As HBM approaches half the cost of a GPU-HBM CoWoS package, can memory still be viewed as merely a passive component of AI computing? The answer goes back to the "memory wall," a challenge that has shaped nearly half a century of semiconductor development.
TSMC said the next phase of AI performance growth will depend less on shrinking a single chip and more on combining multiple chips, memory, and advanced packaging inside larger systems. The shift could reshape semiconductor competition, as demand for compute, bandwidth, power delivery, and cooling rises in tandem. The company outlined the trend at its heterogeneous integration forum during SEMICON Taiwan 2026, saying AI computing transistors in a single CoWoS package are set to rise more than 48-fold by 2029 from 2024 levels, while total HBM bandwidth is expected to increase more than 34-fold.
DIGITIMES Intelligence observes that AI applications are moving through a clear progression: from early classification AI focused on feature recognition, to generative AI capable of creating content, and toward a broader shift into the agentic AI era in 2027. Agentic AI is expected to multiply compute demand, while multi-step reasoning introduces new risks to task quality. Efficient control of total token consumption and reliable task-quality management will therefore become critical to commercial deployment.
Micron unveiled the Micron Ventures Paradigm Fund, a US$250 million venture vehicle, in mid-August 2026, as the memory maker pairs more than US$25 billion in fiscal 2026 capex with a pledge to invest more than US$250 billion in the US by 2035. The fund is designed to give Micron an early look at technology shifts, position it for next-generation AI architectures, and strengthen its strategy to reduce memory-cycle volatility.
Artificial intelligence is rapidly entering enterprise operations, blurring the lines between cybersecurity and supply chain management. In the AI era, Zero Trust is no longer only about who can access a system. Enterprises must also ask whom and what they depend on, how concentrated those dependencies are, and what happens when a critical link fails.
As chip design enters the AI era, heterogeneous integration is driving foundry giants like TSMC, Samsung, and Intel to bring advanced front-end fab capabilities into back-end packaging. In the second part of his interview with DIGITIMES, Dr. Shin-Puu Jeng, chairman of IMAPS and a former TSMC executive, shared how TSMC transformed CoWoS from a niche lab concept into an industry standard.
In a recent interview on the DIGITIMES podcast, Dr. Shin-Puu Jeng, chairman of the International Microelectronics and Packaging Society (IMAPS) and a former longtime TSMC advanced packaging executive, reflected on the early development of CoWoS. In the first of this two-part series, Jeng details how a critical thin-interposer crisis forced TSMC and Xilinx to innovate, laying the foundation for modern advanced packaging.
Global AI has entered a period of dramatic turmoil, with leading researchers warning that top labs are "gambling with our lives" and one Anthropic safety researcher saying there is more than a 10% chance AI will wipe out humanity within the next 10 years. That pressure has pushed Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman into rare agreement: slow model development.
A robot that can map a factory floor still cannot necessarily find the box it has been asked to move, or remember what it saw on its last visit. For autonomous robots, building the map is increasingly just the beginning. What comes next — searching for objects, navigating to specified locations, remembering what the robot has seen, and retrieving that information for model reasoning in complex environments — is where the new applications and challenges lie.
China's two leading memory suppliers are moving to raise capital and expand output as demand for chips grows, a shift that could reshape the country's domestic supply chain. Their efforts reflect a new stage for the sector, but the long-term test will be whether fresh funding can translate into stronger technology, higher yields, and wider customer adoption.
When companies discuss ESG, they often emphasize environmental issues and compliance while paying less attention to employee health, a key component of the social dimension, according to Saxon Chen, founder and CEO of Taiwanese digital health technology company H2U Corp.
GMI Cloud is making Taiwan a key part of its global AI infrastructure expansion, with a US$500 million investment in an AI factory powered by Nvidia's Blackwell GPUs. Announced in late 2025, the facility is expected to deploy roughly 7,000 GPUs and serve customers including Trend Micro and Wistron.
India's new semiconductor incentive regime is putting more capital at risk for manufacturers, but the first wave of investment commitments suggests lower subsidies have not stopped companies from considering fresh projects.
After more than a decade away, Apple may be preparing to return to the enterprise server market. According to The Information, as reported by Reuters, the company is developing an enterprise AI server powered by its planned M8 Ultra chip and has discussed adopting Nvidia's NVLink Fusion interconnect technology.
When Donald Trump and Xi Jinping discuss artificial intelligence, the deeper divide may not be over whether AI needs rules, but over what those rules are designed to protect.
