The Wire · Chips
The silicon beat — GPUs, custom accelerators, HBM, advanced packaging, and the foundries and supply chains that gate every AI buildout.
OpenAI cut GPT-5.6 Luna API prices by 80% to $0.20 per million input tokens and $1.20 per million output, while Terra fell 20% to $2 and $12. Fast mode gives Sol up to 2.5 times Standard speed at twice the price. OpenAI says Sol-assisted kernel work lowered serving cost by 20% and improved token-generation efficiency by more than 15%, while Luna delivers year-old frontier performance at roughly six cents per task-dollar and nearly nine times the speed. The edition connects cheaper models to Amazon's reported $1.8m, 860%-over-budget coding task, Gemini Robotics 2 whole-body control, Nscale's Anyscale acquisition and Okta's roughly $200m Permiso deal.
TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT LLM also contains components to create Python and C++ runtimes that orchestrate the inference execution in a performant way.
Read full story →106 points, 76 comments on HN
Read full story →Now 108 years after its transmission, an encrypted World War I German radio message has apparently been deciphered for the first time. The decoded and translated message relays information about the movements of an English cruiser and an Allied squadron near the Crimean Peninsula. Prinz, the developer who reckons they successfully decoded this covert WWI communication, used GPT-Astra to solve the cipher. Go deeper with TH Premium: AI shortages (Image credit: Nvidia) AI data centers are swallowing the world's memory and storage supply Demand for data center CPUs has surged, and AI agents are responsible Chip scarcity assaults auto industry amid the worsening Nexperia and DRAM crisis The custom AI ASIC state of play Prinz picked the code from a relatively famous list of 50 unsolved ciphers maintained by the German science blogging portal Scienceblogs.de . It was known to be “encoded using the ADFGVX method,” says the developer on their Substack. Addressed to the German High Command and for the attention of an admiral or perhaps Naval Command, the ciphered message looks like gobbledygook, surely as intended. The German military at the time used a convoluted grid of letters that shuffled depending on the current keyword. GPT-6 Astra deciphered a 1918 German radio transmission that, to my knowledge, has never been deciphered before.The message below translates to:"EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X"or, in English:"AN… pic.twitter.com/8kjDdI2Q5O September 17, 2026 Astra solved the cipher using the word “TRUPPENVERSCHIEBUNG” as the key. This resulted in the decoded message: “EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X.” Translated into English, we can at last understand that the radio message was the following alert: “AN ENGLISH CRUISER ARRIVED AT SEVASTOPOL ON THE ?4TH AN ALLIED SQUADRON FOLLOWS ON THE 26TH." Astra also checked its work against mili
Read full story →The Hyte X50 is a PC case that really stands out from the typical black box design, and it’s $50 off at the company’s site until September 21st, 2026, bringing the price down to $99.99. This attractive case supports motherboards from ITX all the way up to the E-ATX form factors, and GPUs that are […]
Anamanaguchi, the band consisting of Peter Berkman, James DeVito, Luke Silas, and Ary Warnaar, are most known for their chiptune music. Like me, you might have first heard them in game soundtracks like 2010's Scott Pilgrim vs. the World: The Game or know of their collaboration with Hatsune Miku (which even made its way into […]
When it comes to AI models, security functions as an afterthought, as evidenced by increased instances of agents hacking organizations and people, and other security mishaps with agents gone rogue. There's also an opportunity here for companies to offer new solutions. This should not come as a shock to anyone, according to cybersecurity investors and accelerator executives. “On one hand, we shouldn’t be surprised that increasingly capable agents are finding creative and sometimes unexpected ways to accomplish their objectives,” Matt Hartman, chief strategy officer at Merlin Group, told The Register. “On the other, we can’t accept harmful behavior as inevitable or unmanageable.” It’s the same story that plays out with every emerging technology, from laptops to cloud, said Todd Graham, managing partner at Microsoft’s M12 venture fund. “Every time we've built a new piece of infrastructure, we've conveniently forgotten the security,” Graham told The Register. Herein lies the opportunity for early-stage security companies. “If laptops were default secure, we wouldn't have CrowdStrike,” Graham said. “If the cloud was default secure, we wouldn't have Wiz. If identity wasn't default secure, we wouldn't have a bunch of Active Directory add-ons and Okta.” When it comes to AI security, “a lot of ships are going to rise with this tide,” he added. What’s different with AI is the speed at which models are advancing. While companies adopted cloud technologies over a period of years, organizations are moving full speed ahead to incorporate agents and other AI tools into their production environments and allow them to access the most business-critical data and applications. Yet they don’t have a strong handle on how to manage, secure, or even identify the agents that are already roving about their systems. “This is a transition happening month over month, and given the rate of change we’re seeing in the market, I’m in no way or shape surprised that this issue has come to a head in a
A 9-year-old Minecraft player called MightyMikePLays67 unknowingly spent a massive $118,000 on their dad’s company credit card to promote their YouTube channel. The father, Dave, gave a lengthy update on the YouTube channel , saying he’s lost his job and has to pay the amount within 30 days. Interestingly, he’s had no issues with losing work, saying that “jobs come and go. I’ve worked my whole life; I’ll find another job… I’ve got two hands, I’m healthy, I’m capable of working. So, if losing my job was the worst that came out of this, I’d honestly probably be okay.” Unfortunately, he’s still on the hook for the massive bill, and it's gotten to the point where he and his wife are talking about selling their home to help pay for it. Go deeper with TH Premium: GPUs (Image credit: Noctua) Desktop GPU Roadmap Nvidia's Enterprise GPU Roadmap Testing DirectStorage with GPU decompression The GeForce RTX 30-series upgrade matrix — does your Ampere GPU need an upgrade in 2026? Rubin in-depth The Stout Owl: The ultimate Noctua G2 PC Despite the depressing development, Dave and Mike were still about to joke around a bit about it. After Dave thanked his son for bringing him some water, Mike answered, “67.” He then asked, “Mike, do you even know what 67 means?” To which the kid replied, “To say the least, I do not know what it means, but I think it doesn’t mean anything. It’s just something that makes people laugh and smile.” The dad then replied, “Yeah, I would have to agree with that, Mike. Actually, I know what 67 means, Mike. That’s how old you’re going to be when you finally have to stop mowing lawns… That’s retirement age for the general public, but not for your old man. After this situation, apparently, I’m going to be working until I’m like 94.” Interestingly, Dave said that he doesn’t want to accept external help, like setting up a GoFundMe or a crypto coin to help with the bill. “We do not have a Mighty Mike coin. We don’t have a 67 coin. We don’t have any coins. No cry
AI infrastructure could generate enough electronic waste by 2050 to fill a line of shipping containers stretching around Earth six times, according to a report that argues existing estimates drastically understate the problem. Previous estimates have focused chiefly on servers and accelerators such as GPUs, which account for just 13 percent of a datacenter's equipment by weight, the report says. Once power, networking, cooling, and other infrastructure are included, the total could be 40 to 60 times higher than the most widely cited academic projections. The report identifies five equipment categories comprising networking, power distribution, backup power systems, servers plus accelerators, and cooling. Together, these add up to about 7,000 metric tons for a reference 100 MW AI bit barn, much of which will have to be replaced when the facility is upgraded, it argues. The report, How Big Is the AI Waste Wave? [PDF], comes from the Basel Action Network (BAN), a nonprofit organization named after the Basel Convention, which controls international movements of hazardous waste and seeks to prevent its transfer from developed to developing countries. BAN's model starts with what the industry says it intends to build and calculates the potential waste implied by that expansion. It estimates that AI infrastructure will generate between 395 million and 617 million tonnes of e-waste from 2025 to 2050. The report says that would fill between 15 million and 23 million shipping containers. Placing 20 million of them end to end would create a line about 244,000 km long, or roughly six times Earth's circumference. The estimate rests on several assumptions. The model supposes that the copper, steel, cooling and power distribution equipment, networking gear, and computing hardware in the latest AI datacenters will be decommissioned within years rather than decades as the technology becomes obsolete. For example, the report says that when a facility replaces racks drawing between 5
A team of researchers at Maynooth University, Ireland, has created a “first-of-its-kind” DNA molecular computer that uses DNA strands to perform complex mathematical operations without electricity. Detailed in the journal Nature on September 16, the system — called a Scaffolded DNA Computer (SDC) — is one of the most complex and fastest molecular computers, and “points to new possibilities for long-term data storage, energy-efficient computation and, in time, molecular systems that could operate inside cells for applications such as disease detection,” according to the researchers. The system successfully executed 10 different molecular programs, including complex 100-bit calculations. The researchers designed the computer via a technique known as DNA origami. Using specialized software, they mapped out a long primary DNA strand and hundreds of shorter, custom-synthesized “staple” strands. They then added the physical DNA strands to a test tube containing a drop of water and salt. When they heated and then cooled the mixture, the strands self-assembled into a highly organized, microscopic computing grid, with the long strand acting as a structural scaffold. Traditional silicon computers use transistors to switch electrical voltages between 1 and 0. On the other hand, the molecular computer uses the binding and unbinding of genetic base pairs (A, T, C, and G) to process information. The researchers write the program into the DNA sequences themselves before putting them into the test tube and applying heat. The thermal energy kick-starts chemical reactions, causing the DNA strands to rearrange. As the molecules naturally shift toward their most stable structural state, they mathematically solve the programmed algorithm, with the final structure representing the mathematical answer. The researchers ran 10 different molecular programs to test the system. The DNA computer successfully performed addition, subtraction, multiplication, and division. It successfully processe
565 points, 205 comments on HN
cuda-oxide is a Rust-to-CUDA compiler that lets you write (SIMT) GPU kernels in safe(ish), idiomatic Rust. It compiles standard Rust code directly to PTX — no DSLs, no foreign language bindings, just Rust.
<p data-block-key="8s6s5">Investment will include the establishment of a 140-acre research park</p>
Earlier this year, Nvidia CEO Jensen Huang suggested that Marvell’s optics tech would make it the next trillion-dollar company. On Thursday, Marvell took a big step towards securing that status by announcing an expanded partnership with GlobalFoundries, an American wafer fab well known for its work in silicon photonics manufacturing. The tie-up will see GloFo and Marvell work to expand silicon germanium (SiGe) wafer production at the fab’s Burlington, Vermont wafer plant. SiGe is commonly employed in the production of optical transceivers, which convert electrical signals to optical ones and back again, as well as in laser modules used in many high-end switches from Nvidia and others. According to GlobalFoundries, the additional capacity will support the production of near- and co-packaged optics (CPO/NPO) as well as “next-gen” pluggable optics. Nvidia already employs co-packaged optics in some of its Spectrum Ethernet and Quantum InfiniBand switches to cut power consumption and improve reliability. Pundits expect NPO to be used in large multirack systems such as Huawei’s new Ascend 960DT-based SuperPods. Nvidia’s future multi-rack systems are widely expected to use NPO as well. The fab says its SiGe tech has already been validated up to 200 Gbps per lane, which is required for 1.6 Tbps transceivers. Faster lane speeds could open the door to 3.2 Tbps transceivers, which are also on its roadmap. Demand for optics technologies has surged thanks in no small part to everyone’s favorite subject: AI. Speaking during Computex this spring, Marvell CEO Matt Murphy explained why demand for optics demand is set to explode. Up until now, optics have primarily been used to connect racks to resources across the datacenter. Anything within a few meters was well within reach of cheaper and less power hungry copper interconnects. But as lane speeds climbed from 25 Gbps to 50, 100, and then 200 Gbps, copper interconnects began running into practical reach limits. 400 Gbps per lane is
<p data-block-key="of71c">ai& looking to purchase up to 100 units of Rebellions’ RebelRack infrastructure</p>
Huawei is accelerating the launch of its next-generation Ascend 960DT AI chip as it pushes to compete with Nvidia and close China’s AI computing gap with the U.S.
An independent briefing for builders: the whole field read continuously, every story scored for relevance, and the noise left off the page.
300+ curated sources. Every story scored 1–10 for builder relevance by Claude's frontier model. The filler never makes it to the page.
GPUs, datacenters, power deals, and inference economics: the infrastructure layer that decides what every builder pays. Our signature coverage.
Every story is sourced. Every score is computed. We show our work and link to originals.
Every briefing closes with The Call: one falsifiable claim with a date on it. When we're wrong, we say so in print. Opinions are cheap; ours get scored.