Scientists in Chinadeveloped an error-aware probabilistic update (EaPU) to improve neural network training on memristor hardware. The method tackles accuracy and stability limits in analog computing.
Training inefficiency caused by noisy weight updates has slowed progress beyond inference tasks. EaPU applies probabilistic, threshold-based updates that preserve learning and sharply reduce write operations.
Experiments and simulations show major gains in energy efficiency, accuracy and device lifespan across vision models. Results suggest broader potential for sustainable AI training using emerging memory technologies.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
MIT researchers have demonstrated a faster, more energy-efficient cooling technique for scalable trapped-ion quantum chips. The solution addresses a long-standing challenge in reducing vibration-related errors that limit the performance of quantum systems.
The method uses integrated photonic chips with nanoscale antennas that emit tightly controlled light beams. Using polarisation-gradient cooling, the system cools ions to nearly ten times below standard laser limits, and does so much faster.
Unlike conventional trapped-ion systems that depend on bulky external optics, the chip-based design generates stable light patterns directly on the device. The stability improves accuracy and supports scaling to thousands of ions on a single chip.
Researchers say the breakthrough lays the groundwork for more reliable quantum operations and opens new possibilities for advanced ion control, bringing practical, large-scale quantum computing closer to reality.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
AI workloads are placing unprecedented strain on system on chip interconnects. Designers face complexity that exceeds the limits of traditional manual engineering approaches.
Semiconductor engineers are increasingly turning to automated network on chip design. Algorithms now generate interconnect topologies optimised for bandwidth, latency, power and area.
Physically aware automation reduces wirelengths, congestion and timing failures. Industry specialists report dramatically shorter design cycles and more predictable performance outcomes.
As AI spreads from data centres to edge devices, interconnect automation is becoming essential. The shift enables smaller teams to deliver powerful, energy efficient processors.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
OpenAI has agreed to purchase up to 750 megawatts of computing power from AI chipmaker Cerebras over the next three years. The deal, announced on 14 January, is expected to be worth more than US$10 billion and will support ChatGPT and other AI services.
Cerebras will provide cloud services powered by its wafer-scale chips, which are designed to run large AI models more efficiently than traditional GPUs. OpenAI plans to use the capacity primarily for inference and reasoning models that require high compute.
Cerebras will build or lease data centres filled with its custom hardware, with computing capacity coming online in stages through 2028. OpenAI said the partnership would help improve the speed and responsiveness of its AI systems as user demand continues to grow.
The deal is also essential for Cerebras as it prepares for a second attempt at a public listing, following a 2025 IPO that was postponed. Diversifying its customer base beyond major backers such as UAE-based G42 could strengthen its financial position ahead of a potential 2026 flotation.
The agreement highlights the wider race among AI firms to secure vast computing resources, as investment in AI infrastructure accelerates. However, some analysts have warned that soaring valuations and heavy spending could resemble past technology bubbles.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
The US administration has approved the export of Nvidia’s H200 AI chips to China, reversing years of tight US restrictions on advanced AI hardware. The Nvidia H200 chips represent the company’s second-most-powerful chip series and were previously barred from sale due to national security concerns.
The US president announced the move last month, linking approval to a 25 per cent fee payable to the US government. The administration said the policy balances economic competitiveness with security interests, while critics warned it could strengthen China’s military and surveillance capabilities.
Under the new rules, Nvidia H200 chips may be shipped to China only after third-party testing verifies their performance. Chinese buyers are limited to 50 per cent of the volume sold to US customers and must provide assurances that the chips will not be used for military purposes.
Nvidia welcomed the decision, saying it would support US jobs and global competitiveness. However, analysts questioned whether the safeguards can be effectively enforced, noting that Chinese firms have previously accessed restricted technologies through intermediaries.
Chinese companies have reportedly ordered more than two million Nvidia H200 chips, far exceeding the chipmaker’s current inventory. The scale of demand has intensified debate over whether the policy will limit China’s AI ambitions or accelerate its access to advanced computing power.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
The United Arab Emirates has joined Pax Silica, a US-led alliance focused on AI and semiconductor supply chains. The move places Abu Dhabi among Washington’s trusted technology partners.
The pact aims to secure access to chips, computing power, energy and critical minerals. The US Department of State says technology supply chains are now treated as strategic assets.
UAE officials view the alliance as supporting economic diversification and AI leadership ambitions. Membership strengthens access to advanced semiconductors and large-scale data centre infrastructure.
Pax Silica reflects a broader shift in global tech diplomacy towards allied supply networks. Analysts say participation could shape future investment in AI infrastructure and manufacturing.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
TSMC reported a strong fourth-quarter performance, posting a 35 percent rise in profit to a record level, supported by sustained demand for advanced chips.
The company forecast robust growth for 2026, citing continued customer interest and tight capacity, while highlighting expectations for a significant increase in revenue in the first quarter of the year.
The Taiwanese semiconductor manufacturer confirmed that capital spending reached US$40.9 billion in 2025, slightly above earlier guidance, and indicated further increases ahead, with investment potentially rising to as much as US$56 billion in 2026 and accelerating later in the decade.
Ongoing projects include additional manufacturing capacity in the US, expansion in Japan, and continued investment in Taiwan.
TSMC also signalled that more US facilities may be planned, following earlier commitments to large-scale investment in Arizona.
Developments come amid discussions between Taiwan and the US on trade and tariffs, as well as broader policy efforts in Washington to encourage domestic semiconductor production.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
Gadget makers face rising costs as AI drives intense demand for memory chips. Supplies of DRAM and storage components have tightened across global markets.
Manufacturers have shifted production towards AI data centres, squeezing availability for consumer devices. Analysts warn the memory shortage could extend well into next year.
Higher prices are already affecting laptops, smartphones and connected devices. Some companies are redesigning products or limiting features to manage the costs of chip components.
Industry experts say engineers are writing leaner software to reduce memory use. The AI surge is marking the end of an era of cheap and abundant memory.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
Canadian quantum computing company Photonic has raised $130 million in the first close of a new investment round led by Planet First Partners. New backers include RBC and TELUS, alongside returning investors.
The funding brings Photonic’s total capital raised to $271 million and supports the development of fault-tolerant quantum systems. The company combines silicon-based qubits with built-in photonic connectivity.
Photonic’s entanglement-first architecture is designed to scale across existing global telecom networks. The approach aims to enable large, distributed quantum computers rather than isolated machines.
Headquartered in Vancouver, Photonic plans to utilise the investment to accelerate key product milestones and expand its team. Investors see strong potential across finance, sustainability, telecommunications and security sectors.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
China’s AI sector could narrow the technological AI gap with the United States through growing risk-taking and innovation, according to leading researchers. Despite export controls on advanced chipmaking tools, Chinese firms are accelerating development across multiple AI fields.
Yao Shunyu, a former senior researcher at ChatGPT maker OpenAI and now Tencent’s AI scientist, said a Chinese company could become the world’s leading AI firm within three to five years. He pointed to China’s strengths in electricity supply and infrastructure as key advantages.
Yao said the main bottlenecks remain production capacity, including access to advanced lithography machines and a mature software ecosystem. Such limits still restrict China’s ability to manufacture the most advanced semiconductors and narrow the AI gap with the US.
China has developed a working prototype of an extreme-ultraviolet lithography machine that could eventually rival Western technology. However, Reuters reported the system has not yet produced functioning chips.
Sources familiar with the project said commercial chip production using the machine may not begin until around 2030. Until then, Chinese AI ambitions are likely to remain constrained by hardware limitations.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
