Primis Protocol
232 posts
@primisprotocol
The pricing layer for compute
primis is the pricing layer for compute. we help builders access compute through one unified layer instead of dealing directly with fragmented providers, unstable pricing, and infra complexity. this is our @BagsHackathon demo submission.
most teams don’t overspend on compute because they’re careless. they overspend because the market is broken. prices are fragmented. availability changes fast. providers expose different rates. spot moves overnight. and builders are forced to make infra decisions with incomplete information. @primisprotocol helps teams save by turning that chaos into workload-specific rates. query rate → reserve rate → attach workload → reconcile usage in the back, we score routes across price, availability, latency, reliability, and confidence and we also hedge prices for our builders. so builders don’t just get “cheap compute.” they get compute they can actually plan around. that’s the unlock. predictable compute economics for the compute markets. primis mode.
where the fuck are the creative devs btw how are kintara & papertrade the only cool things upcoming onchain rn
Lots of buzz recently on compute capital markets. But what might these markets actually look like? A few thoughts on its market structure from first principles: > First, almost everyone agrees that compute has a nonfungibility quality. It behaves closer to electricity (temporal, nonfungible) than corn, oil, and gold. > This nonfungibility creates several downstream corollaries: (1) Reservations/capacity forwards are almost always bilateral OTC trades on particular SKUs and params (I want X hours of H200s in us-east-1 running Y model at 12pm on 8/1/2026) (2) There is no transparent "one-size-fits-all" pricing model for "generic H200s" like there is for corn/oil/gold, hence no proper futures market used for hedging (3) Most of the teams building in the space (eg. Silicon Data, Ornn, Compute Desk) are focusing on "standardization" indices/benchmarks, in preparation to create a liquid futures market. > The short-side of compute markets fundamentally comes from neoclouds (Coreweave, Nebius, Lambda) and indepedent data centers (people with GPUs), while the long-side of compute markets comes from inference dev platforms (Fireworks, Modal, Baseten) and the agentic applayer (Cursor, Perplexity, Suno, Rime) that do not run datacenter fleets > But these principals will never directly trade on general compute exchanges (eg. an H200 basket) because they require specific SKUs. Instead, they'll make their reservations/capacity forwards for specific SKUs with OTC dealers. > These dealers in turn can "hedge" particular SKUs with exposure to the underlying generalized basket exchanges. So the folks actually using compute futures exchanges are going to be MMs/OTC desks/compute dealers on both sides. This creates an endgame market structure like below:
last thursday we opened the gates to the @primisprotocol beta by sending 150 invites to teams and individuals on our waitlist. you can expect an in depth update tomorrow of how this week went with concrete numbers. and the next step will be a full on open beta to everyone. primis mode
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: anthropic.com/news/fable-myt…
been a little quiet, not because nothing is happening but the opposite. first the market has been moving exactly where we thought it would: gpu availability tightening, prices repricing overnight, compute futures appearing and builders realizing “access” is not enough. the missing layer is becoming obvious. and that’s what we’ve been building with @primisprotocol. our closed beta that has been running for months is already processing ~$100k/monthly. and allowed us to understand what needed to improve for scale. which is what we’ve been working on. all this, to say open beta has never been closer as we estimate we’re 1 big sprint away from opening the gates. primis mode
Sam Altman said AI budgeting has recently become a "huge issue" for some companies, something that "never came up" earlier this year. bit.ly/4uxIGnv
So what Big Tech did in the USA is they bought up all the computer hardware needed for AI, including HBM (High Bandwidth Memory), driving prices up 400+%. Nvidia, meanwhile, doubled the pricing on its medium-range GPU, the GeForce 5090, which now retails at near $5000. What this did is it made local AI hardware unaffordable for the vast majority of American consumers who might have been interested in running local AI. Now, instead, they have to "rent" AI through Big Tech and pay the rising monthly access fees to access overpriced AI models offered by the same tech giants that bought up all the hardware. Fortunately, China came along and released DeepSeek, Xaomi, Qwen, etc., with per-token costs absolutely SLASHED to a fraction of U.S. tech company tokens. (Literally costs only about 1-2% to use DeepSeek, compared to using Anthropic Opus.) And at the same time, China is beginning to catch up on microchip fabrication and High Bandwidth Memory manufacturing, which will result in their products driving down prices for consumers sometime in 2027 (on the hardware). So U.S. Big Tech companies tried to create a near-monopoly to enslave the American people, and China actually broke those shackles and gave American access to affordable cloud-based AI right now, with more affordable local AI inference hardware coming soon. Go figure.
fyi only 4 big sprints left ! primis mode
fyi only 6 big sprints left ! primis mode
Compute is the new currency
