AnaRack · The AI-Native Runtime Platform

The heterogeneous AI rack for loosely-coupled enterprise AI.

AnaRack is a heterogeneous rack chassis for loosely-coupled enterprise AI — inference, fine-tune, MoE serving, agentic chains. CPU controls. GPU serves. Standard Ethernet between them. Multi-generation GPUs and mixed accelerators run together on the same rack, with rack-wide workload placement below the application layer. Your CUDA stack, your K8s, your models — unchanged. AnaROS attaches as a runtime overlay; no application rewrites. Physical modularity. API-first. Cloud-native agility. The same abstraction extends to a virtual rack on public cloud — see Chapter 06.

BUILT FOR

Loosely-coupled enterprise AI.

  • Inference serving
  • Fine-tuning · LoRA · PEFT
  • Agentic workflows
  • MoE serving
  • RAG / retrieval pipelines
  • Multi-tenant inference

These workloads benefit from stage independence — different stages on different capability subsystems, refreshed on different clocks, scheduled per pipeline. Heterogeneous beats homogeneous when the bottleneck moves stage-to-stage.

01 · CHASSIS
Heterogeneous rack chassis
Standardize disparate hardware — multi-generation GPUs, mixed accelerators — into one foundational architecture.
02 · ONBOARD
SDI capability registration
Subsysstems drawer detect, register, abstract — then enveloped by the governed perimeter.
03 · INTERSECTION
Compute × pipeline flow
Horizontal drawers · vertical pipelines. AnaRack routes one onto the other.
04 · ADVANTAGE
Blueprint vs siloed
Minutes to onboard. Unified governance. Elastic scaling. Rack-wide resource map.
05 · SCALE
Limitless scale
Many racks. One logical chassis. Cross-rack spine or virtual link to cloud. No rewrites.
06 · ABSTRACTION
Physical and virtual rack
AnaRack extends to the cloud-composed rack (EC2 + GPU + VPC + S3). AnaROS governs both.
07 · GOVERNED
Enterprise-grade serving
Multi-tenant by contract. Guaranteed SLAs, total isolation, unified visibility.
08 · PCIe SHELF
AnaRack PCIe expansion shelf
4U Gen5 chassis. Anavec branded. Available today for design partner pilots.
09 · NETWORK
Anavec Ethernet switch family
Spine to edge. Multi-generations. SONiC + AnaROS-native. 51.2 T in pilots.
CHAPTER 01

A heterogeneous rack chassis, not bespoke infrastructure.

AnaRack standardizes disparate hardware components into a single foundational architecture. Multi-zones, one chassis, one runtime — and any compliant drawer slides in.

Multi-zones. One foundational architecture.

Today, enterprise AI infrastructure is bespoke — every workload assembles a different pile of servers, switches, and storage. AnaRack makes that the wrong question. The chassis is the answer; the modules are the question.

Multi-zones standardize the surface that AnaROS schedules onto. Every zone accepts any third-party drawer that conforms to its SDI contract. The drawers can change. The chassis doesn't.

[ZONE_A]

High-Density CPU Tier

x86 / ARM compute sleds carrying the control plane, services, and data-prep stages of the pipeline.

1U / 2U profiles
[ZONE_B]

Accelerated GPU Compute

PCIe shelves carrying heterogeneous GPU payload — drawers swap on their own clock, vendor-neutral.

multi-vendor
[ZONE_C]

Networking & Switching Layer

Leaf/spine fabric · out-of-band management · scale-up + scale-out + scale-cross under one plane.

SONiC · multi-vendor
[ZONE_D]

Memory & Storage Tier

NVMe-oF shelves, memory pools, and object tiers — drawer-class storage and AMSF Tier 0.5 staging surfaced through SDI.

hot · warm · cold
AnaRack heterogeneous rack chassis · modular drawers and orange/blue pipeline flows
CHAPTER 02

Heterogeneous capability drawers.

Hardware components — GPU servers, CPU nodes, switches — are not independently installed. They onboard and register into the platform as drawers, through SDI, vendor-neutral.

Components don't just get installed. They get onboarded.

The moment a subsystem on a rack slides into a chassis, AnaRack's SDI layer register it, fingerprints its capabilities, runs a class probe, and onboard it onto AnaRack. From that point on, the subsystem is no longer a piece of hardware — it's a logical capability the platform can schedule, below the application layer.

3rd-party hardware integrates through the SDI contract for its class. Onboarding is minutes, not weeks. Removal is a hot-swap, not a maintenance window.

>INITIALIZING CAPABILITY ONBOARDING
>DETECTING:8× H100 TENSOR CORE GPU | 10-Slot PCIe expansion shelf
>CLASS:GSL · accelerator | pcie-shelf
>FINGERPRINT:0x9F2A:H100:80GB:NVLINK·OK | 0x16B8:L40S:PCIe·OK
>STATUS: READY FOR REGISTRATION
ONBOARDING · LIVE
Capability drawer onboarding · INITIALIZING CAPABILITY ONBOARDING · 8× H100 Tensor Core GPU · READY FOR REGISTRATION
01 · DETECT

SDI auto-discovers

Drawer slots in
AnaRack's SDI layer auto-discovers the drawer the moment it lands in a slot. Fingerprints model, count, firmware, NVLink topology, NIC ports.
capability_class=GSL · fingerprint=0x9F2A
02 · REGISTER

Governor admits it

Capability registered
AnaROS Governor validates the drawer against its class registry, runs the probe, then registers the capability as available — bound to this rack, this slot.
probed_ok · lifecycle=deployed_full
03 · ABSTRACT

Raw silicon → logical pool

Physical complexity hidden
AnaRack translates the drawer's physical surface into logical capabilities the platform can schedule: TFLOPS, aggregated vRAM, network bandwidth, IOPS.
logical · TFLOPS · vRAM · bandwidth
04 · GOVERN

Enveloped by perimeter

Inside the boundary
The drawer is enveloped by the AnaRack software perimeter. Unified abstraction, zero-trust boundary, global pool visibility. Pipelines can now claim it.
governance_signal=admitted · ledger=claim_live
SDI · SOFTWARE-DEFINED INTEGRATION

Heterogeneous hardware. Integrated platform.

Every third-party drawer that lands in an AnaRack slot walks the same four-step journey through the SDI contract for its class. Vendor-neutral on the hardware side; uniform contract on the platform side. The chassis stays the same — the drawers do not have to.

Detect · Register · Abstract · Govern
CHAPTER 03

The intersection of compute and flow.

Horizontal drawers carry potential energy. Vertical pipelines carry kinetic energy. AnaRack hosts the runtime where the two meet — pipelines agnostic to hardware, routed onto the right drawer on demand.

Horizontal capacity. Vertical pipelines.

If the horizontal drawers represent the system's potential energy, the vertical pipelines represent its kinetic energy. These are the live AI workflows coursing through the chassis — model weights streaming in, inference requests streaming out, telemetry rising back to the control plane.

Pipelines abstracts the underlying hardware. AnaRack routes each workflow onto the most efficient governed drawer, on demand, and reshapes the topology as the workload mix shifts. The same chassis serves a thousand different pipelines.

Vertical axes of execution · data pipeline flow and execution pipeline flow, 400 Gbps, latency < 1ms
↓ THE DATA & CONTROL PLANE 400 Gbps · vertical flow

Orchestration flows down.

Model weights, RAG context, telemetry, state synchronization — anything orchestrative moves through the data & control plane. AnaROS Governor lives here; AnaRack ferries its decisions to the drawers.

  • Model weight loading
  • RAG context ingestion
  • Telemetry collection
  • State synchronization
  • Policy distribution
↑ THE EXECUTION PLANE 400 Gbps · < 1 ms · vertical flow

Execution runs across.

Live inference requests, dynamic compute scaling, response generation, parallelized batches — anything productive moves through the execution plane. Sub-millisecond round-trip across drawers, governed end-to-end.

  • Live inference requests
  • Dynamic compute scaling
  • Response generation
  • Parallelized batches
  • Egress streaming
01 · REQUEST INBOUND

Pipeline enters the platform

A pipeline carrying an inference workload — agentic call, RAG query, fine-tune step — arrives at the AnaRack runtime. It carries intent, not a hardware target.

02 · OPTIMAL ROUTING

Control plane directs flow

AnaRack's control plane reads the registered drawer pool, applies the pipeline's SLO and tenant policy, and directs the flow to the most efficient governed drawer available right now.

03 · EXECUTION

Drawer provisions exact compute

The selected drawer provisions exactly the compute the pipeline needs — no more, no less — and the workflow proceeds. Telemetry returns up the data & control plane.

CHAPTER 04

The blueprint advantage.

Four dimensions where the AnaRack blueprint changes the equation versus siloed infrastructure — velocity, governance, scaling, and resource utilization.

DIMENSION

TODAY

Siloed Infrastructure

ANRP BLUEPRINT

AI-Native Runtime Platform

Deployment Velocity
How long to land new capacity
Weeks manual provisioning · rack, cable, configure, validate
Minutes drawer onboarding · SDI auto-discover + register
Governance
Where the perimeter sits
Highly fragmented per-node security · stitched policy · per-domain tools
Unified perimeter one bounding box · one control plane · ledger-backed
Scaling
How capacity grows under load
High friction hardware-bound · forklift upgrades · re-architect
Elastic & dynamic capability registration · slot more drawers · stack more racks
Resource Utilization
What fraction is actually working
Low stranded compute · workload-locked silos · idle GPU dollars
High logical pools · scheduled per-pipeline · multi-tenant
CHAPTER 05

Scalable. Many racks, one logical chassis.

Slot in additional hardware and AnaRack takes care of the rest — modular drawers extend into modular racks, all behaving as one logical chassis. Stackable across N racks per pod with a single contract and one stack of operation.

Many racks. One logical chassis.

Slot in additional hardware and AnaRack takes care of the rest. The SDI layer discovers the new drawer or rack the moment it lands, abstracts its raw silicon into logical TFLOPS, vRAM, and bandwidth, and utilizes it inside the same governed runtime as the original chassis. Cross-rack spines between racks. No re-architecture. No application rewrites. The pipelines never see new hardware — they just get more of it.

Every property of the single-chassis runtime extends cleanly across racks. Pipeline X-Ray, POFC fabric, telemetry, and the resource ledger correlate one stack of operation with full visibility, traceability, and governance — at any scale. Expandable, stackable to N racks per pod. The chassis grows; the contract doesn't.

N racks per pod · cross-rack spine · 1.6 Tbps · scalable capacity 3× shown
Limitless modular scalability · Rack 01, Rack 02, Rack 03 connected by cross-rack spine at 1.6 Tbps · scalable capacity 3×
CHAPTER 06

Virtual rack — same abstraction, cloud-composed.

AnaRack is, first, an abstraction — heterogeneous compute, fabric, memory, storage, governed perimeter. The hardware in Chapters 1–7 is one substrate. Public cloud GPUaaS is another: EC2 instances + GPU instances + VPC + EBS/S3 + IAM policies compose a virtual rack with the same five elements.
ANRP · ONE ABSTRACTION · TWO SUBSTRATES AnaRack · RACK-AS-A-SYSTEM compute · fabric · memory · storage · governed perimeter · SDI PHYSICAL RACK silicon · AnaRack chassis CPU sleds · control GPU shelf · heterogeneous NVMe + memory Ethernet fabric · SONiC governed perimeter SDI · POFC · AAIF ≡ VIRTUAL RACK cloud-composed · AWS / GCP / others EC2/GCE/VM sleds · control GPU instances · P5 / H100 EBS · S3 · Cloud Storage VPC · ENI/VNIC · TGW/VM IAM · SG · NACL SDI · POFC · AAIF same five elements · same SDI contract · same AnaROS governance

AnaRack is the abstraction. Hardware is one substrate; cloud is another.

Every cloud-composed AI environment lines up against the same five elements an on-prem AnaRack chassis exposes: EC2 sleds for control · GPU instances for compute · EBS+S3 for storage · VPC for fabric · IAM/SG for the policy plane. The SDI contract that AnaRack defines holds across both. The POFC fabric correlation and AAIF verdict engine attach the same way. Cloud GPUaaS isn't a separate product category — it's a virtual rack you assemble per-workload.

PHYSICAL RACK · ANRP CHASSIS
Heterogeneous hardware drawers.
  • CPU sleds · 1U/2U
  • GPU shelf · 4U, heterogeneous
  • NVMe · hot · warm · cold
  • SONiC switch · 51.2 Tbps
  • Governed perimeter · OCP-open
VIRTUAL RACK · CLOUD-COMPOSED
Same five elements, cloud-rendered.
  • EC2 instances · m6/m7/c7
  • GPU instances · P5 · G6 · A100
  • EBS · S3 · object + block
  • VPC fabric · ENI · TGW · DX
  • IAM · SG · NACL · policy plane
SEE OPERATOR PAIN The operational story — split-brain governance across physical and virtual racks — is told in Use Cases · Case 09 · Physical + Virtual Rack.
Pipeline continuum · physical and virtual

The workflow continues — from your rack into the cloud.

AnaRack makes physical and virtual racks one substrate for one pipeline. The workflow that runs on your hardware today runs on cloud capacity tomorrow — same SDI contract, same governance, same audit trail.
AnaRack hybrid pipeline continuum — an AI workflow flows from physical on-prem racks into the virtual rack composed on hyperscaler clouds (AWS, Azure, GCP) and external LLM endpoints. The pipeline remains one unit of work, governed end-to-end by AnaROS.
FIGURE · ONE PIPELINE · PHYSICAL TO VIRTUAL · ONE UNIT OF WORK
The pipeline does not stop at the rack edge — it continues into the cloud, observed and governed as one.
CHAPTER 07

Governed enterprise serving. Multi-tenant by contract.

Modular physical capabilities dynamically fueling governed, logical AI pipelines — multi-tenant by contract, cryptographically isolated, single-pane-of-glass observable. Whether substrate is a physical AnaRack chassis or a virtual rack composed on cloud GPUaaS, the governance contract is the same.

Multi-tenant by contract, not by configuration.

Every enterprise pipeline carries three things the chassis has to honor: a tenant identity, an SLO budget, and a chargeback line. AnaRack makes all three first-class properties of the runtime — workload prioritization happens at the drawer level, not in a side-car queue. Noisy neighbors get rate-gated before they ever touch a paying tenant's p99, and every SLA is observable, attributable, and chargeback-friendly out of the box.

Multi-tenant pipelines execute behind strict cryptographic separation, with tenant identity carried into every fabric hop — no shared-fate at the GPU, no implicit memory crossing. And every drawer, every pipeline, every hop reports into the same telemetry plane: pipeline X-Ray, POFC fabric, and resource ledger all stitched into one pane of glass for the entire modular runtime.

Governed Enterprise Serving · Guaranteed SLAs — workload prioritization prevents noisy neighbor disruptions · Total Isolation — multi-tenant pipelines execute with strict cryptographic separation · Unified Visibility — a single pane of glass for all telemetry across the modular runtime · shield enveloping rack with secure pipeline flow
CHAPTER 08

Anavec PCIe expansion shelf.

Seven chapters of architecture. The Anavec PCIe shelf is the AnaRack-compliant chassis — available today for design partner deployments.
anavec PCIE · 4U · GEN5 GPU-A · 01
FRONT FACEPLATE · 4U · GEN5

10-Slot programmable PCIe expansion shelf

4U PCIe Gen5 expansion Internal PCIe Gen5 switch fabric provides programmable host:slot partitioning — not a fixed ratio chassis. BMC integration native to AnaROS; hex-cell filter slides out from the front; SDI onboarding on first power. Available for design partner pilots.

FORM
4U · 19"EIA-310
SLOTS
8× dual-width x16 FHFL + 2× single-width x16 FHFL or 16× SW x8
BANDWIDTH
256 GB/sPCIe Gen5
POWER
675W / slotper accelerator
BMC
IPMI + AnaROSdrawer-native
ONBOARD
SDIzero-touch
PARTITIONING Programmable PCIe Gen5 switch fabric — host:slot fan-out is configurable, not fixed. Supports 1, 2, or 4 host uplinks (64 / 128 / 256 GB/s aggregate) and per-host ratios set per backplane at boot or runtime via SDI. Partition topology details on request.
NEXT Own-engineered chassis on roadmap — deeper telemetry, faceplate · filter · drawer ID optimized around partner duty cycles. Same programmable promise. Same standard interfaces.
CHAPTER 09

Anavec Ethernet switches. One stack. Every tier.

Same chassis story, different layer. A family of switches — spine, aggregation, ToR, and edge — built on a curated mix of silicon, current and next generation. SONiC as the network OS for the data plane; AnaROS-native control plane for SDI onboarding, POFC, and AAIF verdicts. One vendor of record from silicon to SLO, across every tier.
anavec SW-A PORTS 1 – 32 · OSFP800 PORTS 33 – 64 · OSFP800 CON MGMT USB STATUS
2U · 51.2 T · SHOWN

64-port OSFP800 fabric switch · SONiC + AnaROS-native.

The top of the family — a 2U Ethernet switch at 51.2 Tbps non-blocking, available today, anchors the rack spine for scale-up within the rack and scale-out across the pod. SONiC as the open, hardened network OS for the data plane; AnaROS-native control plane for SDI onboarding, switch-port-to-pipeline correlation (POFC), and AAIF verdicts. The rest of the portfolio — aggregation, ToR, and edge — ships with SONiC + AnaROS stack, and tracks the next silicon generation on the same compatibility envelope. One vendor of record from silicon to SLO, across every tier. Available for design partner pilots.

FORM
2U · 19"EIA-310
PORTS
64× OSFP800or 256× 200GE breakout
BANDWIDTH
51.2 Tbpsnon-blocking
SILICON
TH5 · SP4TH6 / SP6 on roadmap
NETWORK OS
SONiC + AnaROSdata + control plane
INTEGRATION
SDI + POFCAnaROS-native
FAMILY SPINE · 51.2 T · 102.4T AGGREGATION · 32× 800G · 25.6 T ToR · 48× 100/400G EDGE · 48× 25/100G
All Anavec-branded. Curated silicon across Standard networking. Same SONiC + AnaROS-native stack across every tier.
NEXT Own-curated SONiC distribution + deeper AAIF integration — switch-port-level verdicts, evidence, audit. Same standard interfaces, same single point of accountability — applied uniformly across future generations land.
STARTER KIT · DAY-1 SKU

The 9U entry-level rack.

The PCIe shelf (Ch 08) and the 32×400G Ethernet switch (Ch 09) ship together as the minimum Anavec rack — four chassis, AnaROS pre-loaded, the smallest path into the full platform.
BILL OF MATERIALS Mgmt switch · Ethernet switch (32×400G) · CPU server · PCIe expansion shelf — wired with 1G OOB, 2×200G + 100G Ethernet, and PCIe Gen5 CDFP. AnaROS pre-loaded; pilot SKU available today. See the Starter Kit →

The future Enterprise AI infrastructure.

AnaRack delivers the physical modularity of hardware combined with the unyielding agility of cloud-native software. The same governed engine runs on your homelab today — bring us a workload, we'll instrument it across the drawers and propose a chassis profile.

Next step · Instrument · Profile · Pilot Request a briefing hello@anavec.ai