Samsung to Supply HBM4 Chips to OpenAI: What’s Behind the Deal and Why TSMC and Broadcom Matter

If you thought the AI race was just about algorithms and computing power, I’m here to disappoint (or delight) you. Beneath all that grandeur lies a simple, very expensive piece of silicon. More precisely — memory. And today, March 20, the Korean press broke a story that forces us to look at the balance of power in the AI hardware market from a new angle.

For US-based AI startups and tech giants alike, this deal signals a major shift in the semiconductor supply chain. With the US government actively pushing for domestic AI infrastructure and reducing dependency on a single supplier, Samsung’s emergence as a key HBM4 player offers American companies more leverage and diversification options.

According to Reuters, citing the Korean Economic Daily, Samsung Electronics plans to supply OpenAI with its latest HBM4 memory chips. And not just anywhere — for the first proprietary AI processor being developed by the creator of ChatGPT. There are at least three reasons this is worth discussing.

The Numbers That Matter

Let’s start with the dry but impressive numbers. The deal involves supplying up to 800 million gigabit 12-layer HBM4 chips. For the uninitiated: HBM (High Bandwidth Memory) is a type of memory that runs an order of magnitude faster than conventional DRAM and is packed so densely it takes up minimal space. Without it, modern AI chips are just piles of metal.

Shipments, according to leaks, will begin in the second half of this year. And this is not just a commercial contract. It’s a bid for leadership in a segment where, until recently, SK Hynix was considered the dominant player, with Samsung, shall we say, playing catch-up.

The Dream Team: Samsung, OpenAI, Broadcom, and TSMC

Now for the most interesting part. OpenAI’s proprietary AI processor is not a solo project. It is being developed in collaboration with Broadcom, and will most likely be manufactured by TSMC (starting in the third quarter, with plans to ramp up by the end of the year). So we have a single chain consisting of: the customer (OpenAI), the custom chip designer (Broadcom), the manufacturer (TSMC), and the memory supplier (Samsung).

Recall that last year, OpenAI signed a preliminary agreement with Samsung for memory supply for data centers as part of the Stargate project. Back then, it seemed like a statement of intent. Now — we have concrete volumes and a specific product.

And here’s another detail that adds fuel to the fire. Two days ago, on March 18, Samsung and AMD signed a memorandum to expand their partnership. Under the agreement, Samsung becomes a key supplier of HBM4 for AMD’s future AI accelerators. So the Korean giant is simultaneously supplying memory to both OpenAI and NVIDIA’s main competitor in the AI chip market. That’s what you call playing all the fields.

Why This Matters

The HBM memory market is arguably the hottest battleground in the semiconductor industry right now. Without it, building large language models, training them, and performing inference at scale for millions of users is impossible. Until recently, SK Hynix dominated this space, having been the first to master HBM3 and supply it to NVIDIA. Samsung was considered the laggard.

The contract with OpenAI and the partnership with AMD send a clear signal to the market: Samsung has not just caught up, but has moved to the next level. HBM4 is the next generation, and if the Koreans manage to establish themselves here, the balance of power will shift dramatically.

OpenAI, for its part, is solving its own strategic challenge: reducing dependence on NVIDIA. A proprietary chip not only saves on procurement costs but also allows the company to tailor the hardware to its own algorithms. Broadcom here acts as the classic design house, helping to customize the architecture. And TSMC, as always, takes on the most complex part — manufacturing.

The Bottom Line

At the time of this publication, neither Samsung nor OpenAI is officially commenting on the details. But the structure of the deal appears both logical and inevitable. We are witnessing the formation of a new ecosystem: AI developers (OpenAI, Anthropic, Google) are increasingly thinking about their own hardware, while memory and chip manufacturers are forced to balance between different customers to avoid becoming too dependent.

For the average user, this could mean one thing: AI services will become cheaper and faster as competition in the hardware space reaches its peak. For investors, it’s another reminder that the stakes in semiconductors are rising every quarter. And for tech intrigue enthusiasts, it’s a pleasure to watch Samsung — a company that until recently was seen as a follower — seize the initiative in a segment that will determine the future of artificial intelligence.

Time will tell. And if the announcements hold, we’ll know more in the second half of this year.