Google's TPU expansion opens a second HBM buyer for Samsung and SK hynix
Google's expansion of its custom TPU business to external customers could open a second giant buyer for high-bandwidth memory: Morgan Stanley projected in September that Google will absorb 36.0% of global HBM demand by 2027, just behind Nvidia's 37.3%, with Samsung pursuing the opportunity through Broadcom and SK hynix proposing custom HBM for future TPUs.
📊 Scorecard: our Bullish call on $GOOGL is scored from $293.45, the last close before it ran. Its first result comes after the next close.
The News
Google's expansion of its custom TPU business to external customers could widen the high-bandwidth memory market beyond Nvidia's GPU ecosystem, according to The Korea Herald, as reported on November 11. Google unveiled two new TPU systems on November 9 — the confidential-computing TPU 8tx, built on the same silicon as the training-oriented TPU 8t (216GB of HBM per chip), and the inference-focused TPU 8i Max, a dual-chip package with 384GB of HBM and 17.2 TB/s of memory bandwidth — with external commercial deployment scheduled for early 2027.
Morgan Stanley projects Google will absorb 36.0% of worldwide HBM demand by 2027, trailing Nvidia's 37.3% by just 1.3 percentage points — figures from the bank's estimate reported by ChosunBiz in September, repeated in coverage of this week's TPU launch. Samsung is pursuing the opportunity through Broadcom, Google's TPU design partner — the two signed a memorandum of understanding in July covering HBM, foundry work and advanced packaging, with potential cooperation valued at more than $200 billion through 2030 — a figure that covers multiple businesses and does not represent confirmed HBM orders. SK hynix has proposed jointly designing customised HBM for future TPUs.
Counterpoint Research forecasts HBM bit demand for custom AI server chips will rise 35-fold between 2024 and 2028. No vendor allocations, contract values or unit pricing were disclosed — the demand is projected, not yet contracted — but the direction is clear: the HBM market is about to have two giant buyers instead of one.
For the first time, the HBM market has a second giant buyer — and memory makers can play them off each other.
Why It Matters
The HBM market's structure is changing, and structure matters more than any single order. Until now, Nvidia has been the buyer that mattered — the customer every memory maker had to win. Google at a projected 36.0% of demand by 2027 in Morgan Stanley's September estimate, against Nvidia's 37.3%, breaks that single-buyer pattern. Two giants competing for the same scarce memory means the sellers — Samsung and SK hynix — gain pricing power they did not have.
The hardware numbers back the projection. A TPU 8i Max carries 384GB of HBM and 17.2 TB/s of memory bandwidth, and Counterpoint forecasts HBM bit demand for custom AI server chips will rise 35-fold between 2024 and 2028. That is not a niche product line — it is a second AI hardware ecosystem growing at the same scale as the first, and it needs the same memory.
The race to supply it is already on. Samsung is working through Broadcom, Google's design partner, under a deal worth more than $200 billion through 2030; SK hynix has proposed jointly designing custom HBM for future TPUs. Neither has disclosed an allocation yet, which is the honest caveat — but the competition itself is the point. When two memory giants bid for a second giant buyer, the memory cycle gets longer.
The two TPU systems show why the demand is real and not just a forecast. The TPU 8tx is built for confidential computing on the same silicon as the training-oriented TPU 8t — with 216GB of HBM per chip — while the TPU 8i Max is built for inference, the work of running finished models for users, as a dual-chip package with 384GB of HBM and 17.2 TB/s of memory bandwidth. Both need vast amounts of high-bandwidth memory, and both are aimed at external customers from early 2027. That is a second AI hardware ecosystem with the same appetite as the first.
Analyst's Take
Our call is Bullish. A market with one giant buyer is a market where the buyer sets the price; a market with two giant buyers is a market where the sellers gain leverage. Morgan Stanley's September projection — Google at 36.0% of HBM demand by 2027 against Nvidia's 37.3% — turns Samsung and SK hynix from suppliers of one customer into suppliers of a contest, and that is worth more than any single order.
What to watch: who actually wins the TPU memory orders. Samsung is pursuing the opportunity through its Broadcom partnership, SK hynix has proposed jointly designing custom HBM for future TPUs, and no allocations have been disclosed. The first disclosed contract will tell us which memory maker the second buyer chose — and whether the 36% projection turns into real revenue.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is a TPU?
A TPU is a Tensor Processing Unit — Google's own custom chip for training and running AI models, the equivalent of Nvidia's GPUs but designed by Google for its own systems. Until now TPUs were mostly used inside Google; the expansion opens them to external customers, which means Google needs far more of the memory that goes with them.
What is HBM?
HBM stands for high-bandwidth memory — a specialised memory chip stacked in layers and placed next to an AI processor, feeding it data far faster than ordinary memory. Nvidia's GPUs and Google's TPUs both use it, and it is the most fought-over part of the memory market.
Why does a second buyer matter?
Until now, Nvidia has been the dominant buyer of HBM, which gave it enormous leverage over memory makers. A second giant buyer — Google at a projected 36.0% of global demand by 2027 in Morgan Stanley's September estimate, against Nvidia's 37.3% — gives Samsung and SK hynix a second customer to sell to, and pricing power they did not have when there was only one.
What is Broadcom's role?
Broadcom is Google's TPU design partner — it helps design the chips. Samsung is pursuing the TPU opportunity through Broadcom: the two signed a memorandum of understanding in July covering HBM, foundry work and advanced packaging, with potential cooperation valued at more than $200 billion through 2030 — a figure that covers multiple businesses and does not represent confirmed HBM orders.
When do external TPUs arrive?
Google unveiled two new TPU systems on November 9 — the confidential-computing TPU 8tx, built on the same silicon as the training-oriented TPU 8t (216GB of HBM per chip), and the inference-focused TPU 8i Max, a dual-chip package with 384GB of HBM and 17.2 TB/s of memory bandwidth — with external commercial deployment scheduled for early 2027. No vendor allocations, contract values or unit pricing have been disclosed.
👁
Source: The Hindu BusinessLine