Google Broadcom Sourcing Shift and $50B Estimate Analysis
Google Broadcom Sourcing Shift and $50B Estimate Analysis
I'll check the current GOOGL quote and the original report in parallel.
Conclusion: The reported change is plausible, but the $50B figure is a secondhand Bernstein estimate that I could not verify against the original report. The arithmetic behind it is also stretched. The per-chip gap implies a mechanism larger than the 15–20% HBM markup that Counterpoint cites. I would treat $50B as a high-end, undiscounted lifecycle scenario, not a base case. A markup-only view gets to about $20B at the same chip count.
Market reaction: GOOGL closed at on Friday 2026-11-20 (16:15 ET), down from a previous close. The reports are dated November 18–20, but my evidence does not establish that they caused the move. Broadcom (AVGO) fell in the same session to . I did not verify the cause of that move either.
How the sourcing change would work
- Today, Broadcom sells Google a bundle covering TPU design, packaging and HBM procurement. Google's memory purchases pass through Broadcom as a middleman.
- Under the reported model, Google would buy HBM directly from Samsung, SK hynix and Micron for chips built with MediaTek. The first chip is the TPU 8t (code-named Zebrafish). Reports say Google would design the compute die, MediaTek would handle the I/O die, and Google would place the HBM orders itself. Electronics Weekly adds that Google is also negotiating wafers directly with TSMC.
- Samsung is staying inside the Broadcom channel under a July memorandum with Broadcom. SK hynix has proposed co-developing custom HBM for future TPUs.
- Google and Broadcom have not commented, and none of the reports confirm timing or scale. The change appears limited to the MediaTek-built line, not all TPUs.
The $50B estimate and its assumptions
- The figure traces to Bernstein. VLSI Korea says Bernstein's report is subscription-only and that its numbers come from Exhibit 1 as shared on social media. Those numbers are a TPU v8t unit cost of via Broadcom versus via MediaTek. The per-chip gap is , about of the Broadcom-route cost.
- The lifecycle total assumes 6 million chips. That gives about . Each million chips changes the total by roughly $8.6B, so the chip volume is a key sensitivity.
- The savings are not all HBM. BigGo's summary of Bernstein says they come partly from lower design-partner fees and partly from removing HBM intermediary costs. VLSI says the gap is mainly in the HBM line.
- The markup check is the main problem. Counterpoint estimates Broadcom's turnkey model adds 15–20% to HBM costs. Even if HBM were the entire Broadcom-route cost, a 20% markup is at most per chip. Over 6 million chips that is about . Reaching $50B would need savings that are not explained by the Counterpoint markup. Design fees, packaging, or different HBM pricing would have to account for the rest, and the sources do not show how.
- VLSI notes that the per-chip HBM saving (USD 8,561) nearly equals the total partner gap (USD 8,562). It calls this unlikely to be coincidence but does not explain it. I would not rely on that match.
- The savings are not all cash to Google. BigGo reports that the cost reduction could feed higher TPU volumes, so the lifecycle figure is a gross cost number, not a P&L line.
Risks to the savings case
- Memory pricing: HBM remains short, and suppliers currently hold the upper hand. Some long-term contracts already have price ceilings. Direct buying would shift bargaining power only gradually.
- Execution: Google would take on HBM qualification, allocation, logistics and inventory. Packaging responsibility is not clearly described in the sources, so I could not confirm it.
- Partner dependence: MediaTek's capacity and the TPU 8t schedule are unverified. A slip would delay any savings.
- Competing supply paths: Samsung's Broadcom tie and SK hynix's custom-HBM proposal could leave some supply and pricing locked in other channels.
- Broadcom response: Broadcom could cut prices or lose design volume. The sources do not show this, so treat it as a risk, not a finding.
What would change the view: a Google or Broadcom statement, the full Bernstein exhibit, memory supplier contract disclosures, or TPU 8t shipment and timing data.
Data as of: GOOGL and AVGO quotes are from the Friday 2026-11-20 close (16:15 ET). News items are dated 2026-11-18 to 2026-11-20.
Sources
- Driven quote data: GOOGL and AVGO, 2026-11-20 close.
- TrendForce, "Google Reportedly Eyes Direct HBM Purchases for TPUs," 2026-11-18 (summarizing Chosun Biz): https://www.trendforce.com/news/2026/11/18/news-google-reportedly-eyes-direct-hbm-purchases-for-tpus-reducing-broadcom-related-costs/
- DIGITIMES, "Google looks to bypass Broadcom with direct HBM purchases," 2026-11-19: https://www.digitimes.com/news/a20261119VL217/google-hbm-broadcom-mediatek-2027.html
- Chip Dispatch, "Google's Direct HBM Move for MediaTek TPU 8t," 2026-11-19 (cites Counterpoint's 15–20% markup): https://chipdispatch.com/articles/google-to-buy-hbm-directly-for-mediatek-built-tpu-8t-bypassing-broadcom
- Electronics Weekly, "Google's opportunity to rejig TPU supply chain," 2026-11-20: https://www.electronicsweekly.com/news/business/google-rejigging-tpu-supply-chain-2026-11/
- AIB, "Google Weighs Direct HBM Purchases for TPUs," 2026-11-18: https://www.aib.vote/en/news/google-direct-hbm-purchases-tpu-costs
- BigGo, "Bernstein Estimates $50 Billion in Savings Over TPU Lifecycle," 2026-11-20: https://finance.biggo.com/news/b4d1fa73-eef1-4cc3-b0af-0ce914aa1470
- VLSI Korea, "Google TPU Unit Cost: MediaTek vs Broadcom Breakdown," 2026-11-18 (summarizes Bernstein's 2026-11-12 Exhibit 1, unverified): https://www.vlsi.kr/en/google-tpu-unit-cost-mediatek-broadcom-en/