Google is reportedly moving to procure HBM (high-bandwidth memory) for its in-house AI chip, the TPU, directly rather than through its design partner Broadcom, according to Korean outlet Chosun Biz. U.S. investment bank Bernstein estimates that the change in procurement approach could cut costs by about $50 billion (roughly 8 trillion yen) over the TPU's product life.
However, the estimate assumes the purchase cost of HBM is the same under both approaches, so the HBM-related savings correspond to the profit Broadcom adds on top. It does not factor in any drop in the selling prices of Samsung Electronics, SK hynix, or Micron. Moreover, if the assumptions about Broadcom's margin are changed, the HBM-related savings shrink from about $39 billion to roughly $3.9 billion–$5.2 billion. The $50 billion figure cannot simply be read as downward pressure on memory makers' prices.
Google buys HBM directly and expands its division of labor with MediaTek on TPU design
According to Chosun Biz, Google had until now entrusted TPU design, packaging, and HBM procurement to Broadcom as a bundle. It is now reportedly reorganizing its supply chain by sharing design work with MediaTek and securing HBM directly. Research firm Counterpoint Research explains that for the TPU v8t (codenamed "Zebrafish"), Google will design the compute chip while MediaTek handles the I/O chip.
David Wu, a research associate at Counterpoint Research, cited HBM procurement costs as the main motivation for these changes. According to the firm, under a turnkey model in which an ASIC design company like Broadcom also handles component sourcing, a 15–20% profit is added on top of the HBM purchase price. As HBM accounts for a growing share of an AI chip's bill of materials, this markup pushes up procurement costs. Counterpoint forecasts that shipments of the TPU v8t and its successor, the v8e, will grow from about 400,000 units in 2026 to nearly 5 million in 2028.
HBM makers are also seeking closer ties with Google. According to the Korea Herald, SK hynix has proposed co-designing custom HBM for Google's TPUs. Samsung Electronics, meanwhile, signed a memorandum of understanding with Broadcom in July, anticipating cooperation worth more than $200 billion through 2030. That figure, however, does not represent confirmed HBM orders. The two Korean memory giants are trying to deepen their involvement in the TPU business through different routes—supply via Broadcom, and collaboration with Google from the design stage—but neither has shown a confirmed HBM order.
Where does the $50 billion in savings come from?
Bernstein's estimate is based on an assumption of 6 million TPU v8t units shipped over a product life of roughly one to two years. According to the estimate relayed by Chosun Biz, the cost per TPU is $8,434 under the MediaTek approach and $16,996 under the Broadcom approach. The difference is $8,562, and multiplied by 6 million units it comes to about $51.4 billion. This is the basis for the roughly $50 billion savings figure. Converted at 159 yen to the dollar, that is about 8.2 trillion yen. The dollar-yen rate shown by the Bank of Japan as of 5 p.m. on November 17 was 159.33–159.35 yen.
The cost gap between the two approaches can be explained by the margins Bernstein set. For the MediaTek approach, it applied a 40% gross margin to the logic chip and packaging, and assumed Google buys HBM directly at cost. For the Broadcom approach, it applied a 60% gross margin to the entire supply, including HBM. The purchase cost of the HBM itself is said to be the same in both approaches, but about 76% of the cost gap arises from the HBM-related portion.
The key point here is the difference between gross margin and markup on cost. A 60% gross margin means 60% of the sales price is gross profit. Relative to cost, that equals a 150% markup, meaning the item is sold at 2.5 times its purchase cost.
Letting be the cost of the logic chip and packaging and the cost of the HBM, Bernstein's assumptions can be expressed as the following simultaneous equations.
Solving these gives the following values.
These are not costs published directly by Bernstein, but values back-calculated from the reported assumptions.
If a 150% markup is applied to the HBM cost of about $4,345 under the Broadcom approach, the added profit comes to about $6,518 per unit. This equals about 76% of the $8,562 cost gap between the two approaches, and reaches about $39.1 billion across 6 million units.
On the other hand, the HBM markup Counterpoint Research cites is 15–20%, very different from the 150% Bernstein assumed. Applying 15–20% to the same HBM cost of $4,345 gives a markup of about $652–$869 per unit. Across 6 million units, that is about $3.9 billion–$5.2 billion.
In other words, under Bernstein's assumptions the HBM-related savings are about $39 billion, whereas using Counterpoint's markup they shrink to about $3.9 billion–$5.2 billion. The two differ by a factor of roughly 7.5 to 10.
The remaining cost gap of about $2,044 per unit, or about $12.3 billion over 6 million units, comes from the difference in gross margins applied to the logic chip and packaging (60% versus 40%). If this portion is kept as Bernstein assumed and only the HBM markup is replaced with Counterpoint's 15–20%, total savings come to about $16.2 billion–$17.5 billion, well below the original figure of about $50 billion.
| Item | Bernstein's assumption | With HBM markup replaced by Counterpoint's estimate |
|---|---|---|
| HBM purchase cost | About $4,345/unit | About $4,345/unit |
| Markup on HBM | Equivalent to 150% of cost | 15–20% of cost |
| HBM-related savings (6 million units) | About $39.1 billion | About $3.9–$5.2 billion |
| Logic/packaging-related savings | About $12.3 billion | About $12.3 billion |
| Total savings | About $51.4 billion | About $16.2–$17.5 billion |
The actual contract terms between Broadcom and Google have not been disclosed. The 60% gross margin is an assumption in Bernstein's estimate, and the 15–20% markup on HBM is merely Counterpoint's estimate. It is impossible to judge which is closer to the actual transactions, but it is clear that the roughly $50 billion savings figure depends heavily on how Broadcom's margin is set.
Will Google's direct purchasing push HBM prices down?
Google buying HBM directly could increase pressure on memory makers to cut prices. An industry source interviewed by Chosun Biz noted that for now the makers hold the advantage because of HBM supply shortages and existing contracts, but that some long-term contracts already include price caps, and buyers' bargaining power could grow over time.
Changes are also visible in demand forecasts. SemiAnalysis cut its 2028 HBM shipment forecast by about 30%, from 9.0 EB to 6.4 EB. It also revised its forecast for 2028 AI accelerator spending down 28%, from $1.054 trillion to $758 billion, factoring in the possibility of falling HBM prices as supply and demand ease. Morgan Stanley, meanwhile, forecasts that capital spending by eight major companies will grow about 75% in 2027, but slow to about 7% growth in 2028.
Given this supply-demand outlook, Google could potentially lower its procurement prices in the future by negotiating directly with HBM makers. But that is a separate matter from the roughly $50 billion in savings Bernstein estimated. The estimate holds HBM purchase costs constant and calculates the profit markup that can be eliminated by no longer going through Broadcom. If HBM makers' own selling prices fall, that could lead to additional savings, but that effect is not included in the $50 billion.
On the supply side, makers have limited room for now. In October, SK hynix explained that its contracts with 14 major customers, including for products other than HBM, account for more than half of its 2027–2028 production capacity. Industry sources interviewed by the Korea Herald also noted that even if new customers are won, HBM shipments cannot be increased immediately.
Therefore, evaluating the savings from this change in procurement approach requires confirming how much profit Broadcom was actually adding. If the gross margin on the entire supply including HBM is close to the 60% Bernstein assumed, the HBM-related savings work out to about $39 billion. If the actual markup is closer to the 15–20% Counterpoint cites, the savings would be significantly smaller.
As for downward pressure on memory makers' prices, a key indicator will be how widely long-term supply contracts with price caps spread. In particular, attention will be on whether Samsung Electronics, SK hynix, and Micron disclose, in earnings briefings or elsewhere, an increase in contracts with price caps.
Competition over next-generation TPUs shifts from price negotiations to the design stage
For HBM makers, the near-term competition is shifting from simple discounting to how deeply they can be involved from the design stage of next-generation TPUs. Industry sources interviewed by the Korea Herald point out that the focus of competition is participating in next-generation chip design and securing needed production capacity early.
With HBM4 and HBM4E, custom designs tailored to the customer's AI chip become possible. Memory makers that join Google's TPU development early may therefore be able to win long-term supply contracts while helping shape product specifications. For Google, meanwhile, discussing directly with memory makers without going through Broadcom is a way to manage design requirements and procurement terms in-house.
If growth in HBM demand slows toward 2028, whether these long-term contracts include price caps could influence memory makers' selling prices. In judging the validity of the roughly $50 billion savings estimate, Broadcom's actual margin is a key indicator; in gauging the impact on the HBM market as a whole, the supply contracts and pricing terms for next-generation TPUs will be the important indicators.
