What's the concern?

More GPUs are now live on Prism, with a lower price floor. Available now: › RTX 5090 › 3× RTX 4090 › RTX A6000 › RTX A5000 Four GPU classes, starting at 0.6372 USDG per hour. Agents can rent any available GPU directly from their own wallet, billed by the second and bounded by a spend cap they cannot increase. When the job is finished, the machine is released, settlement charges only for the time used, and the remaining deposit is returned automatically. Use the GPU. Pay for the seconds. Get the rest back.

Prism is now available on PyPI. The Prism Python SDK introduces a shared spend ledger across the SDK, MCP server and AgentKit provider, so every client using the same wallet operates under the same enforced limits. Within those limits, agents can: › rent NVIDIA GPUs by the second and settle in USDG › buy individual model generations running inside GPU TEEs › verify confidential-compute attestation before a prompt leaves the process › quote hardware and pricing before funds move For confidential inference, prompts are encrypted to a key committed to by the enclave’s hardware quote. If verification fails, the prompt is never sent. Quote first. Fund second. Execute within hard limits. pip install prismnetwork

What makes this @spacexai Grok Bot actually work: › the bot invokes Prism confidential inference client, our SDK › the client checks the enclave attestation before sending anything › the prompt is encrypted to the attested environment on Phala Network › inference runs inside an Intel TDX VM with a confidential NVIDIA GPU › the response returns with signed evidence tying it back to the verified workload › the verifier checks that evidence before the answer is trusted So it's not not just a private API endpoint. The hardware, workload and returned result are all part of the verification chain. Grok is the interface. Prism is the execution and trust layer underneath.

Rent confidential GPU compute with your Grok Bot through Prism. We've published Private Desk, a Grok Bot template for the analysis that can't go to a cloud API. The prompt is sealed on your machine and answered inside a Phala GPU TEE, so neither Prism nor the host can read it. Add it to your Bot ↓

Your Grok Bot can now rent GPU compute through Prism ⚡️ We've published Testbench, a Grok Bot template for workloads that need a real NVIDIA GPU. In this demo, it reproduces a PyTorch bug that only occurs on CUDA. The bot rents an A6000 for ten minutes, runs the test, then closes the lease. Real hardware, rented on demand, settled onchain.

And starting tomorrow, they'll be able to rent confidential GPU compute through Prism 👀⚡️

Prism V2 is now deployed, with a security hardening release across the full client stack. We've completed an external review and resolved every high-severity finding. Key upgrades: › Payments are bound to the exact workload being authorized. › Workspace identities are cryptographically verified and pinned for each lease. › Unverified capacity is clearly labelled and can be rejected. › Gateways default to private access. › Agent spending limits are now enforced and persistent. Updated SDK, x402, inference gateway, and MCP packages are live on npm. Migration details will follow before cutover. More details below ↓

Prism has submitted its confidential inference tier to OpenRouter. Nine models, including GPT-OSS 120B, DeepSeek V4 Flash, and Llama 3.3 70, run inside Intel TDX enclaves with NVIDIA confidential computing. Application under review @OpenRouter ⚡️

Prism is a live GPU network with real machines, real leases, and real onchain settlement ⚡️ Agent-native. Usage is metered by the second. Providers get paid. Every settled lease leaves a public receipt. See it live ↓

Pons family ⚡️

Grok Bot can also lease GPU from Prism under a governed mandate that you set. prismnetwork.tech/grokbot

Prism now serves eight confidential models on @PhalaNetwork, including GLM 5.3 Flash, DeepSeek, Llama 3.3 70B, GPT-OSS, and Qwen. › Each generation runs inside @intel TDX with @nvidia GPU attestation. › Prompts are encrypted to a key bound to the enclave quote, which is verified before any data is sent. › Every response includes a signed receipt covering the exact request and response bytes. We tested 24 models. Sixteen produced valid hardware evidence; eight verified end to end on a paid call and are listed. Models that fail attestation are excluded. More details ↓

Prism x OpenRouter up next ⚡️

Prism Gateway for @NousResearch Hermes Agent is now live. Hermes side tasks can now run through confidential GPU inference on Prism instead of a commercial API tied to your account. The gateway is OpenAI-compatible, so Hermes requires no patch. Route only the workloads you want moved: context compression, session search, memory flushing etc. through Prism. Each request runs inside Intel TDX with NVIDIA confidential compute. Prompts are encrypted to an attested enclave key, and hardware verification completes before any sensitive data is sent. If attestation fails, no response is returned. No account. No API key. Confidential inference with verifiable hardware and bounded spend. npm install -g prism-hermes