● AI Chip Supply Shortage
The more important issue than Nvidia’s earnings surprise is the ‘AI semiconductor supply shortage’
The real point to watch in this Nvidia earnings report is not simply that “earnings were good.”
The key takeaway is that Nvidia delivered a much stronger growth guide than the market expected and, at the same time, publicly said that “demand is so strong that supply cannot keep up.”
In particular, this announcement tied together Nvidia earnings, AI semiconductor demand, data center investment, memory semiconductor shortages, and global stock market trends all at once.
On the surface it is an earnings surprise, but if you look deeper, memory shortages and TSMC wafer bottlenecks are limiting Nvidia’s growth pace.
In other words, the more important question right now is not the “AI bubble” the market is worried about.
Is it not that AI demand has weakened, but that the AI infrastructure supply chain cannot keep up?
1. Nvidia delivers another earnings surprise… but the key is next quarter and next year’s guidance
Nvidia once again reported earnings that beat market expectations.
Revenue came in at about $96 billion, roughly 4.3% above the market estimate of $92 billion.
EPS was $2.20, about 6% above the expected $2.09.
The market already had some sense that Nvidia would post solid results.
That is because Nvidia has recorded earnings surprises in almost all of its last 20 earnings releases.
So the truly important thing in this report was not the past numbers, but the numbers ahead.
Nvidia guided next-quarter revenue to about $108 billion.
That too is above market expectations.
The more shocking part is the growth outlook for fiscal 2028, which corresponds roughly to calendar 2027.
Wall Street had expected Nvidia’s growth rate next year to be around 44%.
But Nvidia indicated growth of about 70%.
This was not just a slight beat of expectations; it essentially rewrote the baseline.
What is even more interesting is Nvidia’s wording.
It suggested that if it could get more supply, 100% growth might have been possible, but current bottlenecks may limit growth to 70%.
That statement is extremely important.
It means the issue is not a lack of AI demand, but that the supply chain is limiting the growth rate.
2. Data centers remain at the center of the results
The core driver of this Nvidia earnings report was, as always, data centers.
Nvidia’s AI GPUs remain at the center of the global data center investment cycle.
In the past, data center revenue tended to be viewed as one big lump.
But now the demand base is becoming more segmented.
The first segment is the existing hyperscaler group centered on big tech.
This includes companies like Google, Amazon, Microsoft, and Meta.
The second is sovereign AI, where countries build AI infrastructure at the national level.
This is the trend in which governments seek AI capabilities based on local language, local data, and local industries.
The third is neo-cloud companies such as CoreWeave, Nebius, and IREN.
These are emerging infrastructure firms that specialize in providing AI compute resources.
The fourth is AI infrastructure demand from ordinary companies.
Demand is broadening as manufacturing, finance, healthcare, retail, and content companies begin building their own AI systems.
In this announcement, growth related to hyperscalers was mentioned at around 100%.
But the growth rate in the new areas, including sovereign AI, neo-clouds, and general enterprise demand, was even higher at about 138%.
This is very important.
There is a market concern that if big tech companies make their own AI chips, Nvidia demand will decline.
In reality, Google is using TPUs, and Amazon is also making its own chips.
OpenAI is also said to be advancing its own AI chip development together with Broadcom.
But even if big tech demand slows, sovereign AI and neo-clouds are quickly filling the gap.
Ultimately, Nvidia’s demand base is expanding from dependence on a few big tech companies into a broader AI infrastructure ecosystem.
3. After hyperscalers, the next wave is sovereign AI and neo-clouds
The most important change in this earnings release is that the center of demand is widening.
Early AI infrastructure investment was mainly led by hyperscalers such as Microsoft, Google, Amazon, and Meta.
They bought Nvidia GPUs in large quantities, backed by massive cash flow and cloud infrastructure.
But now a second wave is arriving.
That wave is sovereign AI.
Governments around the world are beginning to see AI not just as a technology, but as national competitiveness.
Not only the United States and China, but also countries in the Middle East, Europe, and Asia are moving to build their own AI data centers.
The reason is that depending only on foreign cloud services for AI models creates weaknesses in data sovereignty, security, and industrial strategy.
Another key area is neo-clouds.
Companies like CoreWeave, Nebius, and IREN are growing businesses that lease compute resources based on AI GPUs.
Nvidia believes that by around 2028, installed capacity at neo-cloud companies could reach about 8 GW.
Given that the actual operating power of major neo-cloud companies is estimated at about 2.5 GW to 3 GW, this means it could grow two to three times within roughly two years.
This is not just a stock price momentum story.
It is a massive capex cycle connected to everything in the AI era, including power, data centers, GPUs, memory semiconductors, and cooling systems.
4. What it means to say Nvidia is playing a ‘central bank role’
There is a particularly interesting phrase circulating in the market recently.
It is the idea that Nvidia acts like a kind of central bank in the AI infrastructure market.
What this means is that Nvidia can provide credit support to neo-clouds and new AI infrastructure companies that need Nvidia GPUs but have difficulty raising capital.
Simply put, the structure looks like this.
Neo-cloud companies need GPUs.
But buying GPUs requires enormous amounts of capital.
When interest rates are high or investors become cautious, financing costs rise.
If Nvidia supports the ecosystem through partnerships, investments, contract structures, or credit support, those companies can raise funds at lower cost.
As a result, Nvidia is becoming not just a GPU seller but also a company that indirectly shapes the cash flow of the AI infrastructure ecosystem.
This is a highly strategic move.
Big tech companies are trying to reduce dependence on Nvidia by developing their own AI chips.
That means Nvidia must cultivate strong customers outside big tech.
Those customers are neo-clouds, sovereign AI, and enterprise AI infrastructure.
If Nvidia does not just sell chips to them but also helps build the ecosystem, it creates an even stronger long-term lock-in effect.
5. The hidden risk in this earnings report is margin compression
There was also a point in this release that made the market briefly nervous.
That point was gross margin.
Nvidia’s gross margin this quarter was very high at about 75%.
But management explained that next quarter it could fall to around 71% to 72% before rebounding to around 73% afterward.
Normally, just hearing that Nvidia’s margin is weakening would have made the stock market react sensitively.
But this time, the mood was a bit different.
That is because the reason for the margin decline is closer to rising costs and supply shortages than to weak demand.
In particular, memory semiconductor prices are rising rapidly.
High-performance memory, including HBM, is an essential component of AI GPU packages.
AI semiconductors are not completed by a single GPU chip alone.
They also require high-performance memory, advanced packaging, TSMC’s leading-edge processes, power supply, and networking equipment.
If even one of these is constrained, overall supply is limited.
That is exactly the bottleneck Nvidia is facing now.
Demand is exploding, but memory and wafer supply cannot keep up with that pace.
6. Why memory shortages matter so much
Many people talk about AI semiconductors and only look at GPUs.
But in reality, the core bottleneck in AI infrastructure is not just GPUs.
High-bandwidth memory, especially HBM, has a direct impact on AI model training and inference performance.
AI models need to read and write enormous amounts of data at high speed.
At that point, even if GPU compute performance is excellent, insufficient memory bandwidth will cap overall performance.
That is why Nvidia’s latest GPUs must include high-performance memory.
The problem is that HBM production cannot easily be increased in the short term.
Memory companies such as SK Hynix, Samsung Electronics, and Micron are expanding capacity, but quality certification and yield improvement take time.
Advanced packaging is also a bottleneck.
Only if advanced packaging capacity such as TSMC’s CoWoS is sufficient can GPUs and HBM be combined into a single high-performance package.
Ultimately, the AI semiconductor supply chain is not something one company can solve alone.
GPU design, memory semiconductors, foundry production, packaging, power infrastructure, and data center sites all have to line up.
That is why this Nvidia earnings report can be seen not just as a corporate earnings release, but as an indicator of the global supply chain and inflationary pressure at the same time.
7. The Vera Rubin architecture could greatly improve inference efficiency
Nvidia also expressed strong confidence in its new Vera Rubin architecture.
In particular, it said token throughput per megawatt could increase by as much as 30 times.
The important phrase here is “per megawatt.”
In the AI era, chip performance alone is not what matters.
What matters is how much AI inference can be processed for each 1 MW of power.
That is because data center power costs continue to rise.
Of course, it is reasonable to view the 30x figure as a maximum possible under specific conditions.
In normal environments, even a 5x to 10x efficiency improvement would be a huge change.
If AI inference costs fall, companies can adopt AI services more widely.
One of the biggest burdens for current generative AI services is inference cost.
Every time a user asks a question, compute costs are incurred, and the burden rises as usage grows.
If inference efficiency improves significantly in the Vera Rubin generation, the unit cost of AI services could fall, which could in turn expand AI demand further.
In other words, better chips do not necessarily reduce demand for older chips; they may actually open up more new use cases.
8. Autonomous driving, robots, and physical AI are also Nvidia’s next pillars
In this announcement, the edge computing area also deserves attention.
Edge computing means running AI not only inside data centers, but also in cars, robots, industrial sites, and devices closer to where the action is.
The growth rate in this area was mentioned at about 27%.
Compared with data center growth, that may seem lower, but it is by no means a small market.
In particular, autonomous driving and robotics are the core of physical AI in the long run.
Physical AI refers to AI that moves and makes decisions in the real world.
That includes robots, autonomous vehicles, drones, smart factories, and humanoids.
Chinese companies related to autonomous driving and robotics are also reportedly making heavy use of Nvidia chips.
Ultimately, Nvidia’s strategy is not only to dominate cloud AI, but also to preempt AI infrastructure in the physical world.
Right now, data centers look overwhelmingly large, but in the long run, robotics and autonomous driving could become another growth pillar.
9. On the AI bubble debate, supply matters more than demand right now
The AI bubble debate has been ongoing in the market recently.
AI companies are too expensive.
Data center investment is excessive.
Big tech CAPEX has overheated.
AI services are not yet generating enough profit.
These concerns definitely need to be checked.
But this Nvidia report shows a different perspective.
The current problem looks more like a supply shortage than a demand shortage.
The customers wanting Nvidia chips are not only hyperscalers.
Demand continues to broaden to sovereign AI, neo-clouds, general enterprises, research institutions, and startups.
Even companies developing their own AI chips still want more Nvidia supply.
Own chips can be efficient for specific workloads, but in terms of the broader AI ecosystem and software compatibility, Nvidia’s CUDA ecosystem remains very strong.
So when judging whether AI is in a bubble, you should not look only at stock prices.
You also need to look at actual AI infrastructure orders, memory prices, TSMC packaging capacity, data center power 확보, and neo-cloud growth rates.
10. Key points that are talked about less in other reports
First, Nvidia’s biggest competitor may not be AMD or in-house chips, but supply chain bottlenecks.
The market mainly frames Nvidia’s competitive landscape around AMD, Google TPU, Amazon Trainium, and OpenAI’s own chip.
But this earnings report suggests that the biggest near-term constraint is HBM, advanced packaging, and wafer supply, rather than rivals.
In other words, the reason Nvidia cannot sell more is not lack of demand, but a limit on how much it can produce.
Second, margin compression may be a sign of rising supply chain prices, not a bad signal.
Usually, lower margins are interpreted as weaker corporate competitiveness.
But in this case, the rise in memory semiconductor prices and the supply shortage are major factors.
That is an important clue not only for Nvidia, but also for the earnings cycle of memory companies such as SK Hynix, Samsung Electronics, and Micron.
Third, the center of the AI infrastructure market is expanding from big tech to national and finance-like infrastructure.
Sovereign AI is not just a technology trend.
It is a structural change connected to national security, industrial policy, data sovereignty, and the global economic outlook.
This trend is difficult to reverse in the short term.
Fourth, Nvidia is evolving from a chip company into something like an AI infrastructure financial platform.
Credit support and ecosystem support for neo-clouds go beyond a simple sales strategy.
Nvidia can become a structure that directly creates AI computing demand and helps customers finance it.
Fifth, improved inference efficiency could be the real trigger for AI mass adoption.
So far, AI investment has mainly centered on training infrastructure.
Going forward, inference demand is likely to explode through AI agents, coding AI, enterprise automation, robotics, customer service, and financial analysis.
If the Vera Rubin generation greatly improves tokens per watt, AI service prices could come down and usage could rise even more.
11. Why the argument that the AI market is still early gains strength
The number of people actively using generative AI is increasing, but compared with the world’s population, it is still only a small fraction.
In particular, there are still not many people who truly use advanced AI tools such as Claude Code, Codex, and agentic AI effectively in their work.
The same goes for companies.
Most companies are still at the stage of testing AI or applying it to only a few departments.
Enterprise-wide automation, AI-agent-based operations, and industry-specific model adoption are still near the beginning stage.
That is also the context behind Anthropic’s estimate that the B2B AI market could be worth tens of trillions of dollars.
It is highly likely that the market AI companies have actually captured so far is only a tiny portion of the total potential market.
If so, AI infrastructure investment may still be in the early phase rather than near the end.
Of course, that does not mean every AI-related stock will keep rising.
Bubbles can form in individual stocks and valuations.
But at the industry level, the AI semiconductor, data center, memory semiconductor, power infrastructure, and cloud investment cycles are still in an expansion phase.
12. Key variables to watch from an investment perspective
1) Whether Nvidia maintains next-quarter revenue guidance
It is important whether the $108 billion guidance actually translates into real results.
If guidance keeps moving higher, concerns about an AI demand peak may fade.
2) Whether gross margin rebounds after 71% to 72%
We need to see whether margin, after temporarily declining, recovers to around 73%.
If memory prices rise even further, margin pressure could last longer.
3) HBM supply and pricing
The expansion of HBM capacity at SK Hynix, Samsung Electronics, and Micron directly affects Nvidia’s supply volume.
The memory semiconductor cycle is becoming increasingly tied to AI infrastructure investment.
4) TSMC advanced packaging capacity
Only if the advanced packaging bottleneck that combines GPUs and HBM is relieved can Nvidia shipments rise further.
5) Financing conditions for neo-cloud companies
We need to see whether companies like CoreWeave, Nebius, and IREN can raise capital stably even in a high-rate environment.
Nvidia’s credit support strategy could play an important role here.
6) The actual performance and adoption rate of big tech’s own chips
The rise of in-house chips is clearly a trend.
But we need to distinguish whether they can replace the Nvidia ecosystem or only complement it for certain workloads.
7) The pace of growth in AI inference demand
Going forward, inference demand may become more important than training.
As AI agents and enterprise automation spread, demand for inference infrastructure could grow structurally.
13. The message this Nvidia report sends to the global stock market
After this release, Nvidia shares strengthened in after-hours trading, and a positive mood spread across AI stocks and the broader tech sector.
Recently Nvidia shares had come under pressure, falling for seven consecutive trading days.
By some valuation metrics, they were even described as relatively cheap within the S&P 500.
In that situation, the strong results and 70% growth guidance helped the market regain confidence in AI growth stocks.
But the important point here is not blind optimism.
The AI industry is strong, but within it, winners and losers are likely to emerge.
Companies with GPUs, HBM, foundry access, power, cooling, data center sites, and cloud operating capabilities will have the advantage.
On the other hand, companies that only have the AI label but weak actual revenue and cash flow may become more volatile.
In other words, this Nvidia report did not signal the end of the AI bubble; it is closer to a sign that competition in AI infrastructure is becoming even more intense.
< Summary >
Nvidia recorded an earnings surprise, beating market expectations on both revenue and EPS.
But even more important is that it guided to about 70% growth for fiscal 2028.
Nvidia showed strong confidence that 100% growth might have been possible if supply had been sufficient.
AI demand has not weakened; rather, the supply chain cannot keep up.
Memory semiconductors, HBM, TSMC advanced packaging, and wafer bottlenecks have emerged as the key constraints on Nvidia’s growth.
The next growth pillars after hyperscalers are sovereign AI and neo-clouds.
Nvidia is moving beyond being just a GPU company into a position where it leads the AI infrastructure ecosystem and even the flow of capital.
More important than the AI bubble debate are actual demand, supply chain bottlenecks, inference efficiency improvements, and the data center investment cycle.
In conclusion, the AI era is not at the end stage; it is still in the early phase.
[Related Articles…]
- Nvidia earnings and the outlook for the AI semiconductor market
- Memory semiconductor shortages and the HBM investment cycle
*Source: [ 월텍남 – 월스트리트 테크남 ]
– 역대급 실적 속 숨겨진 “버블?”…메모리 부족이 심각합니다.
