High-voltage direct current (HVDC) is currently one of the hottest areas of technology and product development for AI data centers. It is not only the first layer of what Nvidia CEO Jensen Huang has described as the industry's "five-layer cake," but also a concrete manifestation of the idea that computing power ultimately depends on electrical power.
One word in Marvell Technology's August 19 filing decides how the Google agreement should be read. The custom silicon programs, it says, attach to the tensor processing unit (TPU) ecosystem. Attach, not replace.
Apple kicked off the tablet OLED era in 2024, introducing Tandem OLED in the iPad Pro to overcome bottlenecks in display lifespan and brightness. The iPad mini followed, set to adopt OLED in the second half of 2026, though with annual shipments hovering at just 2 million units, its impact is largely symbolic. Market research firm UBI Research noted that tablet OLED growth has largely stalled through 2024 and 2025, with OLED costs remaining the primary hurdle: panels currently run roughly twice the price of LCD alternatives, and narrowing that gap is widely viewed as the primary prerequisite for expanding OLED adoption in IT applications.
Unitree Robotics' August 19 debut on Shanghai's STAR Market did more than crown China's first mainland-listed humanoid robot maker. It marked a shift in the industry from competition between robot brands to a harder contest over supply-chain cost, AI capability, and mass production.
Marvell Technology is set to deepen its role in Google's AI infrastructure through an expanded agreement to develop custom chips, in a deal that could generate roughly US$120 billion in revenue for the chipmaker through fiscal 2033 if Google meets purchasing targets tied to the warrant, according to Reuters. The scale of that figure is the story: it reframes Marvell not as a niche supplier, but as a structural bet by Google on custom silicon over merchant GPUs.
Tsinghua Unigroup's former "chip-to-cloud" expansion strategy unraveled after a debt crisis in 2020 and subsequent bankruptcy restructuring. Several projects originating in the group's earlier era have since entered disposal proceedings, and one of the former group's largest planned investments has now reached a formal end.
Scale and growth have decoupled. TSMC still supplies more than half the sector's monthly sales, but the growth table now belongs to DRAM, NOR flash and controller names.
Ardentec credits a new fab entering volume production for the sharpest monthly jump in the sub-sector; ELASER's revenue nearly doubled year on year on customer demand.
China's semiconductor self-reliance campaign is moving beyond lithography machines and advanced processors into less visible materials that can be just as difficult to replace. Ajinomoto Build-up Film, or ABF, has become the latest pressure point.
On February 20, 2026, Canadian AI chip startup Taalas unveiled its HC1 inference chip. Taalas said the chip ran the Llama 3.1 8B model at a single-user inference speed of 16,960 tokens per second, roughly 48 times the inference speed of an Nvidia B200 in the company's testing.
Around 2019, a friend working at a large cloud services provider excitedly shared that his team had trained a CNN model for document text recognition with good results, but the computing cost was too high to deploy directly. They then used distillation to train a smaller model for service, and it turned out to be highly effective.
Amid global supply chain restructuring and geopolitical rivalry, precious metals are no longer merely financial hedging tools but have become core assets supporting high-tech manufacturing and strategic material reserves.
As physical AI and robotics draw unprecedented market attention, a key bottleneck is clogging the deployment pipeline: quality, real-world physical data. To move applications from the lab to real-world deployment, companies are getting creative in overcoming this obstacle while balancing affordability, stability, and data volumes.
