4d chess of how to loose 300+M fastest on things you don’t understand! He is like VC that gets negative power law in politics. Well played
Prashant retweeted
Call one provider and you get their prices, their rate limits and their 3 a.m. outage.
OpenRelay puts one endpoint in front of GPUs across providers and routes every request to the best silicon. Provider down? We reroute.
Change one line in your SDK.
openrelay.inc/inference-deli…
Prashant retweeted
The biggest lie in AI is that the future belongs to one winner.
Every technology wave starts with one company promising that one walled garden is all you'll ever need. And every time, market forces and creativity break the walls down.
I cofounded OpenSea and now OpenRouter around the opposite bet: humanity will refuse to be locked in.
Intelligence was never meant to belong to one lab, one model, or one gatekeeper.
The future of AI is open, plural, and yours to choose.
We are seeing so much software gains left on the table on these supposedly high optimized inference engines! Just by doing context aware routing and fused mega kernels we are getting 3-4x perf improvements in matter of hours. Future for inference is bright and increasingly low cost
Prashant retweeted
OpenRelay is a delivery network for AI inference.
We pool GPUs across providers and locations and route every request to the best silicon. Provider down? We reroute.
We bill flat rate, not per token. Just create an API key and change one line.
openrelay.inc/inference-deli…
Prashant retweeted
good morning to everyone
especially the orange line below
that’s context.dev, now ahead of both Exa & Parallel on NEEDLE’s scholar index
NEEDLE was built by @KeenableAI, it re-tests providers daily using real production traffic. you have to keep earning your spot
shoutout @styskin for building a beautiful benchmark
Prashant retweeted
Reserved GPUs sit idle most of the month.
Bring Your Own Capacity: onboard your fleet with one command, no open ports. We serve inference on the idle hours and pay you for them. You want a node back, it's yours again.
And we run it for free.
openrelay.inc/bring-your-own…
Deepseek v4.1 Abliterated and GLM 5.3 Abliterated is live on @OpenRelayInc
Dm me for access; we already have many organizations using these models to protect themselves from cyber threats.
Agents should have full access to GPU compute with a budget just like humans! Future of HPC compute with agents in the loop is bright and cost efficient
Your agent can manage its own compute on OpenRelay.
Ask for an H100 and it deploys the VM. Ask for a batch and it runs it overnight. It watches what it spends, live, and stops at the budget you set.
Give your agent an OpenRelay key, then let it cook.
docs.openrelay.inc/docs/cli/…
Prashant retweeted
1) Today we're announcing Atomwell, our state-of-the-art foundation model for predicting experimental outcomes, alongside scaling laws for molecular discovery.
We trained models from 800M to 9B parameters on ~80M experimental measurements, and found clean scaling laws in both model size and data.
Prashant retweeted
We just hit 800 customers!
A huge thank you to everyone who’s chosen context.dev and trusted us as a part of your infrastructure.
We’re grateful to have you here.
Prashant retweeted
A 3-year GPU commit costs more than owning the servers.
Most teams we talk to can't get GPUs on demand anymore, so they sign an offtake to reserve capacity. Decent rates only come with commitments of three years or more. Ask for six months or a year and the hourly price jumps.
When we run the numbers, owning beats a three-year commit, including colo (rented rack space) and power. At the end, you still have a server you can sell.
At $500K to $650K for a new 8x B300 server, owning comes to $3 to $3.70 per GPU-hour over three years before resale. A three-year commit costs $4.40 to $5.30 per GPU-hour, with 30% prepaid. Owning pulls ahead between months 10 and 26.
The hardest part is figuring out what you can sell the hardware for. Used GPU transaction prices aren't public. At Stoa Intelligence, we price five @nvidia platforms every week using executed trades and firm dealer quotes.
The A100 80GB came out in November 2020. One bought at launch hit $0 on the books in November 2025 on a five-year depreciation schedule. Our weekly price for a used 8x A100 system stayed between $51,763 and $58,314 from July 10 to October 2, 2026.
Owning is more doable than it sounds. A complete used 8x A100 server costs under $100K and gives you 640GB of GPU memory in one machine. Used 8x H100 systems got 12% cheaper over the same period. Suppliers on Stoa quote new B300 and B200 systems with firm prices and delivery dates, and used H100 and H200 systems available now.
Renting makes more sense if your usage is spiky or you need a few weeks of extra capacity. Buying takes cash up front, and we help with financing options.
The math and price charts are in the note below.
If you're about to sign an offtake, DM me your quote. We'll price out owning the same hardware.
Already own GPUs? Stoa Intelligence can value your fleet every week. DM me to set it up.
Future of inference is open weight and increasingly customized!
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Prashant retweeted
We built an 8000 sqft. high-throughput wet-lab in 30 days, to solve molecular discovery - starting with pesticides.
Most teams in drug-discovery don’t receive target-organism verification until clinical trials. Pesticides allow us to rapidly validate our model in the target organism directly. With the help of our high-throughput lab, we were able to generate compounds that killed a $13 billion pest, the Fall Armyworm.
We believe that the bitter lesson is coming for biology.
Prashant retweeted
Guys we're stupidly close to a major ARR milestone and i refuse to wait until monday
so i'm bribing the internet
Next 10 people to subscribe to context.dev get 50% off ANY plan for an entire year (yes, year = 12 months)
Code: SODAMNCLOSE
Let's push this thing over the line
