Where Is the AI Semiconductor Supply Chain Still Constrained?

HBM (high-bandwidth memory) remains the clearest constraint on AI silicon.

The majority of design and sourcing experts identified memory or HBM as the industry’s core constraint. A smaller group pointed first to advanced packaging, although most expect the bottleneck to shift further toward memory as capacity expands.

“Right now, HBM allocation is very limited. We are short on what we actually need, and what we are getting is probably 80% of what we really want. The rest is essentially rationed.” — Semiconductor sourcing decision-maker, IDM

Download the research report.

‍Current research

AI Silicon Design and Sourcing: 14 interviews across fabless designers, IDMs, hyperscaler custom-silicon teams, and sourcing organizations.

Semiconductor Fab Equipment and Materials: 7 interviews spanning foundries, HBM and DRAM production, advanced packaging, equipment, semiconductor buyers, and industry analysis.

What the research examines

• Where custom silicon is displacing merchant GPUs, and where it is not
• Vendor dynamics across EDA, foundry, packaging, and HBM
• The capex outlook for wafer-fab equipment, test, and materials
• Which constraints could ease or worsen over the next twelve months
• Where experts believe current market consensus is wrong
• How AI is changing semiconductor design cycles and EDA spending

“Earlier we used to execute approximately two to three SoCs in 12 to 14 months. Now we are executing three to five SoCs in eight to ten months.” — Principal engineer, semiconductor design house

Expert interviews across the semiconductor value chain

Participants include professionals working in fabless design, IDMs, hyperscaler custom silicon, foundries, HBM and DRAM production, advanced packaging, equipment, and sourcing.

Companies discussed include Synopsys, Cadence, Siemens EDA, Arm, TSMC, Samsung, Intel Foundry, SK Hynix, Micron, ASML, Applied Materials, Lam Research, KLA, Advantest, and Teradyne.

“We’re approaching 100%, which is why we’re buying more tools. We can’t meet demand. Every wafer that we start has been purchased.” — Lithography process engineer, foundry

Access the full semiconductor research

Qualitate continuously interviews professionals across the semiconductor value chain to track how AI is changing demand, spending, capacity constraints and vendor dynamics.

Explore the executive briefings, respondent summaries and underlying interview evidence with Qualitate. 

Request a demo for full access.

FREQUENTLY ASKED QUESTIONS

How does Qualitate create its reports?

Qualitate conducts studies across industries on a regular cadence, interviewing technology buyers and experts using our AI Moderator about their forward-looking purchasing plans. The platform quantifies the outputs.

What is Qualitate’s AI Moderator?

Qualitate’s AI Moderator is a purpose-built voice agent equipped with 500,000+ minutes of expert discussions. It conducts structured one-on-one interviews with verified buyers and experts, following a consistent question framework with research-grade probing.

How is this different from a survey or expert network call?

Traditional surveys capture checkbox responses. Expert network calls give depth but at limited interviews per project they’re not statistically meaningful. Qualitate runs in-depth discussions at scale, structured consistently, so the output is both deep and quantifiable.

Who are the experts in these discussions?

Qualitate screens for seniority, active buying authority, and relevance to the specific market being studied. Discussions that don’t meet quality thresholds are excluded before analysis.

‍