As one senior executive at Microsoft put it to the New Yorker: “I think there’s a small but real chance he’s eventually remembered as a Bernie Madoff- or Sam Bankman-Fried-level scammer.” #AI #lol futurism.com/artificial-i...

Any email or meeting that starts with "I'll keep this short" has already lost.

Maine is about to lead the way in banning new datacenter buildout (www.wsj.com/us-news/main...). Except it doesn't have any datacenter buildout going on (www.visualcapitalist.com/ranked-us-st...). What a loss.

Interesting that the #SC26 panel submission process 1. Only provides a template in LaTeX, and 2. The template implies that panelists should all have University affiliations I don't think this reflects how most of the world works these days.

Had an interview yesterday after a journo found my website via Gemini. Not the first time this happened. He said to keep doing whatever SEO I've been doing for AI-based search. Truth is, I do zero SEO; I just don't use LLMs to write. Wonder if Gemini etc are getting good at finding human content.

So anyway, who is going to #ISC26 this summer? #HPC

In case AI wasn’t slow enough or costly enough for you, throw some crypto in there because…well, I don’t know why. Chris is more generous with his time/words than I would be. This is 100% blockchain and 0% AI regardless of what the creators hope and dream.

I first noticed Claude and ChatGPT talk about "pressure testing" ideas ~six months ago, and only now am I hearing humans use the phrase in conversation. Is this a case of AI successfully training humans? Or is this just a coincidence?

Tender launched for procurement of #CASPIr, a new #EuroHPC #supercomputer in Ireland by the @eurohpc-ju.bsky.social & @uniofgalway.bsky.social 💻 15+ petaflops 🤖 #AI & data workloads 🌍 For research & industry 📅 Apply by 5 May 2026 👉 link.europa.eu/GdpCrv#CASPIr#EuroHPC#supercomputer#AI#HPC #Supercomputing #TechSovereignty #Europe

Sooo are #HPC centers now having to deal with students trying to run models for OpenClaw on their systems? Genuine question.

Some notes I wrote up from last week at #GTC26: blog.glennklockwood.com/2026/03/gtc-...

Speculative decode is an inferencing optimization that was mentioned a few times at #GTC26. I'd heard of it but didn't know how it worked, so I spent some time figuring it out. Some notes (and toy code that illustrates its benefits!) are here: glennklockwood.com/garden/specu... #AI

Nobody from VAST is on Bluesky, so I'll share with you all that my favorite official swag from #GTC26 this year was this awesome shirt by my friends at Quobyte. I'm going to start wearing this on Zoom calls.

These new NVIDIA STX storage node implementations from Quanta and Supermicro appear identical even down to the LED and USB-C placement. I wonder if the drive carriers are differentiated. #GTC26

Weird stuff you see at #GTC26: cops about to rough up a joke-cracking bipedal robot that runs around like a toddler (cops were actually just as curious as anyone else about how real this thing is, and the robot had its human minder always within 20 feet)

Look at all the unique differentiation that NVIDIA enables in their ecosystem. #GTC26

NVIDIA’s Kyber rack combines NVLink switch ASICs into what used to be the cable backplane. It’s physically huge—height of the whole rack. I wonder what the FRU is on this monstrosity. #GTC26

The lights meet the legal definition of "blinking."

Sooo is Rubin CPX cancelled then? Gone the way of the Rubin NVL144 branding?

Anthropic represents new demand to NVIDIA in the coming two years. Cool. But also a tacit admission that you can make a world-class AI lab (like Anthropic) entirely without NVIDIA, and only go to NVIDIA when you need capacity. Weird value statement about NVIDIA for frontier labs. #GTC25

Went down a rabbithole of looking at what MLPerf Storage Training benchmarks do; turns out they're not very meaningful IMO. Alarming how many practitioners (mostly storage vendors) run them and don't understand what the results do/don't represent. glennklockwood.com/garden/MLPer... #AI #storage

Interesting architecture - one DAOS node paired with five HDD JBODs and VDURA layered on top.

Training with NVFP4 will probably be hot at GTC next week. NVIDIA reports a 1.59x speedup over BF16 (developer.nvidia.com/blog/using-n...), but context: the speedup is on a tiny (8B), dense transformer with minimal high-latency comms in the loop. 1.59x will go way down for modern big, sparse models.

This is funny. Microsoft passed on Stargate, so Stargate went to Crusoe. Then Stargate passed on Crusoe, so now Microsoft is going to Crusoe. These AI infrastructure love triangles keep getting more complicated.

Since I no longer work directly w/ model trainers, I rely on public info to understand the infrastructure reqs of newer model architectures. This paper is a great explainer of how MOE taxes compute/memory/network: arxiv.org/abs/2603.076... My notes here: glennklockwood.com/garden/exper... #AI

Review a paper analyzing lots of HPC system telemetry. Found new ways to solve challenges in the data. Well written, reflects work and care. But forgets to explain why any of the results actually matter to #HPC users/operators/designers. Because they aren’t actually very actionable. What do?

It actually really isn't: arxiv.org/abs/2507.00217 :-)

Suggests that concentrating gigawatts of GPUs in a single low-latency geo (i.e., one allreduce domain) isn't as important today as it was thought to be when this project broke ground a few years ago. www.datacenterdynamics.com/en/news/orac... #AI

Kind words from VAST users. And without being prompted to say these things no less! I organized one of the sessions quoted, and I asked the speakers to be honest. After all, if the audience wanted to be marketed to, they had plenty of other chances during the week. www.hpcwire.com/2026/03/04/v...

I think some types of SSDs have become more expensive than GPUs. What does this say about the relative value of compute and data being driven by AI?

Surprised to see investors continue to pour money into aging startups that are struggling to find relevance in the AI accelerator market. www.eetimes.com/sambanova-ab...

I ran into my pal CT Sun who had a Bluefield-3 self-hosted flash node on display. Same one used for VAST’s Ceres DBox. But there were no SSDs in the display model; they were too valuable to waste on just looking pretty.

Context: consumer/client devices only. For example, one argument hinges on people’s home internet speeds being limited to 2.5Gbps. #storage #shortage

I’m in Salt Lake City this week for VAST’s inaugural user conference. Location isn’t as nice as other such conferences (like CUG), but the money saved on location definitely shows in the dress-up around the venue.

Food for thought: B200 GPUs appear to have lost 73% of their value in the last nine months.[1] At this rate, the Solstice system at Argonne (100K B200) will have no commercial value when it comes online. Maybe that’s what makes it affordable. [1]: julsimon.substack.com/p/two-market... #HPC #maybe

I’m at an NSF/CRA-I workshop today, brainstorming how industry and philanthropy can play a role in developing an AI-ready workforce. Not my day job, but a nice break from the grind; interesting to hear how huge philanthropic orgs interact with government around education.

List of accepted workshops at #ISC26 came out today. Lots of good topics. Now to figure out which one to submit to… isc-hpc.com/submissions/... #HPC

Flash is sold out. Hard drives are sold out. Great time to invest in pencil and paper I guess. wccftech.com/western-digi... #storage

Going through old photos this morning and found this one from an SC17 presentation. Wish I had more context! #HPC

I got to visit the early access system for NERSC’s upcoming Doudna supercomputer. A TALL order of GB200s. Surprisingly noisy. #HPC

@obsidian.md I noticed that multiple example sites (mostly English-language ones) on the Obsidian Publish page (obsidian.md/publish) link to sites that no longer use Obsidian Publish.

This is prefill

Zooming right in, you can see that I am part of a cluster called 'HPC and Supercomputing professionals'. Zooming out, it's fun to see what communities are clustered near us such as 'Ceramic artists and potters'

Dan Reed published an essay on what #HPC needs to do to remain competitive in the age of #AI yesterday (hpcdan.org/2026/02/06/h...). It was so engaging that I jotted notes as I read and figured I might as well publish them: blog.glennklockwood.com/2026/02/hpc-...

I asked Gemini to create a visual guide that describes multi-agent workflows. Grim.

$1200/lb is the lowest cost to put anything in orbit. A GB200/300 NVL72 rack weighs 3000 pounds. It will cost $3.6M to put 72 GPUs in orbit. It will cost $5B to put 100K GPUs in orbit. That does not include any supporting infrastructure. This is not hard math.

To be fair, this is how all the biggest AI infrastructure providers are bankrolling their growth: Google/Meta fund capex through their ads, Microsoft through Windows licenses. Now xAI has that sweet Starlink revenue. But I don't think Starlink has the margins that software/ads do.

Intel CEO says they'll make GPUs for datacenters. Are they just pretending that they didn't try this five years ago??? www.reuters.com/business/int...

Why would any AI practitioner rely on a storage company to tell them what a KV cache is? And why do storage companies keep insisting that they’ve got all the answers to inferencing at scale?? 😤

I am glad to see the leading minds of traditional scientific computing have moved on to the "acceptance" stage of grief and are advocating that others do the same. But existing in this new upside-down pyramid world will require unwinding a decades-long superiority complex. #HPC