flash
Flash fabricator Sandisk stock price was up on news of a leaked memo stating it had set up a multi-year flash supply agreement with Meta Platforms.
The source is Reuters which cites internal documents from Meta discussing the manufacture from September of an AI chip called Iris as part of its strategy to expand its AI computing infrastructure. Sandisk, Samsung and Sumitomo Electric have roles in the infrastructure buildout with long-term agreements (LTA) for NAND (Sandisk), DRAM (Samsung) and fibre-optic kit (Sumitomo Electric).
Such LTAs have become more common in the DRAM and HBM area as consistently growing-and AI-drive demand has met long-term supply constraints resulting in sharp price roses by an oligopolistic group of suppliers. They can involve supply commitments with price floors and ceiling and advance customer financial commitments. These provide revenue receipt and manufacturing capacity certainty to the supplier and budget predictability and supply certainty to the customer.
This is the first time that we have become aware of an LTA in the NAND area, and it signals how important flash is to Meta's AI infrastructure plans.
The details of Meta’s LTAs with the three suppliers are not known but we can assess that the components involved are critical for Meta’s AI infrastructure plans. Meta has a four types of its Meta Training and Inference Accelerator (MTIA) chip, with MTIA 300, 400, 450 and 500 generations. It is partnering Broadcom in areas such as chip design, packaging and network links. Meta said these new chips have either already been deployed or are scheduled for deployment in 2026 or 2027, expanding workload coverage from ranking and recommendation (R&R) inference to R&R training, general GenAI workloads, and GenAI inference with targeted optimizations.
It said: “Given the rapid pace of AI innovation, we have built the capability to ship a new chip roughly every six months.” And: “We achieve high velocity through a reusable and modular design across all levels: chiplets, chassis, racks, and network infrastructure. We architect our accelerators as systems of chiplets — discrete, reusable building blocks for compute, I/O, and networking. Because each chiplet can be upgraded separately, we can implement improvements in months rather than years.”
A third point: ”At the system level, MTIA 400, 450, and 500 all utilize the same chassis, rack, and network infrastructure. Therefore, each new chip generation can be dropped into the same physical footprint, accelerating the transition from silicon to production deployment.”
We understand Iris is the MTIA 400 chip and production is slated to start in September. The MTIA 450 is code-named Arke, with the MTIA 500 being Astrid. Our sister publication, The Next Platform, discusses the MTIA chip architectures and features here.
The Reuters report says Meta intends to deploy 7GW of compute infrastructure this year and has a 14GW target for 2027. Meta could spend up to $145 billion on AI infrastructure this year.
We don’t know what kind of flash Sandisk will supply to Meta and whether it will be chips or drives. Sandisk refused to comment when Reuters asked for details. We might expect to hear more at Sandisk’s next quarterly results announcement as analysts will be agog for details. We might also expect other hyperscale IT service providers offering AI services to be pursuing similar deals with NAND manufacturers.