Never doubt the Matthew Prince #-of-Tweets Signal

Seems like DeepSeek v4 mini is out, though no official confirmation yet. Context window has increased from 128k to 1m tokens

V4 is rumoured to have a new architectural design relating to how the LLM handles data storage and retrieval. The engram conditional memory module is expected to drive NAND demand as it employs a tiered memory approach that utilises SSD to handle storage overflow from long-context agentic workflows, instead of relying exclusively on expensive HBM/DRAM.

Coincidentally $SNDK is trading at local lows around $545 from investors front-running $WDC share sale… Seems too good to be true but what if?

https://x.com/HAOHONG_CFA/status/2021630653226455243?s=20

Just 4 months later, this is an output by the frontier video model. If this was 90 minutes long, most people would have thought that it's a proper film. Progress is exponential, now project that rate of change 5 to 10 years into the future.

Factoring in an exponential increase in multi-modal outputs + NVDA ICMSP + DeepSeek Engram (may encourage other LLMs to also adopt this architectural design *if it works*) + rise of agentic AI workloads, I dont think the market has fully priced in this sea change yet. Can the shortage really ease by 2027 when there's not much greenfield/expansion capacity coming online? Samsung and SK Hynix are even cutting NAND starts to prioritise HBM/DRAM.

Napkin maths:
- currently we're around 10% undersupplied in global NAND demand
- NVDA ICMSP adds +6 to +10% of global NAND demand
- consumer destruction, assuming 10-20% of a 60% market, equates to -6 to -10% headwind
- new fabs/expansions meets +6 to +9% of global demand in 2027
- net net we're still -2.5% undersupplied, with conservative estimates on NVDA ICMSP demand, consumer destruction, and assuming no delays to all planned greenfields/expansions
- some might bring up china players as a supply risk, but I think they wont have enough NAND for themselves to engage in a price war
- we're also not considering the optionality of HBF (binary case), surge in multimodal outputs, deepseek's effect, rise of inference and agentic AI workloads
- while being -2.5% undersupplied is less than -10%, which might translate to pricing power, I still think many are underpricing the through-cycle earning power of these companies

$SNDK $285A.JT $PSTG

First 3 quarters 20% adj ebit, but FY guides for 68.2% midpoint adj ebit.

First 3Qs $1.33tn yen revenue, but FY guides for $2.17tn yen revenue.

So Q4 alone = $0.89tn rev and $0.48tn ebit? Q4 = 67% of first 3Qs rev and 175% of first 3Qs ebit? Q4 ebit% = 54%? In-line with $SNDK implied ebit% guide for next Q

Kioxia went HARD after the LTA expiration. $SNDK $MU being pulled up in overnight session. Would imagine $AAPL is a short from this

$285A.JT

https://x.com/mattshumer_/status/2021256989876109403

Good read; feels like early COVID when 99% of the population still has no idea that the world is about to be turned upside down.

I suggest everyone read this, then take an hour, turn up the AGI-pilled lever to the max, and wargame out all the potential states of the world in 5 years, in 10 years, and in 20 years. I think you’ll find that there really aren’t that many endgame outcomes - and it’s either really good, or really really bad.

Today it’s SaaS under attack, tmr it’s ad agencies, then it’s movie studios, game studios, software engineers, … and every other industry.

I can see a “AI development is recessionary narrative” slowly forming..

Closed $NET long @ $185-$193 for slight loss, think too much risk in the valuation given this environment

This “AI is recessionary” narrative is starting to grow on me. CPI was expected to be a hot print but came out ice cold. Though is a cold CPI really positive?

If AI is a structural deflationary force like we’ve never seen before, isn’t the US30Yr Treasuries at 4.73% a steal now? Assuming Bessent is able to grow their way out of debt or at least convince the market that they can (big if!)

Longed short and long-end treasuries $SHY @ $83 & $TLT @ $89.2. Disclosure: I’m kinda bad at macro

“Where’s the compute demand and ROI going to come from?”

https://x.com/pingtoven/status/2023186798927892811?s=46

Btw no idiosyncratic reasons for cutting $NET, just prefer not to hold stocks w high multiples now.

High multiples = majority of valuation lies in terminal value. And there’s no way one can be certain of terminal value in today’s ai-accelerative environment

IMO there’s better ways to play the long side on the agentic ai theme, such as volume plays like $DOCN / $RPI (?). Reasonable valuation, pure plays, relatively beaten down pre-AI, new tailwinds that contribute directly to rev, decent operating leverage

https://x.com/jaredkubin/status/2023586485744681214?s=46

Shifted 70% of my $SNDK position @ $600 into $MU @ $397. Valuation for both is approximately the same. I preferred $SNDK because it’s a NAND pure-play and the lower market cap = higher beta, but it’s nearing $100b mcap now so it’s not exactly that much easier to move anymore vs Micron in the context of institutional money - all while $MU has more enterprise ssd and dram exposure. Using $285A.JT (Kioxia) as my pure-play NAND vehicle.

LITE is actually ahead of its peers in terms of innovation more than the market realizes. First, $LITE and $COHR already maintain a duopoly on InP lasers.

Beyond that, $LITE is likely years ahead of $COHR in producing leading-edge continuous-wave lasers that are ultra-high-power and narrow-line-width with low (RIN). These lasers are an irreplaceable component in NVIDIA’s pursuit of the next networking frontier: scale-up CPO.

This is holy grail technology. It extends the reach of an interface designed for <2 mm of channel length (die-to-die electrical signals) to potentially 100+ meters. This means the unified memory limit for GPUs could jump from 72 GPUs to 576+, and total pooled HBM from 13.5 TB to 165 TB, unlocking a 10x increase in compute capability.

And only ONE company is capable of producing these lasers, the most critical component of this tech. To further contextualise this, normal lasers used in optical transceivers are already hard to produce (fragile, low InP yields). But continuous wave lasers are even harder (ultra high power, always on, must not have flickers in intensity or frequency). And now, we’re trying to put that fragile, UHP, always-on laser right next to a GPU that can go from 40 Celsius to 85 Celsius in seconds... 1) the laser might actually melt, and 2) for every 1 celsius change in temperature, the laser’s colour shifts by 0.1nm. If it shifts too much, the data/signal is voided entirely. That’s how hard this physics-defying challenge is, and also why it has been delayed for so long despite the obvious benefits.

However, NVIDIA’s ISSCC presentation of their optical NVLink solution this week may have been the catalyst that helps the market realize the importance of this tech. Much like what NVIDIA’s CES speech did for SNDK and the role of NAND in AI infras and context management.

https://x.com/claudeai/status/2024986293248127452?s=46

As a comment succinctly summed up, winners of the hackathon are…:

- a personal injury attorney
- an interventional cardiologist
- an electronic musician
- an infrastructure/roads systems worker
- a software engineer

If held just a year ago, winners would have likely been all software engineers.

From our thesis above: “the best applications can now come from the most creative individuals, rather than the most well-funded engineering teams. It can come from ANYONE with an idea that doesn’t even need to be original!”

Gone is the day of general-ish SaaS, and in is the age of highly customisable & specialised niche software - built just for yourself, by yourself.

https://x.com/qwqiao/status/2019199428578459996?s=46

Adding on to this commentary, it’s time to treat coding/agentic coding/agentic AI seriously and compulsory. You need to at least develop a 6/10 expertise in it to function in the world in future, just like English I think.

Then, build a 10/10 expertise in your chosen domain outside of coding (investing, engineering whatever) that no one else can replicate, go really deep and get really good at it.

Choose your domain wisely though as some might cease to exist in future. My rule of thumb: domains resembling poker games (imperfect information, probabilistic thinking, human psychology, optionality, payoff and risk centric thinking) are safer than domains resembling chess games (perfect information, deterministic outcomes, rigid rules, solvable, do x get y).

Use your 6/10 coding skills to complement and take your domain skills to the next level. That’s my framework for future proofing as hyper-specialisation accelerates.

I think being 10/10 domain and 6/10 coding > 8/10 domain and 8/10 coding, because specialised domain knowledge will be worth a lot more in future and agentic coding is going to get so good. So you don’t need to be top decile at coding skills to extract top decile value from coding

Lastly the most value adding thing you could possibly do this weekend, professionally speaking, is to take 2 hours to set up Claude code/openclaw. Lots of guides online. Then try creating a simple calendar app with a 3 sentences prompt. Then expand from there. That’s literally it, it’s so much easier than people think.

And that will flip a switch in you, I assure you. You will finally understand the AI optimism/excitement with true conviction. You can’t grasp it by reading about it in headlines. Trust me, this will be more value-adding than the hours that you will spend being buried in an excel model.

This best be HBF-related and not mis-messaging about the company’s 1yr spinoff anniversary from $WDC… 😤

$SNDK