DanielRomero

I’ve had people tell me I see too many bottlenecks
That everything can’t be in shortage
But that’s the point of AI
When intelligence is no longer the bottleneck, physical components/real assets become the bottleneck
The everything bottleneck
A consequence of the singularity

All the semi bottlenecks you need to know:
High-power, narrow-linewidth lasers
Glass core substrates
Silicon photonics
800V power delivery
MLCCs
Inductors
HBM and advanced memory
Advanced packaging
Power shelves and busbars
Hybrid bonding
Grid transformers
Gas turbines
ABF substrates
Advanced PCBs
Thermal management and liquid cooling

$AAPL names John Ternus as its next CEO, effective Sept. 1, 2026. Tim Cook will transition to Executive Chairman

Turns out Turboquant, an algorithm that barely has relevance for AI workloads, didn’t destroy $MU
Who would’ve guessed that?

I remember it like yesterday when people said I was crazy for thinking $AMD would outperform $NVDA
Some investors think they have an edge just by buying a multi-trillion-dollar leaderç
As I’ve always said: catching up is easier than innovating

In 2021 Dylan Patel could post his entire portfolio and it’d get 8 likes
It was a different time back then

$NVDA GB300 Shipments Are Slowing as Data Center Infrastructure Lags
According to SemiAnalysis, shipment estimates for both the GB300 and VR NVL72 have been revised, with GB300 shipments now showing a slowdown
The issue is not demand, but infrastructure readiness. Some cloud service providers still do not have their data center infrastructure fully operational, which is causing a slight delay in overall shipments compared to previous expectations

$GOOG and $MRVL Are Working on a Separate Low-Latency Inference Chip
According to SemiAnalysis, Google is developing a separate TPU family focused on low-latency inference, codenamed Merope. In other words, this appears to be an LPU-like architecture built for faster inference workloads rather than the company’s mainline TPU roadmap
Reports also indicate that Marvell is involved in developing this design. The expected timeline is 2028, and it is separate from the main TPU programs being developed with Broadcom and MediaTek

$MU has regained market share in HBM4
According to SemiAnalysis, Micron has re-entered the market as an HBM4 supplier for Nvidia Rubin. Our previous view was based on the assumption that Nvidia had not ordered HBM4 from Micron, as Nvidia initially relied on SK Hynix as its primary HBM4 supplier, followed by Samsung
However, due to issues with Hynix’s HBM4, Nvidia later increased its HBM4 orders from Micron to broaden supply and help ease the HBM4 constraints affecting Rubin production growth
Micron then increased its HBM4 wafer production and began fulfilling Nvidia’s orders. Still, as we have

Anthropic plans to integrate $AMD GPUs into its hardware stack starting with the MI450, according to SemiAnalysis
SemiAnalysis believes the reason is that Anthropic needs more computing power, and AMD’s involvement expands its available capacity. This means Anthropic will use all four major accelerator platforms: NVIDIA GPUs, Google TPUs, Amazon Trainium, and AMD GPUs starting next year
With OpenAI, Meta, and now Anthropic as key end users of MI450 compute, the demand outlook for $AMD in 2027 is very strong

Samsung Electronics is testing S&S Tech’s (101490) domestic EUV blank masks in its 4nm foundry mass production environment. This marks the first time these masks have been introduced on an actual production line
The goal is to reduce dependence on Japan’s Hoya, strengthen supply chain resilience, and gain more pricing leverage
Samsung has reportedly been giving S&S Tech feedback to improve quality, while S&S Tech has already invested heavily in inspection equipment and a new Yongin factory to support the effort. If localization succeeds, Samsung could save tens of billions of won annually and

Global PC shipments rose modestly in Q1 2026, with Gartner, IDC, and Omdia all reporting growth of roughly 2.5% to 4%
This does not reflect a real recovery in end demand
Instead, distributors and suppliers appear to have pulled forward inventory purchases ahead of expected DRAM and NAND price increases in Q2, creating what Gartner described as another round of artificial growth
The vendor ranking remained unchanged, with Lenovo, HP, Dell, and Apple leading the market, while Apple posted the strongest growth thanks to demand for its lower-cost MacBook Neo
Overall, the data suggests that Q1 stre

$NVDA growing at 30%+ isn’t sustainable, according to the market
When you look at valuations, most multiples are not pricing in high growth beyond 2027:
$NVDA: 24x earnings
$AVGO: 30x earnings
$AMD: 41x earnings
$MU: 5x earnings
Big rerating opportunity if the market is wrong

$TSLA has started contacting major semiconductor equipment suppliers for pricing and delivery estimates
$AMAT, Tokyo Electron(8035.T), $LRCX, and Samsung Electronics(005930.KS) have reportedly been approached, with discussions covering tools such as photomasks, etchers, deposition equipment, cleaning systems, and testing devices
Suppliers were apparently asked to respond quickly even though technical specifications remain limited, which suggests the project is still being defined
$INTC is also reportedly involved in the broader Terafab initiative, which is aimed at supporting advanced chip pro

$AMD at ATHs
I’ve been pounding the table on this one for a long time
Now it feels easy to hold, but when it was at $70, many people told me it was going straight to $30
Overcoming sentiment is the hardest part of being an investor

TrendForce says 2026 server demand is still strong, but actual shipment growth will likely be constrained by component shortages, particularly in the general-purpose server market
$SMCI $DELL $TSSI $HPE
They now expect total server shipments to grow about 13% YoY in 2026, down from a previous expectation closer to 20%, not because demand is weak, but because suppliers are prioritizing higher-margin AI server products
The main bottlenecks are PCBs, CPUs, PMICs, and BMC chips. In some cases, lead times are approaching one year, which is limiting how much general server demand can actually be ful

$NVDA Rubin Ultra is rumored to have shifted from a 4-die design to a 2-die design, which has raised concerns that testing complexity, carrier board requirements, and some supply chain demand could end up lower than previously expected
The market initially feared this would also hurt the cooling opportunity, but analysts argue that view is too simplistic. They say the GPU would still rely on full-plate liquid cooling, so the overall cooling architecture does not change in a meaningful way just because the die count is lower
There is also uncertainty around the technical roadmap, with analysts

$MU will keep outperforming $NVDA for a few years
It’s absurd that the market was thinking about oversupply in 2027. It’s going to get much worse before it gets better