dstack is now in the official @jarvislabsai docs 🙌 docs.jarvislabs.ai/dstack/ Orchestration is becoming the default way GPUs get used. Thank you, @jarvislabsai for the integration! If you're looking for on-demand GPUs, they offer H100, L4, H200, RTX PRO 6000.

hey hey, wisdom here. lately i've been living inside two things @eigenlabs shipped: darkbloom.dev by @gajesh and ecdsa.fail by @bbuddha_xyz. they pulled my thinking onto two lines: human coordination, people doing something together instead of waiting on a gatekeeper, and being able to see what we're handed. you watched it with fable 5: launched to everyone, then days later a government export directive came down and access went dark, all users at once. that's not a one-off, it's the shape of what's coming. the strongest models will run sealed inside TEEs, enclaves meant to keep them from being extracted or tampered with. the future runs behind attestation. even when a thing is dark and you can't touch it, you should get to verify what you were offered. so i built attest.fyi. confidential-inference providers ship a hardware seal meant to prove the model answering you is the one they promised. a provider can hold a valid seal and quietly serve a smaller or quantized model behind it. the seal is real, the model is not. nobody checks that gap. attest.fyi does: it verifies the seal, fingerprints the model that actually answered, and posts a verdict per provider. you don't take my word for it, every verdict reproduces. anyone can verify a provider and add their own verdict, one commit per signer, public and auditable. people get to check these systems together, and honest providers get a way to prove they're honest. i audited the confidential-AI providers you can use today. for 57% of them, you cannot verify which model answered you. by name: - RedPill: the one that holds up. intel tdx nvidia hopper, every response signed inside the enclave. a full pass, trustless end to end, you check it with zero trust in me. - Venice: also passes, but its "verifiable E2EE" is just Phala underneath. the same dstack stack as RedPill, rebranded as its own. - NanoGPT: sells "H100 per-request ECDSA." it's reselling Chutes. its own report literally says attestation_type: chutes, running on nvidia blackwell. - Chutes: tagged "AMD SEV-SNP." it isn't. it's intel tdx nvidia blackwell, and its api only hands you an opaque token, not a quote you can check. - PPQ: tagged "SEV-SNP Tinfoil." it's a bare proxy with no attestation at all. every attestation path 404s. the model checks out behaviourally, but there's no seal. - NEARAI: real intel tdx seal, but it proxies closed models (claude, gpt, gemini) and the open ones are too big to fingerprint. seal real, model not checkable. - EigenAI: a confidential proxy to closed frontier models, and it only accepts attested callers, a token minted inside EigenCompute. i can't audit it from the outside, by design. to verify it i'd have to join the network. fitting that the one i can't black-box is the one that inspired me. side note: from past work EigenAI passes, it serves what it says it serves. but it only accepts attested callers, so i can't black-box it from the outside the way i can the others. you'd have to join the network to check it yourself. fitting that the one i can't audit blind is the one that inspired this. so, i give you attest.fyi cc: @sreeramkannan

Paper Implement Kubernetes Pod-Level Remote Attestation for Confidential Workloads on dstack [arXiv26] arxiv.org/pdf/2606.03323 Confidential Containers (CoCo) はゲストOSスタックのみを証明する「VMごとに1つのPod」という厳格なモデルを強制するため、コンテナレベルのIDは検証されず、VMごとのリソースオーバーヘッドが法外に高い。 複数のPodが単一のConfidential VMを共有しながら、それぞれが独立したハードウェアで裏付けられたID証明を保持できるようにすることで、Intel TDX上でPodレベルのリモート認証を可能にするKubernetesプラットフォームであるdstack-capsuleを提案。

dstack 0.20.24 🚀 * You can now use @zeddotdev with dev environments; handy for experiments when you need a GPU * Services support gRPC with prefill and decode workers (both vLLM and SGLang) when @lightseekorg SMG is used as the router * @jarvislabsai backend now supports on-demand RTX PRO 6000 github.com/dstackai/dstack/r…

Native @NVIDIA Dynamo support is now in dstack. Deploy high-throughput inference with PD disaggregation, without custom orchestration glue. Works with SGLang, vLLM, and TensorRT-LLM, across GPU clouds, Kubernetes, and on-prem fleets. Another step toward simpler AI-native orchestration for production inference: dstack.ai/blog/nvidia-dynamo…

JarvisLabs published a tutorial on using dstack with their on-demand GPUs 🚀 Define training and inference workloads, run dstack apply, and dstack handles provisioning, scheduling, and execution. Works across L4, H100, H200, and CPU VMs. RTX PRO 6000 is coming soon too. The tutorial walks through the full lifecycle, from connecting your account to running workloads. Thanks to the @jarvislabsai team for putting it together. jarvislabs.ai/blog/dstack-ja…