Insights · Technology

Custom AI chips: what Broadcom, MediaTek, Amazon and the chip designers' filings show

Broadcom's revenue from AI chips rose from USD 4.1 billion in the first quarter of its fiscal 2025 to USD 16.7 billion in the third quarter of fiscal 2026, and the company guides to USD 21.7 billion for the quarter now ending. Most of it now comes from custom accelerators: chips designed for one customer, such as Google's TPU, rather than sold to everyone like a graphics processor. This page puts together what Broadcom, MediaTek, Marvell, Alchip, Global Unichip, Amazon and China's two cloud chip units have disclosed, and shows why their numbers do not add up to one market size. Every figure links to its source.

Broadcom's AI chip sales quadrupled in six quarters

Broadcom reports AI semiconductor revenue every quarter. It was USD 4.1 billion in fiscal Q1 2025 (November 2024 to January 2025), USD 4.4 billion in Q2, USD 5.2 billion in Q3 and USD 6.5 billion in Q4, about USD 20 billion for the year (Q4 FY25 release). In fiscal 2026 it was USD 8.4 billion in Q1 (release), USD 10.8 billion in Q2 (release) and USD 16.7 billion in Q3 (release), and the company expects USD 21.7 billion in Q4. It now expects about USD 58 billion for fiscal 2026 (Zacks). The fiscal 2025 quarters before Q4 are calculated from the year-on-year growth rates Broadcom gave for the matching 2026 quarters.

Bar chart: 4.1, 4.4, 5.2 and 6.5 billion dollars in the four quarters of fiscal 2025; 8.4, 10.8 and 16.7 billion in the first three quarters of fiscal 2026; 21.7 billion guided for the fourth.
Broadcom AI semiconductor revenue by fiscal quarter, USD billion; Q4 FY26 is guidance. Source: Broadcom results releases and calls.

The mix has shifted towards custom accelerators, which Broadcom calls XPUs. In fiscal Q2 2025, networking chips were 40% of AI revenue, and chief executive Hock Tan said he expected networking to be "less than 30%" going forward (Q2 FY25 call). In December 2025 he said that of USD 73 billion of AI backlog, "maybe $20 billion of it is everything else. The rest, is XPUs" (Q4 FY25 call). In fiscal Q3 2026, XPUs were 73% of AI revenue, and Broadcom said it was shipping Google's Ironwood TPU in high volume to Anthropic and Google, had started production of Google's next TPU and had shipped OpenAI's first custom accelerator (Zacks). Tan said "Anthropic is on track to become our largest XPU customer in 2027 and sustain that in 2028" (Fierce Network).

Why the numbers do not add up

On 31 October 2026, MediaTek chief executive Rick Tsai said the company's first AI accelerator chip, which started shipping in September, should bring in about USD 2.5 billion in the fourth quarter, and that he expects about USD 16 billion in 2027, about 20% of the market (Focus Taiwan). Sixteen billion at a fifth of the market implies a market of about USD 80 billion in 2027 (calculated). Broadcom, meanwhile, says it has secured supply for about USD 115 billion of AI revenue in fiscal 2027 (Fierce Network). If accelerators stay at 73% of that, the figure for its fiscal Q3 2026, Broadcom's accelerators alone would come to about USD 84 billion (calculated).

Bar chart: about 80 billion dollars for the whole market implied by MediaTek's 2027 target, and about 84 billion for Broadcom's accelerators alone at 73% of its fiscal 2027 AI revenue.
Two company views of 2027, USD billion, both calculated by Trifacta Data from the companies' own statements.

Four things explain the gap. First, Broadcom sells some customers complete systems rather than chips, and its finance chief said on the December 2025 call that it would be "passing through more components that are not ours" in those sales, so its revenue counts parts that a chip-only market would not (Q4 FY25 call). Second, a custom accelerator is sold with its high-bandwidth memory inside the package; Broadcom has said its mix is shifting "toward XPUs with higher memory content" (Zacks), and memory is a large part of each chip's price. Third, MediaTek has not said how it defines the market it measures its 20% against. Fourth, Broadcom's fiscal 2027 runs from November 2026 to October 2027, not the calendar year. A market estimate depends on which of these it counts.

The other design partners: a lumpy business

Behind Broadcom, a group of companies turn customers' designs into finished chips and manage their production. Their filings show how much depends on a few programmes. Alchip's revenue fell from USD 1.6 billion in 2024 to USD 992 million in 2025, which the company put down to limited production shipments; one customer had been 60.16% of its 2024 sales (2024 annual report), design fees were over half of its Q4 2025 revenue, and it expects a ramp in 2026 as a 3nm accelerator for a North American cloud customer enters production (presentation). At Global Unichip (GUC), cloud products rose from 66.8% of revenue in January to September 2025 to 86% in Q3 2026, mainly CPUs, AI accelerators and server management chips (presentation). Marvell said its custom chip revenue doubled to USD 1.5 billion in fiscal 2026 and should grow more than 20% in fiscal 2027; its largest direct customer, which it does not name, was 14% of revenue (10-K).

CompanyRoleLatest disclosureSource
BroadcomDesign partner (US)AI revenue USD 16.7bn in fiscal Q3 2026; 73%% custom accelerators; Q4 guide USD 21.7bn8-K, Zacks
MediaTekDesign partner (Taiwan), newFirst AI accelerator chip shipped Sep 2026; about USD 2.5bn in Q4 2026, about USD 16bn in 2027Focus Taiwan
MarvellDesign partner (US)Custom revenue USD 1.5bn in fiscal 2026 (Feb 2025 to Jan 2026), doubledQ4 FY26 call
AlchipDesign partner (Taiwan)Revenue USD 992m in 2025 (USD 1.6bn in 2024); 3nm ramp from 2026Results
Global Unichip (GUC)Design partner (Taiwan)Cloud 86%% of revenue in Q3 2026 (66.8%% in Jan to Sep 2025)Statements
Amazon (Annapurna Labs)Designs its ownOver 1.4 million Trainium2 chips landed; over USD 225bn of Trainium revenue commitmentsQ4 2025 call, Amazon
Alibaba (T-Head)Designs its own (China)560,000 Zhenwu AI chips shipped by Apr 2026; Hong Kong listing filedEastmoney, Reuters via ET
Baidu (Kunlunxin)Designs its own (China)2025 revenue expected above RMB 3.5bn, over half outside Baidu; Hong Kong listing filedReuters via The Standard

The customers that build their own

Some of the largest buyers do much of the work themselves. Amazon chief executive Andy Jassy said in February 2026 that Amazon had "landed over 1.4 million Trainium2 chips, our fastest ramping chip launch ever" (Q4 2025 call), and in April that Trainium had over USD 225 billion of revenue commitments, that Trainium3 had started shipping at the start of 2026 and that Amazon Bedrock runs most of its inference on Trainium (Amazon). These are Amazon's own figures for chips and commitments; it does not say what each chip costs or how much of the value passes through its design partners.

In China, the two largest cloud chip units are heading for the stock market. Alibaba said T-Head had shipped 560,000 Zhenwu AI chips by April 2026 and that T-Head's annualised revenue, across all its products, had reached about RMB 10 billion (Eastmoney); T-Head filed confidentially for a Hong Kong listing on 11 September 2026 (Reuters via Economic Times). Baidu's Kunlunxin filed for a Hong Kong listing in January 2026 (MediaPost); Reuters reported that it expected 2025 revenue above RMB 3.5 billion, more than half from customers outside Baidu (The Standard). MediaPost reported that Kunlunxin's chips handle the vast majority of Baidu's search inference, and that Baidu finds its own hardware more cost-effective than expensive, restricted foreign GPUs (MediaPost).

What is still uncertain

First, Google does not disclose how many TPUs it buys, and Broadcom does not name its customers' shares, so the largest single programme is invisible in the filings. Second, Broadcom gives the accelerator share of its AI revenue only on some calls, so the quarterly split cannot be drawn for every quarter. Third, fiscal years differ across the companies, so their figures cover different months. Fourth, MediaTek's and Broadcom's figures for 2027 and 2028 are forecasts and supply plans, not sales. Finally, T-Head's and Kunlunxin's audited revenue by product will be public only when their Hong Kong listing documents are released.

Frequently asked questions

What is a custom AI chip (AI ASIC)?

An application-specific integrated circuit designed for one company's AI work, such as Google's TPU or Amazon's Trainium, rather than a general-purpose graphics processor sold to everyone. Broadcom calls the ones it builds XPUs.

How much does Broadcom make from AI chips?

Broadcom's AI semiconductor revenue was USD 16.7 billion in its fiscal Q3 2026 (May to July 2026), up 221% on a year earlier, and it guides to USD 21.7 billion in Q4. Custom accelerators were 73% of Q3 AI revenue.

Who makes custom AI chips besides Broadcom?

Design partners include MediaTek, Marvell, Alchip and Global Unichip (GUC). Amazon designs its Trainium chips through Annapurna Labs, and in China Alibaba's T-Head and Baidu's Kunlunxin design their own.

How big will the custom AI chip market be in 2027?

Company statements differ. MediaTek's target of about USD 16 billion at about 20% share implies about USD 80 billion, while Broadcom's secured supply of about USD 115 billion of fiscal 2027 AI revenue, at a 73% accelerator share, would be about USD 84 billion on its own (both calculated). The gap reflects systems sales, memory content and definitions.

How many Trainium chips has Amazon deployed?

Amazon said in February 2026 that it had landed over 1.4 million Trainium2 chips, and in April that Trainium had over USD 225 billion of revenue commitments.

Sources

  1. Broadcom, fourth quarter and fiscal year 2025 results, SEC 8-K exhibit, Dec 2025
  2. Broadcom, first quarter fiscal 2026 results, SEC 8-K exhibit, Mar 2026
  3. Broadcom, second quarter fiscal 2026 results, SEC 8-K exhibit, Jun 2026
  4. Broadcom, third quarter fiscal 2026 results, SEC 8-K exhibit, Sep 2026
  5. Broadcom Q2 FY2025 earnings call transcript (Motley Fool), Jun 2025
  6. Broadcom Q4 FY2025 earnings call transcript (Motley Fool), Dec 2025
  7. Zacks, Broadcom Q3 earnings coverage, Sep 2026
  8. Fierce Network, Broadcom predicts AI chip revenue to double in 2027, Sep 2026
  9. Focus Taiwan (CNA), MediaTek expects 20% share of AI accelerator ASICs in 2027, 31 Oct 2026
  10. Marvell Q4 FY2026 earnings call transcript (Motley Fool), Mar 2026
  11. Marvell Technology, Form 10-K for fiscal 2026
  12. Alchip Technologies, Q4 and FY2025 results, Mar 2026
  13. Alchip Technologies, Q4 2025 investor presentation, Mar 2026
  14. Alchip Technologies, 2024 annual report (major customers)
  15. Global Unichip, Q3 2026 consolidated financial statements
  16. Global Unichip, Q3 2026 presentation
  17. Amazon Q4 2025 earnings call transcript (Fortune), Feb 2026
  18. Amazon, Andy Jassy on Amazon's chips business, Q1 2026
  19. Eastmoney, T-Head spin-off and Zhenwu shipments, Nov 2026
  20. Economic Times (Reuters), T-Head files confidential Hong Kong IPO application, Nov 2026
  21. The Standard (Reuters), Kunlunxin eyes Hong Kong IPO
  22. MediaPost, Baidu's AI chip unit files for Hong Kong IPO, 2 Jan 2026

Related reports

This page is market information, not investment advice. Figures are taken from company filings, results releases and earnings calls, and from news reports where stated; forecasts are the companies' own. Broadcom's fiscal year ends in early November (fiscal Q1 FY26 = November 2025 to January 2026); Marvell's ends in late January (fiscal 2026 = February 2025 to January 2026); Alchip, GUC and MediaTek report by calendar year. Figures marked as calculated are worked out by Trifacta Data.

More insights

All insights →

Questions about
these numbers?