Advanced packaging and high-bandwidth memory are both in short supply, with new AI accelerator companies facing an estimated 18-to-24-month wait for capacity, according to State of the Future. The analysis says production is fully utilised and largely allocated to established chipmakers and hyperscaler programmes.
Watch Desk analysis
What happened
The State of the Future analysis, published on 16 November, identifies constraints in TSMC’s CoWoS-S advanced packaging and HBM3e memory. It says capacity is primarily going to Nvidia, AMD, Broadcom and hyperscalers’ custom-silicon programmes.
The article argues that these constraints create a substantial barrier for companies trying to enter the AI accelerator market. It also suggests that alternatives such as physics-native computing could have a market opportunity through 2027, before the bottleneck eases.
Why it matters
A new accelerator is no use to customers if its maker cannot secure the packaging and memory needed to build it. A queue lasting up to two years could therefore protect established suppliers and delay challengers, even when those challengers have a promising design.
That is a claim about the market outlook, not a guarantee that every entrant faces the same wait or that the bottleneck will persist on this timetable. But it puts an important constraint on the AI chip race: access to manufacturing capacity can matter as much as the chip architecture.
Our read
The analysis makes a useful case for looking beyond the headline processor. Packaging and memory are part of the supply chain, and when both are constrained, a start-up cannot simply code its way around the queue. The forecast for alternative approaches is more speculative; the test will be whether they can secure production and win customers before the window closes.
What to watch
- Whether TSMC or memory suppliers provide evidence of capacity easing.
- Whether new accelerator companies disclose delays or secure meaningful production slots.
- Whether alternative computing approaches attract customers while the constraints remain.
Discussion spark: If packaging and memory capacity are the bottleneck, should AI chip investors prioritise companies with manufacturing access over those with the most promising designs?
Sources and evidence
- HBM / CoWoS Bottleneck (16 November 2026, 07:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.