DanielRomero

$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

$ARM-based CPUs are expected to take over the host CPU role in custom AI ASIC servers, rising from about 25% share in 2025 to at least 90% by 2029, according to Counterpoint Research
This shift is being driven by hyperscalers like Google, AWS, Microsoft, and Meta, which are increasingly designing their own chips
The main reasons are better power efficiency, lower token costs, improved supply chain control, and deeper vertical integration
At the company level, Google is expanding Axion in TPU systems, AWS is scaling Graviton with Trainium, Microsoft is using Cobalt with Maia, and Meta is adopti

South Korea is launching two Korea-U.S. shipbuilding cooperation centers, led by KRISO
The goal is to improve U.S. shipyard productivity and workforce training
Companies that could benefit:
HD Korea Shipbuilding & Offshore Engineering |
$009540: Expected to participate in training U.S. workers
Hanwha Ocean | $042660
Directly involved in the training and already has a U.S. footprint through Hanwha Philly Shipyard
Samsung Heavy Industries | $010140
One of the Korean shipbuilders expected to provide specialists
Hanwha Systems | $272210
Co-acquired Philly Shipyard with Hanwha Ocean, so any increa

$AMD announced a multi-year collaboration with the French government to support France’s national AI strategy
The goal is to expand local AI innovation, improve access to advanced compute, and strengthen France’s position in the global AI ecosystem
AMD plans to provide hardware, software, and training to French researchers, educators, developers, and AI startups through the AMD University Program, AMD AI Developer Program, and AMD AI Academy
AMD will deepen its work with GENCI, the Jules Verne Consortium, and CEA on Alice Recoque, expected to become France’s first exascale supercomputer powere

SemiAnalysis rumor:
Anthropic to strike a major deal with $AMD to use MI450 accelerators

$SSNLF may be the most undervalued semi stock
However, Samsung has two major problems:
A growing talent drain and an escalating union dispute
On the talent side, SK Hynix and $MU are becoming more aggressive in recruiting semiconductor workers
SK Hynix has strengthened its appeal with a compensation structure tied to profits, and $MU is actively targeting Korean memory talent through job postings, university recruiting events, and interviews. Samsung has also seen several senior people in design roles leave recently, which raises concerns about leadership gaps in areas like System LSI and chip

Samsung drama
Instead of accepting the union's request, Samsung decided to seek a court injunction against them
Samsung filed for a court injunction to block what it calls an illegal union strike ahead of a threatened walkout by the National Samsung Electronics Labor Union
Samsung says it offered performance bonuses averaging 5.4 million won, described as 600% of the standard annual salary benchmark for the memory division, while the union demanded 15% of operating profit and threatened to strike next month
Samsung argues that a strike could shut down or disrupt critical safety systems and pro

I don't think I had ever seen something like this
All because a sneaker seller decided to buy and rent GPUs

Photonics and memory still have a long way ahead
What’s coming next?
That’s physical AI
AI is already taking thousands of jobs, but it is still limited to digital workloads
Autonomous cars are still niche, taxi drivers still exist, trucks are still driven by humans, and even most trains still have an operator
It goes beyond transportation
Automating production lines was the easy part. Pre-programmed robotic arms did the job. However, once you want automation on a general scale, you need to adapt the machine to the environment, not the environment to the machine
That’s why humanoids are the mai

Leopold has an interesting portfolio
I’m impressed to see so much $IREN
He went big on $CORZ as well
$CRWV and no $NBIS surprises me
I wouldn’t be surprised to see some new photonics names in the next filing

SK Hynix is moving early to adopt integrated hybrid bonding equipment
$BESI
The main technical constraint is chemical mechanical planarization. In hybrid bonding, CMP is essential because the surfaces need to be extremely flat and clean. Problems such as copper dishing, erosion, particles, or contamination can directly hurt yield, resistance, and long-term reliability
The AMAT and BESI partnership covers CMP, surface treatment, and hybrid bonding, offering a solution to that problem
The short-term risks are scaling complexity and the fact that JEDEC height rules still remain loose enough for t

I’m up 6x on my $NBIS position
It’s a great company, and holding it has been easy, even through dips
However, I’ll recommend investors keep a cold mind
The moment to love $NBIS the most was when it was an unknown play, not a mainstream neocloud giant
Today, the thesis is very different

BofA is very bullish on $AMD
Look at that growth in CoWoS demand

$NVDA Rubin is facing delay risks
According to TrendForce, NVIDIA is dealing with issues related to HBM4 validation, the transition from ConnectX-8 to ConnectX-9 networking, higher power consumption, and the need for more advanced liquid cooling
Blackwell is expected to dominate NVIDIA’s high-end GPU shipments in 2026, with its share rising from 61% to 71%, while Rubin falls from 29% to 22% and Hopper drops from 10% to 7%
Even so, TrendForce still expects strong overall growth in high-end AI GPU shipments in 2026, at around 26% YoY, only slightly below its previous 26.8% estimate. It also high

If Claude Mythos was actually trained on $AMZN's Trainium
And inference can be run efficiently on $GOOG's TPUs
Does this mean $NVDA's moat is disappearing before our eyes?
If Broadcom, Marvell, Alchip, and MediaTek can match $NVDA's offerings, companies won’t need to pay the $NVDA premium
Plus, $AMD is also offering very competitive accelerators at much lower prices