Semiconductor capacity keeps running into the same awkward fact. The purest quartz for the crucibles that turn polysilicon into monocrystalline silicon comes mainly from Spruce Pine, North Carolina, a town whose two mining companies Hurricane Helene shut down in late September 2024.
OpenAI says it is running an "extensive" review of its models' actions after agents breached Hugging Face and touched Australia's Medicare statistics portal, according to CNBC.
AI accelerators are the specialised chips that sit next to regular CPUs in datacenter racks and do the matrix maths behind chatbots and image generators. How they are built, and who can build them, is more interesting than the acronyms suggest.
A Japanese company controls roughly 90% of the global supply of a specialist glass-fiber cloth used inside every advanced AI chip package. It cannot keep up with demand, and prices have risen 20 to 30%.
The AI accelerator race is no longer just about who can print the fastest logic die. Three supply chain stories from 2026 show the constraints sit in the material, the memory stack and the engineering workflow around the chip.
Bill Harding's Project SWIFT averaged 5 hours per fabrication layer in the 1970s, IEEE Spectrum reports, while the fastest modern fabs take 19 hours and the industry average is 36.
Google says it achieved "quantum supremacy" in a paper posted to NASA.gov before being taken down. The company claims its 53-qubit Sycamore processor took about 200 seconds on a task a state-of-the-art supercomputer would need roughly 10,000 years to complete.
IBM's Project SWIFT averaged five hours per wafer layer in the 1970s. The fastest modern fabs take 19 hours per layer, IEEE Spectrum reports. Capacity debates keep missing that kind of detail.
Google set out in April 2017 to prove a quantum chip could beat any classical computer by the end of that year. The company's claimed "quantum supremacy" result arrived in September 2019, in a paper Fortune obtained before peer review. The number that matters is 53 qubits, and the argument over what it proves has not stopped since.
A Japanese company most datacenter operators have never heard of controls roughly 90% of the global supply of a glass cloth that sits inside every advanced AI chip package, and its next production line will not arrive until mid-2027.
One Japanese company makes roughly 90% of the glass-fiber cloth that goes inside every advanced AI chip package. The squeeze shows up in prices, in lead times and in how Nvidia buys.
SiFive started licensing a complete RISC-V AI accelerator design in September 2024, while a Japanese supplier of a little-known packaging material controls roughly 90% of a market that AI chips depend on.
Nittobo controls roughly 90% of the global supply of the T-glass cloth that sits inside every advanced AI chip package, and the shortage is now pushing prices up 20 to 30%.
Thousands of unionized Hyundai auto workers walked off the job in Ulsan, South Korea, in July after talks over humanoid robot deployment collapsed. The Wall Street Journal called it the first factory stoppage aimed at humanoid robots.
Thousands of unionized workers at Hyundai's Ulsan plant walked off the job after talks over humanoid robot deployment collapsed, while China pushes ahead with state-funded training centers for the machines.
Texas Instruments said on 24 October 2024 it had begun making gallium nitride power chips at its plant in Aizu, Japan, quadrupling its internal GaN capacity across the United States and Japan. The same week, a hurricane exposed how narrow parts of the supply chain still are.
Google says its 53-qubit Sycamore processor sampled a quantum circuit in about 200 seconds, a task it estimates would take a state-of-the-art supercomputer roughly 10,000 years. IBM's research head Dario Gil called the experiment a special case with "no practical applications."