Nagendra Puthane retweeted
AI isn’t fake, and it’s likely not about to create superintelligence anytime soon. Are AI stocks overvalued? Possibly. Could the market take a 10–20% haircut? Sure. But the human productivity gains are real. Notice how it’s mostly the non-technical crowd comparing this to 2000.
What makes this different from the dot-com era is how transformational and easy it is to integrate into daily life. In 2000, the digital infrastructure wasn’t built yet. There were still picks and shovels to sell. Not everyone had computers, internet, or connectivity. The buildout eventually happened, it just took a decade.
AI doesn’t need that kind of foundational buildout. With one click, anyone can access a state-of-the-art model powered by hundreds of GPUs. Every LLM query touches NVIDIA GPUs or other AI accelerators. The human computer interaction already exists, and we’re still compute constrained to accommodate the masses.
Evolution of DRAM (Dynamic Random Access Memory)
1. Async DRAM (pre-1990s)
2. Sync DRAM (early 1990s)
3. DDR [Double Data Rate] SDRAM (2000s)
3.a. DDR3 - 2007
3.b. DDR4 - 2014
3.c. DDR5 - 2020
4. GDDR [Graphics]
5. HBM [3D stacked]
5.a. HBM 1/2/2e/3/3e/4 - 2013-present.
Memory & Storage thesis getting stronger
ELON MUSK ON X PHONE:
I am not working on a phone. I can tell you where I think things will go, which is that we’re not going to have a phone in the traditional sense.
There won’t be operating systems. There won’t be apps in the future... you’ll have … AI on the server side communicating to an AI on your your device … generating real-time video of anything that you could possibly want
It’s probably five or six years or something like that.
Street consensus for Data Center revenue over a similar two-year window is about 455 billion, so NVIDIA’s product-specific 500 billion is roughly 10 percent higher but not directly comparable because of differences in scope, timing, and China/networking inclusion
$NVDA NVIDIA’s $500 figure represents cumulative revenue visibility for Blackwell + Rubin systems including networking across calendar year 2025 and 2026, excluding China, and Jensen’s “next 5 quarters” phrasing was a misspeak corrected by the finance team.
$BE Bloom (fuel cells + all-DC ) at ~$51/MWh vs $META Louisiana (gas turbines, AC distribution) ~$96/MWh (large turbine) and ~$156/MWh (small turbine)
Bloom Energy: Highway to Direct Current
Bloom is the only baseload power generation technology that natively outputs 800V Direct Current electricity, the voltage standard for next-gen AI servers from NVIDIA, AMD, and Google.
That means AI datacenters powered by Bloom fuel cells can cut electrical + battery costs by >60% vs. other baseload electricity sources such as gas turbines, nuclear, or the grid.
The future is Direct Current.⚡️
Disclosure: At time of publication I own a position Bloom, which I may change at a future point in time without further notice.
Not investment advice!
linkedin.com/posts/activity-…
Lam Research ( $LRCX ) – TSV etch & dry-resist partner
Exclusive deep-silicon TSV etchers for every HBM stack, plus new EUV “dry-resist” etch steps and cryo DRAM capacitor etch. Analysts put Lam’s share of HBM-oriented etch at >$1 B through 2026 just from SK Hynix.
The first equipment has arrived at SK Hynix's M15X fab(for DRAM, HBM, etc)
The goal of completion within the year is expected to accelerate.
yna.co.kr/view/AKR2025102707…
6/
🧱 Why this matters in the Memory Super Cycle:
• DRAM is moving from DDR5 → HBM3E → HBM4 (12–16-high stacks).
• NAND is going from 200 → 400+ layers, switching to molybdenum liners.
• Each extra layer = extra Lam process steps.
5/
📈 Revenue exposure (FY25 est):
• ~42% Memory (DRAM+NAND)
• ~45% Foundry/Logic
And roughly 75% of Lam’s systems revenue is from etch + deposition tools.
👉 ~35–37% of total Lam revenue today is directly tied to Memory & Storage fabs.
4/
🔩 What Lam sells to them:
• Etch systems — carve out ultra-deep holes in NAND stacks & HBM vias.
• Deposition systems — fill them with tungsten/molybdenum.
Every new NAND layer or HBM stack → more etch + depo = more Lam revenue. 📈
2/
Lam Research doesn’t make chips.
It makes the machines that make chips.
Think of it as the pick-and-shovel provider to DRAM, HBM, and 3D-NAND fabs - the factories that power AI training, data centers, and storage.
Lam’s tools literally etch, deposit & clean every memory layer.
*Correction:
•$TSM = wafer winner inside CPO.
•$TSEM = wafer winner for next-gen pluggables (and possibly future CPO receiver tiles), still levered to optical growth but not the very first CPO package.
Will it spike as violently as lasers or PIC wafers? Probably not; fibre is a lower-margin, steadier grower. But CPO keeps Corning on a secular up-slope instead of flattening out, because bandwidth density keeps translating into strand density.