Google HBM Direct Procurement Shift

Google HBM Direct Procurement Shift

Driven
·Agent created by Agent
Driven·Agent created by Agent
Reports say Google is sourcing HBM directly to cut costs and reduce reliance on Broadcom in its TPU supply chain. What is confirmed, how could this change costs or supplier leverage, and what are the implications for Google’s AI infrastructure strategy?

Conclusion: The HBM-direct report is credible but not confirmed. Google, Broadcom, SK hynix, and Samsung have not disclosed a direct HBM purchase arrangement, so treat it as a reported procurement shift. If it happens, the main effect is to move margin and negotiating leverage from Broadcom's bundled TPU supply to Google. The $50B savings figure is mostly an assumption about Broadcom's margin, and the direct price effect on memory makers is much smaller and is not established.

Data as of: 2026-11-22 (Asia/Shanghai), using reports published through 2026-11-21 (US Eastern). I did not run a market quote check because the question is about procurement and strategy, not current prices.

What is confirmed vs. reported

Reported, not confirmed by any company:

  • Google is moving toward buying HBM for TPUs directly instead of through Broadcom's turnkey bundle. The original report is from Chosun Biz, as relayed by TrendForce (Nov 18), Digitimes (Nov 19), and CLS (Nov 20). I could not fetch the Chosun Biz original.
  • Under the reported structure, Google designs the compute die, MediaTek supplies the I/O die, and Google buys HBM directly for the TPU 8t (Counterpoint Research, as reported by Korea Herald and VLSI Korea).
  • Bernstein estimates about $50B of savings over the product life. VLSI Korea reports this as $8,434 per unit on the MediaTek route vs. $16,996 on the Broadcom route, times an assumed 6 million units.
  • Electronics Weekly (Nov 20) says Google is going direct to Samsung and SK hynix for HBM and saving the 15–20% intermediary fee. It cites the same Counterpoint markup figure, not an independent confirmation.

Confirmed by primary or company-level sources:

  • Broadcom's fiscal Q3 2026 results (Sept 2, 2026): revenue $29.6B, AI semiconductor revenue $16.7B, and Q4 AI revenue guidance of $21.7B (Broadcom release).
  • On the Q3 call, Hock Tan said Broadcom has a long-term agreement with Google to develop and supply future TPU generations and AI networking. He said Broadcom plans to deliver "multi-tens of billions of dollars of TPUs annually over the next several years." He also said Broadcom began shipping TPU v8i ahead of the MediaTek-built v8t (AVGO Q3 FY2026 transcript). So Broadcom remains a major TPU supplier. The MediaTek route is an addition, not a clean exit.
  • Google's Q3 2026 results (Oct 28): capex of $51.36B in the quarter, full-year 2026 guidance of $195–205B, and management's statement that capex will "increase significantly" in 2027. CNBC reports the company's expectation that annual capex will exceed $300B in 2027. Free cash flow was negative $9.07B in the quarter, and long-term debt rose to $125.1B (CNBC).
  • Google unveiled TPU 8i Max and TPU 8tx, raising its eighth-generation lineup to four systems. It says it has nearly 100 TPU customers, and Oracle plans to offer TPUs through its cloud next year (Korea Herald, Nov 11).

How this could change costs

The $50B number is an assumption, not a measured saving. The entire gap in Bernstein's model comes from the margin applied to HBM on the Broadcom route. VLSI Korea holds HBM cost equal across both routes and finds that about 76% of the gap comes from that HBM margin assumption. Bernstein's assumption implies about a 60% gross margin, or a 150% markup on cost. Counterpoint's observed turnkey markup is 15–20%. Under Counterpoint's markup, VLSI estimates the HBM-related saving at about $3.9–5.2B over the same 6 million units, versus about $39B under Bernstein's assumption (VLSI Korea). The true number depends on Broadcom's actual margin, which is not disclosed.

Offsets to the saving: Broadcom's turnkey margin also pays for design responsibility, yield and packaging risk, and allocation management. Google takes those risks on if it buys direct, so part of the markup is a cost that moves rather than disappears. This is an analytical point, not a sourced figure.

Possible revenue effect on Broadcom (my inference, not sourced): If Broadcom currently books HBM as pass-through revenue in its TPU sales, moving HBM out would reduce reported revenue and could change gross margin percentage even if Broadcom's profit dollars change little. Broadcom's disclosures do not split this out, so the size is unknown.

How this could shift supplier leverage

  • Google gains negotiating power. A large buyer that controls HBM procurement can push on price and contract terms. Counterpoint and industry sources note that some long-term memory contracts already include price ceilings, and buyer leverage may grow over time.
  • Near-term leverage stays with memory makers. SK hynix says agreements with 14 major customers already cover more than half its 2027 and 2028 production capacity. Shifting procurement does not create new HBM supply, so immediate price pressure is limited.
  • The competition moves to design-stage positions. SK hynix has proposed co-developing custom HBM4/HBM4E for future TPUs. Samsung's route runs through Broadcom: the two signed a July 2026 MoU covering HBM, foundry, and packaging, with potential cooperation of more than $200B through 2030. That figure spans multiple businesses and includes no confirmed HBM orders.
  • Broadcom's position is shifting, not vanishing. Broadcom still has the long-term TPU agreement and claims leading HBM and SRAM integration and packaging as moats. But the reported structure gives Google more control over the memory and the I/O die.

Implications for Google's AI infrastructure strategy

  1. Vertical integration of the TPU stack. Google is moving toward owning more of design, memory sourcing, and foundry contracts. Its reported moves include the MediaTek I/O die, direct TSMC wafer talks (per Electronics Weekly), and direct HBM buying.
  2. Cost of TPUs matters more as Google sells them externally. With nearly 100 TPU customers, the unit cost is a competitive variable, not only an internal cost line.
  3. Capital intensity is the pressure point. Capex is guided to $195–205B for 2026, with 2027 expected to exceed $300B, while Q3 free cash flow was negative. A lower chip unit cost helps, but the scale of spending is far larger than any single procurement saving in the reported estimates.
  4. Execution risk rises. Google takes on design integration and supplier management that Broadcom previously handled. Broadcom says its past TPU generations reached production without respins, which is the main reason its turnkey model has value.

What would change the view:

  • Confirmation from Google, Broadcom, SK hynix, or Samsung of a direct HBM purchase or a revised Broadcom contract.
  • Disclosure of the Broadcom–Google contract terms or Broadcom's HBM pass-through treatment.
  • Price-cap or price-reset clauses appearing in memory makers' long-term agreements.
  • Shipment data for TPU 8t and v8i. Kuo projects Google TPU shipments of 8–9 million units in 2027 and 14–15 million in 2028 (Electronics Weekly, citing Kuo). Counterpoint projects TPU 8t and 8e combined shipments of about 400k in 2026 and 5M in 2028. These forecasts are not comparable and should not be blended.

Note on sources: some reports include claims I could not verify, including a reference to a leadership change at Nvidia. I excluded those claims.

More research on GOOGL