Q3 Revenue increased 96% from Q2 up to 5.95B. Q4 guide is 8B at the midpoint which is a 34% increase sequentially.
Gross Margins increased 27.5% YoY and have climbed to 78.4%. Net margins has grown to 60.7% this last quarter with 3.615B in net income.
This is my first exposure to any memory companies and I can’t figure out if this is part of the memory cycle or if the cycle is temporary broken and with data centers/AI if the memory pricing will stick around longer than usual memory cycles.
SNDK is still only has a forward pe of 7 which seems very cheap although I admittedly cannot figure out how to valuate memory companies.
Does anyone have more experience with the memory cycle and have an insights on what’s to be expected moving forward?
This is a bit simplistic, but all tech cycles historically reward companies in the following order:
Picks and shovels (think NVDA so far in this round)
Hardware/Infrastructure layer (which seems to be where we are at present)
Software layer (still TBD with some fear the AI revolution might make this layer obsolete)
I’d guess the AI cycle will be no different as more companies enter the space and margins eventually get compressed and/or commoditized. While I wouldn’t say this time is different (because it never is), I’ve heard @GauchoRico suggest the enormity of this buildout could create a higher-for-longer effect for AI’s winners.
Regardless, it certainly seems like we are still early in the AI cycle with the market giving SNDK some love as a potential winner in the memory phase.
Quite! SNDK is up 400% ytd after being up something like 700% or 800% last year. Micron is up 90% ytd after more than tripling last year. I think the OP’s question was how we can know whether this will continue. For that, we should look to the numbers.
The fwd PE is indeed about 7. Last year their “E” was negative (at least on GAAP). This year ~$65 eps is expected. Next year ~$165. Why? Pricing/short supply.
Unlike companies where the revenue growth is at least somewhat predictable based on shipping more units year after year, revenue is spiking for these memory chip makers mostly due to ASP (average sale price) exploding. That’s happening (I think) due to supply constraints. This also makes SNDK’s (and MU’s) margins explode.
I don’t believe anyone can predict these dynamics. Correct me if I’m wrong, but @stocknovice seems to imply that margins will stay high as long as AI demand does. I don’t know enough to confirm or deny that, but I think it’s possibly incorrect. Possibly the pricing will come down even as more units are required in the coming years. Can supply catch up? What other factors will be at play?
Or more specifically I guess, pricing power and margins for this portion of the cycle should in theory hold up as long as supply is constrained.
This. ^^
The second Jensen Huang announces NVDA produced one more chip than they sold in some future quarter is probably the time to head for the exits. Infrastructure, memory, etc should all follow suit from that point even if it takes a quarter or two to trickle down. Just as NVDA has powered this cycle’s climb, it will likely lead the cycle’s rollover and flush. The challenge is having enough returns built up to get off the rollercoaster with a profit when the inevitable haircut arrives.
One of things I’ve learned from Saul over the years is the power of cutting bait entirely when things don’t look right rather than waiting through losses by a thousand paper cuts once a thesis has changed. The AI trade % gives me that type of vibe.
I’m just not sure we’ll see Jensen say that before the memory chip supply loosens up. I think you need another way to know when to sell. Because at some point these will get cut in half and I just don’t see a way to know when.
I don’t have a way to know when to sell. So I’m just not buying the memory chip co’s.
The biggest mistake people have made buying memory companies is on P/E. Typically you sell them when P/E is low and buy when P/E is highest or non existent (losing money).
It’s counterintuitive but when P/E is highest new capacity is not entering the market and when it’s low you can expect capacity flooding. This cycle could be different due to AI. Also, Sandisk (NAND Flash) has many competitors while DRAM is an oligopoly between Samsung, Micron and SK.
Mark Twain said: “History does not repeat itself, but it rhymes.”
The AI boom today has parallels in the Railroad Boom of the late 1800s:
Andrew Carnegie made a ton of money providing steel. The US government had lots of land, but not much money, so Land Grants enabled railroads to cross the country.
Vanderbilt, Stanford, and others built the railroads. Notably, Vanderbilt came from Steamships, so he “made the transition.” Rockerfeller was a successful application, using railroads to transport oil and coal.
Railroad mileage expanded from roughly 45,000 miles in 1870 to over 170,000 miles by 1900, with the financial Panic of 1873 tied directly to over building, high debt, shenanigans and fraud.
The Internet Boom was similar, with a bit less fraud. Telecom companies (eg Worldcom, Global Crossing) laid fiber no one yet needed, hardware companies (Cisco, Nortel, Lucent) ran out of customers. Software Infrastructure companies (eg Yahoo, AOL) struggled, with eventually the Software Application companies (eg Amazon eCommerce, Google) rising and still dominating today.
There were too many miles of railroad track in the late 1800s, and there were too many miles of fiber in the late 1900s. We can look at electrical power plants today as the Land Grants of the 1800s. Running a railroad wasn’t always profitable, and there were lots of buyouts and mergers, just as we saw with Internet providers and “portals” in the 2000’s.
Yet, eventually, all of the laid tracks were needed and used, just as all of the “dark fiber” is now lit up. With even more added since for both. The optimistic rationales for the Booms were eventually realized, just not on the timetable needed for some/many businesses to survive. I think that’s the real issue for us now.
Somehow Cloud Computing didn’t become a bubble, though. Like Vanderbilt migrating from steamships to railroads, Amazon migrated from eCommerce to inventing Cloud Computing. That needed data centers, which needed power and land and fiber, yet somehow all those build-outs went pretty darn smoothly.
The Cloud Computing technological revolution did exactly what those in its forefront expected: Startup companies in the space thrived, application companies that didn’t jump on the bandwagon early got punished. Salesforce went from a little startup to dominating legacy companies like Siebel Systems and SAP. Online database companies like Mongo came about, but Oracle managed to survive, too.
We could expand our scope further and look at both the eCommerce and mobile/smartphone revolutions, too. In those, there were both winners and losers, with startups like Amazon growing to dominate while Walmart and Gap, etc. managed to survive. And mobile hardware came to be dominated by a desktop computer/music player company - Apple, sending Blackberry and Ericson and Nokia to the dust bin (the latter to be acquired by tail-between-its-legs Microsoft and Sony only to have both give up).
So, what’s the rhyme for AI?
In my mind, there’s no doubt that, like railroads, the internet, and cloud computing before it, AI will change the world in both how business is done as well as how people live. Will it become over-built like railroads and internet fiber, or will adoption grow in line with its supply like Cloud Computing and mobile? That’s really the question here for us.
I watched a video from Steve Eisman (famous from The Big Short), and his take was that a major difference is that companies investing in AI today are not borrowing money like the railroads and internet startups, but are investing most of the profits from existing businesses instead. He thinks that both moderates their growth as well as avoiding a leverage-inducing snowball should things slow down.
FWIW, I see the AI development layers as:
• Hardware: Chips, memory, networking, servers
• Hardware Enablement: Chip fabs, Chip design, power plants, data centers
• Software Platforms: Low-level Operating Systems, AI Frontier Models, Agentic Frameworks
• Software Applications: Chatboxes and Agents doing everything all software does today, plus what humans do with software today
I’ll postulate that, in addition to avoidance of financial leverage, the technology revolutions in Cloud Computing and mobile were successful because that they both had applications up and running from the get-go. Amazon literally took the private cloud eCommerce implementation it had and just exposed the infrastructure part to others (since they had done a great job architecting it in the first place), and showed others how they could easily migrate many of their on-premise applications to it. The iPhone literally brought 3 existing applications (a phone, ipod “music”, and internet browser/email) on its initial release, added more, then opened it up for others.
Assuming my grouping/layering is decently appropriate, what are the companies?
• Hardware: Nvida, Credo, Astera, Lumentum, Coherent, Arista, Micron, SK Hynix, Samsung, Dell, SuperMicro (ugh)
• Hardware Enablement: TSMC, ASML, Bloom Energy, Vertiv, CoreWeave, Nebius, Iris (IREN), Eaton, Lam Research
• Software Platforms: Nvidia’s CUDO, Omniverse, Triton, DLSS, Drive (Alpamayo), AI Enterprise, plus Anthropic’s Sonnet & Opus, OpenAI’s GPT, Perplexity, Cursor, Google’s Gemini, Palantir maybe, DeepSeek
• Software Applications: ChatGPT, Claude CoWork, ClawBot, DataDog’s SRE, ServiceNow’s ITSM, Salesforce AgentForce, Adobe Firefly, DeepSeek’s SeeDance, and many others.
Note that many of the AI applications are from legacy SaaS software companies, indicating that Mr. Market’s concern may not be with the products as much as the business models with which they are able to charge for them.
As for the question on Boom/Bust cycles and SanDisk in particular, it’s hard for me to say. A year or two ago some people thought Nvidia would hit its typical boom/bust cycle as it did with gaming and crypto, but that was clearly wrong and looks to be wrong for at least the next year. What’s the competition for NAND and SSDs and HBF, or any new technologies?
CEO David Goeckeler said the company has signed five long-term supply agreements with customers that range between one and five years in length. Three inked during the company’s third quarter ended April 3 were worth $42 billion, while the other two were signed in the current quarter.
“The bane of this industry has been the boom-bust cycle,” Goeckeler told Reuters in an interview. “We want to get out of that. We want consistent, predictable economics.”
Goeckeler said that he understood some investor skepticism about long-term agreements in the memory business, which have been tried before and failed when customers renegotiated them amid slack demand.
Sandisk has avoided those pitfalls, he said, by including a variety of terms such as price ceilings and floors, adjustments based on market demands but also clear terms that do not allow customers to walk away without paying.
“Consistency is very important to me,” Goeckeler said. “We put a financial structure in place that says at the beginning of the contract, if you make a financial commitment to me as the customer, if you walk away from a contract, I get that money.”
So, given their constrained production capacity, how much room is there for production growth? I’m not willing to bet on just price increases for revenue growth, if that’s what’s happening.
For instance, I’ve avoid ASML (only maker of the EUV lithography machines needed for high-end chip production), because the company has said they can only expand production of those machines by a small amount and isn’t even going to raise prices enormously. So, a very essential and high demand company won’t, in my view, see its revenue grow as much as the market is pricing in.
I stumble thru SemiAnalysis semianalysis@substack.com on a weekly basis, not understanding much of the tech info and many of the acronyms, but each week there few nuggets to pick up.
A couple paragraphs from the long article yesterday,
"This rapid pace of AI adoption has created value across the stack, but the unique phenomenon is that the AI labs are capturing all the value now, from almost none last year.
End users are enjoying a productivity bonanza, tasks that used to take tens of person-hours costing thousands of dollars can now be accomplished in minutes with a just a few dollars’ worth of tokens. This huge surge in revenue and margins is because the value of tokens being created is dramatically improving businesses. For example, SemiAnalysis has reached as high as $10.95 million dollar annual spend rate on Anthropic Claude tokens, but the value we derive allows us to outcompete all our competitors and gain market share.
New chips such as Blackwells can generate 30x more tokens per second while running frontier workloads today vs Hoppers a year ago, and ASICs such as TPUv7 and Trainium 3 show similar improvements. Inference providers such as Fireworks, Baseten, Fal, margins are widening while their revenue trends are in hyper growth.
Even parts of the hardware stacks have repriced, with memory prices having gone up 6x in the past year. Neocloud GPU rental pricing is surging as well, up with 1-year H100 rental contract prices up 40% from the bottom in October 2025.
There are two firms in the industry with incredible pricing power that haven’t moved much though. TSMC and Nvidia have not reacted to the recent boom in value generation of AI models."
I have sought to find publicly traded companies with dominant focus in ai labs, and found none, but AI made the following statement.
Here is a summary of what I found.
“Public companies with major AI labs, but none are exclusively focused on AI labs or even AI in general
These companies operate some of the world’s most advanced AI research groups, but AI is not their sole business:
If by “AI labs” you mean what is commonly referred to as “AI frontier labs,” then like everyone else, you’re waiting for OpenAI and Anthropic to go public as the first AI model only/primary companies.
Google, of course, has DeepMind, but Google has lots of other businesses.
xAI has Grok, but that’s part of Space-X now, also IPOing this year.
Finally, Meta (Facebook) has its Muse Spark, but so far that’s for their internal use only.
See my post up-thread for some companies in the deployment space, as well as the feeder companies and downstream consumers.
We are all aware of the major labs and the lab efforts of META, GOOG, etc., but I thought there may be some smaller obscure, public labs with significant upside, as Semi Analysis focus this week was the shift of value flow into the AI lab sector.
Just coming back to your question here. As others have provided excellent input on the AI scene…
Historically yes memory has been a cyclical industry. If you have access to Seeking Alpha - here are a couple of articles from folks who clearly lean into the Memory is still in a boom and bust cycle opinion.
We have seen cycles forever in memory. Typically this is a supply and demand balance creating pricing volatility. As capacity investment comes on stream pricing weakens and value slumps even if volume remains stable. Cyclical maximalists also point to historical “supercycles” just in case you felt that this applies to the current situation - such as the PC boom in the early 90s, the mobile boom in the 2000s and the data centre boom in the late 2010s.
Considerations to help you evaluate: whether we are still in a cyclical situation, if so where and what the risks could be, would include:-
Has memory consolidation help create barriers to entry for new supply?
Have memory makers become more rational and disciplined about Capex driven capacity expansion when facing demand surges?
Have multi year and price flexibility in contracts help dampen the cycle and reduce risk?
How early on are we in the AI driven demand boom (for HBM, NAND and DRAM)?
How much additional demand support can be expected as the crossover happens where flash overtakes spinning disks in total cost of ownership going forwards?
How much demand for memory do the S curves potential have from: autonomous vehicles, Robots, Drones, Smart Devices etc, that are all coming on stream?
To what extent does Say’s law, (supply creates its own demand) or Jevon’s Paradox (as efficiencies are achieved additional uses cases increase demand), apply here?
Whilst it isn’t completely clear cut and it’s possible we face a Cisco saturation moment at some point, I do find it helpful to bear in mind that Nvida and processor chips were considered cyclical and nobody is talking that way about Nvidia right now.
On balance I believe in enough positive answers to the above questions that I re-entered Micron at ~$100, (about where I exited after a previous cycle up from $25) and am happy to hold as it approaches 10% and vying for the #1 position in my portfolio.
If memory demand has grown like that, how is it different this time with AI?
Bear
PS - I don’t think anyone has offered a good reason yet for why supply and demand won’t find equilibrium again, sending ASPs way down. If not simply supply increases (with all the capex MU and competitors are spending) then innovation is another way it could happen. Even if all that takes a while, the market is forward looking.
So, you’re looking for MLMs (Medium Language Models) or SLMs?
There actually are some research and other companies in this space. The thing about LLMs is that they take alot of money to train, so smaller models are cheaper to train, even if they’re special purpose in nature.
One company I know of is SoundHound, which started as music recognition, but has pivoted to AI voice solutions (TTS - Text to Speech, NLU - Natural Language Understanding, and ASR - Automatic Speech Recognition). Some cars use their stuff, as well as things like automated phone ordering systems for businesses. I have not researched the company at all.
I was reading that HBM3 required 3-9x the waifer to deliever the same amount of memory as traditional storage. Because they are layering it they can’t be as dense, they need pass through lanes. They also have lower yields with stacking problems. I was reading that they expected it to get worse when going from HBM3 to HBM4 because they are stacking more.
HBM scaling is a complicated technical subject. This is as good an article (from Tom’s Hardware) as I could find, and it’s not for the technically faint: https://archive.li/mzrM4
“AI compute scaling is driven by the combination of advanced logic, SoIC 3D stacking, and CoWoS technologies,” a statement by TSMC reads.
“HBM bandwidth scaling comes from multiple factors,” said TSMC. “First, there is the memory itself — progressing from HBM3 to HBM4, with higher I/O counts. In addition, we are leveraging more advanced logic technologies for the base die, which allows us to push data rates well beyond 10 Gb/s per pin, something that was unheard of in traditional DRAM. At the same time, our CoWoS technology enables integration of more HBM stacks within a single package. […] All of these factors together — higher data rates, more I/O, and more stacks — contribute to the overall bandwidth scaling.”
Part of what’s going on is the need to keep the HBM memory as near the GPU as possible to reduce latency. That’s where CoWoS comes in, and eventually they’ll stack the GPU and HBM on top of each other, which is as close as you can get, but then there are heat removal issues to deal with.
I don’t know what this means for the boom-bust cycle of memory. That is probably more dependent on the world’s AI needs and whether there are any changes to memory requirements in running the training or inference loads as the software gets better.
% gives me that type of vibe.
