@pequityresearch Reliability higher than pluggables -> double clicking on that -> want to know a) what is a link flap b) the risk is still itโs the weakest link given wear and it failing kills the gpu, so curious that trade off ? C) heat / power benefit?
Great nugget though
Data movement has become one of the largest contributors to latency and power consumption in AI infrastructure. As models scale, the cost of repeatedly moving large datasets between processors and memory is increasingly difficult to absorb.
Marvell and @SKhynix are addressing this through a collaboration that combines Marvell Structera A CXL near-memory acceleration with SK hynix advanced memory technology. Structera A moves processing closer to where data resides, with 16 Arm processor cores, up to 200 GB/s memory bandwidth, and CXL connectivity that enables low-latency communication with host processors and accelerators. The result is a memory subsystem that functions as an active computing element rather than a passive storage resource.
The joint CMM-Ax solution, co-developed by Marvell and SK hynix, targets long-context LLM inference specifically, where GPU memory capacity limits create a critical bottleneck that conventional architectures cannot efficiently resolve.
Associate Vice President Khurram Malik and SK hynix Vice President of System Architecture Kangkyu Park detail the technology and its implications for memory-centric AI infrastructure: https://t.co/sLBtivarfb
#FMS2026
A year ago I was uber bullish $GOOG
The thesis was a) leading AI capability, b) differentiated datasets (e.g. search, youtube), c) TPU IP -> you could imagine in 3 years it was like a $NVDA in terms of TCO per unit output and thus hidden value, d) other bets e.g waymo, $SPCX
Interesting that the downloads for Nvidia Cosmos 3 -> foundations for World Action Models suggest developer focus is wary of hardware costs
Mostly focused at the 16B param level where useful inference is more likely the cost base is more commercialisable
NVIDIAs Alpamayo 2 Super is an opensource vision-language-action (VLA) model designed to accelerate autonomous vehicle development
It can take 7 camera feeds, analyse 360 degrees synchronously with improved trajectories modeling, Chain-of-Causation reasoning traces
Onyl prob is its only been tested on US$30k H100s
https://t.co/pupiNK8pSY
@InvestingVisual I am a long term holder of $NBIS but used to run a long short with a PM thats also short this - his thesis isn't just valuation is stretched but that the company will need to raise billions to fund and achieve those revenue forecasts and he'll cover into those raises
I take short seller predictions with a grain of salt
they swing long if their thesis of share prices falling on the catalyst of failing to meet lofty expectations and momentum trade reversing
fine if you are steadfast long - not so much if you're a retail punter long using margin
@capitalsignalss@RJCcapital see my other post - this might be Burrys short
when I was running a long short fund, our prime broker was long the stock as they had custody e.g. a super fund posted GS their $NBIS and we would borrow it. You don't see the net position in these filings
@RJCcapital sometimes these are really low signal i.e. can be the prime brokerage side reflecting how many clients have lodged stock to the prime side can lend to hedge funds to short
VCโs who donโt know how to drive a forklift telling us at @GalvanickCo we should โpivot to AIโ to defend industrial facilities, please walk into a chemical plant and talk to people first. Or an automotive factory. Or a drug manufacturing facility. Or water treatment plant.
@bubbleboi does this imply you need to reboot and reload HBF frequently or risk corrupted data points? Prob not eevry 24hrs, but is the risk of data loss like a half life concept??