Kush Bavaria
@bavaria_kush
helping finance compute @ornnexchange | CS @MIT
397 Following 2.4K Followers
Older NVIDIA GPUs can remain productive and revenue-generating long after newer generations arrive.
@OrnnExchange's analysis of NVIDIA A100 rental pricing offers third-party evidence of a multi-year earning life for prior-generation NVIDIA infrastructure.
CUDA’s support for diverse workloads, combined with continuous software optimization across generations, helps keep that hardware productive longer.
Learn more:
Show more
Live rental prices, by accelerator:
B200 $7.09
H200 $5.40
H100 $2.72
RTX PRO 6000 $1.50
A100 $1.05
RTX 5090 $0.68
Each GPU is a separate market with its own supply and demand.
An hour of B200 costs 10.4 times an hour of RTX 5090.
See more:
Show more
Today, we are introducing the Ornn B300 Payback Forecast.
A GPU pays for itself through rental income. Until now, nobody could tell you how long that takes for a chip that only started shipping this year.
Built on OCPI forward curves, this forecast projects how long a B300 takes to earn back its purchase price at market rental rates, giving operators, lenders, and investors a repayment timeline they can underwrite against.
It's live now on Ornn Data, see it here:
Show more
consensus take now
BREAKING: Trump says AI data centers are the oil of the next 50 years
Good group of people
Today, we release the first series of our findings.
1 - We cut pretraining costs by 62% at frontier scale (8x Chinchilla), and we further cut all AI inference costs by 30% (Faster decode).
2 - Introducing our model ‘Feather’, which decisively overtakes Qwen 3 as the leading frontier 1.7B math model, using 180x fewer total training tokens, even matching Qwen 4B at potent benchmarks in math.
3 - With Anvil II, our state of the art LLM Optimiser, we record the largest pre-training efficiency jump on the NanoGPT Speedrun, leaping the previous by 34seconds - Our record singularly, is a greater percentage drop than the past 45 World Records combined.
The website :
Show more
Today, the @OrnnExchange team published a research paper on The Economics of Open-Weight Inference.
We measure cost per completed task across eleven open-weight and eight closed models, the cost of a million output tokens on rented @NVIDIA A100 through B200, and rental term prices from one month to five years.
On the same model, the A100 can produce a million output tokens for less than half the cost of an H100. A new GPU generation does not make its predecessor economically obsolete.
GPU financing is changing.
Show more
Even the oldest GPUs might have a longer useful life than 6 years.
Our team published a paper, The Economics of Open-Weight Inference which asks whether a new NVIDIA generation makes the previous one economically obsolete. 🧵
Show more
Today I learned you just have to ask
Dell just did $16.4B of AI servers in a quarter, double last year and ISG operating income up 225%. A lot of new capacity heading for the rental market and every one of those boxes needs an hourly rate once it's racked.
Show more
Anthropic has signed two six-year compute agreements this month, $45b with Nscale and $35b with Lambda.
Six years is the whole depreciable life of the hardware. Most of what got committed to at that length is the building and the power behind it.
Show more
Whenever rents from this campus get quoted as comparables, remember how they were assembled. The tenant holds warrants in the landlord, and Nvidia backs defined portions of the lease and power payments.
Show more
SB Energy gave OpenAI warrants valued at $5.5B to help secure it as a major data-center tenant, according to draft IPO documents reviewed by WSJ.
The SoftBank-backed company is preparing a potential $5B-$7B IPO and says it has nearly 9 GW of contracted compute capacity, despite having no data-center revenue yet.
Its data-center backlog is expected to exceed $400B, with OpenAI signing 17 leases for ~8 GW at its planned Ohio campus.
OpenAI invested $500M in SB Energy, while Nvidia has committed $3B and is also providing a residual-value guarantee tied to financing the Ohio buildout.
Show more
“Faster” and “better” are meaningless without context.
Physion Labs would be happy to sponsor an independent evaluation of both models—not to pick a winner, but to map the real quality–speed frontier and show which tradeoff is worth making for each use case.
Would both (@haozhangml, @isidentical) be interested in collaborating on this?
Show more
Jensen seems to be running the Elon playbook: own every layer you depend on
> $12.9B for Hugging Face
> $6B for Poolside's model factory
> Perplexity talks at $30B+
> ~$70B of equity out to OpenAI, xAI, Anthropic, SSI.
Chip company stops describing this fairly soon.
Show more
$1.1B for the Machine Age.
Most of the capacity being built right now is already leased before it opens and where a site needs new transmission you're looking at two to four years before anything runs. Doesn't really come across in the announcements.
Show more
The Hub counts downloads, and their own report says downloads miss private deployments. Nobody knows how much compute those weights burn.
BREAKING: NVIDIA agrees to buy Hugging Face for $12.9B, per The Information
"When you buy compute inference you're buying short term rentals based off demand" says @bavaria_kush
Kush says that's why inference providers are a natural early customer to compute indices such as @OrnnExchange's H100 volatility index.
Show more
Pretty sure all this flexible data center load being promised to utilities is inference because nobody is pausing a training run for the grid.
Compute got called the most important commodity of our time at the White House and America was told to go and win the market for it.
A commodity market runs on a unit people quote, a price built from transactions and contracts that move risk. Compute has all three of them now and the contracts written on it trade.
OCPI is the price, settling the GPU hour off executed transactions. Memory, tokens and power each have an Ornn index of their own.
So the market America is being told to win has actually been open for a while and it has already price.
Welcome to @OrnnExchange
Show more
American Compute Markets 🇺🇸 Never been more proud to be an American
