One assistant
Spark or a 32–72GB GPUA planner breaks down requests. Lightweight specialists start on demand; heavy work queues or uses authorized external services.
01 / PRIVATE ARCHITECTURE
A companion up front.
Private compute behind her.
Echo receives your request and stays with it. A control node routes specialist work to your private compute pool.
Your continuous point of contactEcho
Conversation · Presence · Review
Control node / SparkPlanner & Router
Plan tasks · Allocate resources · Track progress
Your private boundary
Private Compute Pool
Research / Code / Video / Review
Memory · Permission-scoped context
Private files and retrieval libraries stay outside tenant workloads
Private files and retrieval libraries stay outside tenant workloads
Optional external services
Named provider, task, data and budget.
Only approved context leaves after consent.
External workloads are separate from personal Memory.
Explore companion interactions and privacy controls
The companion comes first.
Still here, even when the answer takes time.
Echo acknowledges you, keeps the conversation moving, and shows what is happening while deeper work continues.
Presence is real-time; intelligence can take time.
Illustrative interaction · product target
“Can you work through these notes?”
Echo
I’m here. Let me work through that.
One experience, across the proposed range.
Local first. Your choice, always.
Your context. Your decision.
Personal memory stays local by default. Cloud assistance shares only the context you explicitly approve.
- Choose which files, apps, microphones, cameras, and screens Echo may access.
- Review and delete saved memories; pause observation and background work.
- Inspect the exact proposed cloud payload before sharing.
- Disable cloud inference and continue with local capabilities.
More capacity. The same boundaries.
When outside help would be useful, Echo explains why, shows what would be shared, and asks first. You can always keep the task local.
- 01Know who, and why.
Named provider and model, why outside expertise would help, and the proposed task.
- 02Review the exact context.
Messages, excerpts, files, images, and tool context prepared locally. Nothing is pre-uploaded.
- 03Decide for this task.
Review retention and training terms, charges or a spending cap, and any unknown completion time. Consent applies only to the disclosed task and payload.
Local means your configured device or private deployment. Remote hosting, backups, telemetry, connected apps, and web requests each need clearly disclosed data boundaries; local inference alone does not make those activities offline.
Policy explanation only. These controls do not connect to an AI service or send any context.
The consent contract, in full
Before anything leaves
Bind consent to a task, provider, disclosed payload, and cost limit. Any expansion requires new consent. Declining or not responding never authorizes upload.
Prepare the smallest useful payload locally. No background pre-upload, full memory export, or automatic fallback to a second provider.
If you decline
Continue locally where feasible, queue the work, narrow its scope, or explain that it cannot be completed locally.
If you change your mind
Cancel unsent work and stop future transmissions. Explain that already transmitted data cannot be recalled by Echo.
After outside help
Identify when an answer used external assistance and preserve the task’s sharing record.
Disclose the selected provider’s retention and training terms before consent. Never imply zero retention unless the selected service and configuration support it.
Permissions stay yours
- Require user-granted access to each data source and integration.
- Keep delegated work within an approved task scope, schedule, and resource budget.
- Pause at approval checkpoints for publishing, sending, purchases, destructive changes, or other consequential actions unless explicitly pre-authorized within that scope.
- Make active work inspectable and cancellable; keep a record of actions and exceptions.
What “local” means
Local means your configured device or private deployment. Remote hosting, backups, telemetry, connected apps, and web requests each need clearly disclosed data boundaries; local inference alone does not make those activities offline.
02 / AGENT SCALING
One point of contact.
A team for the task.
From planning and execution to review, Echo organizes complex work into results you can inspect.
Personal team
96GB or 2× PRO 6000Planning, execution and review have distinct roles. Two GPUs can separate conversation from heavy video jobs.
Private work team
4× PRO 6000Specialist tasks can run in parallel and create subtasks within depth, concurrency and budget limits.
Private production
8× PRO 6000Keep interaction services resident and separate from burst capacity for multiple workflows and batch work.
Agent factory
1× B300 → 8× B300 HGXEvaluate larger model services and heavy workloads with more memory and high-speed multi-GPU interconnects.
Echo→Planner→Researcher / Coder / Video Worker→Critic→Delivery
Resident roles stay available. Burst roles release resources when done.
Running: 3 demo tasks · Queued: 2 demo tasks
Demo data: the planner splits research, code and video tasks. Explore the next step; no real agents execute.
Agent creation is a software capability. GPU count is not an agent count. Recursive work needs depth, concurrency, budget and timeout limits, plus cancellation propagation. Concurrency requires measurement.
03 / VIDEO AI
From a single shot
to your private production team.
Echo coordinates storyboards, generation, review and editing to turn an idea into work you can review.
01 / Audio + videoMiniMax H3Local + hosted hybrid
NM integration planned
H3-Base / Context-IR / Regenerate-2K
Text, first/last frames and multimodal references; 4–15 seconds, 24 FPS and native stereo audio.
Open Base defaults to a 768p short edge. The full official 2K workflow includes hosted Context-IR and Regenerate-2K.
Official documentation & examples ↗02 / Local creationWan 2.2Local workflow candidate
NM integration planned
TI2V-5B
Text-to-video and image-to-video at 720p / 24 FPS.
Configure each variant separately. Do not assume the native audio features of H3.
Official documentation & examples ↗03 / World modelsCosmos 3Simulation + data generation
NM integration planned
Evaluate by variant
Multimodal world generation and action conditioning; primary generation settings up to 720p / 30 FPS.
For physical AI. Check Edge limitations separately; this is not a real-time avatar service.
Official documentation & examples ↗04 / Cloud creationVeo 3.1External cloud service
NM integration planned
Cloud API / service
Native audio, reference controls and enhanced consistency.
Access through cloud services. More local GPUs do not make Veo locally deployable.
Official documentation & examples ↗- DirectorScript
- StoryboardShots / References
- Video WorkersGenerate candidates
- CriticSelect / Rework
- EditorEdit / Audio
- YouReview final cut
A 60-second ad is a multi-shot production goal. Offline video is separate from real-time avatars; character consistency and automated review are workflow goals. Linked examples are from official providers, not footage of an integrated NM service.
04 / AUTONOMOUS COMPUTE ECONOMY
Yours when you need it.
Useful when you do not.
Reserve capacity for yourself. Put permitted idle resources to work, and expand beyond local compute within your budget.
Protect Echo and your personal work
External renting is off by default. Reserve interaction capacity and see personal jobs alongside available headroom.
Owner-first
Stop taking new external jobs first. Drain, preempt or wait according to each contract. A UI switch cannot promise instant recovery.
Zero-trustExternal jobs require separate execution environments, networks, storage, credentials and permissions. Clear tenant data afterward and retain audit records.
8× B300 / AllocationDemo data
GPU 01Echo reserveInteraction
GPU 02Echo reserveInteraction
GPU 03Personal agentPersonal use
GPU 04Video jobPersonal use
GPU 05IdleEligible · Off
GPU 06IdleEligible · Off
GPU 07IdleEligible · Off
GPU 08IdleEligible · Off
Reserved 2 GPUs · Personal use 2 GPUs · External use 0 GPUs · Idle 4 GPUs
2 queued · No recovery needed · External budget: $0
Demo data. Resources stay local and nothing is sent externally. Four GPUs represent eligible headroom only.
01
GPU-hours
Offer authorized GPU capacity by the hour
02
Inference / tasks
Deliver inference or video tasks
03
Agent outcomes
Deliver completed agent outcomes
Three commercial stages, not three markets launching at once. After reserving two GPUs, the remaining six are not a complete eight-GPU HGX. No simulated live income is shown.
05 / CAPABILITY LADDER
From your first Echo
to parallel production.
Start with DGX Spark for Echo Core, then explore nine dedicated GPU configurations for generation and multi-GPU production. Size for the workload; memory alone does not determine visual quality.
Showing 10 / 10 reference configurations
01Planned · validation required
NVIDIA DGX Spark
128GB unified memory
Shared CPU + GPU memory
Echo Core · Personal AI entry
Start with a compact private node for everyday assistance, document search, drafting and task planning. Add a dedicated generation node as your workload grows.
128GB unified memory is shared by the CPU, GPU and operating system. Heavy tasks may queue; model fit and application compatibility require validation.
Official specifications ↗02Planned · validation required
RTX 5090
32GB per GPU
1 GPU × 32GB per GPU
Optimized generation entry
Evaluate local generation with quantization, CPU offload or staged loading.
Heavy jobs share resources with Echo. Full component residency is not guaranteed.
Official specifications ↗03Planned · validation required
RTX PRO 5000
48GB per GPU
1 GPU × 48GB per GPU
H3 Creator Node
More loading headroom than 32GB, designed around an optimized pipeline.
Do not assume every component stays in memory at once. Validate the exact model variant.
Official specifications ↗04Planned · validation required
RTX PRO 5000
72GB per GPU
1 GPU × 72GB per GPU
Fewer memory compromises
Keep more components resident and reduce transfers between system and GPU memory.
Both 48GB and 72GB variants have 1,344 GB/s bandwidth. More capacity does not imply proportionally faster generation.
Official specifications ↗05Planned · validation required
RTX PRO 6000
96GB per GPU
1 GPU × 96GB per GPU
Dedicated video node
Evaluate an optimized H3 service together with video pre- and post-processing.
Coexistence with a large language model depends on measured peak memory.
Official specifications ↗06Planned · validation required
2× RTX PRO 6000
192GB aggregate
2 GPUs × 96GB per GPU
Room for conversation and creation
Reserve GPU 0 for interaction and GPU 1 for video generation.
192GB is aggregate capacity. Split services across cards; it is not one unified memory pool.
Official specifications ↗07Planned · validation required
4× RTX PRO 6000
384GB aggregate
4 GPUs × 96GB per GPU
Parallel creative studio
One GPU can serve interaction and orchestration while three generate shot candidates.
384GB aggregate. Shot concurrency and generation speed require workload benchmarks.
Official specifications ↗08Planned · validation required
8× RTX PRO 6000
768GB aggregate
8 GPUs × 96GB per GPU
Personal video production pool
Separate interaction, generation, reference images and post-production to scale workflow throughput.
768GB aggregate. Budget separately for CPU, RAM, storage, power and networking.
Official specifications ↗09Planned · validation required
1× B300
288GB per GPU
1 GPU × 288GB per GPU
Large-memory H3 node
Evaluate BF16 and larger working sets with less aggressive quantization and offload.
A server resource tier, not a desktop GPU replacement. No guarantee that every workflow fits.
Official specifications ↗10Planned · validation required
8× B300 HGX
2,304GB aggregate
8 GPUs × 288GB per GPU
Agentic Media Factory
Use high-speed interconnects for multi-GPU work, shot pools, candidate batches and editing.
2,304GB aggregate. Frameworks must explicitly support sharding and parallel execution.
Official specifications ↗Aggregate memory ≠ single-GPU memory. These are deployment recommendations, not measured performance tiers. On 32–96GB, feasibility depends on quantization, offload, input size and software version. Hardware changes loading headroom, latency, concurrency and scaling. Published benchmarks need a model version, precision, resolution, duration and test conditions.
06 / INFRASTRUCTURE
Start with your work.
Plan room to grow.
Plan models, interaction capacity, storage, networking, power, cooling and maintenance together. Pricing requires a defined configuration and a current quote.
Personal node
DGX Spark · Add a generation node when neededValidate your primary workflow first.
Plan this stage ↗Private teams
2× / 4× PRO 6000Reserve resources for interaction and creation.
Plan this stage ↗Production pool
8× PRO 6000Scale shots, review and post-production.
Plan this stage ↗Research clusters
1× B300 → 8× B300 HGXValidate large working sets, interconnects and whole-node scheduling.
Plan this stage ↗Explore the five Echo product tiers and comparison
The five product tiers describe service plans; the hardware configurations are references, not one-to-one equivalents. Explore the original comparison, download it or create a local inquiry brief.
Find your Echo.
Choose your level of local independence.
Every Echo tier can seek outside expertise with your permission. Higher tiers give more of your work room to happen locally, together, and over time.
Proposed range · final capabilities and availability to be validated.
Echo Core · Guided tasks
Your companion starts here.
Start with Echo at your side: remembering what matters, helping with everyday work, and finding a way forward when a task needs more.
Designed forFor individuals who want private everyday assistance in a compact personal system.
Ways to put it to work
- Recall a detail
Find a decision in your local notes and show the supporting passage.
- Prepare your day
Draft a brief from approved local information and flag missing context.
- Work through a request
Summarize a document, create a draft, and ask before sharing it.
Your context, close to you
Bring your notes, documents, and everyday context into one companion.
Help that follows through
Turn a request into a short, visible sequence of actions.
A path forward
Wait locally, simplify the task, or approve outside expertise.
Local capabilities & boundaries
Proposed local capabilities
- Voice interaction, interruption, and lightweight visual context when enabled.
- Personal memory search, document summaries, drafting, and everyday planning.
- Short tool workflows using approved integrations; background indexing scheduled around conversation.
Guided tasks
Completes short, bounded workflows with checkpoints; proposes extra resources when a task outgrows the local system.
Light background work, with heavier tasks queued.
What to plan for
- Limited room for large models and simultaneous heavy tasks; long jobs may queue.
- Shared memory also serves the operating system and CPU; 128 GB is not fully available for model weights.
- Local quality depends on the chosen models. Some tasks may remain unsolved if cloud assistance is declined.
Outside help, by choice
Expect to consider outside help for difficult reasoning, large workloads, or tight deadlines.
Continue locally where feasible, queue the work, narrow its scope, or explain that it cannot be completed locally.
Reference hardware & deployment
~128 GB shared accelerator-addressable memory
DGX Spark-class personal appliance
Single integrated CPU/GPU system with shared memory.
Compact desktop appliance · 1 reference node
NVIDIA DGX Spark system overview ↗- No multi-node cluster required for this reference.
- Leave memory and scheduling headroom for speech, perception, the OS, and runtime state.
- Validate the full conversation stack under concurrent indexing load.
Hardware references are illustrative, not certified Echo configurations or performance benchmarks. Installed memory is not fully available to models.
Echo Plus · Routine workflows
More of your day, handled locally.
Give everyday work more room. Echo Plus brings more reasoning and routine workflows onto your own system, with outside help still available when you choose it.
Designed forFor frequent users and creators who want more of their routine work to stay on their own system.
Ways to put it to work
- Build a project brief
Combine an approved document collection into a sourced working brief.
- Develop an idea
Iterate on copy, code, and supporting notes in one session.
- Run a routine
Organize new local files and prepare a reviewable digest.
More work stays close
Give everyday tasks more local reasoning capacity.
Routines that keep moving
Delegate repeatable work with clear checkpoints.
Room to create
Work across documents, code, and ideas with more local headroom.
Local capabilities & boundaries
Proposed local capabilities
- More room for stronger local solvers and resident specialist models.
- Document collections, coding assistance, creative drafts, and richer personal context.
- Routine multi-step workflows and background memory updates while maintaining the presence loop.
Routine workflows
Carries out routine multi-step work within approved tools and scopes, using checkpoints for exceptions.
Routine background workflows alongside conversation.
What to plan for
- Heavy parallel workloads can still compete with conversation and need scheduling.
- Memory spread across GPUs requires supported model partitioning; PCIe topology can limit performance.
- A four-GPU Plus configuration overlaps Pro’s lower hardware range; memory alone does not establish the tier.
Outside help, by choice
Consider outside help for exceptional reasoning, specialist models, or workload bursts; no fixed local completion percentage is promised.
Continue locally where feasible, queue the work, narrow its scope, or explain that it cannot be completed locally.
Reference hardware & deployment
~192–384 GB aggregate VRAM
2–4 × RTX PRO 6000 Blackwell-class GPUs
Validated multi-GPU workstation or server; plan partitioning and GPU communication around the actual PCIe topology.
Workstation or small private server · 1 reference node
NVIDIA RTX PRO 6000 Blackwell Server Edition ↗- Prefer one host for the initial reference; multiple hosts are optional engineering choices.
- Use the correct GPU edition and a chassis qualified for its cooling and power needs.
- Size host RAM, NVMe storage, and data ingestion to the intended models and personal corpus.
Hardware references are illustrative, not certified Echo configurations or performance benchmarks. Installed memory is not fully available to models.
Plus and Pro intentionally overlap at 4 × 96 GB. Assign the tier by a validated workload envelope, including reserved presence capacity, solver workload, concurrency, and supported workflow software. A GPU count alone cannot resolve the overlap.
Echo Pro · Delegated projects
Keep talking. Let deeper work continue.
Stay in conversation while Echo works through the bigger picture. Pro adds local capacity for deeper analysis, longer projects, and more work happening together.
Designed forFor developers, researchers, and professionals with sustained, demanding personal workloads.
Ways to put it to work
- Investigate a codebase
Trace a problem, propose a patch, and run approved checks while you discuss tradeoffs.
- Compare a research archive
Analyze local papers or reports and build a source-linked synthesis.
- Carry a project forward
Prepare drafts and intermediate outputs across a longer scoped workflow.
Depth without disappearing
Follow a demanding project while continuing the conversation.
Projects with continuity
Keep plans, intermediate results, and next steps together.
More work in motion
Coordinate multiple approved workflows with visible progress.
Local capabilities & boundaries
Proposed local capabilities
- Deeper local reasoning with more room for model state, context, and specialist workers.
- Concurrent coding, document analysis, and research over approved local sources.
- Longer tool workflows with resumable state and dedicated capacity for conversation.
Delegated projects
Plans and executes longer projects within a user-approved scope, using persistent state, checkpoints, and explicit exceptions.
Longer projects and multiple scheduled workflows.
What to plan for
- Larger context and more parallel workers consume memory; concurrency requires explicit resource limits.
- Four GPUs qualify as Pro only if the configured software and measured workload envelope support its intended service level.
- Long-running work still needs reliable tools and user review where appropriate; hardware cannot ensure correct reasoning.
Outside help, by choice
Offer external specialists or overflow when the selected local models, available capacity, or desired turnaround are insufficient.
Continue locally where feasible, queue the work, narrow its scope, or explain that it cannot be completed locally.
Reference hardware & deployment
~384–768 GB aggregate VRAM
4–8 × RTX PRO 6000 Blackwell-class GPUs
Multi-GPU server with validated PCIe communication paths and explicit workload placement.
Dedicated private compute server · 1 reference node
NVIDIA RTX PRO 6000 Blackwell Server Edition ↗- Start with one qualified server; do not assume eight GPUs fit or operate adequately in a desktop chassis.
- Reserve presence capacity through scheduling or a separate local endpoint, then benchmark under solver load.
- Multi-node expansion needs a supported distributed runtime and suitable networking; memory totals alone do not predict latency.
Hardware references are illustrative, not certified Echo configurations or performance benchmarks. Installed memory is not fully available to models.
Plus and Pro intentionally overlap at 4 × 96 GB. Assign the tier by a validated workload envelope, including reserved presence capacity, solver workload, concurrency, and supported workflow software. A GPU count alone cannot resolve the overlap.
Echo Max · Ongoing operations
A private system for sustained ambition.
Build a private foundation for the work that keeps growing. Echo Max gives demanding projects and ongoing workflows room to run, with one familiar companion at the center.
Designed forFor owners who want a personal compute installation built around demanding, continuous workloads.
Ways to put it to work
- Maintain a private research desk
Keep several local investigations running with source records and periodic briefs.
- Coordinate ongoing work
Run approved analysis, drafting, and code workflows on separate schedules.
- Understand an archive
Build and refresh a searchable local knowledge collection across substantial private material.
Keep projects moving
Maintain ongoing work with checkpoints and clear ownership.
More depth on your infrastructure
Bring demanding analysis to a dedicated private system.
One companion, coordinated work
See progress across multiple workers through Echo.
Local capabilities & boundaries
Proposed local capabilities
- A high-memory, tightly interconnected platform for large local solvers and multiple specialist workers.
- Sustained analysis across extensive approved archives, with longer-lived project state.
- Scheduled perception, memory maintenance, planning, and execution alongside reserved conversational capacity.
Ongoing operations
Coordinates persistent project workers and recurring workflows within explicit permissions, schedules, and resource budgets.
Persistent workers across several demanding projects.
What to plan for
- Requires a rack/server installation with engineered power, cooling, networking, and operational support.
- A single eight-GPU node is not a high-availability deployment; failures and maintenance can interrupt work.
- High memory capacity does not guarantee frontier-model parity, unlimited context, or accurate autonomous decisions.
Outside help, by choice
Cloud remains optional for distinctive external capability, unavailable models, or bursts beyond the private system’s capacity.
Continue locally where feasible, queue the work, narrow its scope, or explain that it cannot be completed locally.
Reference hardware & deployment
~1 TB HBM-class; reference total 1,128 GB
8 × H200 in an HGX/DGX-class system
One eight-GPU H200 SXM server with NVLink/NVSwitch fabric.
Private rack installation · 1 reference node
NVIDIA DGX H100/H200 system introduction ↗- Use a qualified HGX integration or complete DGX-class system, not eight standalone cards in a generic host.
- Provide local storage, backup, monitoring, and a responsive user endpoint.
- Multiple nodes are optional for scale or redundancy, with separate networking and failover design.
Hardware references are illustrative, not certified Echo configurations or performance benchmarks. Installed memory is not fully available to models.
Echo Infinite · Custom ongoing operations
Your personal AI infrastructure.
Shape a personal AI environment around your ambitions. Echo Infinite offers the most local headroom in the range for demanding projects, persistent workflows, and a deeply provisioned private deployment.
Designed forFor owners building a personal AI compute installation with extensive local workload capacity and custom operations.
Ways to put it to work
- Build a personal research environment
Coordinate multiple demanding investigations over private datasets.
- Run a creative and engineering workspace
Schedule compatible analysis, coding, and generation workloads on dedicated infrastructure.
- Maintain long-lived project context
Keep local knowledge and project state current across a broad set of approved activities.
Room for what comes next
Provision the most demanding local workloads in the Echo range.
Infrastructure shaped around you
Plan the deployment around your projects and privacy boundaries.
Continuity across your work
Keep approved workflows coordinated through a single companion.
Local capabilities & boundaries
Proposed local capabilities
- The greatest model-residency and concurrent-workload headroom in this proposed lineup.
- Large compatible local solvers, specialist workers, and substantial context or cache allocations within measured limits.
- Continuous scheduled knowledge maintenance and complex project coordination with reserved presence capacity.
Custom ongoing operations
Coordinates a broad set of persistent, permissioned workflows with custom resource allocation, oversight, and operating policies.
Broad persistent workloads with custom capacity planning.
What to plan for
- Infinite is a tier name; compute, context, storage, and throughput remain finite.
- Requires substantial facilities planning and ongoing operations; deployment location determines the actual privacy boundary.
- Does not include every frontier model, confer general intelligence, or remove the need for consent and human oversight.
Outside help, by choice
Outside help stays available for unique services, unavailable models, and overflow. Even this tier can benefit from external expertise with consent.
Continue locally where feasible, queue the work, narrow its scope, or explain that it cannot be completed locally.
Reference hardware & deployment
~2+ TB HBM-class; reference total 2,304 GB
8 × B300 in a DGX/HGX-class system
One eight-GPU Blackwell Ultra B300 server with a high-bandwidth scale-up fabric.
Personal AI rack infrastructure · 1 reference node
NVIDIA DGX B300 system introduction ↗- The reference is one eight-GPU system; a multi-node cluster is not required by the tier definition.
- Engineer site power, cooling, rack space, storage, and management to the selected vendor configuration.
- Additional nodes, redundancy, and remote endpoints are separate deployment choices.
Hardware references are illustrative, not certified Echo configurations or performance benchmarks. Installed memory is not fully available to models.
Proposed capabilities; final availability and performance depend on the validated configuration. Reference hardware is illustrative. Cloud assistance is optional and may incur separate charges.
Hardware determines how independent Echo is, not how intelligent Echo is allowed to be.
More compute increases capacity to act independently within a defined scope; it never grants broader permissions.
The right kind of room.
Compare your Echo.
Compare the intended local work envelope. Responsive presence and control over sharing are common to every tier.
| Designed for | Core | Plus | Pro | Max | Infinite |
|---|---|---|---|---|---|
| Local capability | Assisted local capability | Expanded local capability | Advanced local capability | Dedicated private compute | Flagship private compute |
| Delegated work | Guided tasks | Routine workflows | Delegated projects | Ongoing operations | Custom ongoing operations |
| Background work | Light background work, with heavier tasks queued. | Routine background workflows alongside conversation. | Longer projects and multiple scheduled workflows. | Persistent workers across several demanding projects. | Broad persistent workloads with custom capacity planning. |
| Responsive presence | Same design goal on every tier | Same design goal on every tier | Same design goal on every tier | Same design goal on every tier | Same design goal on every tier |
| Default privacy | Local first; explicit consent for cloud inference | Local first; explicit consent for cloud inference | Local first; explicit consent for cloud inference | Local first; explicit consent for cloud inference | Local first; explicit consent for cloud inference |
| Outside assistance | Optional, with task-specific consent | Optional, with task-specific consent | Optional, with task-specific consent | Optional, with task-specific consent | Optional, with task-specific consent |
Scroll across to compare all five tiers.
Compare reference hardware
| Reference hardware | Core | Plus | Pro | Max | Infinite |
|---|---|---|---|---|---|
| Reference memory | ~128 GB shared accelerator-addressable memory | ~192–384 GB aggregate VRAM | ~384–768 GB aggregate VRAM | ~1 TB HBM-class; reference total 1,128 GB | ~2+ TB HBM-class; reference total 2,304 GB |
| Reference compute | DGX Spark-class personal appliance | 2–4 × RTX PRO 6000 Blackwell-class GPUs | 4–8 × RTX PRO 6000 Blackwell-class GPUs | 8 × H200 in an HGX/DGX-class system | 8 × B300 in a DGX/HGX-class system |
Plus and Pro intentionally overlap at 4 × 96 GB. Assign the tier by a validated workload envelope, including reserved presence capacity, solver workload, concurrency, and supported workflow software. A GPU count alone cannot resolve the overlap.
Aggregate multi-GPU memory is not automatically a single fast memory pool. Fit and performance depend on sharding, quantization, context/KV cache, activations, runtime overhead, and interconnects. Spark’s 128 GB is shared CPU/GPU memory.
NVIDIA DGX Spark system overview ↗NVIDIA RTX PRO 6000 Blackwell Server Edition ↗
NVIDIA DGX H100/H200 system introduction ↗
NVIDIA DGX B300 system introduction ↗
07 / TRUST & FAQ
Move forward.
Know the boundaries.
Clear answers on capabilities, privacy and resource use.
What is available today?
This page supports hardware browsing, workflow illustrations and allocation demos. Echo model services, video generation, capacity markets, automatic purchasing and settlement are planned. Hardware tiers are design guidance, not benchmarks. A saved receipt confirms an inquiry submission.
Does 2×96GB behave like a single 192GB GPU?
No. Multi-card memory figures are aggregate. PRO 6000 configurations prioritize service separation. Even B300 HGX with NVLink / NVSwitch needs explicit framework support for sharding and parallelism.
Do I need B300 to start with H3?
No. Evaluate quantization, offload or staged-loading paths on 32–96GB configurations. Feasibility depends on the variant, precision, input size and software version. There is no universal memory threshold for a complete workflow.
Is the full H3 2K workflow local?
The official complete pipeline includes hosted Context-IR and Regenerate-2K, so it is presented as a hybrid service. Before execution, disclose the materials sent, destination and costs. Users can choose to keep work local.
Can it generate a 60-second ad in one pass, or a real-time avatar?
A 60-second ad is a multi-shot generation, review and editing goal, not a single 60-second H3 output. Offline video and low-latency avatars use different service chains. Character consistency and automated review do not succeed on every attempt.
Can rented GPUs return to me immediately?
Personal capacity should be reserved. When demand rises, stop accepting external work first, then drain, cancel preemptible work or wait for the contract to expire. Show release timing and alternatives when immediate recovery is unavailable.
Will outside tenants see my private memory?
Private memory, retrieval libraries and files are excluded from the proposed tenant data flow. External workloads require separate execution environments, networking, storage and credentials. Isolation must be verified in the backend; a diagram alone is not a security guarantee.
What revenue or payback can I expect?
There is no real settlement data yet, so this page makes no live-income or payback claims. Future estimates must disclose price, hours, utilization, electricity and other costs. GPU-hours, tasks and finished outcomes are separate commercial stages.
Can I rent a full eight-GPU HGX after reserving two GPUs?
No. Six available GPUs are not a complete eight-GPU node. Whole-node jobs require safely migrating personal services or providing separate reserved capacity within your rules.
What is sent with an inquiry?
Only your entered name, email, intended work, scale and selected configuration are submitted to this website for infrastructure consultation. No personal AI memory or private files are sent. Do not include passwords, keys or sensitive business information. Local brief copying and downloading remain available.
08 / LET’S BUILD YOURS
Plan your private
AI infrastructure.
Tell us what you want Echo to accomplish, and the scale you have in mind.
From research and coding to a private video team. Define your workflow, then choose the hardware that fits.
A saved receipt confirms submission. NM / NM Technology uses your inquiry to respond to your needs. Do not include passwords or sensitive files.