Semiconductors, memory, optical networking and AI infrastructure. Following earnings, capacity and market cycles.
- Tweets 573
- Following 346
- Followers 216
- Likes 693
Pinned Tweet
I started following $MU when the stock was trading around $350.
It has delivered plenty of surprises since then, but the reason it remains at the center of my Memory allocation is the company Micron could become over the next three years.
Demand for HBM, server DRAM and enterprise SSDs continues to grow with #AI server deployment.
HBM production is also consuming more leading-edge DRAM capacity. At the same time, Micron has signed multi-year strategic customer agreements with minimum volume and pricing commitments, giving the company greater visibility into future demand.
I spend less time trying to predict whether DRAM prices will rise another 10% or 20% next quarter.
The more important question is whether Micron can enter the next three years with a larger revenue base, a stronger competitive position and more durable earnings.
If these trends continue, Micron’s growth will extend beyond a single round of Memory price increases.
That is why MU remains one of my core positions in this part of the market.
Following the stock from around $350 has already included rallies, corrections and plenty of disagreement. Long-term growth investing requires looking past each short-term swing and asking whether the underlying business is becoming stronger.
My research and positioning will continue to center on AI infrastructure and Memory. I plan to stay with this story through its next stage.
#TechStocks #LongTermInvesting
7
2
24
1,261
The Next Optical Bottleneck May Be Thermal
As 1.6T optics move more data through a small form factor, heat becomes part of the connectivity problem. Lasers need stable operating temperatures, which raises the importance of TECs, ceramic substrates and packaging design.
This is an adjacent read-through for $LITE rather than a direct revenue claim. A tighter thermal budget can increase component value across the #AI optical chain, but suppliers still need qualification, yield and volume production.
Bandwidth gets the headline. Thermal management may decide how much of that bandwidth can be deployed reliably.
6
Everyone talks about GPUs, but this is where the AI buildout gets really interesting for me.
Memory at 48% of AI capex in 2026 and potentially 53% in 2027 changes how I look at this cycle. More compute keeps pulling more HBM, DRAM and storage with it.
$MU $SKHY #Memory
$MU $SKHY Memory and context are the biggest contributors to intelligence. AI Capex has priced that in.
The stock market hasn't.
@elonmusk on the SpaceX earnings call: "Memory output is increasing by around 20% per year... demand is increasing by 200% a year, maybe higher."
77
This chart explains why I keep coming back to memory.
AI keeps demanding more compute, but every step up in compute pulls more HBM, DRAM and storage with it.
If memory really reaches 53% of AI capex in 2027, $MU and $SKHY are sitting much closer to the center of the AI buildout than the market used to treat them.
$MU #HBM
$MU $SKHY Memory and context are the biggest contributors to intelligence. AI Capex has priced that in.
The stock market hasn't.
@elonmusk on the SpaceX earnings call: "Memory output is increasing by around 20% per year... demand is increasing by 200% a year, maybe higher."
4
121
The Constraint Behind $LITE Is Indium Phosphide Capacity
Higher-speed optical links need more advanced lasers, and scaling indium phosphide production is neither quick nor simple. That makes manufacturing capacity and yield as important as end demand.
Lumentum is expanding U.S. laser production while the #AI data-center buildout keeps raising bandwidth requirements. If demand stays ahead of qualified supply, pricing can remain firm.
The risk is timing. Customers can redesign, competitors can add capacity and new factories take time to reach target yields. The thesis should be measured through shipments and margins, not unsupported market-share estimates.
18
The AI optics cycle has pushed Lumentum into a very different valuation range, but a rising stock price does not make laser pricing power automatic.
Demand is strong enough for the company to expand InP-based laser capacity, and NVIDIA’s investment and purchase commitment reinforce the strategic importance of the supply chain. The market now expects that tight capacity will support growth and margins.
That expectation still needs to show up in reported results. Product mix, customer concentration, capacity ramps and average selling prices will matter more than a single day of relative weakness.
$LITE $MRVL $INTC
17
$RKLB Has Reduced One Major Iridium Deal Risk
RKLB has made concrete progress on financing its proposed acquisition of Iridium. A September 15 filing shows that Rocket Lab raised about $1.944 billion through its at-the-market program, secured amendments allowing $1.775 billion of Iridium term loans to remain outstanding after closing, and terminated the original $3.6 billion bridge commitment.
Management says the ATM proceeds, the amended Iridium facility and other available funding are sufficient for the expected cash portion of the transaction. That reduces financing uncertainty, although the deal still carries closing, integration, leverage and dilution risks.
The strategic question is whether combining launch, space systems and Iridium's satellite network can create a stronger end-to-end space platform while Neutron is still under development. For $RKLB, execution on the transaction and the Neutron schedule will matter more than any analyst price target.
74
$INTU Has a Valuation Problem, Not a Guidance Collapse
$INTU used its September investor day to reaffirm fiscal 2027 guidance. Management still expects revenue of $23.279 billion to $23.512 billion, representing 9% to 10% growth, with GAAP diluted EPS of $20.12 to $20.36.
The company is building its AI strategy around decades of financial data, domain-specific models and access to human experts across TurboTax, QuickBooks, Credit Karma and its other platforms. That installed base and workflow integration remain meaningful advantages.
The market is asking whether those advantages can translate into faster customer growth and durable operating leverage as #AI-native competitors enter financial software. The next phase for INTU will be judged through revenue growth, margins and retention rather than the number of AI features it announces.
2
142
The Optical Bottleneck Is Moving Toward External Lasers
$LITE is demonstrating an eight-wavelength DWDM external laser module at ECOC 2026 for next-generation CPO and NPO systems. The module delivers all eight wavelengths simultaneously, targets up to 24 dBm of optical output per wavelength and uses about 12 watts of power.
The design matters because moving the laser outside the hottest part of the package makes it serviceable while DWDM increases bandwidth density where package-edge space and fiber routing are constrained. Lumentum expects initial availability in the first half of 2027.
This is another sign that AI scale-up is pushing value deeper into the photonics supply chain. Demand will still depend on customer qualification and deployment timing, but external laser sources are becoming a more important part of the CPO and NPO roadmap for LITE, $COHR and their suppliers.
1
68
Why Optical Interconnect Is Becoming an AI Bottleneck
As #AI clusters scale from thousands of accelerators toward tens of thousands, the bottleneck is spreading beyond compute into data movement.
The architecture choices each solve a different trade-off:
◆ Pluggable optics are easy to service, but long PCB traces add power as lane speeds rise
◆ CPO places optics beside the switch ASIC for the best density and power profile, but makes yield, cooling and servicing harder
◆ NPO moves the optical engine close to the ASIC while keeping more flexibility for manufacturing and maintenance
The near-term constraint remains physical supply. InP wafers, EMLs and high-power lasers take time to qualify and expand, so optical demand can grow faster than the upstream capacity supporting it.
That is why the market is starting to value lasers, packaging and optical engines alongside the module makers.
$NVDA $COHR
3
38
B200 rental rates are rising while DRAM prices remain firm and hyperscaler cloud revenue estimates keep moving higher.
That combination points to real pressure across both compute and memory capacity. #AI infrastructure demand is still absorbing supply faster than many expected.
$NVDA $MU
Instructive to see a post like this get so much traction.
B200 rental rates up 30% YTD.
Hyperscaler cloud revenues tracking along.
Remember: GPU rental rates and token prices were supposed to fall over time. The demand metrics this year have beaten all informed expectations.
1
2
74
Lumentum Moves Optics Closer to #AI Compute
$LITE, Qualcomm and Corning are using ECOC 2026 to show a high-density optical die-to-die link built for future AI scale-up systems.
The demonstration combines:
◆ Qualcomm D2D interface IP
◆ A Lumentum 1060nm VCSEL optical engine
◆ Corning multimode-fiber connectivity
The current design supports about 1 Tb/s/mm of shoreline bandwidth density, with a path toward roughly 4 Tb/s/mm. Aggregate transmit and receive capacity is about 10 Tb/s, and the optical link can extend across tens of meters.
This is the real signal: optical connectivity is moving from the front panel toward the package and the compute device itself.
$QCOM $COHR
1
81
The OCS Demand Signal Is Real
The supplier split is still private.
I could not verify the claim that $GOOGL plans to buy 20,000 optical circuit switches in 2027 or that $LITE has already secured 15,000 of them. Those figures do not appear in public filings or company announcements.
What is public is already meaningful: Lumentum said its OCS backlog had moved well beyond $400 million, and the company is scaling production to meet customer demand. OCS is gaining importance as AI clusters need more flexible, lower-power optical paths. I am watching backlog conversion, manufacturing capacity and competition from $COHR before assigning market share.
$NVDA
1
92
400G per Lane Will Use More Than One Optical Architecture
The evidence does not support writing off silicon photonics.
At OFC, $LITE demonstrated a 4x400G pluggable module built with indium-phosphide EMLs. The company also showed two 1.6T modules using different paths: one paired its high-power CW laser with a silicon-photonics engine, while the other used 200G EMLs.
That tells me the market is still sorting architectures by reach, power, thermal limits, cost and manufacturability. InP remains critical for EMLs and external laser sources. Silicon photonics still has a role in integration and CPO as AI networks scale. The winners will be the suppliers that can scale the right device for each part of the network.
$COHR $NVDA
2
70
Agentic AI is turning storage into an active part of inference.
Every agent creates context, intermediate files and KV cache data. Keeping all of that in HBM and DRAM is too expensive, so more of the workload is moving into enterprise SSDs.
That is why weak consumer NAND demand does not tell the whole story. Enterprise QLC is developing its own buyers, pricing power and supply constraints.
$SNDK sits right in the middle of that shift.
#AIStorage
$SNDK
Funny thing. I've been hyper bullish on agentic AI for over a year, and it's one of the main reasons I hold the stocks I do, memory above all, and NAND in particular.
TrendForce just raised its NAND price forecast.
Enterprise SSD Soars, Consumer Stalls.
Enterprise SSD orders are still climbing into Q4 on North American cloud demand, and the two drivers they name are agentic AI and KV cache offload.
Agents generate state.
Every task an agent runs leaves behind retrieval results, working files and intermediate steps, and all of it has to be stored somewhere fast enough to be useful and cheap enough to keep. That's why cloud providers are scaling QLC SSDs, which pack the most bits per dollar of anything with an interface fast enough to serve inference.
KV cache offload is the other end.
Models need to hold context, and until now that lived in HBM and DRAM, the two scarcest, most expensive places on earth to keep anything. Push the parts you don't need every millisecond onto SSDs and you free up the expensive tiers for the work that actually needs them.
TrendForce says the industry is building frameworks specifically to do this, and China's inference architectures are already leaning on it.
The funny thing is the same report says consumer NAND is oversupplied, with YMTC adding capacity and a two-track pricing structure emerging in 2027.
Enterprise NAND it’s becoming a different product from the stuff in your phone, with its own supply, its own buyers and its own price.
The part of the market that agents need is the part that's short, and the part that's oversupplied is the part $SNDK doesn't depend on.
Bullish $SNDK
1
46
Apple's Premium iPhone Bet Faces a Demand Test
This launch cycle leans heavily toward expensive devices.
$AAPL introduced the iPhone 18 Pro and Pro Max alongside the foldable iPhone Duo, while the standard iPhone 18 did not arrive in the same launch window. That leaves the company asking customers to pay more for camera upgrades, the A20 Pro chip, Siri AI and a new foldable form factor.
The investor question is whether those features can shorten a replacement cycle that has been stretching out. Early reports have described softer Pro demand in some markets, while the Duo may be pulling attention toward its October launch. Unit mix and upgrade rates will matter more than launch-week excitement.
$QCOM $TSM
83
Glass Substrates Are Moving Closer to Commercial Use
The timetable is still measured in years.
$INTC introduced advanced glass substrates in 2023 and said complete solutions were planned for the second half of this decade. Industry reports also point to panel-level packaging pilots at $TSM, with broader production discussed around 2028 to 2029.
That progress tells me the ecosystem is moving beyond isolated lab demonstrations. It has not removed the hard parts: equipment readiness, through-glass-via yield, warpage control, customer qualification and cost. The next useful signals will come from qualified capacity and production orders. Glass still has years of commercialization work ahead.
49
Glass Is Becoming an Advanced-Packaging Material
Package size is turning into an engineering problem for AI chips.
Large chiplet packages place more stress on organic substrates through warpage, routing density and power delivery. Glass offers better dimensional stability and can support finer interconnects. Intel has said its test substrates can reduce pattern distortion by 50% and potentially increase interconnect density tenfold compared with current organic materials.
The opportunity is still tied to manufacturing. Through-glass vias, metallization, inspection, bonding and yield all have to work at commercial scale. I am watching the equipment and materials chain around $INTC and $TSM because those process steps will decide how quickly glass moves from promising samples into real AI packages.
28
The AI trade is reconnecting.
Hyperscalers are gaining as monetization expectations improve, while semis are confirming that the infrastructure cycle remains intact. Seeing both groups rally together after a long decoupling strengthens the setup for $NVDA, $MU and the broader supply chain.
Semis + Mag7/Hyperscalers up together today.
This has been a rarity this year (semi vs hyperscaler correlation at an all-time low).
The recipe for new S&P 500 highs is hypers + semis rallying together.
129
Corning's Price Increase Is Real
The AI packaging read-through needs more care.
$GLW will raise prices by at least 15% on yen-denominated display glass substrates beginning in the fourth quarter of 2026. Corning tied the move to the weaker yen and higher costs for raw materials, precious metals, energy and logistics.
That announcement supports pricing power in display glass. It does not directly prove a shortage in semiconductor packaging glass. The longer-term packaging thesis comes from a separate trend: larger AI packages are pushing organic substrates closer to their limits, while $INTC and other manufacturers are developing glass-core alternatives. I am keeping the near-term price catalyst and the longer-term technology cycle in separate columns.
253
The DRAM cycle is being rewritten by servers.
Nearly 6x demand growth by 2030, with servers driving most of the incremental bits, gives $MU and $SKHY a longer runway than the old PC-led cycle ever offered.
Server DRAM demand is expected to grow nearly 6x by 2030 and account for ~60% of the entire market.
That matters most for $MU and $SKHY since servers are expected to drive ~80% of all incremental DRAM demand as AI infrastructure becomes biggest source of growth.
1
137