Google rejigs TPU supply chain, cutting Broadcom out of memory buying — but the $50B savings may shrink
Google is moving to buy high-bandwidth memory directly for its TPUs instead of through Broadcom's bundled arrangement, and splitting TPU design work with MediaTek, Korean outlet Chosun Biz reported on November 18. Bernstein puts the savings at about $50 billion — but that estimate rests on Broadcom's margin and could shrink to about $3.9–5.2 billion.

The News
Google is moving to buy high-bandwidth memory directly for its TPUs instead of through Broadcom's bundled design, packaging and procurement arrangement, Korean outlet Chosun Biz reported on November 18, with follow-up reporting on November 20. Under the old arrangement, Broadcom handled the design, the packaging and buying the memory as one turnkey package. Under the new model, Google buys the memory itself, straight from the memory makers.
Google is also splitting TPU design work with MediaTek. For the TPU 8t, codenamed Zebrafish, Google designs the compute chip while MediaTek handles the input-output chip, according to Counterpoint Research. The input-output chip moves data in and out of the processor; the compute chip does the actual AI maths. Splitting the work spreads the design load and reduces Google's dependence on a single partner.
Investment bank Bernstein estimates the procurement shift could save Google about $50 billion over the TPU product life. But that headline number needs a prominent flag: the estimate assumes the memory costs the same under both approaches, and only about 76% of the roughly $50 billion gap — about $39 billion — is HBM-related. If the margin assumptions change, that HBM-related portion shrinks to roughly $3.9–5.2 billion, according to XenoSpectrum's analysis of Bernstein's basis — so the total savings would fall, but not to $3.9–5.2 billion.
The scale of the programme is large and growing. Supply-chain analyst Ming-Chi Kuo projects Google's TPU shipments at 8 to 9 million units in 2027 and 14 to 15 million in 2028, and Counterpoint forecast in April 2026 that TPU 8t and successor 8e shipments would rise from about 400,000 units in 2026 to nearly 5 million in 2028. SemiAnalysis founder Dylan Patel told CNBC that Google is already going to Nvidia's suppliers and asking for capacity, backed by serious cash, as Jensen Huang's exit hits a capacity-constrained AI chip market. On the memory side, SK hynix is said to have proposed jointly designing customised HBM for future Google TPUs, while Samsung signed a July 2026 memorandum of understanding with Broadcom on HBM, foundry and advanced packaging potentially worth over $200 billion through 2030 — a figure that represents no confirmed HBM orders.
The $50 billion headline rests on Broadcom's margin — change the margin, and the savings shrink to a tenth of that.
Why It Matters
This matters because the TPU programme is now big enough for procurement to move the numbers. At 8 to 9 million units in 2027 and 14 to 15 million in 2028, even a small saving per chip adds up — which is how Bernstein reaches $50 billion. But the same scale cuts the other way: if the margin assumption is wrong, the savings collapse to $3.9–5.2 billion, a tenth of the headline. Readers should treat the $50 billion as an estimate built on one assumption, not a confirmed saving.
It matters for Broadcom because the rejig takes away a role, not just an order. Broadcom's bundled arrangement let it earn a margin on the memory as well as on the design and packaging. Direct buying removes that margin stream, and the MediaTek split removes design work too. That is why the AVGO call here is bearish even though the overall story is Mixed — the two companies' interests in this rejig point in opposite directions.
It also matters as a signal about the AI chip market after Jensen Huang's resignation. Patel's comment that Google is already going to Nvidia's suppliers and asking for capacity, backed by serious cash, shows a capacity-constrained market where the big buyers are competing for the same factories. Google building its own supply chain, and SK hynix proposing custom HBM for future TPUs, are part of that scramble. The Samsung-Broadcom memorandum, potentially worth over $200 billion through 2030, represents no confirmed HBM orders — so it should not be read as offsetting the rejig.
Analyst's Take
Our call is Mixed. The rejig is a real strategic shift — Google buying memory directly and splitting design with MediaTek — but the $50 billion savings figure is fragile, and the market has already punished both names in the rout. Alphabet closed Friday at $265.29, down 2.7%, and Broadcom closed at $304.41, down 12.2%. The direction of the story depends entirely on whether the savings are $50 billion or closer to $5 billion.
The Broadcom read-through is the sharper edge. If Google buys memory directly, Broadcom loses the margin on that procurement — the very margin the Bernstein estimate is built on — which is why this post carries a bearish call on AVGO even though the story verdict is Mixed. That is a structural loss of a revenue stream, not a one-off order change, and it lands on top of the reported cut to Broadcom's TPU shipments that the Nov 21 post already covered. This post is the rejig angle only; it does not re-file that shipment cut as the story.
What to watch next is whether Google or Broadcom comments, whether the direct-buying model spreads to more TPU generations, and whether the shipment numbers Kuo projects — 8 to 9 million units in 2027, 14 to 15 million in 2028 — hold up. The origin reporting is November 18, with follow-up on November 20, so this is a dated reach-back filed because the procurement-model angle was never covered. Reader poll: Does Google's TPU rejig hurt Broadcom or help Google more? Vote: Broadcom loses, Google saves big, Both win, Too early to tell.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is HBM?
HBM stands for high-bandwidth memory, a fast type of memory stacked in layers and placed next to AI chips. It feeds data to the chip quickly enough to keep up with AI training and inference. It is one of the most expensive parts of an AI chip package.
What is a TPU?
A TPU, or Tensor Processing Unit, is Google's own AI chip, designed in-house as an alternative to Nvidia's chips. Google uses TPUs in its own datacenters to run AI models, and is now also selling TPU capacity to outside customers.
What is a markup?
A markup is the extra amount a supplier charges above its own cost. The 15–20% figure is Counterpoint Research's estimate of the markup a turnkey design partner adds on HBM. Bernstein's $50 billion estimate rests on different assumptions: a 60% gross margin applied to the whole Broadcom supply including HBM (equal to a 150% markup on cost), versus a 40% gross margin on logic and packaging in the MediaTek approach, with Google buying HBM at cost.
What does rejig mean here?
Rejig means reorganise. Google is changing who does what in its TPU supply chain: buying memory directly, and splitting design work so MediaTek handles the input-output chip while Google designs the compute chip.
👁
Source: Electronics Weekly