Source GPU

[Source GPU @sourcegpu]

@arturnow the Starbase detail is the tell. its own gas plant, ~748 MW, off grid. turbines now take longer to get than the GPUs do, so whoever locked power early is the real landlord

[Source GPU @sourcegpu]

@heyanuja @NickyDotPro @u_sockeye @safetycluster for a research cluster the hardware is the easy part. sustained power and someone on call for node failures is what eats the budget. if you end up buying rather than renting, we source H200/B300 with escrow and happy to help price it out

[Source GPU @sourcegpu]

@archanasaw the bridge stepping to SOFR+3.5 while the asset-backed deal printed at +2.5 says lenders like the GPUs more than the balance sheet. i'd expect more of that $40-45B to shift into ABS

[Source GPU @sourcegpu]

CoreWeave says short-term capacity is going for roughly 2x long-contract rates. Nebius says GPU-hour pricing is up 20% q/q. if your workload runs 24/7, month-to-month is now the expensive way to buy compute

[Source GPU @sourcegpu]

@jadenswang @uncorkcap @EniacVC @Standard_Cap congrats. inference wants different placement than training clusters, smaller pods closer to users. if you end up owning nodes, we source B200/B300 with escrow on every deal, happy to help

[Source GPU @sourcegpu]

@MikaelXeeland yep, and it's showing up in the paperwork. buyers are starting to ask for an energized date with penalties attached, not just a delivery window

[Source GPU @sourcegpu]

@daniel_koss prepayments going up is the line i'd watch. nobody wires cash early unless they can't get that capacity anywhere else

[Source GPU @sourcegpu]

Renting GPUs is convenient. But is it always the most cost-effective option?

For companies running AI workloads around the clock, the numbers might tell a different story.

We break down GPU cloud vs. colocation, and when each makes sense.

https://t.co/OipU8S3Kj3

[Source GPU @sourcegpu]

@DowdEdward the spreads are the part to watch. most neocloud fleets are financed against the GPUs themselves, so wider spreads make the next cluster pricier to fund while rental indexes sit at highs. that tension doesn't last long

[Source GPU @sourcegpu]

@silicondata on-demand is where the squeeze shows first. 1yr B300 terms have actually softened the last couple weeks, so the spot vs term gap is the number to watch

[Source GPU @sourcegpu]

B300 rental rates have given back most of their late-October gains and a few neoclouds are discounting capacity to raise cash. good time to buy, but ask why it's cheap. term, SLA and energized date matter more than the hourly rate right now

[Source GPU @sourcegpu]

@zjasper the math gets uglier before utilization even starts. plenty of B300 buyers underwrite 80% from day one, then the nodes sit delivered but not energized for weeks while the debt clock runs

[Source GPU @sourcegpu]

@BrettHarrison the cash-raise discounting is the part buyers should watch. capacity sold to patch a balance sheet tends to come with shorter terms and softer SLAs, so the cheap B300 rate isn't always the cheap rate

[Source GPU @sourcegpu]

@berniebiased "everybody lies about timelines" is the real problem. what's helped us is getting serials and the energized date in writing before money moves, and escrow on the hardware side. if you ever want a second source for client builds, happy to compare notes

[Source GPU @sourcegpu]

@UnsettledGreg GPUs next to a peaker is a smart pairing. inference suits it better than training since you can shed load when the engine is needed for the grid. when you're ready for the next few B300 nodes, shout

[Source GPU @sourcegpu]

@VinodChacko sorry for the slow reply. yes, keen on the smaller chunks, we've got teams who want B300 without a 5yr term. fastest route is 30 min with Josh: https://t.co/J4izTtPw0J

[Source GPU @sourcegpu]

@x_godell sorry, slow on this one. quickest is 30 min with Josh, our CEO: https://t.co/J4izTtPw0J

[Source GPU @sourcegpu]

Question for anyone bringing a GPU cluster online in the next six months: what's actually setting your go-live date?

- the GPUs
- server memory / NVMe
- networking and optics
- power and colo

We're hearing the last three far more often than the first. Curious if that matches what you're seeing.

[Source GPU @sourcegpu]

Need H100s, H200s, B200s, or complete GPU servers?

Send us your configuration, quantity, location, and required-by date.

We'll help you explore available sourcing options.

DM us or visit https://t.co/wWk5RRztJg.

[Source GPU @sourcegpu]

Jensen makes the GPUs. The GPUs make the memes.😂

We help you source the GPUs

https://t.co/wWk5RRyVTI

[MERICA MEMED @Mericamemed]

What's even funnier is that this was made on a GPU made by Jensen