The South Korean chipmaker said on Thursday that the new chips are already being delivered, though it did not name customers. 

This comes as demand for high-bandwidth memory steeply increases, driven by the global build-out of data centres used to train and run advanced artificial intelligence systems.

HBM is a form of dynamic random-access memory designed to handle very large volumes of data at high speed. It has become an essential component in modern AI processors. 

Samsung, the world’s largest memory chipmaker by revenue, has struggled in recent years to keep pace with competitors in earlier generations of the technology.

The company said its HBM4 chips provide a consistent processing speed of 11.7 gigabits per second, about 22% faster than the previous HBM3E version. It added that the chips can reach a maximum speed of 13 gigabits per second, easing data bottlenecks as workloads grow heavier.

Samsung also said it plans to provide samples of its next version, HBM4E, in the second half of the year.

Sources say Samsung began mass production and shipments of HBM4 chips in February 2026, with Nvidia graphics processors among the key targets.

Those chips are expected to power Nvidia’s upcoming Vera Rubin AI accelerator platform, due for launch in the second half of the year.

The HBM4 chips are built on a 4-nanometre logic process and offer capacities between 24 and 36 gigabytes, with plans to scale up to 48 gigabytes. 

Samsung says the new generation delivers up to 3 terabytes per second of bandwidth per stack, roughly 2.4 times that of HBM3E, alongside a 40% improvement in power efficiency and better thermal control.

Competition in the market is getting tighter. SK Hynix said in January that it aims to maintain its “overwhelming” market share in next-generation HBM4 chips, which it said were already in volume production. 

The company added that it plans to achieve production yields for HBM4 in line with those of HBM3E.

Micron has also moved early. The company’s chief financial officer said it is in high-volume production of HBM4 and has begun shipping the chips to customers.

Samsung’s new focus underlines a catch-up initiative after falling behind in earlier HBM cycles. Investors welcomed the update. Samsung shares ended the day up 6.4%.

Memory bandwidth has become just as important as processing power in modern data centres, and demand is not falling behind. Companies are expanding AI capacity worldwide.

This is an important step for Samsung as it tries to close the gap with SK Hynix, its long-time competition and the top supplier of advanced memory used in Nvidia’s accelerators. 

Delays in Samsung’s earlier HBM programmes weighed heavily on its earnings and share price last year. This time, the company is moving faster and with more at stake.

The initial production run is tied to months of qualification testing with Nvidia and AMD. Samsung cleared those tests late last year, opening the door to commercial shipments in early 2026. The company has not disclosed volumes, and it is unclear how quickly supply will scale.

Samsung declined to comment on the development. Nvidia did not respond to requests for comment.

Following the development, Samsung shares rose more than 2% in early trading, while SK Hynix shares slipped by nearly 3%. Investors are reading this as a sign that competition in the HBM market is about to increase.

HBM4 is the sixth generation of high-bandwidth memory, designed to handle the heavy data loads of artificial intelligence and high-performance computing. 

Compared with HBM3E, it gives higher bandwidth while using less power. That combination is important as chipmakers push systems harder to train and run ever-larger models.

For Nvidia, memory supply is now a strategic issue, not just a technical one. The company’s next platform, Vera Rubin, is already in production and will rely on HBM4 to achieve its performance targets. 

As Nvidia’s chief executive Jensen Huang said earlier this month, the platform is in “full production,” ahead of a launch later this year.

Until now, SK Hynix has carried most of that burden. It supplies the bulk of the HBM used in Nvidia’s current accelerators and has already locked in supply talks with major customers for next year. 

The company is also expanding capacity, with plans to start deploying wafers into its new M15X factory in Cheongju early next month. It has not said whether HBM4 will be part of that first output.

Micron is also pushing into the space, adding pressure to an already tight market. Demand for high-bandwidth memory is surging as data centres scale up, and analysts expect the market to grow sharply over the next few years.

Samsung’s entry into HBM4 production changes the balance. It gives Nvidia and AMD another qualified supplier, reduces the risk of shortages, and may help cool pricing pressures. It is a chance for Samsung to regain ground in a segment it cannot afford to miss.

Both Samsung and SK Hynix are due to report quarterly earnings later this week. Investors will be listening for any detail on HBM4 orders, production timelines and customer commitments.