Tesla and SpaceX Commit $16.8B to Build "Terafab," a 100-Million-Square-Foot Chip Plant in Texas (3 minute read)
Tesla and SpaceX confirmed Terafab, their jointly developed chip factory, will rise in Grimes County, Texas outside Houston, with an initial $16.8 billion investment toward what Musk calls the largest building on Earth by manufacturing footprint. The vertically integrated site will handle manufacturing, packaging, and testing of both logic and memory chips, split between edge-inference silicon for Optimus robots and self-driving Cybercabs and high-power chips for SpaceX's planned space-based data centers, with SpaceX filings suggesting total spend across the project's multiple phases could reach $119 billion.
|
Trump Signs Executive Order Protecting US Polysilicon Industry (3 minute read)
Trump signed a Section 232 proclamation setting minimum import prices on polysilicon and its derivatives (the raw material behind both semiconductors and solar panels) along with a 15% tariff on downstream products, aiming to shield US producers from Chinese dumping. The order also authorizes Commerce to create an incentive program for companies building or expanding domestic polysilicon facilities, part of a broader push to rebuild a supply chain that's shrunk from 50% of global production capacity in 2005 to under 2% today, with China now supplying roughly 90%. The remedies take effect 120 days after signing.
|
|
Engineering and Applications
|
Chinese Chip Toolmaker Claims Wafer Polishing Breakthrough (3 minute read)
Chinese equipment maker Hwatsing Technology has announced China's first in-line metrology system integrated directly into chemical mechanical polishing tools for six-inch wafers. By capturing real-time measurements before and after polishing, the system feeds live operational data back to control systems to improve wafer surface uniformity and minimize material waste. Shipped to a leading domestic optical silicon supplier, this chemical mechanical polishing breakthrough demonstrates how domestic toolmakers are quietly closing critical manufacturing gaps beyond advanced lithography, enhancing overall yield for mature semiconductor production lines across China.
|
The Real Debate in AI-RAN Isn't Whether AI Belongs in 6G, It's Where (4 minute read)
As the 3GPP prepares to set 6G's direction this September, Nokia and Nvidia's $1 billion partnership is pushing to put GPU-grade AI hardware directly inside radio access network equipment, with early trials at T-Mobile already running live 5G connections and AI video tasks on the same server. Skeptics like analyst John Strand and telecom CTO Kim Kyllesbech Larsen argue there's no clear economic case yet for loading every cell tower with expensive AI hardware, especially since wireless efficiency is already near its theoretical limits and some AI gains are achievable in software on existing infrastructure. Researcher Merouane Debbah frames the real architecture question as hierarchical: small, fast models embedded directly in radios for real-time decisions, more capable models at edge sites, and large foundation models running centrally for reasoning and coordination.
|
|
ω-0: One Model Lets Humanoid Robots Walk and Manipulate Objects at the Same Time (3 minute read)
Household chores usually require a robot to move, keep its balance, and manipulate an object all at once, but most humanoid policies still treat locomotion and manipulation as separate problems solved one after another. ω-0 instead predicts whole-body actions directly, using compact learned representations of what the robot expects to see next as a lightweight guide, rather than trying to generate full future video frames, which keeps the model fast enough for real-time control while still reasoning about how the scene will change.
|
Four Decades of In-Fab Metrology and Inspection: Pivotal Milestones and Road Ahead (4 minute read)
Research institute imec published a comprehensive overview detailing four decades of in-fab metrology and inspection evolution across logic transistor scaling roadmaps. The review analyzes four distinct industry eras: predictable happy scaling, initial device material innovations, FinFET architecture transitions, and advanced EUV lithography adoption. As dimensional scaling reaches physical two nanometer limits and artificial intelligence workloads require 3D system stack integration, imec presents a forward-looking metrology framework combining specialized hardware sensor extensions with data-driven inspection algorithms to ensure manufacturing yield and process sustainability.
|
|
SpaceX May Deploy 10GW Datacenters in Space by 2027 (4 minute read)
Research firm SemiAnalysis projects that SpaceX is on track to build 10 gigawatts of AI compute capacity by the end of 2027. Driven by sky-high inference margins generating over $100 billion per gigawatt annually for AI labs, SpaceX's "speed first" infrastructure strategy bypasses traditional data center construction bottlenecks. Offering rapid deployment alongside flexible 90-day cancellation clauses, SpaceX is securing massive off-take commitments with major hyperscalers like Microsoft. This aggressive expansion could drive up to $300 billion in annual recurring revenue.
|
Jane Street's New Puzzle: Reverse Engineer an ASIC From Its Raw Chip Layout (5 minute read)
Jane Street released a public puzzle, handing out only the final GDS file for a chip it designed, the raw geometric layout of every transistor and metal layer that a foundry uses to fabricate silicon, with nothing labeled. Solvers have to recover a netlist (the underlying logic gates and how they connect) directly from that layout, figure out what the circuit actually does, then use it to extract a hidden output value, all using free open-source tools like KLayout or Magic VLSI.
|
|
Chinese Chipmakers Poised to Win (3 minute read)
Chinese AI chipmakers are poised for significant growth as Beijing pushes self-reliance and Chinese companies plan to increase their spending on domestically made AI accelerators from 30% to 46% over the next year.
|
|
|
|
|
0 Comments