The proliferation of custom AI ASICs, or XPUs, from OpenAI, Meta, Microsoft, and others has led many to question Nvidia’s grip on the market. After all, if everybody’s building their own, who needs the GPUs the AI arms dealer has made its fortune on?

But Nvidia isn’t concerned in the least and has instead invited its competition to raid – or, rather, license – the GPU giant’s IP holdings, particularly those related to networking.

On Monday, MediaTek became the latest to embrace Nvidia’s NVLink Fusion high-speed interconnect technology for its fledgling datacenter XPU offering. In exchange, Nvidia has invested $3.5 billion in convertible bonds issued by MediaTek. The GPU giant will also continue to license the Taiwanese SoC provider’s own designs for use in future DGX and RTX Spark systems.

If you’re curious, we explored MediaTek’s datacenter ambitions last month. But the designer, best known for its Arm-based smartphone and tablet processors, is only one of several high-profile chip designers that plan to integrate the tech. Amazon, Fujitsu, Qualcomm, Arm, and Marvell are all on the same list.

NVLink Fusion started as a commercialized version of Nvidia’s high-speed inter-GPU interconnect, introduced early last year. Initially, it was offered in two varieties: a chip-to-chip (C2C) variant for connecting CPUs to GPUs (or XPUs) in a memory coherent fabric, and a switched fabric used to stitch together multiple accelerators into one big logical rack-scale chip. It’s since expanded to become the blanket offering by which Nvidia licenses its semiconductor IP.

Given how long it’s taken Broadcom and others to develop competitive alternatives to NVLink, it’s not surprising companies like MediaTek would license the tech, rather than reinventing the wheel, or piping alternative interconnects, like UALink, over regular old Ethernet.

By licensing Nvidia’s NVLink Fusion tech, companies can focus on building competitive accelerators without worrying about how they’re going to scale in production.

And because Nvidia’s MGX rack designs are part of the Open Compute Project, MediaTek and its partners not only benefit from Nvidia’s scale-up networking tech, but also can essentially take existing NVL72 racks and slot in their compute blades. This should dramatically reduce the system design and mechanical engineering experience required to go from silicon to AI racks.

And this isn’t theoretical. Amazon is doing just that with its Trainium series of AI accelerators. As we wrote at the time, Amazon used Nvidia’s MGX NVL72 reference design for its Trainium3-based rack systems launched last year. Meanwhile Trainium4, expected late this year, will ditch Amazon’s in-house NeuronLink interconnect tech for NVLink Fusion.

Nvidia’s cut

So, what does Nvidia get from letting rival chip designers piggyback off its hard-won networking tech?

A lot more than licensing fees: It’s also a way to keep companies hooked on its other products.

When NVLink Fusion was first announced, customers had two options. They could pair their XPUs with Nvidia’s CPUs or their CPUs with Nvidia’s GPUs. Customers also were on the hook for NVSwitches if they wanted to take advantage of the tech’s scale up networking capabilities.

So, on top of licensing the tech, Nvidia had the potential to sell roughly two GPUs per partner CPU, a Grace or Vera CPU for every two partner XPUs, and up to nine NVSwitches for every 72 or so accelerators, and that’s just the rack.

If you’re already buying NVLink, why not use InfiniBand or Nvidia’s Spectrum-X Ethernet kit to switch those racks together to train and serve even larger models at scale?

Nvidia has since softened its stance and now seems content with the combination of NVSwitch sales and NVLink Fusion licensing. If it hadn’t, Amazon couldn’t have paired Graviton with Trainium.

“The beauty about this platform is it provides the flexibility and the fungibility in order to meet the customers where they are. And so, leveraging the NVLink Fusion platform, they can adopt anywhere from the CPU to the GPU to the scale-up fabric to the scale-out networking across our Spectrum-X and InfiniBand, as well as our DPU within the scale-in fabric to do the north-south storage management as well as security isolation,” Dion Harris, Nvidia’s senior director of HPC and AI Hyperscale Infrastructure, told press on Monday. “So, the five different networking technologies and topologies that we offer in our technology stack are all available to customers via NVLink Fusion.”

Nvidia isn’t quite ready to talk about the licensing structure just yet, but Harris says they won’t be giving it away.

“…It is safe to assume that there will be some licensing elements in place in that Nvidia NVLink Fusion is an Nvidia technology,” Harris said.

It’s not just about networking

Nvidia is already the largest datacenter Ethernet networking provider in the world, thanks in no small part to the AI boom. Its quarterly networking revenue now exceeds $15 billion, and NVLink Fusion licenses and associated NVSwitch sales will only bolster that. However, Nvidia’s licensing ambitions don’t stop there.

Last week, Nvidia detailed its latest IP offering in the form of NVHBM, essentially a custom base with an integrated memory controller.

According to Nvidia, the approach frees up 25 percent of the die for compute while cutting HBM power draw by 15 percent and boosting effective bandwidth by 30 percent.

Amazon’s Annapurna Labs will be among the first to field the tech with its Trainium family of AI ASICs.

But while Amazon’s adoption of NVHBM is a clear win for Nvidia, it should be noted that the idea isn’t unique to Nvidia. As we understand it, both Broadcom and Marvell are developing similar base-die tech as well. But the tech is another example of how Nvidia is leveraging its engineering might to extract revenues from everything associated with AI, including custom ASICs from others.

A deal with the devil or mutually beneficial coexistence

While Nvidia’s fledgling IP offerings are rather narrow for the moment, it’s not hard to imagine Nvidia competing with rivals like Broadcom or partners like Marvell and MediaTek on XPU design before long.

Depending on how things play out, Nvidia, Broadcom, Marvell, and MediaTek could end up looking more like frenemies.

XPUs are quickly becoming big business for both Broadcom and Marvell, but the networking required to scale that compute and system design required to bring it to market remains a competitive advantage for Nvidia.

Sure, both have competent Ethernet switching appliances capable of going toe-to-toe with NVSwitch, but even then you still have to integrate that into a rack. The advantage of being able to slot a custom compute blade into a generic NVL72 system and for it to work out of the box is hard to ignore.

If NVLink Fusion helps the XPU designers sell more chips, it could end up being a net benefit even if it means losing out on switch sales. Nvidia is clearly aware that while the barrier to entry for custom silicon has fallen through the floor, scaling that compute remains the hard part. ®

ai and ml

GenAI now reaches 80% of occupations, but in most of those fewer than half of workers use it

Economists have taken another stab at assessing the impact of generative AI adoption on different occupations and have found that adoption is broad but not very deep.

While generative AI reaches 80 percent of occupations and more than 40 percent of tasks, in most of these occupations fewer than half of workers adopt it.

Occupational use is high in professions where you’d expect – in management and professional occupations, particularly those that are related to finance, business, and computers. It’s lowest in personal service occupations and jobs requiring manual activity or interpersonal interaction.

That in and of itself isn’t surprising – other studies have suggested as much. But AI companies that have looked at occupational exposure to generative AI (Anthropic, Microsoft, OpenAI) based on chat logs come up with results that differ from those described in “What Work Does Generative AI Do?”

The paper [PDF], by authors Alexander Bick (Federal Reserve Bank of St. Louis), Adam Blandin (Vanderbilt), David J. Deming (Harvard), and Tyler R. Schumacher (Vanderbilt), measures how workers actually use generative AI for their jobs by analyzing Real-Time Population Survey (RPS) data.

The measurement discrepancy is the result, the authors say, of chat log classifiers that associate chats with a few tasks that have generic, activity-based descriptions like “Edit written material or documents.” 

The problem with this approach is that when OpenAI data concludes 15 percent of chats involve that sort of editing, those figures don’t match up with the US Department of Labor’s O*NET database, where only 2.4 percent of workers are in jobs that include that task.

The result, the researchers say, is that the exposure figures from vendors likely overstate how much generative AI is actually relevant to people’s jobs.

“GenAI is used in many occupations and tasks, but few of these exhibit very high adoption rates,” the authors state. “For example, four out of five detailed occupations have adoption rates above 20 percent, but only one out of six occupations exceed 70 percent adoption.”

Most occupations exhibit relatively low adoption rates, but for a few, the figure is substantial. About 15 percent of occupations had adoption rates above 70 percent (largely computer-oriented). And in terms of tasks, only 2.8 percent of tasks had adoption rates above 50 percent and none surpassed 70 percent.

Overall, as of May 2026, 55 percent of US adults between the ages of 18 and 64 used generative AI for non-work reasons and 45 percent of US adults used GenAI for work, according to the paper. The overall adoption rate (work or otherwise) is 62 percent.

One thing that drives AI adoption, they say, is experience. Those who start using it in one domain tend to adopt it in other domains.

Thus the researchers argue it’s at least as important to understand why some workers adopt AI and some don’t as it is to understand which tasks it can handle. ®

Actuators account for 40% to 60% of a humanoid robot’s bill of materials and China refines roughly 90% of the rare earth magnets inside them, which makes the components rather than the models the limit on scale. Europe has a 2030 cap on single country supply under the Critical Raw Materials Act, and one company, Schaeffler, building the actuators at volume.

The hardest problem in physical AI is not the AI. Actuators are in short supply from a small pool of qualified manufacturers, and new capacity takes years, not quarters, Nate Evans wrote in Unite.ai.

The argument arrives with an interest attached. Evans cofounded Fictiv, a manufacturing sourcing platform the Japanese components supplier MISUMI bought for $350M, and now heads MISUMI AI.

The column drops the point just as it gets useful. Actuators are 40% to 60% of a humanoid robot’s bill of materials, McKinsey found, and the gearbox is 30% to 50% of the actuator.

That gearbox is where the queue forms. Harmonic and strain wave drives come from a short list including Harmonic Drive and Nabtesco, and qualification cycles are long.

Inside the motor sits the harder dependency. Every high torque joint needs a neodymium magnet, and China holds about 90% of the processing capacity for rare earth magnets.

The mining figure is lower and less alarming. China accounts for roughly 69% of extraction, so refining rather than ore is what binds.

Europe has written a law about exactly this. Light and heavy rare earths sit on the strategic raw materials list, with a 2030 target that no single third country supplies more than 65% of consumption.

The Critical Raw Materials Act also reaches the robot itself. Articles 28 and 29 name industrial robots directly, and from 2028 magnets above 0.2kg need a label, a digital record and a recycled content figure.

Europe’s industrial answer is currently one company. Schaeffler invested in Humanoid, the London startup that raised $152M in July, and Bosch builds the robots.

Schaeffler also sells the joints. It is preferred supplier for more than half of Humanoid’s joint actuators under an agreement running to 2031, covering a seven-digit number of units.

It is buying as well as supplying. Schaeffler will put a four-digit number of the robots into its own plants by 2032, starting at Herzogenaurach in December. TNW has argued Europe could win on deployment despite China holding 63% of the hardware chain.

Which is where the column’s remedy runs out. An actuator assembled in Herzogenaurach still contains a magnet refined in China, and no amount of supplier qualification changes that.

The gap is a policy one rather than a purchasing one. Europe has a 65% cap dated 2030, a labelling regime dated 2028, and almost no refining capacity to meet either.

The average hot water is a relatively simple appliance to understand. It uses gas or electricity to dump energy into water in the form of heat, keeping it at a pleasant temperature for uses like bathing and cleaning. Basic mechanisms are in place to ensure the water stays at a relatively constant temperature, neither too hot where it could cause burns, nor too cold such that it wouldn’t be fit for purpose.

One of the problems, though, is that sometimes storing water at the desired temperature can create the perfect breeding ground for bacteria. However, a neat little trick developed by NIST could solve that problem rather elegantly.

You’re Hot And You’re Cold

One of the interesting problems of our modern era is that changing our water usage has changed the risk profile for pathogens growing in hot water. NIST has noted for some time that while the pipes in our walls haven’t changed size, things like our shower heads have changed their flow rate to save water. Less flow rate in the same sized pipes means that water stays in the pipes for longer, increasing the time in which harmful pathogens have to grow in that environment.

Chief amongst these pathogens? The one most commonly feared is legionella pneumophila, the bacterium responsible for causing Legionnaires’ disease. The severe form of pneumonia comes with a fatality rate of 10%, and infection typically comes from inhaling aerosolized droplets containing the bacteria.

Of course, a great way to create aerosolized droplets filled with bacteria is to spray not-quite-hot water through a shower head. This is why hot water systems, particularly in large multi-occupancy buildings, are a risk for such infections. Legionnaires’ disease is rare, but it’s always out there, with 6000 confirmed cases showing up in the US each year. It’s possible that the true number is up to 10 times higher because the disease isn’t routinely tested for. The disease’s rarity is in part because engineers and tradespeople put in plenty of work to minimize the ability for the bacterium to breed in hot water systems. As NIST’s work demonstrates, though, there is possibly even more that could be done to tackle this problem.

Ideally, your hot water pipes in your house would not host any nasty bacteria. But often, hot water systems sit around 49 °C (120 °F), a temperature that’s hot enough to be useful and pleasant without a major scalding risk. Legionella pneumophila won’t easily grow at that temperature; the bacteria thrives between 20 °C to 45 °C (68 °F – 113 °F). Unfortunately, though, it’s not hot enough to kill the bacteria completely. Furthermore, water cools as it travels through pipes to an outlet, and thus pipes can be a perfect environment to spur further growth.

The simple solution would be to simply jack up the hot water system to store water at a much higher temperature, say around 70 °C (160 °F).  This could be undesirable in many contexts, though, as water that hot can more easily cause burns. It would be ideal to store the water at a higher temperature, while making sure it was delivered at a slightly lower temperature so it didn’t hurt anybody when it left the tap or shower head.

It turns out that there is a remarkably simple way to achieve this. NIST developed a simple heat exchanger which can be mounted atop any old hot water system. It’s designed such that the hot water leaving the tank passes through a heat exchanger where it’s surrounded by cold water which is running to the water heater’s intake pipe. This has dual benefits. The hot water inside the tank can be stored in the tank at 70 °C, which kills almost all legionella peumophila almost instantly. That water, though, is then cooled to a slightly safer temperature as it passes through the heat exchanger and out to the tap or shower. The hot water remains at a safe temperature for the user, and contains far less harmful bacteria, and cooling it doesn’t introduce any pathogens from another source. In turn, the cold water coming into the tank is pre-warmed slightly by the hot water, so the energy isn’t simply wasted.

This trick isn’t just simple, it’s also cheap. NIST engineers were able to fabricate a prototype unit to fit to a residential water heater using just $100 of components. It could be a useful addition for homeowners looking to make their hot water service a little cleaner. However, the real benefits are likely to be for larger operators of multiple-occupancy facilities, like nursing homes and hotels. These facilities often have to take great care to avoid the build up of harmful bacteria in building-wide hot water systems. Having the ability to increase the tank temperature to the point that legionella pneumophila bacteria simply die off would be a huge boon to this effort.

Sometimes, it’s possible to make great gains with a simple hack. Using the cold water service to cool hotter outgoing water is a smart trick that uses only minimal additional equipment to offer a safer hot water supply. It could yet become a common install for hot water systems in order to fight the good fight against a harmful disease that likes to lurk in our pipes. If you happen to see a heat exchanger atop a hot water heater in future, you can smile that someone has done the work to try and make that hot water supply just a little safer going forward.

Android offers tons of options to keep your device secure and private; it’s well worth exploring them all.

Update your software

Enable Android’s basic protections

Turn on anti-scam features

Manage your privacy settings

Thx4stock/Getty Images

Customize your advertisement settings

Protect against harmful apps

Turn on theft protection

Perhaps the most annoying kinds of faults are those that involve expensive equipment that just sit around in a cupboard, only for them to just stop working at all. Such was the case with the £2,000 Amprobe cable trace kit that [Tom] bought for work-related purposes. After sitting around unused for a few years, the signal generating part of the kit refused to power up at all, with Fluke’s service department wanting at least £600 to even attempt a repair after already having had [Tom] cough up £70 to even get this quote.

Subsequently he instead sent it to [Buy it Fix it] on YouTube for an attempted repair, which at the very least would be a less costly option. These systems use a signal generator connected to the cable, with a separate detector wirelessly tracking this signal. By itself that doesn’t sound too complex, but as it turns out it wasn’t quite so straightforward to diagnose.

After eliminating a basic power or display issue, the SoC’s boot sequence was traced, including reading from the 128 MB NAND Flash. At first glance the unit appeared to be trying to boot and work as normal, which was confirmed after finding a serial port and seeing the log output on it. This showed that the NAND Flash’s boot image failed validation due to apparent corruption.

After desoldering the TSOP 48 Flash package and stuffing it into a reader, at first glance the data on it looked fine. After a tragic detour with Google’s Gemini chatbot that led to a lot of wasted time, the solution that the human intelligence came up with was to crack open the tracer unit in the set and ogle at its firmware, in particular the bootloader.

Comparing the two Flash dump files, there were a few flipped bits in the bootloader section, likely due to cells in the Flash having lost their charge. Writing the generator’s Flash with a corrected image led to it booting up happily again into the Linux 2.6-based firmware, seemingly no worse for wear. Of course, one has to consider here that the NAND Flash IC clearly has a few leaky cells in it, so replacing it with a fresh one could be a good idea for a long-term fix.

In a nutshell: The RTX 5090 and its predecessor, the RTX 4090, have been plagued by reports of catastrophic thermal failures that can cause the 16-pin power connector to melt, often destroying the card in the process. A team of Brazilian YouTubers has now seemingly found a “solution” to the problem by soldering the power cables directly to the GPU PCB.

TecLab recently demonstrated a workaround for the RTX 5090 thermal failures by modifying a 5090 to bypass the problematic 16-pin power connectors and route the power wiring directly to the GPU’s PCB instead. For the experiment, they used a Galax GeForce RTX 5090 D HOF OC LAB XOC, which comes with two 16-pin power connectors designed for extreme overclocking.

The modification was carried out in two steps. First, the team used two adapters to accommodate two sets of wires from each 16-pin connector before soldering them to the connector traces on the PCB. Next, they removed the heatsink and the two 120mm fans cooling the GPU and added an industrial-strength water cooler to reduce the GPU’s core temperature.

While the YouTubers did not reveal the GPU’s temperatures after the modifications, the ~50A reading from the power meter connected to the 12V rail suggests that the card was drawing around 600W. For context, the Galax HOF OC LAB XOC can handle up to 2,000W, far more than the 575W TDP of standard RTX 5090 cards, thanks to a custom BIOS designed for professional overclockers.

TecLab acknowledged that its “solution” is not practical for most gamers and insisted that it was intended as a sarcastic take on Nvidia’s repeated failures to fix the long-standing problem, rather than an actual recommendation to disassemble and permanently modify a $5,000 graphics card.

Ever since the RTX 5090 launched in January 2025, dozens of users have complained about melting 16-pin connectors. One of the earliest reported cases surfaced in February 2025, followed by two more reports in April and another in May. The issue was also documented by John Papadopoulos of DSO Gaming, who claimed that the 12V-2×6 cable connected to his RTX 5090 melted while he was playing Wuchang: Fallen Feathers.

Most people already carry a camera in a pocket, so buying a second gadget that only spit out tiny sticky squares can feel like clutter. Liene’s M100, priced at $135.99 (was $220), skips that whole category. You send a photo from a phone, tablet, or computer and get a glossy 4-by-6 print in about a minute, which is the size most frames, albums, and fridge magnets still expect.

The M100 is built around the dye sublimation process. Using a heat press, it transfers the yellow, magenta, and cyan dyes from a specific ribbon to their own patented paper before applying a clear protective coat for the fourth time. The resulting print measures 3.94 by 5.83 inches on a somewhat larger piece of paper, with enough room for the image to run all the way to the edge once any excess tabs are removed. Resolution is a decent 300 x 300 dpi, with a respectable 256 shades per color, which should be plenty for casual printing into albums and small frames, without the irritating grain that many pocket printers produce.

Sale

Liene M100 4×6” Photo Printer, Phone Printer 100 Sheets & 3 Cartridges, Full-Color Photo, Portable…

• User-Friendly Portable Photo Printer for iPhone, Android & PC – Get beautiful and colourful memories with Liene 4×6 photo printer! After opening the…
• Superior Quality 4×6 Photo Printer – Print photos like no other thanks to the thermal-dye sublimation tech adopted by the picture printer for iphone…
• Wi-Fi Photo Printer for Smartphone – The built-in hotspot of the photo printer makes it easy to print pictures from your phone, avoiding outside…

The device’s exact body dimensions are 8.07 by 4.91 by 3.36 inches, and it weighs around 2.65 pounds. It has a back cassette that can contain 20 sheets of paper and a magnetic tray that snaps on top when you’re ready to pack it away. The official print time is 59 seconds, but you’ll most likely hear the sheet rattling in and out of the machine four times as it applies each color layer and protective coating.

Connecting it to your home network is simple because it establishes its own Wi-Fi hotspot, eliminating the need for you to be connected to the internet at home. Simply launch the free Liene Photo app on your iOS or Android device, join the network, select your JPEG or PNG file, add a border or collage if desired, and submit it to print. USB-C is compatible with both Windows and Mac computers, and AirPrint is available for Apple devices. The fact that numerous individuals may connect to the same hotspot is a wonderful addition, especially if you’re at a family gathering and everyone wants to print their Christmas photos on the same table.

The prints themselves come out nice and glossy, all laminated up, which is exactly what photographers who have compared this model to the Canon Selphy and Kodak dock printers have said, as they rave about the skin tones, fine detail, and (most importantly) the lack of those annoying inkjet speckles you get from cheaper printers. However, strong sky gradients can cause some banding, but the final quality is much closer to what you’d receive from a real-life photo studio than from a conventional ZINK pocket printer. That protective layer is a significant bonus because it prevents fingerprints, light wetness, and fading, which is wonderful if you plan on storing those photos in an album for years.

A ribbon cartridge lasts for about 40 prints, so you’ll need to replace them on a regular basis. It goes without saying that this is a photo printer, so you won’t be printing letters, homework, or oddly shaped documents on it, as that’s not its intended use.