I got 37% on ARC-AGI-2 Kaggle without LLMs!
100M looped transformer trained from scratch
- Likely highest score without an LLM
- Peaked at 6th place
- Beats last year's winning LLM
- Looping, added in the last week: 33% → 35% → 37%
thanks to @arcprize@fchollet@mikeknoop and @GregKamradt for the competition!
Link: https://t.co/iDBUrHhmdI
(writeup for the paper track)
The H200 is a weird chip.
Not enough compute to be used by training labs
Not enough bandwidth to be worth all the extra memory (compared to an H100)
So its best in a pretty narrow use case of “memory-bound inference.”
But still, demand and prices have been steadily increasing…
#fyp #computeshortage
@atlasberry008 speed dials our co-founder & CEO @bavaria_kush to talk trading AI compute like oil, why older NVIDIA chips are getting more expensive, and what’s next for Ornn as we grow our team.
Watch the full conversation ↓
Good morning!!! Let's check on A100s ⤵️
A100 is running at 80–90% utilization on @OrnnExchange. Little idle capacity. This chip is six years old and still booked and busy!
Rental rates up ~26% YoY -> "aging like fine wine" https://t.co/0QXZFdL6dO
09:30 ET | 29 OCT 2026: A100 on-demand trading at $1.08/GPU-hr
Anthropic’s S-1 reveals more than $700 Billion in future compute commitments, CoreWeave’s stock fell 21% in just three days, and B200 rental prices are approaching $8.50 an hour. What is happening in AI infra?
TLDR:
→ CoreWeave has a levered up balance sheet. It has ~$10 of debt for every $1 of equity. Amazon, Microsoft and Alphabet owe 20 to 40 cents
→ $10.6B of payments comes due by the end of 2027, with only $5.5B in the bank
→ DC delays squeeze CoreWeave through both potential penalties to customers and delayed revenue, while interest keeps accruing on debt-funded GPUs
→ The loan runs longer than the customer contracts underlying it, meaning that lenders are now underwriting residual value
→ After the dust settled, CoreWeave’s borrowing costs stayed high even as the broader bond market recovered.
Through 2030, $7 Trillion is going into AI data centers, all on opaque guesses about future GPU rents. Ornn’s H100 and B200 futures on ICE will give lenders and operators a public price to lend and hedge against.
Anthropic will be a big compute trader.
Effective capacity planning will involve balancing and hedging their active compute book.
It is hard to see a world where secondary compute trades don't exist at scale.
Last night's sunset got me thinking about how life is full of opportunities.
Nature gives us so much, it is only right that we give it back things just as beautiful.
Like the NVIDIA GB300 NVL72.
If you want to experience it for yourself, come rent out capacity at @ornnexchange!
This week in compute:
1. H100 closed at $2.81/hr, down $0.09 (-3.1%) on the week, after swinging from a low near $2.27 (Oct 11) to a high near $3.67 (Oct 15). Utilization ended at 86.3%, ranging roughly 78% to 99%.
2. B200 closed at $8.53/hr, up $0.97 (+12.9%), climbing in two step-ups (Oct 13 and Oct 16) from a low near $7.27 to a high near $8.63. Utilization fell to 81.4% after a late-week dip near 61%, so price kept rising even as usage softened.
3. CoreWeave ($CRWV) fell 23.7% on the week ($80.44 to $61.39), including a 15.5% drop on Oct 14, and is now 57% below its 52-week high. Credit is flashing distress, with 2032 notes yielding ~15% and 5Y CDS at a record above 1,000bps, as Q2 interest expense ran 5x operating income.