OpenAI and Anthropic could start buying part of the memory chips they need for their AI infrastructure directly from manufacturers starting in 2027. The forecast, reported by DIGITIMES based on South Korean industry sources, points to greater involvement by AI companies in the high-bandwidth memory (HBM) and other DRAM supply chain, a shift that could change the customer mix at Samsung Electronics, SK hynix, and Micron.
OpenAI and Anthropic’s memory purchases in 20 seconds
- OpenAI and Anthropic could expand direct purchases of HBM and DRAM starting in 2027.
- The goal would be securing memory in advance for their growing computing infrastructure.
- Samsung, SK hynix, and Micron could add the major AI labs as direct customers.
- Some memory would still arrive through server makers, chipmakers, and cloud providers.
- The shift could bring memory purchasing closer to custom AI accelerator development.
The available information does not yet describe specific contracts signed by OpenAI or Anthropic with the three major memory manufacturers. DIGITIMES describes a market expectation around an expansion of direct purchases starting in 2027. The scenario should therefore be read as a supply-chain forecast, not a confirmed purchase announcement from either company.
The shift makes sense within an AI industry where access to memory has become an increasingly important issue. Accelerators need HBM to keep their compute cores fed, and servers also require large amounts of conventional DRAM. As data center deployments grow, securing memory capacity well in advance can become as important as reserving chip manufacturing capacity.
OpenAI and Anthropic want to get closer to memory
For much of the expansion of generative AI, the major labs haven’t necessarily been the direct buyers of all the memory used in their systems. A significant part of the chain runs through server makers, accelerator vendors, data center operators, and cloud service providers.
That model could change as companies like OpenAI and Anthropic get more directly involved in designing their own infrastructure.
DIGITIMES notes that both companies could expand direct purchases of HBM and other DRAM products starting in 2027 to secure resources for their computing infrastructure. The forecast comes from South Korean industry sources and is presented as an expected development, not a confirmed contractual timeline.
For memory manufacturers, the difference matters. Instead of selling mainly to server, GPU, and CPU makers or large data center operators, they would also have as direct customers the companies that develop AI models and increasingly control their own computing architecture.
That shift could reshape the business relationships in a chain that’s already under heavy demand.
HBM has become a strategic piece
High-bandwidth memory, known as HBM, is used alongside AI accelerators to provide high bandwidth between memory and processor. Producing it requires advanced stacking and packaging technologies, which limits the number of manufacturers capable of producing it at scale.
Samsung, SK hynix, and Micron are currently the three major memory manufacturers in the scenario described by DIGITIMES. The outlet suggests that greater direct purchasing by OpenAI, Anthropic, and other AI companies could diversify these companies’ customer base.
The pressure on memory isn’t limited to HBM alone. AI infrastructure also needs conventional DRAM for servers and other system components.
This helps explain why the decision to buy directly could have a broader scope than simply acquiring HBM for GPUs.
The next step could be linking memory and custom chips
The shift could also bring together two parts of the chain that have so far been managed more separately: memory and AI accelerators.
DIGITIMES notes that a closer direct relationship between AI companies and memory manufacturers could open up new collaboration opportunities in custom memory and chips.
Developing proprietary accelerators is already part of the strategy at several major tech companies. OpenAI, for example, is working on its own accelerators and maintains a relationship with Broadcom to develop custom computing infrastructure. Broadcom has indicated that OpenAI plans to deploy 1.3 GW of its accelerators in 2027 and that both companies are also working on later generations.
Anthropic is also increasing its reliance on specialized infrastructure. Broadcom expects the company to deploy 5 GW of TPU v8i in 2027, following roughly 1 GW of Ironwood in 2026, according to forecasts shared by its management.
None of this means OpenAI or Anthropic will manufacture their own memory chips directly. The possible change sits at a different level: getting more directly involved in purchasing and planning the memory component their accelerators and servers need.
For a company designing its own computing architecture, knowing in advance how much HBM and DRAM will be available can matter for planning how many systems it will be able to deploy.
Direct purchasing wouldn’t eliminate intermediaries
One of the most important consequences is that a potential move to direct purchasing wouldn’t necessarily mean abandoning current channels.
AI companies would still need servers, boards, accelerators, power systems, networking, and other components. Some memory could also continue arriving already integrated into platforms supplied by other manufacturers.
Some analyses drawn from an Edgewater report suggest, for example, that Anthropic could directly acquire around 20% of its memory needs in 2027 and increase that share afterward. The same analysis estimates that its direct DRAM purchase volume could reach a scale comparable to NVIDIA’s. However, these figures come from a secondary estimate and do not appear in the DIGITIMES report as data confirmed by Anthropic, so they shouldn’t be treated as purchase commitments.
The difference between the two models would therefore be one of degree, not necessarily complete replacement.
| Supply model | Role of the AI company |
|---|---|
| Traditional purchase | Receives memory integrated into servers or platforms |
| Purchase through chipmakers | Memory arrives bundled with the accelerator or system |
| Direct purchase | The company negotiates part of the supply directly with the memory manufacturer |
| Capacity agreements | May try to secure future production before it’s needed |
The interest in securing capacity also has a market context. Memory manufacturers are increasing capacity dedicated to HBM and other AI-related products, while data center demand keeps growing.
Micron, for example, is making major investments to expand its memory production, including HBM, ahead of expected growth in AI infrastructure demand.
Memory manufacturers would gain customers with more decision-making power
For Samsung, SK hynix, and Micron, a greater presence of AI labs as direct buyers would mean a change in the makeup of their customer portfolio.
The advantage for these companies would be having a closer business relationship with the parties that directly determine how much computing power they need to build and which architectures they’ll use. At the same time, that relationship could demand more forward planning around capacity, specifications, delivery schedules, and memory generations.
DIGITIMES specifically frames the shift as an opportunity for the three manufacturers to diversify their customer base and develop new collaborations around custom memory and chips.
For OpenAI and Anthropic, the motivation would be different: securing access to a component that could otherwise limit how much their systems can expand.
The issue is especially relevant because AI models don’t depend on compute capacity alone. Accelerator performance is closely tied to the memory system feeding them and to the ability to move large amounts of data fast enough.
That’s why this possible move toward direct purchasing represents another step in the verticalization of AI infrastructure.
Until now, much of the conversation about this integration has centered on custom accelerators, data centers, and energy. Memory is starting to occupy a similar position in that planning.
2027 will be the year that shows whether this forecast actually turns into direct supply contracts. For now, the available information points to an industry-expected trend, not public agreements confirmed by OpenAI or Anthropic with Samsung, SK hynix, or Micron.
Frequently Asked Questions
Will OpenAI and Anthropic buy memory directly in 2027?
The forecast reported by DIGITIMES indicates both companies could expand direct purchases of HBM and DRAM starting in 2027. No specific contracts confirming these purchases have been announced in the available information.
Which memory manufacturers could benefit?
DIGITIMES points to Samsung Electronics, SK hynix, and Micron as the three major manufacturers whose customer base could diversify if AI companies increase their direct purchases.
Why do AI companies need so much memory?
Accelerators used for artificial intelligence rely on memory like HBM to provide high bandwidth, while servers also use large amounts of conventional DRAM. The growth of AI infrastructure is increasing demand for both types.
Does direct purchasing mean OpenAI or Anthropic will manufacture their own memory?
No. The available information points to greater direct purchasing from memory manufacturers, not to these companies producing their own DRAM or HBM chips.
