect better thermal management.

As for the battery, this is the first time Samsung is using the silicon-carbon anode technology. It has allowed the phone-maker to upgrade the cell capacity from 4,400 mAh to 5,000 mAh on the new model. The Ultra also supports faster wired charging at 45 watts (the Fold 7 offered 25 watts).

On paper, it is still behind Chinese rivals like Oppo and Honor, which include over 6,000-mAh batteries in their respective folding phones, but I’m hoping this spec bump will help me get through a day with ease.

Both phones have Qi wireless charging, but no inbuilt magnets. You’ll have to buy a separate magnetic case to add compatibility with Qi2 wireless chargers and other magnetic accessories.

Camera, software and AI

The Galaxy Z Fold 8 Ultra’s camera array is mostly unchanged from last year’s model, with one exception: the ultrawide sensor. You get a 200-megapixel main camera and a 10-megapixel telephoto sensor with 3x optical zoom on both phones. However, the 12-megapixel ultrawide camera has been upgraded to a 50-megapixel sensor. You can expect higher-quality images, but I’ll wait for CNET’s tests to identify the real differences.

Both foldables shoot video in 8K at 30 frames per second, but the Fold might appeal to more creatives. The new model adds support for the APV codec and adds a Cine LUT to give cinematic color and tone control directly from your phone. This adds greater flexibility for editing and production workflows for professional creators.

The Galaxy Z Fold 7 launched with Android 16 onboard last year, and the Fold 8 Ultra is coming with Android 17 (with One UI 9 on top). You get a similar software experience on both phones, and the Z Fold 7 will get the update to Android 17 with One UI 9.

The two devices support the full suite of Samsung’s Galaxy AI features alongside Gemini access. There’s AI Select for on-screen search support, an object eraser tool in the Gallery app and contextual AI features like Now Nudge for first-party apps. The latter displays contextual actions based on on-screen information. Like, if you’re having a chat in the Messages app, and they send you an invite with the date, time and address, Samsung’s Now Nudge feature will show a pop-up to add it to Calendar.

The phone-maker also demoed a new agentic AI feature at my demo, which wasn’t very impressive. Ideally, the new Gemini feature would help book a hotel just by looking at an on-screen photo of the Eiffel Tower, but in its current state, it was unreliable. But since the Z Fold 8 Ultra gets seven years of software updates and the Z Fold 7 has six more years to go, there’s plenty of time for both phones to continue to improve on these features.

To see more differences between the Galaxy Z Fold 8 Ultra and the Z Fold 7, check out the specs chart below.

Samsung Galaxy Z Fold 8 Ultra vs. Z Fold 7: Specs and prices

Prakhar Khanna is a Writer at CNET with bylines in ZDNET, Forbes, and other major outlets. A digital nomad at heart, he thrives in chaos–mixing deadlines with travel, often writing stories on planes, trains, and cafe tables while hunting down the best pizza spots wherever he lands. His backpack usually carries one phone too many, and his playlists are stacked with 2000s Bollywood. Find him across social media @ParkyPrakhar.
Contact: parkydoesstuff(at)gmail(dot)com

See full bio

Advertisement

Recent coverage of Flock Safety, the fast‑growing US surveillance technology company, reveals how opposition to automated license plate readers is playing out at the local level.

In South Carolina, officials have been sparring over whether to install 25 new cameras, a dispute that reflects broader questions about cost, oversight, and how much surveillance communities are willing to accept. The city of Los Angeles recently broke its contract with Flock Safety, though Flock’s CEO argues that the suspension should be temporary while the police department revises its rules.

These cancellations are in large part driven by public concern around a system designed to enable mass surveillance, which is susceptible to serious abuses, including biased policing and privacy violations.

To track such surveillance, journalists and researchers are increasingly turning to a free tool called the Atlas of Surveillance, a collaboration between the renowned digital privacy group Electronic Frontier Foundation and the University of Nevada, Reno. The site has become a go‑to source for mapping police technology — which also includes drones, facial-recognition cameras and body cameras — with a searchable database that’s often more dependable than alternatives.

With the Atlas of Surveillance tool, you enter a city, county, state, law enforcement agency or vendor in the US and select which technology you’re interested in tracking. You’ll get a list of details based on your search, including links to public records or news articles.

“We frequently run into situations where local councilpeople don’t even know that a certain tool is in use,” said Beryl Lipton, Electronic Frontier Foundation senior investigative researcher. “Learning about a technology we have logged in the Atlas of Surveillance is meant to help people have conversations [about] whether that tool is appropriate for local use, start asking questions, and advocate for standards around how their police can or can’t use this equipment.”

Surveilling surveillance

While other tools for tracking Flock cameras and similar surveillance systems exist, I prefer the EFF’s approach, which uses reliable crowdsourced data from specific volunteers, including University of Nevada students and staff, as well as experienced researchers. These volunteers use a variety of information, including government documents, press releases, news stories and confirmed social media posts, to create lists of exactly which surveillance technologies local police or sheriff departments are using, as well as when they were adopted.

Deflock, another popular crowdsourced app for reporting Flock camera locations, has an open-source design that leaves it vulnerable to false reports from people who only think they’re seeing a surveillance device. Self-reports are also becoming less reliable as surveillance cameras are being hidden in speed signs and cacti.

I tried the Atlas tool with several location searches, including Albany, Georgia, where five police officers were fired and arrested in July on charges of misusing the city’s Flock system. Sure enough, I could see exactly when Albany began its Flock program, among other details about the technology it uses.

“While we may know that a police department has a tool, we may not know that they don’t have a governing policy or what is in the policy they have, like how long the data is retained,” Lipton told me.

Lipton recommends that users pair the Atlas tool with the EFF’s Street-Level Surveillance hub, which offers deeper documentation, including details on third‑party vendors that collaborate with law enforcement.

The Atlas database isn’t perfect, and the volunteers can do only so much. In my own searches, I found that my home city offered accurate listings for traffic cameras, but didn’t mention the period when it adopted Flock automated license plate readers last year, before removing them after public complaint. I’ve seen similar patterns in other cities with comparable histories, suggesting the database may emphasize active programs or exclude portions of older records.

Resources like the Atlas of Surveillance give residents a clearer view of the technologies their police departments use, information that can otherwise be difficult to obtain. As I’ve covered before, some cities will carefully omit the name “Flock” when announcing new surveillance or monitoring programs, even if they are using the brand’s drones or ALPR cameras.

If you’d like to volunteer for the Atlas of Surveillance, you can select the Collaborate option to submit a datapoint for consideration.

Tyler Lacoma

Editor / Home Security and Smart Home

Tyler has worked on, lived with and tested all types of smart home and security technology for over a dozen years, explaining the latest features, privacy tricks, and top recommendations.

With degrees in Business Management, Literature and Technical Writing, Tyler takes every opportunity to play with the latest AI technology, push smart devices to their limits and occasionally throw cameras off his roof, all to find the best devices to trust in your life. He always checks with the renters (and pets) in his life to see what smart products can work for everyone, in every living situation.

Living in beautiful Bend, Oregon gives Tyler plenty of opportunities to test the latest tech in every kind of weather and temperature. But when not at work, he can be found hiking the trails, trying out a new food recipe for his loved ones, keeping up on his favorite reading, or gaming with good friends.

See full bio

Advertisement

Scientists have been trying to figure out how kids pick up language so fast for decades, and a new study out of the Okinawa Institute of Science and Technology (OIST) might have cracked part of the puzzle: curiosity.

Researchers built a virtual robot with a brain-inspired neural network and set it loose in a simulated 3D world full of shapes, colors, and simple commands like “push left magenta dumbbell.” 

Some robots were rewarded only for completing tasks correctly. Others got an extra reward for curiosity, essentially getting a little internal high whenever they encountered something that challenged their existing understanding of the world.

How does curiosity help a robot learn language?

The curious robots didn’t just edge out their indifferent counterparts; they blew past them. According to the study, published in Science Advances, curious robots reached a genuine understanding of language in about half the time. 

Study author Theodore Tinker compared it to trying white chocolate for the first time even though you already love dark chocolate. You take the risk anyway, and you walk away knowing more about chocolate in general.

Things got even more interesting halfway through training. The curious robots started knocking things over and experimenting with actions nobody asked for, basically playing. Nobody programmed that behavior in. It just showed up on its own.

Do robots really make the same mistakes as toddlers?

The robots also mimicked a well-known quirk in how children learn language. Kids often get certain verb forms right at first, then start applying grammar rules too broadly and make mistakes on verbs they’d previously used correctly, before eventually sorting out the exceptions and correcting themselves. The robots followed the same U-shaped dip in performance.

It’s also a nice contrast to how today’s chatbots learn. Large language models like ChatGPT train on massive datasets and spit out the statistically likely next word. This robot’s brain works more like ours, prioritizing accuracy while trying to keep its beliefs intact, only updating them when something surprises it enough to be worth the trouble.

None of this means robots understand language the way we do. But it does suggest that curiosity paired with a wide variety of experiences might be a big part of how toddlers crack the language code with so little to go on.

AMD and Anthropic announced a strategic partnership on Tuesday that commits AMD to invest up to $5 billion in Anthropic and deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in Helios rackscale solutions to run Claude. Deployment of the first gigawatt begins in the first half of 2027. Anthropic is already using AMD’s MI355X GPUs and will now scale to MI455X accelerators paired with EPYC “Venice” CPUs and Pensando networking.

The engineering collaboration may matter more than the hardware. AMD and Anthropic will use Claude to optimise workloads for AMD Instinct GPUs and accelerate ROCm software development. AMD will also adopt Claude broadly across its engineering and product development teams. ROCm is AMD’s answer to Nvidia’s CUDA, and its software gap has been the primary reason AI developers default to Nvidia hardware even when AMD’s specs are competitive. If Claude can materially improve ROCm’s developer experience, AMD addresses the problem that has held it back for years, using its customer’s AI to fix its own software.

Anthropic’s compute strategy is now genuinely multi-vendor. Anthropic signed its biggest compute deal with Google and Broadcom, and has separate arrangements with Amazon (Trainium chips, 5GW), CoreWeave (Nvidia GPUs), and SpaceX (Colossus data centres). Adding 2GW of AMD Helios gives Anthropic what Tom Brown, its chief compute officer, called the ability to “map the right workloads to the right hardware.” Diversifying away from any single chip vendor reduces dependency risk at a time when compute is the binding constraint on frontier model development.

For AMD, the $5 billion equity investment mirrors Nvidia’s playbook. Nvidia invested $2 billion in Nebius and has taken stakes in multiple AI infrastructure companies to lock in hardware demand. AMD is doing the same: investing in a customer to guarantee that its chips, not Nvidia’s, run a meaningful share of the world’s most capable AI models. Lisa Su called it “a major platform for the next generation of AI infrastructure.” Whether Helios can compete with Nvidia’s NVL72 at production scale is the question the first gigawatt will answer.