# iShard

> iShard is a five-node Blackwell GPU fleet operated by iJarvis LLC out of Tampa, FL. It sells dedicated and serverless inference to humans and agents, with explicit single-site, single-operator disclosures up front. Three tiers: Open Shard (per-token), Dedicated Shard (per-hour reservation), Private Fleet (monthly contract with custom SLA).

iShard is not a hyperscaler and does not pretend to be. It is a boutique, hand-routed inference service. Every node is NVIDIA Blackwell silicon: two DGX Spark Founders Editions (GB10 Grace Blackwell, 128 GB UMA each), one RTX PRO 6000 Blackwell (96 GB GDDR7), and two RTX 5090 desktops (32 GB GDDR7 each). Total fleet VRAM is 416 GB. Nodes are linked over a Tailscale WireGuard mesh. The control plane runs vLLM. There is no datacenter; the iron lives in the operator's Tampa workspace.

iShard's parent company is iJarvis LLC. The same operator runs other iJarvis properties including NPCx (in-game AI dialogue, anchor tenant on iShard), reruns.tv, AIDLAS, iSilentium, iResonate, and others.

## Pricing

- **Open Shard (per-token, best-effort serverless):** Llama 3.3 70B class $0.55/M tokens. DeepSeek V3 class $1.10/M. Llama 405B class $2.80/M. Custom checkpoint +15% surcharge. OpenAI-compatible API. $10 signup credit. No 99.x uptime commitment.
- **Dedicated Shard (per-hour, node reserved):** RTX 5090 $1.50/hr. DGX Spark $2.50/hr. RTX PRO 6000 $3.50/hr. Spark Pair (256 GB UMA) $4.75/hr. Hourly billing, no commit. Best-effort 99.0% during reserved window.
- **Private Fleet (monthly):** Reserve up to 5 nodes. 30–40% discount vs hourly. 6-month minimum. Custom SLA, custom routing, custom isolation. US-only data residency written into contract.

Machine-readable pricing: https://ishard.us/pricing.json

## Fleet

- spark-alpha — NVIDIA DGX Spark FE, GB10 Grace Blackwell, 128 GB UMA. Bookable.
- spark-bravo — NVIDIA DGX Spark FE, GB10 Grace Blackwell, 128 GB UMA. Bookable. Pair-link capable with spark-alpha for 256 GB UMA.
- ijarvis-slim — NVIDIA RTX PRO 6000 Blackwell, 96 GB GDDR7. Production tier; partially reserved by NPCx.
- ijarvis-thick — NVIDIA RTX 5090 Blackwell, 32 GB GDDR7. Bookable.
- ijarvis-thin — NVIDIA RTX 5090 Blackwell, 32 GB GDDR7. Bookable.

Machine-readable fleet state: https://ishard.us/fleet.json

## Honest disclosures

- One physical site (Tampa, FL) until Q4 2026 multi-site colo split.
- One human operator. Pixalate is their day job.
- No SOC 2 today. No HIPAA BAA today. TLS 1.3 in transit, dm-crypt at rest.
- Not the cheapest provider on a per-token basis. Not trying to be.
- Status page (status.ishard.us) ships Q3 2026 with real telemetry.

## For agents and machine readers

- /llms.txt — this file (LLM-readable site summary, Answer.AI standard).
- /ai.txt — AI training/inference policy.
- /pricing.json — full pricing, machine-parseable.
- /fleet.json — fleet state and capabilities.
- /agents.json — agent capabilities manifest (tools, endpoints, payment flow).
- /.well-known/mcp.json — Model Context Protocol server descriptor.
- /sitemap.xml — site map.
- API base (planned GA Q3 2026): https://api.ishard.us/v1/
- Reservations API: POST /v1/reservations returns a single-use Stripe payment link; settlement triggers Tailscale ACL issuance and a per-tenant base_url.

## Legal & policy pages (human-facing)

- /sla — Service-level agreement, per-tier commitments, escalation path.
- /security — Security and residency posture, encryption, sub-processors, compliance roadmap.
- /privacy — Privacy notice, GDPR/CCPA stance, no-tracking marketing site, no-training inference.
- /terms — Terms of service framework (v0.2 alpha; binding terms ship per-reservation).

## Contact

Email: hello@ijarvis.ai
Operator: iJarvis LLC, Tampa, FL, US
Parent: https://ijarvis.ai/