// AI BLUEPRINT

Build AI you own, sized to the way your organisation works.

This blueprint explains, in plain language, what it takes to run enterprise-grade AI inside your own walls. Our foundational philosophy is that AI should augment human professionals as a decisive force multiplier: absorbing the heavy cognitive burden of data synthesis, code analysis, and document review so your people can focus their expertise where human judgement, ethics, and accountability matter most.

Try the sizing tool Request a briefing →

// 01 // THE AI STACK

Models, interfaces, and agents are not the same thing.

These terms are often used interchangeably, but they are separate layers of the stack. Knowing where each boundary sits is the first step to sizing hardware correctly and enforcing data security.

Layer 1 // Core engine

The model: the brain

The neural network itself, billions of mathematical weights such as our Delta Babylonia series. It has no memory, no interface, and no awareness; it receives text as tokens and predicts the most likely next tokens.

Layer 2 // Experience

The interface: the window

The application wrapped around the model. It handles sign-in, role-based access, chat history, formatting, rate limits, and sanitising prompts before they reach the model.

Layer 3 // Autonomy

The agent: the force multiplier

A system that pairs a reasoning model with tools, memory, and feedback loops to act as an operational force multiplier. Rather than replacing human professionals, an agent acts as a tireless specialist co-pilot: it reads thousands of matter files, runs regression suites, drafts structured findings, and surfaces anomalies so human experts can make informed decisions in a fraction of the time.

Agent Planning, tool orchestration, and human augmentation Sandboxed execution, retrieval, human-in-the-loop validation

Interface Identity, presentation, and access Authentication, RBAC, audit trails, rendering

Model Sovereign neural weights Stateless computation on private hardware

// 02 // TOKENS

What is a token?

A token is the unit an AI model reads and writes. It is usually a whole word or part of one. Almost every figure on this page, from cost to speed to memory, is measured in tokens, so it helps to know roughly what one is.

// HOW A SENTENCE IS SPLIT

Sovereign AI keeps your records onshore.

6 words, 8 tokens. Common words are usually one token; less common words are split into pieces. Each model splits text slightly differently, so treat this as an illustration.

1 token ≈ three quarters of a word

100 tokens is about 75 words of everyday English.

1,000 tokens ≈ 1.5 pages

About 750 words, or a short memo.

128,000 tokens ≈ 200 pages

A policy manual or a contract bundle, read in one go.

1 million tokens ≈ 1,500 pages

The unit most cloud AI services bill by.

Cost

Tokens are what you pay for

Cloud AI services charge per million tokens, counting both what you send and what comes back. A long document read many times a day adds up quickly. A private deployment has no per-token charge; you pay for the hardware and power instead.

Speed

Speed is tokens per second

People read at roughly 5 to 8 tokens per second. A model writing at 25 to 60+ tokens per second acts as an immediate cognitive multiplier: it reads, cross-references, and drafts faster than any human can, leaving your professionals to focus on high-value review and judgement.

Memory

Context is measured in tokens

The context window is how many tokens a model can hold at once: your question, the documents it is reading, and its answer. Every token held uses graphics memory, which is why long documents need more capable hardware.

Estimate the tokens in your own text

Calculated in your browser. Nothing you paste is sent anywhere.

Words: 0
Estimated tokens: 0
Pages: 0
Share of a 32k window: 0%
Share of a 128k window: 0%

// 03 // WORKLOAD PROFILES

Different workloads stress different hardware.

Spending well starts with matching your main objective to its real bottleneck. Chat, document search, and fine-tuning each push on a different part of the system.

Tier 1 // Moderate

Chat and everyday coding

Dialogue, drafting, syntax review, and summaries across small context windows of 4k to 16k tokens.

Bottleneck: Memory bandwidth
Why: Generation speed depends on how fast weights move from memory to the chip.
Suits: Unified memory desktops or a single 16GB to 24GB GPU.

Tier 2 // High

Heavy coding and document search

Whole repositories, 500-page manuals, and discovery sets across 64k to 128k+ token contexts.

Bottleneck: VRAM capacity and storage speed
Why: Holding 128k tokens of conversation memory (the KV cache) can need 16GB to 32GB on its own.
Suits: One or two 96GB professional GPUs, or 128GB+ unified memory, with fast NVMe storage.

Tier 3 // Extreme

Fine-tuning and training

Adapting open-weight models to proprietary legal, medical, or defence data using LoRA or QLoRA.

Bottleneck: Tensor compute and GPU interconnects
Why: Training calculates and applies gradients across GPUs millions of times.
Suits: Professional workstation or server GPUs, with power and cooling planned for long runs.

// 04 // THE DEPARTMENT THRESHOLD

Around 200GB of VRAM, the question changes.

Reaching roughly 200GB of usable VRAM moves you out of consumer computing and into workstation and server architecture. Below it, the question is whether a model fits. Above it, the questions become how fast frontier-class models can run, and whether you can fine-tune them privately. For organisation-wide use, the conversation moves again, to multi-node platforms in a data centre.

Capability 01

Frontier-scale inference

Run 400B-class reference builds such as Delta Babylonia Omni 400B, Gov 400B, Legal 400B, or Code 400B at about 3-bit quantisation in roughly 170GB to 190GB. At 4-bit they need about 230GB, which suits a four-GPU server.

Capability 02

Lossless 70B to 90B models

Run Delta Babylonia Omni 70B and Delta Babylonia Code 70B unquantised at FP16 or BF16, with long context windows held in memory.

Capability 03

Private fine-tuning

Enough headroom to fine-tune 70B models on your own contracts, discovery material, and code without any of it leaving your network.

// 05 // HARDWARE PATHWAYS

Three ways to reach 200GB, each with trade-offs.

Cost, speed, power, noise, and setup effort pull in different directions. Here is how the three main routes compare at the 192GB level, followed by the full range from a simple upgrade to a cluster.

Option A

Unified memory desktop

System memory doubles as graphics memory, so one compact, quiet desktop can hold very large models.

Spec: Mac Studio, Ultra-class chip, 256GB to 512GB unified memory
Usable: About three quarters of memory for the model
Power: ~150W to 270W under load; any standard power point

Verdict: the simplest way to run a very large model at a desk, but slow (about 2 to 4 tok/s on a 400B model) and not suited to fine-tuning or many users at once.

Option B

Multi-GPU consumer rig

Six 32GB consumer graphics cards combined to reach roughly 192GB.

Spec: 6x RTX 5090 class GPUs on a workstation or server board
Needs: Several 1,600W+ power supplies and custom cooling
Power: ~3,500W to 4,000W; more than one circuit can supply

Verdict: not the cheaper option it appears. Six consumer cards cost about as much as two professional cards, draw around three times the power, and need electrical work and a dedicated room.

Recommended Option C

Professional workstation

Two 96GB professional GPUs with a workstation-class processor and error-correcting memory.

Spec: 2x NVIDIA RTX PRO 6000 Blackwell (96GB each, 192GB total)
System: AMD Threadripper or EPYC, or Intel Xeon; 256GB to 512GB ECC memory
Power: ~1,150W with 300W cards, which suits a standard power point on its own circuit; ~1,900W with 600W cards, which needs a dedicated circuit

Verdict: the best balance for most organisations: about 30 to 40 tok/s on a 70B model at 8-bit, room for 400B-class models at 3 to 4-bit, full fine-tuning support, and enterprise warranties.

Hardware pathway comparison
Metric	Unified memory desktop	Professional workstation (2x 96GB)	Consumer rig (6x 32GB)
Indicative hardware	~$10k to $20k	~$62k to $80k	~$70k to $110k
Power under load	~150W to 270W	~1,150W to 1,900W	~3,500W to 4,000W
Plugs into	Any standard power point	A standard power point on its own circuit, or a dedicated circuit for 600W cards	Two or more dedicated circuits; electrician required
Noise and form factor	Compact, near silent	Large tower or 4U rack, audible under load	Open frame, very loud
70B model at 8-bit	7 to 10 tok/s	30 to 40 tok/s	20 to 35 tok/s
400B model, quantised	2 to 4 tok/s	10 to 15 tok/s	8 to 15 tok/s
Fine-tuning	Limited	Full native support	Possible, less stable
Support	Manufacturer cover	OEM workstation warranty	Component warranties only

From an upgrade to an enterprise platform

Not every organisation needs 192GB. Many start with a single upgraded desktop and grow from there, while the largest organisations run multi-node platforms in a data centre. Indicative Australian hardware prices, including GST:

Indicative hardware tiers
Setup	Graphics memory	Runs well	Indicative hardware
Upgrade an existing desktop	16GB to 32GB	8B models; 25B to 32B models at 4-bit for one or two people	$2,000 to $10,000
New workstation, one 32GB GPU	32GB	25B to 32B models for a small team	$12,000 to $16,000
Unified memory desktop	256GB to 512GB shared	Very large models for one or two people, slowly	$10,000 to $20,000
Workstation, two 32GB GPUs	64GB	70B at 4-bit, quickly; 32B with long documents	$28,000 to $40,000
Workstation, one 96GB professional GPU	96GB	70B at 8-bit for a team	$35,000 to $45,000
Workstation, two 96GB professional GPUs	192GB	70B at full precision; 400B-class at about 3-bit for one or two people; fine-tuning	$62,000 to $80,000
Rack server, four 96GB professional GPUs	384GB	400B-class at 4-bit for a department; many people at once	$120,000 to $160,000
Two four-GPU servers	768GB	Division-wide use with headroom	$250,000 to $350,000
Enterprise AI platform: two or more 8-GPU data-centre nodes	1.4TB+ per node	Bespoke 1T+ class models with million-token context for thousands of people, across sites	From about $1.4M, or reserved sovereign capacity

Hardware figures are indicative, include GST, and are based on Australian retail pricing at the time of writing. They exclude advisory, implementation, and build services, which we scope separately. Enterprise platform figures cover compute nodes only; networking, storage, and facility costs are scoped in discovery. GPU prices move with supply, and upgrading equipment you already own usually costs less than a new build.

Not sure which pathway fits? We size hardware to your people, workloads, and power, from one upgraded desktop to a multi-site enterprise platform.

Try the sizing tool Request a briefing

// 06 // POWER, HEAT, AND NOISE

Will it run from a normal power point?

Most AI hardware plugs straight into the wall. The larger builds do not, and it is far easier to plan for that before anything is ordered. Here is how the numbers relate to an ordinary Australian office or home.

2,300W one standard 10A power point

Australian power points supply 10 amps at 230 volts: about 2,300 watts at most.

~1,800W comfortable all-day load

For equipment that runs for hours, plan on about 80% of the maximum.

1 circuit is often shared

Several power points usually share one circuit breaker. A heater or kettle on the same circuit counts too.

3,450W one 15A power point

Recognisable by its wider earth pin. Needs its own circuit, installed by a licensed electrician.

Power needs by hardware setup
Setup	Draw under load	Share of one 10A power point	What it needs
Unified memory desktop	150W to 270W	~12%	Any power point. Sits on a desk and is near silent.
Desktop with one 16GB GPU	350W to 550W	~24%	Any power point.
Workstation, one 32GB GPU	700W to 900W	~39%	Any power point. Audible when busy.
Workstation, one or two 96GB professional GPUs (300W cards)	650W to 1,200W	~52%	A standard power point, ideally on a circuit with little else on it.
Workstation, two 32GB GPUs or two 600W professional GPUs	1,500W to 1,900W	~83%	At the all-day limit of a 10A point. Plan for a dedicated circuit or a 15A outlet, and a room with airflow.
Rack server, four 96GB GPUs	2,400W to 3,200W	~140%	Too much for one power point. A dedicated 15A, 16A, or 32A circuit in a comms or server room with cooling.
Two four-GPU servers	5,000W+	220%+	A server room with dedicated circuits, often three-phase, or space in an Australian data centre.
Enterprise AI platform, 8-GPU data-centre nodes	About 14,000W per node	~600% each	Data centre power and cooling, usually high-density air or liquid. Hosted in your data centre or an Australian sovereign facility.

// HEAT, NOISE, AND BACKUP POWER

Every watt a computer draws ends up as heat. A 1,500W workstation running all day warms a room about as much as a fan heater, so a small office needs good airflow or air conditioning. Professional cards are generally quieter at a desk than consumer cards, but anything above about 1,000W is noticeable. A battery backup (UPS) should be sized to the load: around 1,500VA for a 1,000W workstation, and 2,200VA to 3,000VA for 1,500W to 2,000W. During discovery we check your switchboard, circuits, and cooling, and arrange a licensed electrician where needed.

// 07 // MODEL CLASSES

Match the hardware to the model class.

The right hardware depends on the size of model you deploy. Each Delta Babylonia family is shown in more than one reference size, so the same capability can run on a workstation or a rack. For large organisations we design bespoke enterprise models well beyond these sizes.

7B to 8B

Gov 8B, Legal 8B, Icon 8B

Quick answers, triage, and drafting for a single team on modest hardware.

Hardware: Mid-tier GPU or 16GB unified memory
VRAM: 8GB to 16GB
Speed: 40 to 80+ tok/s

25B to 32B

Omni 25B, Core 32B, Code 30B

Longer documents, structured extraction, and multi-step reasoning for a department.

Hardware: One 32GB GPU or 64GB unified memory
VRAM: 24GB to 36GB
Speed: 18 to 32 tok/s

70B to 90B

Omni 70B, Gov 70B, Legal 70B, Code 70B

Detailed analysis, policy review, and whole-codebase work with long context.

Hardware: Two 32GB GPUs, one or two 96GB GPUs, or 128GB+ unified memory
VRAM: 48GB to 96GB
Speed: 7 to 40 tok/s, depending on hardware

400B+

Omni 400B, Gov 400B, Legal 400B, Code 400B

Our largest reference build, for the hardest reasoning and multi-agent work, with no cloud exposure.

Hardware: A four-GPU server, two 96GB GPUs at about 3-bit, or 256GB+ unified memory
VRAM: About 190GB at 3-bit, 230GB+ at 4-bit
Speed: 2 to 25 tok/s, depending on hardware

1T+ // BESPOKE

Enterprise builds for large organisations

For banks, government departments, and national enterprises we design a bespoke model on a trillion-parameter-class foundation, trained further on your own material and run across a resilient multi-node platform.

How enterprise builds work →

Hardware: Two or more 8-GPU data-centre nodes, across sites
GPU memory: 1.4TB+ per node
Context: 1M+ tokens, with private search
Speed: 25 to 45 tok/s per user, thousands at once

// 08 // OPTIMISATION

Getting more intelligence per gigabyte.

Hardware is an investment. These techniques let large models run on realistic budgets without giving up meaningful quality.

Compression

Quantisation

Storing weights at 8-bit or 4-bit precision instead of 16-bit cuts memory needs by 50% to 75%. A 70B model drops from about 140GB to about 40GB, typically with only a 1% to 2% change on reasoning benchmarks.

Memory

KV cache management

In long sessions, the conversation memory can outgrow the model itself. Paged attention allocates that memory like an operating system does, and flash attention computes it in fast on-chip memory, cutting waste and speeding up long documents.

Context

Context window scaling

Rotary position scaling stretches native windows to 128k tokens and beyond while keeping retrieval sharp, and the long-context architectures used in enterprise builds reach a million tokens, so a model can read a full discovery set or repository in one pass.

// 09 // SOFTWARE STACK

The software that makes the hardware useful.

Production AI needs an orchestration layer that serves many users at once, enforces limits, and routes document lookups safely.

Serving

Inference engines

High-throughput servers that batch requests, manage memory, and expose industry-standard APIs, from enterprise-grade engines such as vLLM to lightweight runtimes for single workstations.

Orchestration

Agent frameworks

Libraries that connect reasoning models to internal databases, tools, and APIs, with schema validation so outputs arrive as clean, predictable data and reflection loops that check results before they are returned.

Retrieval

Private knowledge bases

Vector databases on local NVMe act as long-term memory for your documents. Hybrid search blends meaning and keywords for precision, and offline reference libraries provide verified background facts without any web access.

// 10 // TOTAL COST OF OWNERSHIP

Buy the hardware, or rent the API?

Every executive evaluates this trade-off. Cloud APIs offer zero friction to test an idea, but once work transitions into everyday team operations and multi-step agent loops, per-token billing compounds rapidly. Here is how the two operating models compare across key dimensions.

Executive comparison between sovereign private infrastructure and commercial cloud APIs
Dimension	OWNED INFRASTRUCTURE Sovereign Private Compute Dedicated on-premise hardware or private cloud	RENTED SERVICES Commercial Cloud APIs Public multi-tenant endpoints and per-token billing
Upfront Capital Initial setup & hardware	$2,000 to $80,000+ (CapEx) Upgrades from ~$2,000; dedicated 192GB workstations $62,000 to $80,000; 4-GPU server nodes $120,000+ (indicative, inc GST). Equipment is a depreciable institutional asset.	$0 initial capital Zero hardware procurement or installation costs. Teams start immediately with an API key and monthly credit card billing.
Operating Profile Monthly run costs	Predictable fixed OpEx Electricity, cooling, and scheduled maintenance. Processing 10 queries or 10,000,000 tokens carries virtually identical marginal operating run costs.	Metered variable billing Billed per million tokens across input and output. Invoices climb each month as staff adoption, context windows, and automations expand.
Data Custody Privacy & compliance	100% sovereign & air-gapped Prompts, documents, and telemetry never leave your premises. Aligns with Australian Privacy Principle 8, APRA CPS 230, and SOCI controls.	Shared multi-tenant cloud Prompts transit public infrastructure and provider logs. Subject to foreign cloud provider terms and cross-border regulatory scrutiny.
Scaling Economics Unit cost at high volume	Cost per token trends to zero Higher utilisation drives effective unit costs down to ~$0.30 per million tokens. Zero financial friction for running continuous multi-step agent loops.	Costs compound continuously Multi-step autonomous agent loops, retrieval lookups, and wide context windows multiply token consumption, creating unpredictable cost spikes.
Strategic Fit Recommended deployment	Production, sensitive IP & 24/7 agents Daily team workflows, confidential legal matter files, financial models, government and defence holdings, and always-on autonomous agents.	Exploratory prototyping & light tasks Proof-of-concept testing, occasional ad-hoc research on non-sensitive data, and early-stage experimentation.

// THE AUTOMATION INFLECTION POINT

When autonomous agents act as a force multiplier, per-token billing breaks down.

In manual chat interfaces, human reading speeds keep token consumption modest. Once you deploy autonomous agents to augment your workforce, generating test suites, refactoring legacy codebases, triaging inbound procurement tenders, or auditing compliance registers, token volume can surge from thousands to tens of millions per day.

At that scale, per-token cloud invoices force organisations into artificial rationing, blunting the productivity gains of AI. Dedicated private hardware removes that ceiling. At sustained high volumes it can pay for itself within a year, freeing your teams to automate with confidence while keeping proprietary intelligence inside your perimeter.

Model your organisation's numbers in the Token Economics Calculator →

10M+

Daily Token Inflection

Around this daily volume, dedicated hardware typically pays back within two to three years. At 50M+ tokens a day it is often within a year.

~90%

Effective Unit Cost Savings

In one industry case study, unit costs fell from leading cloud rates to about $0.30/M tokens on dedicated hardware at high utilisation.

Zero

Token Rationing

Teams run massive context windows and multi-step agent loops without budget anxiety.

// 11 // GOVERNANCE

Security and compliance, built in from the start.

For regulated organisations, governance often decides the architecture. Running verified weights on your own infrastructure removes many of the risks that come with shared cloud services.

Human agency

Amplification, not substitution

AI is an amplifier for human expertise, not a substitute for institutional judgement. Every workflow maintains clear human checkpoints where professionals review, adjust, and approve outputs before decisions take effect.

Isolation

Air-gapped by design

Models run disconnected from the internet, with telemetry disabled and no dependency on a third-party provider. Sensitive IP, records, and designs never cross an external network.

Australian regulation

Privacy, CPS 230, and SOCI

Keeping processing onshore supports Australian Privacy Principle 8 on cross-border disclosure, removes offshore single points of failure under APRA CPS 230, and helps protect critical infrastructure under the SOCI Act.

Global frameworks

EU AI Act, GDPR, and ISO

Deterministic logging, meaningful human oversight, and in-flight anonymisation of personal information align with high-risk obligations, and map cleanly to ASD Essential Eight, NIST, and ISO 27001 and 42001 controls.

// 12 // MEASURING SUCCESS

How we know it is working.

Production AI needs clear service levels and continuous automated testing, so quality is measured rather than assumed.

Responsiveness

Time to first token

How long a user waits before the answer starts to appear.

Chat< 450 ms

Document search< 1,200 ms

Throughput

Tokens per second

How quickly the answer is written once it starts.

Interactive chat25 to 35 tok/s

Automated pipelines50 to 80+ tok/s

Accuracy

Automated evaluation

Golden test suites score faithfulness and catch hallucinations whenever weights change.

Faithfulness target> 98.5%

Regression gateEvery release

// 13 // SIZING TOOL

Find your starting configuration.

Choose your main workload, model class, and where the hardware will live to see an indicative setup, its power needs, and its price range. Every engagement is then tailored to your team size, usage, and existing equipment.

Main workload

Model class

Space and power

Indicative only. Prices include GST and exclude advisory, implementation, and build services. They vary with existing hardware, upgrade versus new build, and supply.

// Recommended topology Calculated

Recommended setup Workstation with two 32GB GPUs

Graphics memory needed About 40GB to 48GB This setup provides 64GB

System memory 128GB to 256GB ECC At least equal to graphics memory

Speed per user 25 to 40 tok/s 4-bit

Indicative hardware $28,000 to $40,000 Inc GST, hardware only

Power under load 1,500W to 1,900W At the all-day limit of a 10A point How power works →

// NEXT STEP

Let us design the right stack for your organisation.

We benchmark your workloads, size hardware to your people and usage, from one team to a whole enterprise, and deploy audited models with governance built in.

Request a briefing