Foundryecosystem Report: Nvidia GPU delays; Tools; IC Prices
The ramp of Nvidia’s new Rubin GPUs is slightly delayed. Plus, Applied's new GAA tools, chip price hikes and SUMCO's new plans
By Mark LaPedus
The foundry industry is an important part of the semiconductor business. AI chips, memory and IC packaging are also important.
Nearly every week, there are several new and major announcements in these segments.
To help the industry, Semiecosystem has released the latest edition of “The Foundryecosystem Report.” This report provides a snapshot of the latest announcements in the foundry and other segments. (The report is free for readers.) Here’s what this report covers:
1—Nvidia’s Rubin GPU delays
2—Applied’s two new GAA tools
3—Silicon photonic interposers
4—More price hikes for chips
5—SUMCO’s revised wafer manufacturing plans
Nvidia’s Rubin GPU delays
The ramp of Nvidia’s new Rubin GPU chip line has been slightly delayed, due to qualification issues with the HBM4 memory devices, according to KeyBanc Capital Markets.
The delays are expected to be a minor issue—it could last only a handful of months. Still, Nvidia is targeting production of 1.5 million Rubin GPUs this year, down from 2 million in the previous forecast, according to KeyBanc Capital Markets, an investment banking firm. Nvidia did not return e-mails regarding the delays.
Based in Santa Clara, Calif., Nvidia is the leader in the AI chip market. In this segment, Nvidia’s market share is roughly 80% or more.
In recent times, Nvidia has introduced its latest chips for the AI market, including the Rubin GPU, Vera CPU, and the Groq 3 LPU. Rubin is a new GPU device, while Vera is a CPU. The Groq 3 LPU is an inference AI chip.
These chips are sold as standalone devices or in a system. In a system, Nvidia incorporates these devices into a rack of servers. These server racks, which are manufactured by various vendors, are sold to data center operators.
Meanwhile, Nvidia’s new Rubin GPU is a complex chip, which is designed to process the latest AI algorithms in data centers. Rubin also incorporates eight HBM4 memory stacks. Based on a 3nm process from TSMC, the Rubin GPU is in production. Rubin-based products, such as servers and related systems from various vendors, are supposed to be available in the second half of 2026.
However, Nvidia has encountered a slight delay here. “The ramp of NVDA’s Rubin GPU has been delayed due to issues related to the qualification of HBM4 at SK Hynix and to a lesser extent MU (Micron Technology). NVDA had initially planned to ramp volume production of Rubin with SK Hynix, but issues with SK’s base die has led to delays of qualification, with ODM mass production being delayed from June to September this year. We believe NVDA is targeting production of 1.5M Rubin GPUs, down from 2M originally as a result of this delay,” said John Vinh, an analyst with KeyBanc Capital Markets, in a new research note.
“Previously, SK Hynix was expected to be in the leadership position in terms of HBM4 qualification for Rubin. However, performance issues have resulted in SK Hynix respinning its base die. This has resulted in NVDA reducing the scale of its ramp plans this year and has resulted in Samsung taking over the leadership position on Rubin. We believe it’s likely NVDA will launch and ramp Rubin with Samsung’s HBM4. MU has largely resolved its issues with its logic die but is working through minor performance issues with its base die. It’s unclear what the issue is, but it appears that it has to do with low manufacturing yields, as we hear that HBM4 yields for MU are less than 30%. Ultimately, it’s expected that all three memory producers will be qualified as there is not enough HBM4 capacity to support Rubin if all three are not,” Vinh said.
Nvidia’s Rubin also faces delays amid ongoing supply chain adjustments, according to TrendForce. “In addition to the time required for HBM4 validation, challenges include transitioning network interconnects from CX8 to CX9, managing significantly higher power consumption, and optimizing performance under more advanced liquid cooling solutions,” according to TrendForce, a research firm.
KeyBanc, meanwhile, is expecting Nvidia to ship 60K+ NVL72 server racks and 6K Vera Rubin server racks this year. “We’re still expecting NVDA to ship 60K racks this year, while supply has the potential to upside to 70K racks. We had previously anticipated that Vera Rubin (VR) would represent 12K-14K of the 60K racks, but with the delay, we are now expecting 6K racks,” Vinh said.
What about the server racks based on the Groq 3 AI chip? “Currently, there’s demand for ~10K LPU racks with much of the volume expected to ship next year. We’re expecting mass production of these LPU racks in 4Q26 with ~1-2K racks expected to ship this year. With 256 LPUs in a rack, we’re estimating the cost of the rack at ~$5M, which is comparable to the ASP of a VR200 NVL72,” he added.
Applied’s two new GAA tools
Intel, Samsung and TSMC are separately ramping up their respective 2nm-class processes. These 2nm processes will enable the next wave of chips, which are built around a new transistor type called gate-all-around (GAA).
GAA transistors enable new and faster chips, but they are also more difficult and expensive to manufacture in the fab. “Building the complicated 3D structures inside a GAA transistor takes more than 500 process steps, many of which require entirely new ways of depositing materials with precision, repeatability and control--all within tolerances approaching the size of individual atoms,” according to Applied Materials, a Santa Clara, Calif.-based supplier of semiconductor equipment.
To address these challenges, Applied has unveiled two new systems, which can create some of the most complex features associated with GAA transistors. The new systems enable the deposition of metals and insulating dielectrics – essential materials that impact the performance and power efficiency of advanced chips.
Applied’s first new tool, called the Producer Precision Selective Nitride PECVD system, is designed to address the challenges with the shallow trench isolation (STI) process in devices.
The Precision Selective Nitride PECVD system uses a selective bottom-up deposition process to place silicon nitride only where it’s needed in the trench. “It deposits a dense silicon nitride layer on top of the silicon oxide, which helps the isolation withstand later processing steps that would otherwise recess the STI material. The process operates at low temperatures to avoid any damage to the underlayer film or structure. By preserving the original shape and height of the isolation trench, Precision Selective Nitride helps maintain consistent electrical behavior, reducing parasitic capacitance, lowering leakage, and boosting overall device performance,” according to Applied.
The Precision Selective Nitride PECVD system is now being adopted by logic chipmakers at 2nm and below.
Applied’s second new tool, called the Endura Trillium ALD system, is designed to precisely deposit metals in the most complex GAA transistor gate stacks. “By integrating multiple metal deposition steps in a single platform, Trillium gives chipmakers the flexibility to tune threshold voltage across different transistors,” according to Applied.
Silicon photonic interposers
CEA-Leti and CEA-List have announced a strategic collaboration with Powerchip Semiconductor Manufacturing Corp. (PSMC), a Taiwan foundry vendor.
The collaboration will leverage CEA-List’s RISC-V design expertise and CEA-Leti’s silicon photonics expertise to introduce high-bandwidth communication and high-efficiency computing technologies into PSMC’s 3D stacking and interposer platforms.
France’s CEA-Leti is an R&D organization. CEA-List is a specialist in smart digital systems.
More price hikes for chips
Memory suppliers have recently raised their prices. So have CPU suppliers. What’s next?
“We’re seeing indications of broad-based price increases across many analog semiconductor companies including IFX, LSCC, MPWR, NXPI, ON, Renesas, STM, with most of the price increases expected to go into effect at the beginning of 2Q26. The key factors driving the price increases include: 1) Tight supply of power management components due to strong AI server demand and increasing power requirements; 2) Higher supply chain input costs including foundry, substrate, lead frames, etc.,” KeyBanc’s Vinh said.
“There’s a range of how much ASPs are increasing. More specifically, we’re hearing power management ASP increases are averaging 5-10% for PCs and 10-15% for servers,” Vinh said.
SUMCO’s revised plans
Japan’s SUMCO, the world’s second largest silicon wafer maker, has revised its manufacturing plans.
Originally, SUMCO was supposed to construct new silicon wafer plants in two locations--Imari City, Saga Prefecture and Yoshinogari-cho, Kanzaki-gun, Saga Prefecture.
Going forward, the company has decided not to build the new plants. Instead, it will upgrade its existing plant in Imari City. SUMCO is still assessing its plans in Yoshinogari-cho.
Accordingly, the government has reduced its subsidy for SUMCO from 75.0 billion yen to 19.3 billion yen.
“The semiconductor market has recently undergone structural shifts; while demand for applications such as PCs and smartphones has stabilized, demand for generative AI is surging,” according to SUMCO. “In the leading-edge field of the 2nm generation and beyond, quality requirements are becoming increasingly stringent, and intense technological competition is anticipated. In light of these structural changes, the company has determined that, at this stage, focusing management resources on upgrading manufacturing equipment to reliably capture rapidly growing demand for leading edge products, rather than pursuing volume expansion, is the optimal course of action from the perspectives of economic rationality and competitiveness.”