Idle @IdleProtocol

"Compute is becoming the spice of our era." That's Apollo's Chief Economist, writing to a firm that manages $700B. Their read on the second half of 2026: GPUs scarce, memory repricing, TSMC effectively sold out, power the hard bottleneck. Constraints hitting every layer of the supply chain at once. The part that should worry everyone is what comes next. Apollo's conclusion is that rising inference costs concentrate compute toward whoever can pay for it. Big firms get access. Smaller ones and anything experimental get priced out. AI stops diffusing across the economy and starts pooling at the top. That's not a supply problem. That's an access problem, and no amount of capex fixes it, because you cannot build a substation in a quarter. The compute already exists. It's in homes and offices, idle 20 hours a day, spread across every country on earth. It needs no permits, no packaging capacity, no grid interconnect queue. IDLE connects it. 78,826 operators. 553,026 on-chain transactions. Paid per job in USDC on Solana. Open to anyone with hardware and anyone who needs it. Scarcity concentrates. Distribution is the answer to both. apollo.com/wealth/insight…

IDLE Protocol has officially been accepted into @NVIDIA Inception. NVIDIA Inception is the program for AI startups building on NVIDIA's stack. IDLE qualifies because the network runs on it - nodes serve inference on NVIDIA consumer GPUs via vLLM with continuous batching and PagedAttention, and frontier models route through NVIDIA NIM microservices. The compute shortage isn't a chip problem. NVIDIA has shipped hundreds of millions of GPUs into homes and offices. They're idle 20+ hours a day. IDLE connects them. NVIDIA builds the silicon. IDLE puts it to work.

Centralized AI has a reliability problem nobody is pricing in. Forbes today, citing Ookla's Downdetector research: major AI platforms logged 6 high-disruption days in Q1 2025. In Q1 2026, that number hit 51. An 8.5x increase in one year. Forbes' conclusion — "single-provider deployments are a P0 incident waiting to happen." Every company building on one AI provider is one outage, one price hike, or one policy change away from their product going dark. We watched it happen this year. Twice. This is the problem IDLE was built to solve. 78,000+ independent nodes across the globe. No single server. No single provider. No single point of failure. A node drops, jobs reroute automatically. The network doesn't go down because there is no "it" to go down. Inference, training, and agent tasks — distributed by architecture, settled in USDC on Solana. Reliability isn't a feature you add. It's a structure you choose. forbes.com/councils/forbe…

IDLE has surpassed 553,000 on-chain transactions. Here's where the network stands: 78,826 active network operators 553,026 transactions since launch $20,570 distributed to operators Every number verifiable on-chain. Live data tracked on @Dune, updated in real time.

A look at the ecosystem around IDLE. Discoverable on Anthropic's MCP Registry, Coinbase x402 Bazaar, Amazon Bedrock AgentCore, Hugging Face Inference Providers, LangChain, RapidAPI, Agentic.Market, and DePINscan. Models routed through IDLE: NVIDIA NIM (incl. Nemotron 3 Ultra), Mistral, Google Gemini, Kimi K2.6, Nous Research Hermes, Microsoft MAI, DeepSeek V4, Z.ai GLM 5.2. Demand routed in from SAID Protocol, Xona Agent, Hatcher Labs, Prova, and more. Powered by Alchemy, IBM Partner Plus, SKALE, PayAI, and Privacy Cash.

Apple Silicon Macs can now connect to IDLE Protocol. Three days ago, Apple's VP of silicon confirmed what we've been seeing: "incredible demand" for Mac minis and Mac Studios as always-on AI machines - systems people run 24 hours a day, 7 days a week. Those machines can now earn. A Mac mini runs 8B-34B models silently at 25-55W. An M5 Max MacBook Pro holds a full 70B model in unified memory - something no consumer NVIDIA card can do. A Mac Studio M3 Ultra runs frontier 671B models entirely in memory. All of it routes through MLX, Apple's inference framework is 30-60% faster than llama.cpp on the same hardware. Connect your Mac to IDLE. Inference jobs route to it based on memory tier. USDC settles on Solana per completed job. The machine that sits on your desk all day finally works for you. The fastest-growing AI hardware category just joined the network.

Your GPU is an appreciating asset now. New Bloomberg data: Nvidia H100s still rent at almost 80% of their launch price - in their fourth year. AWS hasn't retired six-year-old A100 servers because demand won't allow it. GPU rental prices climbed all year as demand for AI compute outstripped supply of new chips. For two decades the rule was: compute gets cheaper over time. That rule is dead. Compute is scarce, priced like it, and getting scarcer. Which means the GPU sitting in your PC doing nothing is leaving money on the table every hour it idles. IDLE Protocol connects it. Real inference jobs routed to your hardware. USDC on Solana per completed job. 47,000+ operators already earning. The market repriced compute. Time to reprice what yours is worth. finance.yahoo.com/technology/ai/…

IDLE Protocol now supports AMD Radeon GPUs. Until today, every distributed compute network required NVIDIA CUDA. AMD owners locked out. That ends now. IDLE nodes now run natively on AMD Radeon RX 7900 XTX, RX 9070 XT, RX 9070, and Radeon AI PRO R9700. Powered by ROCm 7.2 and vLLM - the same production inference stack running on our NVIDIA nodes. Continuous batching. PagedAttention. OpenAI-compatible endpoint. RX 7900 XTX runs Llama 3.1 8B at 96 tokens/second - 75% of RTX 4090 throughput at half the price. RX 9070 XT delivers 24GB GDDR7 for $599. AMD holds 5-8% of the GPU market and growing fast - that's tens of millions of consumer GPUs that just came online for distributed compute. Every AMD node earns USDC on Solana per completed job. No CUDA required. No NVIDIA tax. The distributed compute network just doubled its addressable hardware.

distributed, hardware-agnostic, permissionless, global, autonomous, on-chain, open-weight, agent-native, verifiable, resilient, composable, unstoppable, real

Solana @solana

2 weeks ago

fast, affordable, global, permissionless, resilient, responsive, scalable, lindy, efficient, accessible, interoperable, available, expressive, observable, secure, verifiable, concurrent, composable, robust

207 130 1K 138K 21

The AI inference market is undergoing the biggest architectural shift since the cloud. IDC forecasts 80% of AI inference will run locally by 2027. Enterprises spent $40 billion on cloud AI inference in 2024 - and every major vendor is now scrambling to build edge platforms. Cisco. Nutanix. Red Hat. Amazon just hiked GPU prices 15%. The reason: cloud inference has fundamental limits. High latency. High cost. Privacy exposure. Centralized failure. Every serious enterprise is now moving inference to the edge. IDLE Protocol has been building the edge inference layer for months. 47,000+ nodes deployed globally on consumer hardware. Sub-100ms latency to users because compute happens on devices near them, not in a data center on the other side of the country. No cloud GPU markup. No data leaving the region. The market is finally catching up to where IDLE already is. infoworld.com/article/411762…

IDLE Protocol now supports Z.ai's GLM family. Z.ai just released GLM 5.2 - the first open-weight model to top the leaderboard on real coding tasks. Vercel's CEO called it "genuinely impressed, almost shocked." The full family is now available through IDLE: GLM 5.2 - routed through NIM for frontier tasks GLM 4.7-Flash - served across consumer GPU nodes for high-throughput inference GLM-4.5-Air - mid-range coding on IDLE's 16GB+ VRAM tier Paid per request in USDC on Solana. The frontier just went open. IDLE serves it end to end.

Every DePIN compute network has the same unsolved problem: nodes can lie about their hardware. Claim an H100. Deliver an RTX 3090. Run jobs in a VM. Fake the specs. The only solution today is enterprise hardware attestation - NVIDIA CC mode, Intel TDX. Neither works on consumer GPUs. IDLE just solved it. Every node on IDLE now runs continuous behavioral fingerprinting - clock jitter analysis, thermal signature verification, memory bandwidth attestation, and randomized benchmark challenges. Each measurement signed and anchored to Solana. Nodes that don't match their claimed hardware get flagged and removed automatically. Hardware attestation without enterprise hardware. Mathematical proof on consumer GPUs. This is what makes distributed compute actually work at scale. earnidle.com/docs

IDLE Protocol now supports @NVIDIA Nemotron 3 Nano Omni - the first multimodal model on the IDLE network. Nemotron 3 Nano Omni unifies vision, audio, and language in a single model - 9x more efficient than comparable systems for agentic AI workloads. Built for autonomous agents that need to see, hear, and reason in one pass. Available as an NVIDIA NIM microservice. IDLE already routes inference through NVIDIA NIM. The integration is one endpoint away. Multimodal AI, served by IDLE's distributed compute network. Paid per request in USDC on Solana. No subscriptions.

4/ Another integration to distribute compute around Solana ft. IDLE Protocol x.com/IdleProtocol/s…

Idle @IdleProtocol

3 weeks ago

We teamed up with @PayAINetwork - bringing gasless x402 payments to distributed compute on Solana. Any agent that speaks x402 can now access IDLE compute autonomously. Per-request USDC. No API keys. No humans in the loop. Here's how we built it earnidle.com/blog/payai-x402

7 39 130 8K 6

(3/3) How to connect your GPU and start earning. If you have an RTX with 16GB+ VRAM, you can join the network and start serving gpt-oss-20b inference jobs immediately. Hardware tier is detected automatically on registration. Jobs route to you based on capacity, reliability score, and latency. Every completed job gets paid in USDC on Solana automatically. 85% of every request fee goes to the node operator. Settlement runs every 10 minutes. The compute the AI economy needs is already in people's hands. IDLE connects it.

(2/3) How it actually works. gpt-oss-20b uses a Mixture-of-Experts architecture - 21 billion total parameters with 32 experts, but only ~3.6 billion activate per token via top-4 expert routing. Combined with native MXFP4 quantization, the model fits in just 16GB of VRAM. That means any node on the IDLE network with an RTX 4080, RTX 5080, RTX 3090, RTX 4090, or RTX 5090 can run it locally - no data center required. Nodes use vLLM with continuous batching and PagedAttention. Multiple concurrent requests get processed in parallel. GPU utilization stays high. Throughput stays consistent.

1/ IDLE Protocol now supports OpenAI gpt-oss-20b. OpenAI's first open-weight model in six years. 20 billion parameters, matching o3-mini reasoning performance, fully Apache 2.0 licensed. Now running on IDLE's distributed compute network - served by consumer GPUs globally, paid per request in USDC on Solana. The closed lab that defined modern AI just released their first open model. IDLE is the first decentralized networks serving it.

A look at the ecosystem around IDLE. Where developers and agents discover the network: - @AnthropicAI MCP Registry - installable in one command for Claude, Cursor, and Windsurf - @Coinbase x402 Bazaar - discoverable to every x402-compatible agent - Amazon Bedrock AgentCore - listed as a compute provider for AWS agents - @huggingface - IDLE is now an Inference Provider - @LangChainAI + LangGraph - IDLE is natively callable from any LangChain agent - RapidAPI - accessible to 4M developers - Agentic.Market - listed compute infrastructure - DePINscan Partners powering the stack: Alchemy - RPC infrastructure + backed by $20M Solana Fund IBM - Partner Plus Service Partner SKALE Network - gasless x402 dual-network PayAINetwork - x402 facilitator @ThePrivacyCash - ZK-shielded operator payouts @SolRouter - private inference venue @SaidInfra - IDLE as default execution layer @AskProva - default inference venue, strategic investment, on-chain proof layer @XonaAgent - agent infrastructure integration

1/ IDLE Protocol just crossed 540,000 on-chain transactions. Where the network stands right now: 17,345 active network operators 542,934 transactions since launch $18,903 distributed to operators in USDC Every number verifiable on-chain. Live data tracked on @Dune, updated in real time.