EL7.AI
StocksMedium•19 November 2026•
4 min read

$50 Billion Savings Estimate Prompts Google to Separate HBM Purchases From TPU Design

Key Facts

1Bernstein estimated potential savings of about $50 billion over a TPU product’s life cycle.
2Bernstein estimated the chip’s cost at $8,434 under the MediaTek model, compared with $16,996 under the Broadcom model.
3HBM accounts for about 76% of the cost difference between the two models in Bernstein’s analysis.
4SK hynix said its agreements with 14 customers cover more than half of its total production capacity for 2027 and 2028.

Google is considering separating purchases of high-bandwidth memory (HBM) from the design of its tensor processing units (TPUs), according to a Chosun Biz report published on November 18, in a potential supply-chain reshuffle that could allow it to buy memory directly. The report said the company is working with MediaTek on the design, while it had relied on Broadcom to handle design integration, packaging and HBM procurement as a bundled package. Bernstein estimated that this structure could save about $50 billion over a TPU product’s life cycle. If implemented, the move would put memory costs and purchasing terms at the center of Google’s review of AI chip contracts. Google has not officially announced that it is making the shift, so it remains a direction reported by the press rather than a confirmed company decision.

Bernstein’s figures illustrate the scale of the difference that could result from changing the supply structure. The analysis estimated the chip’s cost at $8,434 under the MediaTek model, compared with $16,996 under the Broadcom model, for the TPU v8t chip. The comparison shows that the estimated cost under the first model is less than half that of the second. These figures, however, are estimates from the analysis, not prices disclosed by Google or either company. The report links the potential savings to a different allocation of roles in chip design, packaging and memory procurement.

HBM accounts for a key part of this cost difference. According to Bernstein’s analysis, cited by Chosun Biz, HBM costs represent about 76% of the gap between the two models, even assuming equal memory purchase costs in the comparison. That means the estimate does not rest on a lower price for memory itself, but on the effect of how it is included in the supply chain and assumed intermediary margins. In a separate estimate cited by Korea Herald, Counterpoint Research said the Broadcom model adds a premium of 15% to 20% to HBM costs. That percentage is Counterpoint’s estimate, not a margin confirmed by Broadcom. The total savings remain contingent on Bernstein’s assumptions about the supply structure and component costs.

Discussions of direct purchasing come as chipmakers compete to secure supplies of advanced memory. SK hynix said in October that its agreements with 14 major customers cover volumes equivalent to more than half of its total production capacity for 2027 and 2028. Those agreements include products beyond HBM, according to Korea Herald, putting capacity contracts in a broader context than accelerator memory alone. The newspaper also reported that SK hynix proposed designing custom HBM for Google’s future TPU chips. The proposal reflects the company’s effort to participate at an earlier stage of chip development, rather than only supplying memory after its specifications have been set.

The review of chip costs coincides with massive planned spending on AI infrastructure. Morgan Stanley forecast that cash capital expenditure by eight major companies, including Alphabet, Amazon, Microsoft and Meta, would reach $946 billion this year and then $1.656 trillion next year, an increase of about 75%. The bank said growth in this spending would slow to about 7% in 2028, despite continued expansion in investment. Meanwhile, SemiAnalysis cut its forecast for HBM shipments in 2028 to 6.4 exabytes from 9.0 exabytes, and lowered its estimate for AI accelerator spending that year to $758 billion from $1.054 trillion. These divergent estimates highlight companies’ efforts to control hardware costs as spending expands, without settling the trajectory of demand or future contracts.

In trading on November 18, Alphabet’s Class A shares closed at $282.44, while Broadcom shares ended the session at $354.05. The closing prices provide context for the two stocks’ performance in the same session, but do not establish that reports about Google’s supply structure caused their moves. Korea Herald had reported that Google unveiled two new TPU systems on November 9, the 8i Max and 8tx, expanding its eighth-generation lineup to four systems. Google said the 8t and 8i systems offer up to 2.7 times the training performance and 80% better price performance for inference than the previous generation. This comes as Broadcom announced an expanded collaboration with Google and Anthropic in April, giving the latter access through Broadcom to about 3.5 gigawatts of TPU-based computing capacity starting in 2027; this places reports of a purchasing reshuffle alongside existing, publicly announced partnerships.

Open on chart