The Heterogeneous Blueprint.

The Blueprint.

A structural breakdown of the BetterAI ecosystem. Throughput is a scheduling problem, not a procurement problem — NativeOS pools the accelerators you already own, across generations and across vendors.

Models reason; code decides.
Fig 0.1 // Cluster Execution
Heterogeneous Fleet · 10 racks · NVIDIA B300 + A100 + AMD ROCm
scheduling live
Model
initialising cluster
Scheduler log
  • awaiting placement…
Hot
0
Resident
0
Dispatches
0
GLM-5.3 — Dense · every shard resident, spread wide across A100 tiers
Pooled VRAM
2,048 GiB
Shards
64
Precision
FP8
Nodes engaged
74 / 80
Prior-gen share
78%
B300 / A100 / ROCm
16 · 36 · 22

Layout figures are illustrative of NativeOS scheduling behaviour, not published benchmarks.

L0 // Base OS

NativeOS

NativeOS is an AI appliance operating system for server farms built on the hardware you already have. It pools heterogeneous accelerators — prior-generation cards, mixed vendors, uneven memory — into one addressable substrate, so capability scales with scheduling rather than with procurement. A deterministic controller owns every side effect; the model only proposes.

ROCm (AMD) SupportNVIDIA CUDA IntegrationMixed-Generation FleetsBare-Metal OrchestrationTopology-Aware SchedulingMemory-Tiered Sharding
Heterogeneous Pooling

One scheduling plane across ROCm and CUDA, spanning generations. A workload is placed by measured topology and free VRAM, never by a hardware SKU.

Memory-Tiered Sharding

Weights are partitioned to fit the cards you own. Older accelerators with modest VRAM carry real shards instead of sitting idle.

Sealed Side Effects

Every proposed action is validated before execution. Agents run without ambient credentials; capability grants are scoped and revoked per task.

Explore NativeOS
Fig 1.0 // Layered Architecture
L3 // AGENTIC SHELLDataReactor · ROMAL2 // AUTONOMOUS ENGINEAuto ResearcherL1 // CONTROL PLANEDeterministic ControllerL0 // BASE OSNativeOS · ROCm / CUDA / Bare-Metal
L2 // Autonomous Engine

Auto Researcher

Autonomous research engine for specs, plans, and codebases. A bounded loop of context retrieval → patch → validate → LLM-Critic evaluation → keep-or-rollback, executed inside an isolated git worktree.

Go CLI Controller

Side-effect ownership and loop management.

VCS Isolation

Linked git worktrees for safe execution.

Institutional Memory

Persistent .agent_kb/ recording.

Explore Auto Researcher
Fig 2.0 // Workflow Loop
SYSTEM BOUNDARYCONTEXTPATCHVALIDATELLM-CRITICROLLBACKKEEPDECISION
L3 // Agentic Shell

DataReactor

DataReactor facilitates logic-unit synthesis, transforming raw user inputs and metadata into frozen, executable code agents. When standard processing paths reach ambiguity, the ROMA Escalation Module triggers a priority system override, ensuring complex orchestration is handled with deterministic precision.

Logic-Unit Synthesis

Proprietary synthesis of immutable agentic logic from high-entropy data.

ROMA Escalation

Automated priority escalation for edge-case resolution and system overrides.

Explore DataReactor
Fig 4.2 // System Architecture
USER INPUTSMETADATAREACTOR CORELOGIC UNITSROMA ESCALATIONSTANDARD OUTPUT
Fig 5.0 // Comparative Throughput

Benchmarks

44,000-token prompt · GLM-5.3 · FP8 · Distributed A100 + B300 + AMD ROCm

Mode
Chat — batch size 1 · interactive
+24%tokens/s vs best
−24%time to first token

NativeOS lands 15–40% ahead of the best available alternative, because the layout spreads across a distributed A100 tier instead of queueing on a handful of current-generation cards.

RuntimeGen tokens/s TTFT 
01SGLang
42.1
2,180 ms
02vLLM (clustered)
38.7
2,460 ms
03TensorRT-LLM
47.9
1,870 ms
04NativeOSours
59.4
1,421 ms

Based on client usage and monthly reports. Aggregated across production fleets on a 44,000-token prompt. Figures are medians; per-fleet results vary with topology, quantisation, and accelerator mix.

Fig 6.0 // Cost of Service

What you pay now, versus ours

Annual run-rate for an equivalent serving workload

Scale

Production — 32 current-gen accelerators equivalent · 5,000 h/yr utilisation

Baseline A

On-demand cloud

$1.57M/ year
32 accelerators engaged
On-demand B300 · 32 × 5,000 h$9.80/GPU-hour
$1.57M
32 × current-generation
Baseline B

Owned B300

$494k/ year
32 accelerators engaged
B300 capex · 32 × $42,000amortised over 3 years
$448k
Power + hosting$1,450/GPU/yr
$46k
32 × current-generation
Distributed A100

NativeOS

$367k/ year
26% lower annual cost
Used A100 capex · 86 × $6,500amortised over 3 years
$186k
Power + hostinglower TDP per unit
$77k
NativeOS licence$1,200/accelerator/yr
$103k

Estimates, not a quote. Derived from published on-demand rates and typical used-accelerator pricing, amortised over three years. Rate cards vary by region, commitment, and contract. Every figure is computed from the assumptions in lib/pricing.ts - adjust them to your own contracted rates.

Fig 7.0 // Common Questions

Common questions

Heterogeneous fleets, prior-generation silicon, and deterministic execution

Can I run large language models on older GPUs like the A100?

Yes. NativeOS pools prior-generation accelerators — A100-class NVIDIA cards and AMD ROCm cards — into a single addressable substrate, then shards model weights to fit the VRAM each card actually has. Capability scales with scheduling rather than with buying current-generation silicon.

What is a heterogeneous GPU cluster?

A heterogeneous cluster mixes accelerator generations and vendors in one serving pool — for example NVIDIA B300 alongside A100 and AMD ROCm cards. Most inference runtimes assume uniform hardware and degrade or refuse to schedule across mixed fleets. NativeOS places each shard by measured topology and free VRAM instead of by hardware SKU.

Does BetterAI work with both NVIDIA and AMD accelerators?

Yes. NativeOS provides one scheduling plane across NVIDIA CUDA and AMD ROCm, so a single model layout can span both vendors simultaneously. Workloads are pinned to physical topology rather than to a vendor runtime.

How does NativeOS compare to vLLM, SGLang, and TensorRT-LLM?

On a 44,000-token prompt, NativeOS delivers 15–40% higher generated tokens per second and correspondingly lower time-to-first-token than the best of those runtimes, because the layout spreads across a distributed A100 tier rather than queueing on a handful of current-generation cards. Figures are based on client usage and monthly reports across production fleets.

What is deterministic side-effect control?

Every action an agent proposes is validated before it executes, and agents run without ambient credentials. Capability grants are explicitly scoped and revoked at the end of each task, so a failed run leaves no unauthorised state mutation behind. The model proposes; the controller decides what is kept.

What is ROMA?

ROMA is the Recursive Meta-Agent escalation module inside DataReactor. When a standard processing path reaches ambiguity, ROMA triggers a priority system override, applying dynamic resource allocation, bounded recursion depth, and cryptographic state verification so complex orchestration stays deterministic.

Can BetterAI be deployed air-gapped or on-premise?

Yes. NativeOS is an appliance operating system for server farms and runs bare-metal, air-gapped, or hybrid, with no egress requirement. Licensing is per-cluster rather than per-accelerator.

How much can running on prior-generation GPUs save?

Serving an equivalent workload on a distributed A100 tier typically lands around 25% below the cheaper of on-demand current-generation cloud rates or owned B300 capacity, once capex is amortised over three years and power, hosting, and licensing are included. Actual figures depend on your contracted rates.

L4 // Deployment

Put logic in charge of the loop.

Every proposed action validated before execution. Every decision replayable after the fact.

Models reason; code decides.