Arin

Data @OrnnExchange

[Arin @gpujaff]

Interesting trend I've been watching with compute pricing is the "predecessor/successor" rule.

In a perfect market, on-demand prices/growth rates should have the newest hardware increasing the most, and the oldest remaining stagnant or growing slower. But the B200/B300/H200 have violated this pattern.

IMO, this is due to availability and market fit. B300s are not that saturated on the on-demand marketplace, so people haven't found ways to leverage their value yet.

[Ornn @OrnnExchange]

Most GPU price indexes track listed rates. We track what buyers paid.

Over the past year, transacted prices went up across every generation, led by B200 and H200. While new chips usually push down prices for the one before them, B200 kept climbing even after B300 arrived in June.

Explore our data → https://t.co/oxQ9uR98Bq

[Arin @gpujaff]

I have 128 nodes of Vera Rubins ready to deliver today.
Specs:
Site: Back of a Toyota Prius
Power: Premium Hybrid Electricity
Location: US East/West/Central, will drive where you prefer.
1yr: $20/hr, 2yr, $40/hr. DMs are open.

[Arin @gpujaff]

@andylawk type of description you see on a merged PR containing 4 new blockers

[Arin @gpujaff]

Fable 5.2 is reaching levels of slop we didn't know were possible

[Arin @gpujaff]

@genharris_ great caption wonder who made it

[Arin @gpujaff]

Wow. $SOXL has fully diverged from compute pricing.

Used to move together somewhat but there's been some serious FUD in the past month.

[Arin @gpujaff]

@andylawk I wonder if napoleon inspired his troops in a similar fashion

[Arin @gpujaff]

@ElsorThomas @kushal_mungee luckily they do carry terms, go take a look!

[Arin @gpujaff]

Thanks! I think this is where all eyes need to be: the underlying relationships between companies involved in the compute space.

This is also just step 1 - understanding the basis of these relationships helps us understand how they will progress if dominos begin to fall.

[Arin @gpujaff]

@Leothebayesian @arakharazian Ornn. @OrnnExchange

[Arin @gpujaff]

Ok, $CRWV is now trading under $45.

Reminder that they signed a $1 BILLION equity agreement with Jane Street at $109.00 per share earlier this year.

They also have large compute contracts with Meta, OpenAI, Anthropic, etc.

Who's going down if they keep sliding? Go check it out: https://t.co/g8j9qUr9sp

[Arin @gpujaff]

@OrnnExchange and @SpaceX leading the shortlist of names to follow!

[Ara Kharazian @arakharazian]

NEW: SPACEXAI is the fastest growing software company according to Ramp spend data.

I find the following:
- SpaceXAI's advantage in coding + its new consumer offering could accelerate business adoption of its models, much like what powered Anthropic's rise.
- Continued growth in cybersecurity SaaS.
- Multiple winners in AI dictation (Wispr Flow and Granola both take a spot on our list).

[Arin @gpujaff]

@thenarrator Great choice on compute data source @OrnnExchange

[Arin @gpujaff]

$CRWV and $NBIS both report earnings next week. They've both been cut roughly in HALF in the past month, but they’re fundamentally different businesses financially:

• ~98% of CoreWeave’s revenue is take-or-pay on ~5-year contracts. Only about half of Nebius’s is long-term; the rest reprices within 1–3 years, plus their spot market.

• On ~70% of Nebius’s Q2 deals, customers prepaid 50–60% of capex. CoreWeave’s prepayments run 15–25% of contract value.

• On five-year GPU loans, CoreWeave’s latest priced at SOFR+5.50%. Nebius’s: SOFR+2.50%.

CoreWeave carries more financing risk, but has over a GIGAWATT more running than Nebius (1.5 GW vs. ~0.25 GW). Nebius is behind on power, but its balance sheet is more...balanced.

Take your pick. Battle of the neoclouds is happening next week!

[Arin @gpujaff]

Wild. Weights are becoming something you program, meaning custom self-hosted models are about to get a LOT cheaper.

Compute is king.

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Arin @gpujaff]

Given the current state of AI financing, I wanted to make a simple live tracker for the state of the AI money machine.

Check it out at https://t.co/PH3ut5VrEp. If I missed anything there's a submission form per-company, feel free to contribute, will be updating this live!

[Arin @gpujaff]

Ah the assumption here is that only the racks depreciate over 5 years and the site over ~20, so you can't really multiply the whole $7.9B by 5. Total build is about $52M/MW - ~$35M for the racks and ~$16M for the site, which should be right where you'd expect VR to land if VR runs 50-100% higher than GB.

[Arin @gpujaff]

The Vera Rubin buildout by @nebiusai will bring in a whole new era of neocloud profit margins

[Annie Hage @annie_hage]

Nebius announced 1GW online in 2027 in VERA RUBINS alone -> @gpujaff gives us an earnings breakdown below ⤵️

***Not to mention - this is assuming a 1yr contract , not the additional profit opportunities of renting out on-demand/M2M

[Arin @gpujaff]

@jiahanjimliu @wayne_nelmz @nebiusai This factors in one year of capex ($7.9B depreciation on racks and site) and after $3.3B of interest on the full build cost, but before any repayment of the debt that funds the build! Obviously a range/projection.