N Nvidia has reportedly asked Taiwan Semiconductor Manufacturing Co. to ramp up production of its H200 graphics processing units.

Nvidia Corp. has reportedly asked Taiwan Semiconductor Manufacturing Co. to ramp up production of its H200 graphics processing units. Sources told Reuters today that the move is a response to strong demand from customers in China. According to the report, those customers have placed orders for more than 2 million H200 chips. Nvidia is expected to charge about $27,000 per GPU, which means it could generate more than $54 billion in revenue. The H200 is an upgraded version of a graphics card that Nvidia introduced in early 2022. It features a nearly identical design as the original with the exception of the onboard HBM memory pool, which is nearly twice as large. Nvidia also boosted the GPU’s memory bandwidth, the speed at which data can move between the chip’s HBM and memory circuits, by 40%. A large language model generates prompt responses through a process that involves numerous steps. After each step, the model saves its processing results to the underlying GPU’s memory. It then moves those results from memory to the GPU’s logic circuits so that the next processing step may begin. The movement of data between a GPU’s memory and logic circuits accounts for much of the time required to generate prompt responses. That data movement can be accelerated significantly by boosting a GPU’s memory bandwidth, which is one of the optimizations Nvidia applied to the H200. As a result, the chip can perform some inference tasks twice as fast as the 2022 graphics card from which it’s derived. The H200 is also 6 times faster than the H20, the fastest chip that the company currently sells in China. Nvidia developed the H20 after the Biden administration blocked the export of top-end GPUs to China in 2022. Earlier this year, the Trump administration floated the possibility of extending the ban to the H20. Officials scrapped the idea shortly thereafter and announced that 15% of the proceeds from GPU shipments to China would go to the U.S. government. A few weeks ago, U.S. President Donald Trump raised that percentage to 25% and approved the sale of H200 chips to authorized Chinese companies. According to Reuters, Nvidia currently has 600,000 H200s in its inventory. It also has 100,000 GH200 Grace Hopper Superchips, which each include a pair of H200s and a single 72-core central processing unit. Nvidia reportedly plans to make both standard H200 chips and the GH200 (pictured) available to customers in China. It’s not yet clear whether the company can go ahead with the shipments. While the U.S. has given a greenlight to H200 exports, regulators in China have reportedly not yet given their approval. Nevertheless, Nvidia is said to have asked TSMC to ramp up H200 production. Reuters’ sources said that the manufacturing push is set to begin in the second half of 2026. Most of the H200s destined for China have reportedly been ordered by large tech companies. Nvidia plans to fulfill the first batch of orders by February using its existing inventory of 700,000 HB200s and Grace Hopper chips. The remaining 1.3 million units are set to ship later in 2026. Nvidia’s current flagship data center GPU, the Blackwell Ultra, is more than 10 times faster than the H200. The chip’s speed partly stems from the fact that it includes about twice as much HBM memory. Like the B200, the Blackwell Ultra is available on a standalone basis and as part of an accelerator that combines two GPUs with one CPU. Image: Nvidia

Physicists at the University of Oxford show they can assemble Schrödinger’s cat-like superpositions from far more exotic quantum ingredients.

OneAdvanced has launched what it bills as the UK’s first private, sovereign healthcare large language model trained on NHS primary-care data.

Technology shares faltered amid escalating tensions between the U.S. and Iran.