JM🌱

@HedgieMarkets You can look at the on-demand price at neoclouds for H100 vs B200. The market suggests Blackwell is 2-2.5x more efficient

🦔 H100 GPUs that cost $40,000 new are now selling for around $6,000 on eBay, an 85% drop. The math on why is straightforward: it costs about 11x as much to run an H100 for inference as a B300. Anyone running H100s needs to charge dramatically more than competitors on newer hardware. Upgrading isn't simple either. At a $50,000 price tag for a B200, it takes about 33 months to break even on the upgrade from an H100. And the B300s are already making B200s less attractive. My Take I've been covering the depreciation problem in AI infrastructure for a while now. Companies are booking these GPUs on five to six year depreciation schedules when Nvidia releases new generations every two years. Michael Burry flagged Big Tech lengthening depreciation timelines as suspicious because it hides the real losses. A hedge fund manager I wrote about found that industry insiders estimate actual component lifespans at 3-10 years, but the economics don't work at any of those numbers. The hyperscalers are sitting on hundreds of thousands of GPUs that lose value every time Nvidia announces something new. David McWilliams called them "digital lettuce" because they go stale while you're still installing them. The difference between what the books say these assets are worth and what you could actually sell them for is enormous. At some point that gap has to be reconciled. H100s selling for 85% off on eBay is a preview of the writedowns coming to earnings reports. Hedgie🤗

@SlowestTimelord @gospaceport Data is more valuable when you can analyze it

@gospaceport Can confirm, selling 3-4 year old white label HDDs for more than I paid for them… But why? Why does AI need so much storage?

Happy Valentines day!

@gospaceport @efnats lol I’m building something like this for CDI

Banger new HDD mgmt software from my friend @efnats you likely want to check out! github.com/efnats/diskmind

This seems like a lock, call our boy Elon to confirm 👽

Command the agents or get replaced by them

@lauriewired @tomshardware spdk on Linux was able to hit 10 million IOPS per CPU thread in 2019...yeah

@tomshardware pretty scary there was that much room on the table to improve though...

Windows 11 rockets SSD performance to new heights with hacked native NVMe driver — up to 85% higher random workload performance in some tests tomshardware.com/pc-components/…

@Beelinkofficial Would love to see something like this with E1.S form factor on the SSD. Insta buy. Consumer M.2 NVMe are really really bad for RAID

It's a PC. It's a NAS. ME Pro is coming soon! #Beelink #MEPro #NAS #MiniPC #homelab #Tech

@lauriewired Sigh…ssdcentral.net/waf/ Not quite @lauriewired but let me know if you want to learn about how SSDs 😎

Pro-Tip, you can turn your crummy “consumer” SSD (mostly) into an enterprise drive by formatting it correctly. You can easily get 10x endurance by giving up ~5% of space. Crucial MX500 1TB: 0.2 DWPD Micron 5300 PRO 960GB (same NAND!): 1.5 DWPD

@M_Kirk Would be very happy if I never install another hdd 😎

@LebanonJon I thought you were living the quiet life of GPU server maintenance now? HDD days aren't behind you?

HDDs sold out for next 2 years, XCH 🌱@chia_project could do the funniest thing rn tomshardware.com/pc-components/…

@SemiAnalysis_ God you guys are on absolute fire

>be me, qualcomm >time to enter AI chip market >nvidia making $$$, how hard can it be >spend years developing AI200 chip >finally ready for BIGG announcement >make fancy slide deck >put 768GB memory on there (sounds big) >160kW power consumption (sounds powerful) >add "liquid cooled" (sounds cool) >ohshit.jpg what about FLOPS >hmm >decide to just not mention it >also don't mention price >or how many chips per rack >or actual benchmark numbers >just vibes >launch presentation >"Qualcomm AI200: It exists and uses electricity" >refuse to elaborate >stock goes up 15% 🚀 >tfw investors don't know what FLOPS are either >mfw "greater than 10x" with no baseline >ships in 2026 >AI250 ships in 2027 >still won't tell you the specs by then probably >low TCO trust me bro >confidential computing (the performance is confidential)

@gospaceport lol I have mountains of SAS cables, care package ready

I have hit a SAS cable buying moment of dissonance. I need to buy hundreds of $ more in additional cables and usually that makes me happy, but not this time. Unintended side effect of cable wall is me actually reflecting and... I may be a tad bit insane.

@Mascobot Every week I’m like “you cannot believe this insane meeting I just had”

Opportunity cost is so high right now, never seen anything like this…

🚨 New: We built @a16z's personal GPU AI Workstation Founders Edition - 4x NVIDIA RTX 6000 PRO Blackwell Max-Q (384GB total VRAM) - 8TB of NVMe PCIe 5.0 storage - AMD Threadripper PRO 7975WX (32 cores, 64 threads) - 256GB ECC DDR5 RAM - 1650Watts at peak (runs on a standard 15Amp/120V circuit). For training, AI research, and deploying models locally. A datacenter-class AI rig you can keep under your desk. We are planning to make a limited number of these a16z AI Workstations. Build guide + how you can make your own 👇

@gospaceport NVIDIA asking all vendors for 100M IOPS per SSD 😅

@theo Learn how to do basic debug? Mr gg

Been having PC issues for awhile now. Loud whine sound from my rig when booting some games. Doesn’t happen when running benchmarks. Thought it was the GPU, so swapped from a 4090 to a 5090. Issue persists. Thought it was PSU related (14900k + 5090 is a lot of power). Finally upgraded to a 1500w Corsair. Issue persists. I hate computers so much. I’m never ever building one again. I just want to play some games to relax. I’m so fucking done that I am fighting the urge to throw this $6k+ rig off of my balcony. Is there anywhere in SF that I can drop this off, hand over $1k and get back a working PC?

@jasonfried This is fucking torture to millennials

Just call. Don't schedule, don't send me a link to pick a time, don't show me your open slots, just pick up the phone and call. I'll answer or I'll call you back. I bet we'll naturally find a time to talk that day. It'll all work out.

@dhh @bitsweat Just curious, why Pure? They are like the “AWS” of AFA. Much more flexibility to use standard OEM HW even if you choose licensed software like VAST, WEKA, or MinIO

To move out of S3 and onto our own Pure Storage, @bitsweat built a little app called Nostos to manage the transition. We've been using S3 in some form for over 15 years, so there were a lot of old buckets in the attic to inspect! But well worth the effort to save nearly $1m/year.

Something brewing 🤔

New GPUs are now live on marketplace.silicon.net/drops/silicon-… 156 GPUs available for purchase via ACH, debit/credit card, and USDC on Base. 30-day availability, first-come first-served. What's in the drop? 🧵