Oct 8, 2026 · 51 min
AI demand keeps semiconductors ahead of the cycle
Stacy Rasgon: “Demand Is Off The Charts” in Semiconductors… And Set To Double Again Soon
The discussion tests whether extraordinary AI infrastructure spending reflects durable demand or the familiar excesses of a semiconductor boom.
- 1Chip demand remains exceptionally strong, but limited visibility makes the cycle unusually difficult to forecast.
- 2Memory, power, optics, land, and clean-room constraints could extend the investment cycle while raising the cost of new capacity.
- 3NVIDIA faces credible competitive pressure, yet its software ecosystem and investment scale still reinforce its lead.
Don't miss
Stacy Rasgon contrasts today’s AI infrastructure spending with the dot-com era, emphasizing meaningful returns from neocloud capacity rentals rather than unused dark fiber.
The brief
Bernstein analyst Stacy Rasgon argues that semiconductor demand remains exceptionally strong, even as falling valuation multiples signal investor fears of a cyclical peak.
The cycle’s defining tension is limited visibility: shortages encourage duplicate orders, while analysts and company managers struggle to distinguish durable demand from temporary amplification.
Rasgon points to bottlenecks in memory, power, optics, land, and clean-room capacity, arguing that infrastructure constraints—not weak demand—may determine how fast supply catches up.
The conversation weighs threats to NVIDIA from AMD, Broadcom custom ASICs, and startups against NVIDIA’s software ecosystem, investment scale, and CUDA-related advantages.
The standout comparison comes with 2000: unlike dark-fiber stockpiling, Rasgon says current AI infrastructure is already showing meaningful returns through data-center capacity rentals.
Her broader conclusion is constructive: cycle indicators have not turned decisively bearish, Intel is improving, and semiconductor fundamentals continue to strengthen despite investor nervousness.
What was said on this episode
45 statements · 36 positive · 6 negative · 2 mixed · 1 neutral
Semiconductors probably are not approaching an earnings cycle peak yet
“at least at this point, from an earnings standpoint, we probably aren't”
Listen at 1:38
Semiconductor fundamentals are currently improving
“fundamentals still look like they're getting better, not worse”
Listen at 2:14
The semiconductor cycle peak is unlikely this year through 2029
“it's not this year. It's not that next year. Doesn't look like it's 28, 29.”
Listen at 3:45
Semiconductor stocks and fundamentals likely have further room to run
“my guess is we've still got some room”
Listen at 3:52
Semiconductor demand is exceptionally strong
“Demand is off the charts”
Listen at 4:56
Semiconductor companies’ order visibility is strong and improving
“their order visibility is very, very strong. And again, it seems to be strengthening.”
Listen at 5:18
The top five hyperscalers will probably spend $1.5 trillion next year
“the top five hyperscalers will probably spend a trillion and a half dollars next year”
Listen at 6:08
Demand is strong across most AI-related semiconductor spaces
“demand just looks really strong across the board in most of the AI-related spaces”
Listen at 6:42
Current AI infrastructure demand appears genuine
“the demand itself at this point does really seem to be real”
Listen at 7:47
Supply constraints may extend the strength of the AI buildout
“may actually help this be stronger for longer”
Listen at 8:13
Early evidence indicates AI infrastructure investments are generating returns
“the incremental data points around is there a return or not on these investments, I think is landing pretty much decisively on the side of, yes, there is a return”
Listen at 8:33
Neocloud infrastructure payback may be one to 1.5 years
“the payback at. you know, a year, a year and a half for the infrastructure”
Listen at 8:51
AI investments are likely to generate returns rather than fail to do so
“I don't think that's where we're going, but ultimately every AI bear case that you can think of, ultimately, I think boils down to that. Answering that question, is there a return or is there not? I think there is.”
Listen at 10:03
Semiconductor industry forecasting remains broadly unreliable
“forecasting in general is an unsolved problem in this space. Everybody's bad.”
Listen at 10:30
Double ordering can make actual semiconductor demand appear ten times larger
“that hundred units of actual demand to the industry collectively looks like a thousand units of demand”
Listen at 11:49
AI hardware is generally being deployed immediately rather than stockpiled
“Nobody's really stockpiling AR hardware. Everything that's actually getting built and deployed is being used immediately”
Listen at 12:28
Large warehouses of unused GPUs probably do not exist broadly
“I do not think that there are like huge warehouses all over the place filled with GPUs that are just I do not think that that is true.”
Listen at 13:17
Data center construction progress will probably continue despite backlash
“my guess is progress will not be stopped”
Listen at 13:44
Power availability is a major constraint on AI data center expansion
“Power is a huge constraint”
Listen at 14:40
NVIDIA could earn about $20 per share if revenue doubles
“If they actually double the revenues, they'll probably do $20 in earnings”
Listen at 17:04
NVIDIA and Broadcom stocks would be very cheap if growth is not peaking
“if they can do it. So the question then becomes, again, back to our earlier discussion, does that represent a peak or not? If it's not a peak, the stocks are absurdly cheap”
Listen at 17:29
Semiconductor stocks will not be at a peak if demand proves real
“if they're right, we won't be at a peak”
Listen at 18:20
AI’s scale is currently lifting sectors inside and outside semiconductors
“I think AI has gotten so big, it's dragging everything along with it in and out of semis.”
Listen at 18:27
AI infrastructure drives substantial demand for memory and storage
“AI is dry. It drives a ton of demand for memory and storage.”
Listen at 27:58
Memory comprises about 80% of silicon wafer area in NVIDIA’s AI rack
“memory, like DRAM and NIN, it's like 80% of the silicon wafer area inside that rack is memory”
Listen at 28:22
Material memory capacity additions will begin appearing in 2027–2028
“It'll be 27, 28 before you start to see material amounts of capacity start to come on.”
Listen at 28:39
Global WFE spending is forecast above $150B in 2026, $200B in 2027, and $250B in 2028
“We will probably wind up spending... $150 billion plus on WFE on equipment this year in 2026. We are over $200 billion next year, and I think over $250 billion in 2028.”
Listen at 30:14
WFE spending may reach $300 billion or more by decade-end
“I have no problem thinking that WFE needs to go $300 billion or well over”
Listen at 30:37
Semiconductor capital-equipment stocks are attractive
“because of that, I'm actually quite bullish on the semi-cap names”
Listen at 31:55
Applied Materials is Rasgon’s slight preferred semiconductor-equipment stock
“If I had to lean a little toward one, maybe it's AMAT a little bit. Just because they have more DRAM exposure, we need a lot of DRAM to make the HBM. And it's a little cheaper.”
Listen at 32:22
Lam Research has substantially diversified beyond NAND exposure
“they've done a really good job over the years of diversifying”
Listen at 33:37
NVIDIA’s investment and CUDA ecosystem strengthen its competitive moat
“NVIDIA is investing like crazy and they have software that people call it CUDA.”
Listen at 39:45
The AI semiconductor opportunity remains large next year
“Right now, it's big, and you can look at the numbers for next year.”
Listen at 41:13
AMD could reach $40–$50 billion in AI revenue if it ramps successfully
“AMD, if they can ramp the rats, I mean, they could do $40, $50 billion”
Listen at 41:33
Rasgon is not currently concerned about market-share losses among covered companies
“I actually fundamentally, I don't really worry about market share for any of my guys”
Listen at 42:02
Rasgon currently feels unusually positive about Intel
“I feel better about Intel right now than I have maybe ever”
Listen at 42:27
Strong server demand is causing customers to buy available Intel products
“Demand is so strong that customers are buying anything they can get their hands on.”
Listen at 43:02
Intel will not generate material foundry revenue for several years
“I still think it's not going to be for quite a few years before you see any sort of material foundry revenue”
Listen at 43:32
Intel’s current situation is better than it has been for some time
“I think things are better for them than they have been in quite a while.”
Listen at 44:55
Semiconductor channel inventories have remained high for four or five years
“However, they've been high for like four or five years now. And this is one problem I'm having.”
Listen at 46:00
Historical semiconductor cycle indicators may no longer be representative
“I don't know how representative these historically typical cyclical indicators are anymore.”
Listen at 48:01
Current AI bubble risk is more valuation-driven than earnings-driven
“Like we don't have the same kind of multiple driven, like if you're worried about a bubble, it's definitely on the multiple side.”
Listen at 49:05
Current AI infrastructure demand exceeds available supply
“I think actually right now demand is, nobody has enough.”
Listen at 49:19
Semiconductor fundamentals continue improving overall
“I still think fundamentals overall are getting better not worse”
Listen at 49:53
Rasgon favors traditional AI stocks, semiconductor equipment, and AMD
“we like the AI, the traditional AI names. I also like Semicap. On the CTU side, we've liked AMD.”
Listen at 50:10
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.