AI demand keeps chip capacity tight as Apple warns on memory costs

For chip buyers and device makers, the squeeze is already visible in prices and lead times.

Apple expects memory costs to rise again in the December quarter after calling current prices unprecedented. Suppliers are adding fab and packaging lines, but advanced packaging substrates and some HBM materials remain tight. For buyers, that means less room to negotiate on price or delivery. For device makers, it raises the risk of higher build costs and slower launches. The pressure is not limited to one vendor. The tightness covers both memory and packaging, two parts of the supply chain that device makers cannot easily swap. That combination is what makes the current squeeze hard to plan around.

Factory data shows the strain

The ISM's October reading was 55.5%, a full point above September's 54.5%.

It marked the 10th consecutive month of expansion after a 10-month contraction. Its Production Index climbed 2.3 points to 59.0%. That is the highest reading since November 2021.

Survey comments are blunter than the headline. One machinery executive described "explosive growth in the semiconductor and government/defense spaces" and said lead times that ran four weeks now run 50. The same panelist said a $2,000 item now costs $4,000. Another respondent said any commodity consumed by AI and data centers is "in a crunch, both for supply and pricing". Those comments point to demand outrunning capacity, not a single product line. They also show how quickly costs can move for buyers. It is also a warning for procurement teams.

Apple feels the memory squeeze

Apple is the clearest consumer-facing case. CEO John Ternus told the company's 2 November earnings call that "what we're seeing with memory pricing right now is unprecedented".

Tom's Hardware reported on 3 November that Apple's manufacturing bill has climbed by $33 billion in just a few months. Apple has softened the blow with cheaper inventory, cost cuts and price rises on Mac, iPad and iPhone, the report said. It has not announced another round of increases.

Suppliers are not promising relief. Samsung said memory demand would far exceed supply as AI infrastructure expands, and a senior executive warned the crunch could run to 2028 and beyond. SK Hynix says long-term agreements cover more than half of its 2027 and 2028 production capacity.

Tokyo Electron lifts forecasts

The equipment side is where the money shows up first. Tokyo Electron reported first-half revenue up 41.3% year on year and operating profit up 67.2%, according to a note.com analysis of the company's 30 October results. Operating margin rose to 30.4% from 25.6% a year earlier. For the year to March 2027 Tokyo Electron forecasts revenue up 51.4% and operating profit up 87.1%. That would be its first operating profit above ¥1 trillion. CEO Toshiki Kawai said customers are requesting earlier deliveries and additional purchases on AI server demand, and the company sees the 2027 wafer fab equipment market growing.

New lines, still tight

Capacity is being added anyway. BriefAsia reported on 3 November that several AI supply chain firms have announced fresh investment in chip fabrication, advanced packaging, optical components and data center infrastructure in recent weeks. Several cited order visibility extending into the first half of 2027. Taiwanese packaging houses are expanding interposer and chip-on-wafer-on-substrate capacity. In South Korea, memory makers have pulled forward high-bandwidth memory timelines. Clean-room space is arriving ahead of schedule.

The bottlenecks have not moved. Advanced packaging substrate availability remains tight, lead times for some HBM materials have extended, and PCB makers have flagged shortages of laminates for high-layer-count boards.

What to watch

Two risks sit under the forecasts. Tokyo Electron flagged China export regulations and the duration of AI investment as threats to its outlook. BriefAsia noted that analysts are questioning how long AI infrastructure spending can hold its pace, citing cloud provider utilisation rates and the long lead times between model improvements and revenue.

Apple expects memory purchases to get more expensive into 2027.