Chips

How it works

AI runs on accelerators, the high-bandwidth memory stacked beside them, the packaging that joins them and the networks that link thousands of them. Each step is made by a handful of companies, and several report demand above what they can supply³⁴.

NVIDIA reported data center revenue of $89.0 billion for the quarter ended July 26, 2026, up 117% from a year earlier¹. TSMC's revenue for July to September 2026 was US$46.72 billion, up 41.2%, with high-performance computing 68% of the total². Global semiconductor sales were $159.7 billion in August 2026 on a three-month average, up 144.3% from a year earlier⁸.

Show the data

Fiscal quarter	Months	Data center revenue
Q4 FY25	Nov 2024 to Jan 2025	$35.6 bn
Q1 FY26	Feb to Apr 2025	$39.1 bn
Q2 FY26	May to Jul 2025	$41.1 bn
Q3 FY26	Aug to Oct 2025	$51.2 bn
Q4 FY26	Nov 2025 to Jan 2026	$62.3 bn
Q1 FY27	Feb to Apr 2026	$75.2 bn
Q2 FY27	May to Jul 2026	$89.0 bn

How an AI chip gets made

1. Design. A chip company, or a cloud company working with a custom-chip partner, designs the accelerator.

2. Wafer fabrication. A foundry prints the design onto silicon wafers using leading-edge processes. TSMC has kept its 2026 capital budget at US$60 billion to US$64 billion³.

3. Memory. Memory makers stack DRAM dies into high-bandwidth memory. SK hynix says long-term agreements with 14 customers cover more than half of its 2027 and 2028 production capacity⁴.

4. Advanced packaging. The accelerator and its memory stacks are joined into one package. TSMC said its CoWoS and SoIC packaging capacity is expected to grow more than 80% a year through 2027, as reported by Taiwan's CNA⁵.

5. Servers. Manufacturers build the packaged chips into boards, servers and full racks with power and cooling.

6. Networking. Racks are linked with switches and copper or optical connections. NVIDIA's networking revenue was $14.8 billion in February to April 2026, up 199% from a year earlier⁶.

Show the data

Month	Revenue (NT$ million)	Change from a year earlier
Jan 2026	401,255	+36.8%
Feb 2026	317,657	+22.2%
Mar 2026	415,191	+45.2%
Apr 2026	410,726	+17.5%
May 2026	416,975	+30.1%
Jun 2026	442,680	+67.9%
Jul 2026	467,580	+44.7%
Aug 2026	514,806	+53.3%
Sep 2026	511,857	+54.6%

Terms

Accelerator

A processor built to run AI calculations in parallel, such as a GPU or a custom chip.

High-bandwidth memory (HBM)

Memory chips stacked vertically beside the processor so data moves between them quickly.

Advanced packaging

Methods such as TSMC's CoWoS that combine several chips, like a GPU and its memory stacks, into one package.

Foundry

A company that manufactures chips designed by others.

Wafer

A thin silicon disc, usually 300 mm across, on which many chips are made before being cut apart.

Optical transceiver

A module that converts electrical signals to light and back, so data can travel over fiber between servers and switches.