In the computing accelerator segment, NVIDIA’s main production partner remains TSMC, whose management recently admitted that it will not be able to fully meet demand for such services until the end of next year. NVIDIA isn’t going to sit idly by and says it’s ready to recruit new contractors to solve the problem.
Recall that TSMC CEO Mark Liu recently stated that the company covers only 80% of the current market need for chip packaging services using CoWoS technology, which is also used for the production of NVIDIA A100 and NVIDIA accelerators H100 families are necessary as well as their “Chinese” modifications of the A800 and H800. By the end of next year, TSMC expects to double its capacity in this area and is building a new special factory in Taiwan.
NVIDIA CFO Colette Kress attended two technology conferences last week, hosted by Evercore and Citi, respectively. In both cases, the burning issue of chip packaging capacity constraints was addressed, and a company spokeswoman made no secret of the fact that the company is actively working with all suppliers involved in the production of computing accelerators to overcome existing limitations. “This means we need to add new suppliers, find additional capacity… and also optimize the production cycle to shorten its duration.” – Colette Kress explains.
She immediately added: “Remember that we can involve many suppliers in the process CoWoS with the aim of increasing delivery volumes and we did it. You will see this become part of our efforts to increase delivery volume, the number of suppliers will continue to increase.”. Basically, NVIDIA is completely dependent on TSMC directly in the area of testing and packaging products using CoWoS technology, as this is their expertise. However, in some secondary phases, NVIDIA could certainly attract additional suppliers to accelerate the work of the Taiwanese partner. By the way, the Intel company, which aims to develop services for testing and packaging chips using advanced packaging solutions, has recently almost openly proposed itself as an “assistant” in this area. In addition, Samsung Electronics is now ready to supply HBM3 memory for NVIDIA accelerators alongside SK Hynix.
Confirming the seriousness of working with suppliers, the NVIDIA CFO said conferences Citi showed a slide showing the growth dynamics of the book value of the company’s inventory and purchasing obligations to partners. At the base of the histogram there are light green bars that show inventory costs. Not much has changed in this regard in the last few quarters. The central dark green segment shows an increase in commitments to partners for the purchase of components and has grown significantly in the last quarter. The upper part of the eggplant-colored histogram corresponds to advance payments that ensure the required supply quantities and quotas for the production of chips at contracted companies. Its value has remained almost unchanged in recent quarters.
Colette Kress also reiterated that NVIDIA compute accelerator shipments will increase quarterly over the next fiscal year. This is ensured by working with contractors and attracting additional suppliers for the CoWoS packaging segment. According to a company representative, relationships with contractors in this area are long-term.