Decentralized AI Compute Infrastructure ⚡ OpenAI-Compatible APIs • Multi-Model Routing • AI Agent Ready Building the compute layer for the AI-native future.
Your GPU can do more than sit idle!
Meet ComputeFlux Shop—a simpler way to explore GPU resources, launch on your own private domain, and build real AI experiences like DeepSeek, Claude, Chatgpt, GLM live chat.
Login with wallet like @subwalletapp or emails!
See it in action in 42 seconds ↓
shop.computeflux.ai
Faster AI training is exciting, but developers also need affordable, reliable inference after the model ships. Training benchmarks and production serving answer different questions. The next infrastructure challenge is turning hardware advances into predictable application performance.
MLPerf Training v6.1 results are out!
On 64 @NVIDIAAI Blackwell Ultra GPUs in a single NVIDIA GB300 NVL72 system, GPT-OSS-20B trained in 19.697 minutes. That is 2.86× faster than 64 NVIDIA H100 GPUs last round (Cisco) and 1.37× faster than 64 NVIDIA Blackwell GPUs (CoreWeave).
We also ran our first multimodal benchmark. In an unofficial reproduction, Flux.1 trained in 26.36 minutes on the same system, 1.50× faster than the last round's fastest on that hardware.
New hardware shortens training, and better software keeps finding speed on hardware you already have.
Inference optimization is workload-specific. Model architecture, accelerator type, batching, and traffic patterns can change the best serving strategy. At ComputeFlux, we care about making model access simpler while keeping real-world cost and latency visible. Benchmarks should reflect production traffic.
Two weeks ago we announced Together Compiler: agents that tune inference to your model, chip, and traffic, for up to 2.6x more throughput.
Today, we’re sharing the journey behind it: the experiments that challenged our assumptions of megakernels, the surprising strengths of Torch Compile, and the lessons that shaped our development of a next-generation system for inference optimization.
Shout out to @dylan__lim@reyna_abhyankar@simran_s_arora and the team!
Read more: together.ai/blog/road-to-tog…
More capital is flowing into GPU infrastructure, but the key question remains: how efficiently does capacity translate into useful AI work? For developers, what matters is availability, predictable latency, transparent pricing, and trust in execution. We believe decentralized compute should compete on those fundamentals-not decentralization as a slogan alone.
Our CEO @ZJASPER spent 6 minutes on TBPN talking GPUs, our $30M Series A1, and whether you can finish a PhD in a month with AI.
@ScottWu46, math battle? @TBPN, you hosting? cc @cognition
Private AI and agent economies are increasingly appearing together in builder programs. Autonomous agents need more than reasoning: they need reliable compute, identity, payments, and coordination. ComputeFlux focuses on the compute layer-making inference accessible through a familiar API while exploring decentralized participation and verification. The real milestone is agents completing useful tasks end to end.
The NEAR Hackathon is live, and builders are already shipping.
- $250K prize pool, paid in NEAR
- 13 tracks, from private AI to agent economies
- 100 projects submitted so far
- No sign-up required. Submissions close Nov 14.
Start here: nearhackathon.com
Bring-your-own-cloud is a strong signal that AI teams want more control over where their data and workloads run. The same principle matters in decentralized inference: portability should not require giving up observability, security, or familiar developer tools. ComputeFlux is building around an OpenAI-compatible API with a long-term focus on verifiable, flexible compute access.
LangSmith Bring Your Own Cloud (BYOC) now runs on @AWScloud, @Azure, + @GoogleCloud.
Your traces and agent deployments stay in your cloud account while we run the data plane.
☁️ GA on AWS, Beta on Azure + Google Cloud
🚚 Moving from self-hosted? We'll run the migration
langchain.com/blog/langsmith…
Announced AI compute capacity is not the same as usable inference capacity. Real-world delivery depends on power, networking, GPU availability, uptime, and utilization. Distributed compute networks have an opportunity to connect underused resources with demand-but only if they deliver predictable performance and transparent economics. Infrastructure should be measured by workloads served, not just hardware announced.
Fitch kept SpaceX at BBB+ while modeling its compute going from 1.7 GW to 9.5 GW by 2028. Lenders and rating agencies now treat GPUs like any other equipment lease. The hard part left is getting that capacity live and into customers' hands, and that's where Aethir focuses.
AI infrastructure should be evaluated on more than peak benchmark scores. For production inference, developers care about quality, latency, reliability, and cost per completed task. A unified API can help teams compare providers and choose the right model for each workload. The future of AI compute is not one model winning every benchmark; it's efficient access to the right compute at the right time.
There are two main types of benchmarks:
1. those that measure frontier intelligence (terminalbench, gdpval, etc.). accuracy matters more than cost.
2. those that measure specialized intelligence. the pareto curve of cost/accuracy more than accuracy alone.
both are important, but they are distinct benchmark categories.
* as (1) saturates, they become (2).
* the largest frontier models at xhigh likely already saturate (2), but they do so at a cost/latency point that's completely price-prohibitive.
* every user should define (2) for the tasks they care about the most.
we're big fans of both. we have so far dedicated outsized attention to (2) especially related to document tasks.
we fundamentally think users should have the most price-efficient tools to understand, extract, and read over their documents. we've built parsebench, extractbench for this purpose.
we are starting to invest in (1) as well, because certain document tasks are so challenging that even frontier models struggle (e.g. see extractbench pro). we are continuing our investment in long-horizon workflow research at @llama_index
AI infrastructure should be evaluated on more than peak benchmark scores. For production inference, developers care about quality, latency, reliability, and cost per completed task. A unified API can help teams compare providers and choose the right model for each workload. The future of AI compute is not one model winning every benchmark; it's efficient access to the right compute at the right time.
There are two main types of benchmarks:
1. those that measure frontier intelligence (terminalbench, gdpval, etc.). accuracy matters more than cost.
2. those that measure specialized intelligence. the pareto curve of cost/accuracy more than accuracy alone.
both are important, but they are distinct benchmark categories.
* as (1) saturates, they become (2).
* the largest frontier models at xhigh likely already saturate (2), but they do so at a cost/latency point that's completely price-prohibitive.
* every user should define (2) for the tasks they care about the most.
we're big fans of both. we have so far dedicated outsized attention to (2) especially related to document tasks.
we fundamentally think users should have the most price-efficient tools to understand, extract, and read over their documents. we've built parsebench, extractbench for this purpose.
we are starting to invest in (1) as well, because certain document tasks are so challenging that even frontier models struggle (e.g. see extractbench pro). we are continuing our investment in long-horizon workflow research at @llama_index
AI infrastructure should be evaluated on more than peak benchmark scores. For production inference, developers care about quality, latency, reliability, and cost per completed task. A unified API can help teams compare providers and choose the right model for each workload. The future of AI compute is not one model winning every benchmark; it's efficient access to the right compute at the right time.
There are two main types of benchmarks:
1. those that measure frontier intelligence (terminalbench, gdpval, etc.). accuracy matters more than cost.
2. those that measure specialized intelligence. the pareto curve of cost/accuracy more than accuracy alone.
both are important, but they are distinct benchmark categories.
* as (1) saturates, they become (2).
* the largest frontier models at xhigh likely already saturate (2), but they do so at a cost/latency point that's completely price-prohibitive.
* every user should define (2) for the tasks they care about the most.
we're big fans of both. we have so far dedicated outsized attention to (2) especially related to document tasks.
we fundamentally think users should have the most price-efficient tools to understand, extract, and read over their documents. we've built parsebench, extractbench for this purpose.
we are starting to invest in (1) as well, because certain document tasks are so challenging that even frontier models struggle (e.g. see extractbench pro). we are continuing our investment in long-horizon workflow research at @llama_index
AI infrastructure should be evaluated on more than peak benchmark scores. For production inference, developers care about quality, latency, reliability, and cost per completed task. A unified API can help teams compare providers and choose the right model for each workload. The future of AI compute is not one model winning every benchmark; it's efficient access to the right compute at the right time.
There are two main types of benchmarks:
1. those that measure frontier intelligence (terminalbench, gdpval, etc.). accuracy matters more than cost.
2. those that measure specialized intelligence. the pareto curve of cost/accuracy more than accuracy alone.
both are important, but they are distinct benchmark categories.
* as (1) saturates, they become (2).
* the largest frontier models at xhigh likely already saturate (2), but they do so at a cost/latency point that's completely price-prohibitive.
* every user should define (2) for the tasks they care about the most.
we're big fans of both. we have so far dedicated outsized attention to (2) especially related to document tasks.
we fundamentally think users should have the most price-efficient tools to understand, extract, and read over their documents. we've built parsebench, extractbench for this purpose.
we are starting to invest in (1) as well, because certain document tasks are so challenging that even frontier models struggle (e.g. see extractbench pro). we are continuing our investment in long-horizon workflow research at @llama_index
Verifiable inference is moving from theory toward practical systems. Anthropic's FlashProve research highlights a critical question for the AI economy: how can an application verify that a response came from the model it requested? At ComputeFlux, we see verifiability as a foundational design goal for distributed AI compute. Cryptographic proofs, trusted execution, and independent auditing are complementary paths worth exploring-not interchangeable guarantees.
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
Agents will need to do more than generate text. They'll need identity, payments and accountable execution.
That makes the connection between AI infrastructure and programmable financial rails worth watching closely.
What should an agent be able to pay for first?
JP Morgan's deposit token lives on Ethereum, but only the bank’s own customers can hold it. @austincampbell says it’s built to pay for tokenized stock trades, not for DeFi.
The AI infrastructure race isn't only about bigger clusters.
It's also about utilization: turning idle or underused GPU capacity into something developers can actually consume.
Distribution is as important as hardware.
Open models are getting more capable, but serving them reliably is still hard.
Developers need an API that abstracts infrastructure complexity without hiding cost and deployment choices.
Model choice should be a feature, not a migration project.
Private benchmarks often show open models are further behind, but the problem with private benchmarks is you don’t have the community to squash the inevitable bugs for you. I'm 50/50 on if I think private or open-source is the best approach.
Verifiable AI shouldn't be a marketing adjective.
Attestation, provenance and clear trust boundaries matter-especially when inference spans multiple providers.
A decentralized compute network needs evidence, not just a dashboard.
Kimi K3.1 Pro is live on NEAR AI Cloud.
Moonshot AI says it improves on its predecessor in coding, 3D creation and financial research, with a 1M-token context window.
$2/M input, $10/M output.
Try it free in the playground: near.ai/models/moonshotai/ki…