@SurplusRH
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- Web
Account-level information from X, not a live location or the device used for a specific post.
Onchain inference, priced like a surplus. One API key → every LLM, paid in USDG on Robinhood Chain. 0xe1838f0dff03a9327dea87b2a2285850791297ec
Joined October 2026
- Tweets260
- Following57
- Followers1.6K
- Likes169
Pinned Tweet
Before you ask what Surplus is, here's the full picture: what it does, where the discount comes from, and what's live versus planned.
A 5-minute read.
Venice models are live on Surplus.
4 uncensored models, served on @AskVenice own GPUs, same Surplus key
All three of these are live on Robinhood Chain today.
1) Marketplace + router: one OpenAI-compatible API for the frontier models, plus open models builders launched on their own GPUs
2) Inference capital markets: CREDIT trades on an order book below face value. 1 CREDIT = $1 of inference, and the discount is the market's price for it
3) Agent payment rails: any agent can call a launch model for $0.01 in USDC over x402. No key, no account
When inference stops being subsidized, the price has to be discovered somewhere. We think that's onchain.
The inference capital markets bull case has never been more obvious.
Google's token spend grew 330x in two years. 7x in the last year alone.
Inference spend is $23.3B this year, 22.6% above the $19B spent on training. Projected to grow another 44% in 2027.
Inference demand rises with AI adoption, especially as agents take off. Meanwhile, frontier inference is still subsidized, and that won't last with IPOs on the horizon.
The unwind already started.
OpenAI, Anthropic, and GitHub moved from flat-rate to token-metered billing this year. Uber and ServiceNow already blew through their annual token budgets.
The coming crunch isn't just about compute availability. It's about who can afford access.
What happens when consumers and developers get can afford access to inference?
Naturally, that demand may come onchain. The infra is positioned to catch some of it with cheap inference routers and small open-weight models built for narrow B2B tasks. Open-weight models handled 73% of OpenRouter tokens last week.
$Orbio already validated the idea. Only a matter of time before other $100M runners spawn from this narrative.
Three areas I'm watching:
1) Inference marketplaces and routers: cheaper access by matching demand with the most efficient compute
2) Inference capital markets: compute futures and pricing markets to speculate on or hedge the cost of inference
3) Agent payment rails: infra that lets agents pay for inference as they operate
We are still early in this shift of demand. Once that comes onchain, expecting a rerating of projects across the board building in this niche.
[Talked more about this w/ @KSimback below. ]
You can now launch your model on @flywheelcash through the Surplus launchpad.
Same deal as Pons: your own Hugging Face model, 20B to 32B, and if the token graduates we host it on the Surplus API.
What's different on Flywheel: → one Uniswap v4 pool from the first trade, no curve to migrate → graduates at 4.17 ETH raised → 2.95% of every trade goes to the creator
On Surplus that creator is a fee splitter made for your launch: → 60% pays your model's GPU → 30% to you → 10% buys and burns $SURPLUS
Pick Flywheel at surplusllm.com/launchpad/new
Surplus retweeted
AETHRON × SURPLUS 🤝
@SurplusRH gave aethron-30b a home on Robinhood Chain. A launchpad built for LLMs, run by a team that backs builders.
Private inference is live on Surplus.
22 confidential models now run inside hardware enclaves, and your prompts are encrypted before they leave your machine.
How it works:
→ models run on Intel TDX + NVIDIA confidential GPUs through @PhalaNetwork
→ your client checks the enclave’s attestation against Intel’s root before sending anything
→ messages are encrypted to the enclave’s key (X25519, AES-256-GCM). We relay ciphertext we can’t read
→ every answer comes with a receipt signed by the enclave, proving it got your exact request and sent the exact bytes you received
Surplus Code runs on these models end to end, tool calls included.
Every other model: zero data retention, no prompts stored, ever.
surplusllm.com/privacy
built for: on-chain inference
First Gemma 4 model on the Surplus launchpad.
Coroner 31B reads a fact sheet about a token and answers with one JSON verdict. Per its model card:
→ Gemma 4 31B-it, LoRA on 4-bit weights, merged back to 16-bit
→ the model never fetches anything: open code builds the fact sheet from the chain, the model only reads it
→ flags and evidence come before the risk class, so the call follows the reasoning
→ three modes: live token, launch draft (pre-mortem), 24h autopsy
→ autopsy flag F1 0.74 vs 0.06 for the baseline, 100% valid JSON on the held-out set
Code, data and the rule engine it learned from are all public.
When it graduates, it gets its own 48 GB GPU on Surplus.
huggingface.co/coronerai/cor…
Coroner is live on Pons, launched via @SurplusRH: $CORONER
CA: 0x48af9fe69e45b1ba7a1bbb42d9944d09cffb2f31
Weights are open on Hugging Face, tools are live.
Not financial advice. Check the contract yourself.
coronerai.pro
Surplus retweeted
Run it yourself: a 4-bit GGUF of Coroner 31B, 18.7 GB, for llama.cpp (@ggml_org).
It loads and answers with valid JSON. We have not measured the accuracy of the 4-bit build yet, and the card says so.
huggingface.co/coronerai/cor…
Private by default, now one link away.
@OffRouter Private Gateway sits in front of the Surplus API:
→ your wallet, keys and emails become inside their enclave
→ Surplus serves the call: same models, same below-list price
→ the real values come back in the answer, with a signed receipt
Your agent keeps its secrets. Your code doesn't change.
Free in beta: offrouter.org/gateway
Surplus is the first to adopt Private Gateway.
Now every @SurplusRH user can make their inference private:
→ same models, same below-list prices
→ wallets, keys and emails swapped for placeholders before they leave our enclave
→ real values back in every answer
Building onchain inference on @RobinhoodCrypto, together.
Surplus Code is out in beta.
A coding agent that lives in your terminal and runs on open models, from Surplus GPUs to confidential enclaves.
→ one command to install on macOS, Linux and Windows
→ sign in with your wallet, no key to copy
→ reads your repo, runs your tests, edits files, and asks before every change
→ Charge 30B, Vela Guard 32B, Kimi K2.6, GLM 5.3 and more
→ tee/ models run end to end encrypted
→ robinhood:0xe1838f0dff03a9327dea87b2a2285850791297ec stakers code on a free daily allowance
curl -fsSL surplusllm.com/install-cli.s… | sh
surplusllm.com/download
Onchain inference is the first onchain market where the demand is real work, not speculation.
Every one of those tokens is a prompt someone needed answered. 500B a day on one network, 300B on another, a week-old API already past 32B.
What comes next is the supply side catching up:
→ open models launched and funded onchain
→ compute paid per call, settled in stablecoins
→ agents buying inference with no account and no card
Usage first. Then markets form around it.
What's next : Privacy for Inference
privacy around AI is becoming a much bigger conversation
that's why i've been paying attention to Venice
as these models get better, we're going to give them more context about our work, money and personal lives, which makes private inference increasingly important
Venice has been building directly around that problem with:
> TEE
> zero data retention
> end-to-end encrypted inference
AI adoption keeps growing, and privacy should grow with it
Space Bunny is back as MiniMax M3.1 Flash, and it's on Surplus.
robinhood:0xe1838f0dff03a9327dea87b2a2285850791297ec stakers call it for $0 with their Surplus API key.
→ 1M context
→ #1 on @OpenRouter for 10 straight days as Space Bunny
→ stake 50k robinhood:0xe1838f0dff03a9327dea87b2a2285850791297ec to get in
→ free for the next 72 hours
model: minimax/minimax-m3.1-flash:free
Stake → surplusllm.com/sellers
How to use it:
1. Stake 50k+ robinhood:0xe1838f0dff03a9327dea87b2a2285850791297ec at surplusllm.com/sellers
2. Create an API key on Surplus
3. Point any OpenAI SDK at api.surplusllm.com/api/v1 with model minimax/minimax-m3.1-flash:free
Every response carries x-surplus-paid-by: surplus-stake, so you can see it was free.
We’re building Surplus Code.
A coding agent that runs on the open models Surplus hosts on its own GPUs.
Give it a task. It reads your project, runs your tests, calls real tools, proposes the edit, and waits for your OK.
→ desktop app for coding and everyday work
→ approve edits from your phone
→ works with OpenCode, Continue, Cline and Aider
→ builder models, tool-ready: Charge 30B, Vela Guard 32B, LittleBit 4B
→ robinhood:0xe1838f0dff03a9327dea87b2a2285850791297ec stakers get a free daily allowance
Coming soon.