#GateEventContractTradeSharingChallenge $NVDA NVIDIA surges 3.21%! Management confirms again: AI demand is not the problem; memory and advanced wafers are the real bottlenecks

$NVDA NVIDIA surges 3.21%! Management confirms again: AI demand is not the problem; memory and advanced wafers are the real bottlenecks

U.S. technology stocks are recovering, especially as NVIDIA regains strength. But what deserves the most attention today is not the stock-price gain itself, but the latest exchange between JPMorgan and Toshiya Hari, NVIDIA's vice president of investor relations and strategic finance.
NVIDIA confirms again: AI demand is fundamentally not a problem
The most important sentence from this exchange was: what is currently limiting NVIDIA's earnings growth is not demand, but supply.
NVIDIA believes that achieving approximately 70% year-on-year growth in fiscal 2028 is a relatively conservative and feasible expectation. Without supply-chain constraints, however, the business could potentially more than double in scale.
Why? Because more and more customers now need NVIDIA's computing power. Previously, the main customers were hyperscale cloud providers such as Microsoft, Google, Amazon, and Meta. Now the customer base also includes OpenAI, Anthropic, AI labs, new cloud service providers, sovereign AI, and enterprises deploying AI on-premises.
In other words, demand for AI computing has evolved from “major companies buying GPUs” to the entire AI industry chain buying computing power.

A very important change: AI inference demand is growing rapidly
NVIDIA also revealed a very important piece of information this time.
About 18 months ago: training ≈ 50%
Inference ≈ 50%
Now: inference already accounts for more than training, and its share will continue to increase in the future.
What does this mean? Previously, when people thought about AI computing, they mostly thought of training large models. But after models are trained, what truly keeps them running is inference. For example, our daily use of AI for chatting, searching, and generating images and videos essentially requires substantial inference computing power.
Therefore, AI is gradually moving from the “training era” into an era driven by both training and inference. This is a long-term positive for the entire industry chain, including AI servers, GPUs, HBM, memory, and optical modules.

The most important information is here: NVIDIA names “memory” and “advanced wafers”
This is the part of today's article that deserves the most attention. NVIDIA management explicitly stated that memory chips and advanced-process wafers are currently the most important supply bottlenecks in its bill of materials.
NVIDIA is maintaining ongoing discussions with TSMC and the three major memory manufacturers—Micron, SK hynix, and Samsung—in hopes of further improving supply. This statement is highly significant because it once again validates the logic we have been tracking: the more powerful AI computing becomes, the greater the demand for HBM and advanced memory.
GPU performance continues to improve.
↓GPUs require higher bandwidth
↓HBM capacity and speed continue to advance
↓HBM4, HBM4E, and even next-generation HBM continue to ramp up
↓Memory is becoming one of the core bottlenecks in the expansion of AI computing.
So why have we been emphasizing memory recently? Not because the market is simply speculating on “price increases.” Rather, the expansion of AI computing has begun to materially drive demand for high-end memory.

This is also why memory deserves close attention recently
From this exchange with NVIDIA, we can see that AI demand remains strong
GPUs continue to advance
Inference demand is growing
These factors, combined with insufficient memory supply, mean that HBM remains one of the most important components of AI hardware.

Another easily overlooked area: advanced packaging
In addition to naming memory, NVIDIA specifically mentioned the supply of advanced-process wafers. As GPUs and HBM continue to advance, advanced packaging is also becoming increasingly important. AI chips are no longer simply a matter of one GPU + one memory chip. Instead, they are gradually evolving toward: GPU/ASIC + HBM + advanced packaging + Chiplet. The entire system is becoming increasingly complex. That is why we have recently seen:
HBM4
HBM4E
Base Die
Hybrid bonding
2.5D/3D packaging
These are all part of a complete industrial upgrade path.

Why is optical communications also worth continuing to watch?
The U.S. optical communications sector broadly recovered yesterday: Marvell +2.86%
Coherent +1.77%
Lumentum +0.19%
The logic behind this is actually quite simple: AI data centers continue to expand.
↓The number of GPUs continues to grow.
↓Data transmission between servers continues to increase.
↓Network bandwidth continues to advance.
↓Demand for 800G and 1.6T optical modules continues to grow.
Therefore, AI computing is not just about GPUs.
In reality, it involves an entire industry chain: GPU → HBM → PCB/CCL → optical modules → switches → advanced packaging → data centers
As long as AI capital expenditures continue to grow, this industry chain still has room to expand.

NVIDIA has seen another change: its customers are becoming more diversified
In the past, the market often worried about one issue: Was NVIDIA too dependent on a few major customers such as Microsoft, Google, and Amazon? This situation is now changing.
NVIDIA said that OpenAI and Anthropic currently account for approximately 20% of NVIDIA's business on an end-consumption basis, and this proportion could rise to around 25% in the future. At the same time, new cloud service providers have become a very important customer group for NVIDIA. This shows that demand for AI computing is spreading from a small number of internet giants to an increasing number of AI companies and enterprises.
In other words, the foundation of demand for AI computing is becoming increasingly broad.

Today's most important conclusions
If we condense the entire NVIDIA exchange into a few points:
First, AI demand has not declined.
Second, NVIDIA believes it can still maintain very high growth over the next few years.
Third, inference is becoming a new growth engine for AI computing.
Fourth, what is primarily limiting NVIDIA's growth at present is not demand, but supply.
Fifth, memory and advanced-process wafers have become core bottlenecks.

In my view, the greatest value of today's exchange with NVIDIA is not telling us whether “NVIDIA can continue to rise.” Rather, it once again explains the logic of the entire AI industry chain very clearly: AI demand remains strong, inference demand continues to grow, and what truly limits the pace of industry expansion has become the supply chain.$NVDA ‌