ReportGem ReportGem 中文

REAL-TIME GLOBAL RESEARCH

Diving into ‘28 Compute Capacity, TPU Sales and Revenue to Come

Published: 2026-06-29Institution: Morgan StanleyCompany / ticker: GOOGL.O,AMZN.O,META.OPages: 21Original language: EnglishEvidence page: 2

Research evidence excerpt

Diving into ‘28 Compute Capacity, TPU Sales and Revenue to Come

IdeaMcombining AVGO management commentary about shipping 10 GW in 2027 with MS

estimated 2027 AVGO revenue, implying roughly $12 of ASIC revenue per watt. This

$12/W framework, when applied to their '28 AVGO revenue estimates, implies

approximately 14 GW of compute shipments in 2028, of which we estimate ~60%

(~9 GW) will be TPUs.

We also believe GOOGL remains compute constrained (consider paying $50/W

recently for compute) and therefore estimate that ~60% of AVGO TPUs, or roughly

5 GW, will be retained by GOOGL to support its core business and Google Cloud

workloads. We estimate the remaining ~4 GW of AVGO TPUs will be sold to

external customers as 1P merchant TPU sales.

In addition to AVGO, we model MediaTek (covered by Charlie Chan) to add an

additional 1 GW/1.5 GW of TPU in '27/'28. As such, in aggregate we raise our TPU

capacity added by GOOGL to 5 GW of rack power in '27 (vs 3.3 GW prior) and in

doing the bottom up TPU capacity work for the first time in '28 expect GOOGL to

add ~7 GW of TPU capacity. Lastly, we also model GOOGL to add 2 GW of

compute capacity of NVDA GPUs in '28...for total net additions of ~9 GW of on-

premises compute capacity.

Exhibit 1: We increase our estimated GOOGL capacity in '27 from 5.2 GW to 6.9

GW (rack power), and now expect ~9 GW of rack power to be delivered in '28 (~7

GW of which are TPU).

GPU TPU

Power) 8.8

(Rack 6.9

Type 6

Chip 4.6 5.2 6.6

by 4.3

5.0

3.3 Additions 3.0 3.3

GW 2 1.6 1.6

0.5 0.5

2.2 2.0 2.0 GOOGL

1.1 1.1 1.3 1.3 N/A

Prior Current Prior Current Prior Current Prior Current

2025 2026 2027 2028

Note: these GW additions exclude 1P TPU sales

Source: Company data, Morgan Stanley Research

Monetization of the capacity remains a debate but our base model assumes

The English excerpt is extracted automatically from the cited source page and may contain layout or recognition errors. It is never batch translated.

Open report viewer