AI Chips Haven't Peaked. The Real Risk Sits in 2028 Supply

Two clocks are running at once

There's a reason both the bulls and the skeptics on AI semiconductors are right at the same time. The demand clock is still ticking upward. The supply-investment clock has already started setting up the next downcycle.

AI demand strength → memory shortage → rising ASP and margins → capex expansion → 2028 supply and ROI test

The data available right now backs the first three stages. But investors need to watch the fourth stage, capacity expansion, just as closely. Memory is an industry where high prices summon new supply, and that supply eventually pushes prices back down. The stronger today's boom runs, the bigger the supply risk a few years out could become.

The shortage didn't stop at HBM. It has reached SSDs

<div class="metric-grid"> <div class="metric"><span class="num">+59.5%</span><span class="label">2Q26 DRAM industry revenue, quarter over quarter</span></div> <div class="metric"><span class="num">16 weeks</span><span class="label">Enterprise SSD lead time</span></div> <div class="metric"><span class="num">8 weeks</span><span class="label">TrendForce's benchmark for a balanced market</span></div> </div>

According to TrendForce, DRAM industry revenue in the second quarter of fiscal 2026 rose 59.5% from the prior quarter to roughly $154.7 billion. Supply growth failed to keep pace with demand as rising conventional DRAM contract prices and AI server memory demand moved in the same direction at once.

NAND is the more interesting case. Consumer NAND and client SSDs show price pressure and soft demand. Enterprise SSDs show the opposite. In TrendForce's September 21 check, enterprise SSD lead times ran 16 weeks, double the roughly 8 weeks TrendForce treats as a balanced market. Suppliers are prioritizing enterprise SSD capacity as storage demand tied to AI servers and agentic AI workloads climbs.

So a single line like "NAND is weak" misses the real signal. Consumer NAND softness and an enterprise SSD shortage are happening at the same time, in the same market.

The real risk shows up where CXMT meets capex

The single piece of news that most reshaped medium-term risk this week came from China's CXMT. The company announced it has entered mass production on a fifth-generation DRAM platform, and said the new process lifts gross die output per wafer by at least 50% over the prior generation. It also unveiled a 24Gb LPDDR5X part.

These numbers matter because memory supply isn't set by factory floor space alone. Supply is roughly a function of wafer capacity × yield × die density. Run the same factory, but lift process productivity, and actual bit supply can grow far faster than the physical footprint suggests.

Layer that onto expansion already underway at Samsung, SK hynix and Micron. Through 2026 and 2027, market growth is likely fast enough that all three companies can post strong profits simultaneously. But if new capacity, improving HBM yields, and rising productivity at Chinese manufacturers all land around 2028, today's shortage could flip into tomorrow's oversupply, the classic memory cycle repeating itself.

The bottleneck in AI infrastructure now looks more like electricity than GPUs

<div class="metric-grid"> <div class="metric"><span class="num">470GW+</span><span class="label">Data center projects waiting on Texas grid interconnection</span></div> <div class="metric"><span class="num">~5x</span><span class="label">That backlog versus Texas peak power demand</span></div> <div class="metric"><span class="num">Price hikes</span><span class="label">Recent direction for some GPU cloud pay-as-you-go rates</span></div> </div>

In Texas, new state permits for data centers are on hold until an audit of their grid impact wraps up. Data center projects waiting on ERCOT interconnection now exceed 470 gigawatts, roughly five times the state's peak power demand.

That gap shows the unit of AI infrastructure competition sliding from GPUs to HBM to packaging, and now down to megawatts and gigawatts. GPU rental prices, meanwhile, still aren't signaling a demand collapse. Nebius said it will raise pay-as-you-go pricing on some Nvidia GPUs starting October 1.

Put together, the current bottleneck looks less like "too many GPUs sitting idle" and more like "can operators secure the power and data center space to run them, on time."

ASML and CoWoS are still in "bottleneck easing," not "peak"

Data on ASML's High-NA EUV also leans toward wider adoption rather than a cycle turning down. TSMC has said it plans to use High-NA for advanced-node mass production starting in 2030, and Samsung plans to bring High-NA into future DRAM production. Intel has said it has now processed more than one million wafers cumulatively on High-NA tools.

The key distinction here: easing a bottleneck is not the same as demand slowing down. Even as CoWoS capacity expands and EUV productivity improves, the cycle can keep running as long as utilization and pricing hold. The real peak isn't the capacity expansion itself. It shows up later, when utilization and prices start to fall after that capacity comes online.

By the current numbers, this isn't a peak yet. Worth remembering, though: today's shortage is exactly the kind of structure that calls 2028's supply into being.

Insight Times Editorial Desk