We went through 776 papers, announcements, filings and news items across the computing stack this week. Here’s what made the cut.
-Tara
THE DOWNLOAD
SILICON AND POWER
DDR5 Retail Prices Are Up 485% in Twelve Months
DRAM is the working memory between a processor and storage, and it is built on the same production lines as the high bandwidth memory that goes into AI accelerators. HBM earns more per wafer, so manufacturers have shifted output toward it, putting supply pressure on DRAM. This goes into laptops, phones, etc. Memory makers have reportedly sold their 2027 capacity in full.
SK hynix Publishes a Roadmap for Optical Links Between Compute and Memory
Co-packaged optics puts the optical components in the same package as the chip, so data leaves over light rather than copper. SK hynix has published a roadmap for it in Nature Electronics, written with the University of Virginia, MIT, UIUC, Nanyang and Yonsei, that extends the approach to the memory interface. The mechanism is a photonic interposer between accelerator and memory, which removes the packaging limit on how much memory one chip can reach and lets several accelerators draw on a shared pool. Stated targets are over 100 terabits per second per node and chip-to-chip latency under ten nanoseconds. No product date yet.
TerraPower Signs Hyundai E&C to Build Eight Natrium Reactors
TerraPower and Hyundai Engineering & Construction have signed a framework agreement making HDEC the builder for up to eight 345-megawatt Natrium plants in the US. Natrium is a sodium fast reactor with molten salt heat storage, which lets output rise to 500 megawatts for more than five and a half hours. Meta agreed in January to eight plants, first units as early as 2032.
COMPUTE AND CAPITAL
Nvidia Is Guaranteeing Up to $105 Billion Behind OpenAI’s Ohio Data Center
Nvidia has told the SEC it will provide up to $105 billion of credit support for an OpenAI data center in Pike County, Ohio. SB Energy, a SoftBank-backed developer, builds and owns the campus. Nvidia’s promise to the lenders: if OpenAI stops paying, Nvidia covers the lease and power payments and makes good on what the building and the chips are worth. Nvidia also put $1.5 billion into SB Energy and supplies the chips for the first 800 megawatts, due in 2028.
Groq Is Pivoting From Designing Inference Chips to Operating Data Centers
Groq is pivoting from designing inference chips to operating data centers built on Nvidia hardware. Back in March, Nvidia bought out Groq’s leadership, including founder Jonathan Ross, along with a non-exclusive licence to its inference technology, for roughly $20 billion. The plan now appears to be to drop chip design and build out the existing footprint: thirteen data centers, more than six million developers on the platform, and a move from 54 megawatts to over 200 in 2027.
Unitree Rose 460% on Its First Day of Trading in Shanghai and Anthropic’s Public S-1 Is Expected This Month
Unitree listed on Shanghai’s STAR Market this week, raising $900 million on a book covered more than 8,000 times. Anthropic filed a confidential S-1 on June 1 at roughly a $965 billion valuation and is expected to file publicly as soon as the end of August. OpenAI has not filed.
MODELS AND DATA
DeepSeek Released Its New Vision Model Through the API, With No Weights
DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal model, through its API. No weights. The company says it matches V4-Flash on text and brings multimodal agent performance “close to Opus-4.8” on its own benchmark chart. Images are billed as tokens, up to 384 each at V4-Flash pricing. DeepSeek has released open weights for its main models.
Thinking Machines Is Serving Inkling Free to Agentic Harnesses
Thinking Machines has made Inkling free on OpenRouter and says the offer is limited to agentic harnesses, will run for a few weeks, and that the interaction data, disassociated from accounts, will be used to improve the model’s agentic performance. What the lab can observe: traces from long tool-use loops inside real harnesses, which downloaded weights do not produce.
Replit’s Free Mode Routes Simple Tasks to GPT-5.6 Luna and Stops Metering Them
Replit has made GPT-5.6 Luna the default for chat, ideation and simple coding, and stopped charging those tasks against a subscriber’s token budget. It is called Free Mode and it requires a paid subscription, $20 a month for Core or $100 for Pro. Harder work routes automatically to a larger model, then reverts. “A lot of tasks don’t require the frontier-level intelligence. It just requires smaller models that are faster and more affordable,” said Michele Catasta, Replit’s president and head of AI.
Amazon Is Cutting the Bindings Off Rare Books to Scan Them
404 Media put a tracker in a shipment of rare books and followed it to an Amazon facility in North Las Vegas, where a team cuts the bindings off so the pages feed through scanners faster. The book is destroyed. Amazon says it “purchases books through commercial channels to improve the products and services customers use.”
LABOR & LEARNING
Homework Scores Up 18%, Exam Scores Down 20%, Same 26,811 Students
Researchers at Stockholm University and the University of Hong Kong followed 26,811 Chinese students in grades seven through twelve for two years. Those who took up AI homework help raised homework scores 18% and cut homework time 30%. Within six months the same students scored 20% below classmates who had not used AI on monthly exams, and college entrance results came in 18 to 24% lower. About 80% of the exam gap came from students the authors classify as outsourcing the work rather than working alongside the tool.
40,000 Hyundai Workers Struck Over Wages & Robots Arriving in 2028
Hyundai’s Korean union, about 40,000 members, began its first full strike in a decade on Friday, stopping the Ulsan, Asan and Jeonju plants. Wage talks are the formal deadlock. The union is also demanding protections against automation, tied to Hyundai’s plan to put Boston Dynamics Atlas robots into its plants from 2028. Hyundai Motor Group controls Boston Dynamics. Partial walkouts since July have cost about 55,200 vehicles worth $1.67 billion. “This strike is to protect the roots of Korea’s automobile industry and uphold the dignity of labor,” said Park Sang-man, chairman of the Korean Metal Workers’ Union.
LATEST FROM THE REVIEW
Coding agents are breaking code infrastructure
Coding agents are generating more code than existing infrastructure was built to handle.
EVENTS
We’re hosting a casual happy hour at Hot Chips!
DDR5 prices spiking 485% in twelve months isn't a temporary supply chain blip. It's the physical memory wall catching up to consumer hardware as wafer capacity gets reallocated to HBM stacks for cloud training clusters. When memory makers sell out their 2027 output to hyperscalers, standard computing gets starved of basic working memory. 📉
Look at what happens when working memory gets outsourced higher up the stack. Stockholm University's study of nearly 27,000 students shows homework scores rose 18% with AI assistance, but exam performance plummeted 20%. Cognitive delegation strips out the internal representation loop. The moment you delegate active execution to an external probabilistic model without holding the local state, real performance collapses. 🧠
We're seeing this exact same structural breakdown in hardware architecture. Groq abandoned custom inference chip design to run Nvidia infrastructure, while SK Hynix is pushing photonic interposers just to bypass physical package limits with co-packaged optics. High-level software schedulers and metered cloud models can't patch over physical bandwidth shortages or latency bottlenecks. ⚡
The externalization of working memory, whether from human cognitive routines or local DRAM into centralized cloud HBM, creates an unsustainable thermodynamic tax. True execution stability requires phase-locking critical reflex state directly into spatial SRAM registers and optical chip-to-chip interconnects. If your local edge systems still rely on remote token streams and starved consumer memory lines to execute real-time decisions, how long before physical wafer constraints pull the plug on your cloud-dependent architecture? ⚙️
(ಠ_ಠ)