DanielRomero

Core inflation has held well, although we may see second-order effects in the coming months as higher energy costs start feeding into the rest of the economy, pushing core inflation higher while overall inflation normalizes

Lead times for transformers now extend to 4 years
If you have access to power, whether it is on-site or, preferably, reliable, connected, and cheap power, you have a gold mine

$AMD continues to make progress on the software
The company introduced vLLM-ATOM, a plugin designed to make major AI models run better on $AMD Instinct GPUs, including MI350 and MI400
Developers can keep using the same vLLM commands, APIs, and workflows, while ATOM works in the background to improve performance on AMD hardware, requiring no new tools or complex configurations
It also gives users instant access to AMD’s latest optimizations, including FP4 support on MI355X, rack-scale inference on MI400, fused attention, custom AllReduce, and other kernel improvements
ATOM also acts as an inn

$IREN is now trading at less than 50% of $NBIS’ market cap
$IREN has roughly 2x the secured power and 4x the energized capacity, with higher asset ownership
$NBIS has the talent, software, and partner premium, plus equity stakes in private companies worth nearly $10B
Still, I’m not sure such a wide gap is justified

I must admit I wasn’t expecting $DGXX to sign such a good deal
$2.75M of revenue per MW per year is great
However, the colocation fee is all-in, which means it includes the cost of delivered electrical power up to the contracted IT load
That helps explain the gap versus the $1.90M per MW per year of $HUT’s River Bend deal or the $1.79M per MW per year of $CIFR’s Fluidstack deal

Quantinuum $QNT could become one of the most interesting IPOs in recent years
Rumored $20B+ IPO valuation
$IONQ is currently valued around $18B, with 2025 revenue of $130M and 2026 revenue guidance of $265M
Quantinuum’s revenue in 2025 was only $31M
However, Quantinuum is arguably the leader from a technology standpoint
Quantinuum demonstrated 4 reliable logical qubits, using 30 physical qubits, in partnership with Microsoft, which applied its error-correction software layer on top of Quantinuum’s hardware
$IONQ has not demonstrated any logical qubits yet
Microsoft reported an 800x improvement

DRAM Market Share Trends (by Revenue)
Source: Nomura Asset Management

As per Reuters, hyperscalers are offering to help finance new production lines and expensive equipment, including $ASML EUV tools, for SK Hynix
However, SK Hynix is cautious because these deals could lock capacity to specific customers and force lower long-term pricing
Available capacity is essentially zero right now, meaning there is not even a small portion that can be reserved for one specific buyer
SK Hynix is expanding with its massive P&T7 fab, focused on next-gen HBM, but this facility will not be ready until 2028
DRAM makers are reportedly only able to meet around 60% to 70% of current

$HUT had a 1GW site approved in the ERCOT grid, and nobody was paying attention 😴
That’s why putting in the hours pays off
Up 102% on the position now

$AMD launched the MI350P, its first PCIe Instinct GPU since 2022
It is a smaller and easier-to-install version of AMD’s flagship AI GPU, designed for companies that want AI inference without needing expensive liquid-cooled, rack-scale systems
It fits into standard air-cooled servers, which makes it much easier for enterprises to use in existing infrastructure
$AMD is targeting a market $NVDA has largely left open: modern PCIe data center GPUs. NVIDIA is focused on higher-end systems, while AMD is offering a more practical option for companies that want AI acceleration without rebuilding their

$AMD's MI430X will deliver up to 200 TFLOPs of native FP64 performance, making it potentially the highest-performance FP64 GPU ever built
That would put it far ahead of NVIDIA’s Rubin, which is listed at around 33 TFLOPs of native FP64 vector compute, or up to 200 TFLOPs only through Tensor-Core-based emulation
$AMD is focusing on native FP64 performance, which matters for classic high-performance computing workloads, including scientific simulations, energy research, biology, national security, advanced materials, and manufacturing
MI430X will be part of AMD’s broader MI400 series, while MI45

$MP Materials says the industry is finding ways to reduce heavy rare-earth usage
MP’s CEO, James Litinsky, said magnet producers are increasingly able to make high-performance magnets with "little or no heavies content."
Instead of relying heavily on scarce and expensive heavy rare earths, the industry is trying to redesign magnet production around lower heavy rare-earth content while still preserving performance
Litinsky’s view is that NdPr remains the more important bottleneck, while heavy rare-earth prices may not rise as much

The most concerning bearish case I can find for $AMD is its potential inability to secure what it needs from a tight supply chain, while $NVDA can circumvent this problem with deeper pockets

$RDW's market cap keeps widening the gap with its share price
Promising tech, but the company can’t make money and can’t stop diluting

$INTC server CPU demand is now so strong that chips previously considered low-value or near-scrap edge dies are being sold as lower-spec products instead of being discarded
Intel’s server CPU ASPs reportedly rose 27% YoY, even while unit volumes fell 5% due to supply constraints
Reuters also reported delivery lead times of up to six months for some Xeon processors