DanielRomero

$AMD says unified memory architectures (UMA) are becoming more important for AI PCs, workstations, and future high-performance platforms
The idea is that CPU, GPU, and memory share one large memory pool instead of separating system RAM from GPU VRAM. This helps AI workloads because large models need a lot of accessible memory
David McAfee, VP & GM of Ryzen CPU and Radeon Graphics, said:
"With Strix Halo, with $NVDA entering the space as well, there's going to be a lot of focus on UMA systems, and on identifying the right architecture for what these UMA systems can do."
"This is a totally new w

Most of the quality AI-buildout names, no matter the valuation, are at least a 10x in a decade
Think of any company in the supply chain and 10x the market cap. Does it sound crazy over a 10-year timeframe, considering the power and progress of AI?
The opportunity is insane

I'm hoping for more red on Monday
I can't believe the market still decides to doubt semis earnings after all this time

According to Mercury Research’s Q1 2026 CPU market-share data, $AMD reached a record 32.6% share of the overall x86 CPU market
Intel still leads with 67.4% share, but its position weakened from 68.6% last quarter and 72.9% a year ago. $AMD’s gain was driven partly by Intel’s supply constraints and partly by AMD’s continued share gains across client, server, and semi-custom markets
Excluding embedded, semi-custom, and IoT chips, AMD’s x86 share reached 30%, up from 29.3% sequentially and 24.4% year over year

According to a Wccftech report, CXMT’s DDR5 is not meaningfully cheaper than Samsung, SK Hynix, or $MU DDR5
Multiple memory vendors told Wccftech that CXMT’s real advantage is supply availability, not pricing. Since CXMT is not heavily exposed to premium AI memory products like HBM or SOCAMM, it has more DRAM available for client and mainstream markets
However, CXMT still appears to lag the big three in higher-end DRAM technologies. Its DDR5 is currently more relevant for entry-level and mainstream modules, while it remains behind in premium formats like CUDIMM, CQDIMM, MRDIMM, and CSODIMM

SK Group chairman sees memory bottleneck persisting through 2030
- > Capacity is expected to double by 2030, but demand will likely outpace that massively, probably by over 5x
In the same week, the market decides to dump memory stocks because of a jobs report
Does that sound like a structural bear thesis, or a buy-the-dip opportunity?

$TSM CEO recently said:
“I envy memory makers’ 80% gross margins, but I would never do that.”
The reason for this explains why memory has historically been the cyclical component
Unlike memory, if $TSM hikes prices too aggressively, it encourages clients to go to competitors, even if those competitors have higher costs because of worse yields or other inefficiencies
In memory, it is mostly a commodity. That means you care less about the competition and more about making as much as you can while the opportunity is there
And that opportunity only exists in tight market environments

$META is now flat over nearly two years
Zuck is also thinking about an equity raise
On the one hand, $META is one of my least favorite top 10 megacaps qualitatively
On the other hand, if it went down another 15–20% because of the raise, it would be trading at:
-> 15x NTM EPS
-> with a strong balance sheet
-> 25% projected growth
-> and still 38% margins

Samsung and SK Hynix join strategic investment round in Anthropic

Samsung used Computex to preview HPB, or Heat Path Block, a thermal feature for its next-generation HBM5 memory
This is Samsung’s answer to a cooling approach that SK Hynix already revealed
Both companies are trying to solve the same problem: heat
HBM stacks many DRAM dies vertically on top of a base die, and each new generation keeps increasing capacity and bandwidth by adding more layers and pushing higher data rates. That also raises power density
Heat generated in the middle of a tall stack struggles to escape because it has to move upward through layers of silicon and through-silicon vias

$IREN secures power when no one else can
800MW secured in Australia, IREN’s first announced APAC campus
➯ Located 78 miles northeast of Adelaide
➯ Connection uses four 330kV feeder exits into the utility substation
➯ $IREN says this can support up to 800MW without network upgrades
➯ Energization starts from 2028
➯ The site has submarine fiber connectivity into Singapore, Indonesia, South Korea, and Japan
➯ South Australia targets 100% net renewable energy by 2027, which helps AI infrastructure
Macquarie kept an Outperform rating and $90 PT, saying the site gives IREN direct exposure to APAC

Panel-level packaging will become a bottleneck too
I believe it is actually one of the most undiscovered future bottlenecks for AI
People will start paying attention soon

$NVDA claimed Vera superiority in recent benchmarks
But there are caveats
They compared Vera, a 2026 CPU, against $AMD’s EPYC 9575F, a 2024 CPU
Vera was 10% faster than AMD’s EPYC 9575F on the geometric mean of the tested workloads
However, the real competitor to Vera will be AMD’s Venice / Zen 6
Also, the benchmarks were limited. NVIDIA controlled the initial access, and the workload set was focused on areas closer to Vera’s intended use cases
These were NVIDIA-permitted benchmarks, power consumption could not be measured, and Phoronix explicitly said this was only a subset of its normal full

Barclays analyst Tom O’Malley raised his $AMD price target to $665 from $500, while keeping an Overweight rating
Barclays believes the market is still underestimating how agentic AI changes demand for server CPUs
“CPU-to-GPU ratios are narrowing as CPU demand reaches new levels in the rapidly expanding world of agentic AI,” O’Malley said. “AMD is best positioned to benefit from this transition.”
Agentic AI requires systems that can coordinate tasks, route requests, call external tools, manage memory, and track multi-step workflows over time. That means the workload depends on the full compute

TrendForce released an extremely bullish take on memory
TrendForce says HBM prices have not fully caught up yet because the big suppliers usually negotiate HBM contracts annually. Meanwhile, conventional DRAM prices have been rising sharply since late 2025, making products like DDR5 64GB RDIMMs more profitable than HBM on a per-wafer basis in 1Q26
That creates a major incentive problem. If conventional DRAM becomes more profitable than HBM, suppliers may prefer to allocate more wafer capacity to conventional DRAM unless HBM prices rise significantly. TrendForce expects this to support much hig

This is insane
And I thought I did well buying $MU at $70

Jensen Huang just called $MRVL the next trillion-dollar company