Local AI Hardware Investment validation study

Decision pending: RTX 5090 desktop build ($11-15K) or M5 Max 128GB ($9.5K) — Centre Com quote has critical issues

Why we hold it: Local AI hardware opens a path to sell privacy-compliant AI employee solutions to Australian healthcare, legal, and finance firms — a $52B market by 2030. The $12K budget is sufficient for a capable platform, but the Centre Com quote has compatibility issues that must be resolved before purchase. Two viable paths remain: RTX 5090 desktop (best CUDA ecosystem) or M5 Max 128GB (only device that runs 70B models today). Business demand must be validated before committing capital.

⏱ 12d in stage (since 2026-09-18)🔄 last update 2026-09-24

Recommend RTX 5090 desktop build or M5 Max 128GB. Centre Com R9700 quote rejected: CPU-motherboard mismatch, no Windows support, 2.8x slower than RTX 5090.

Draft structure

Problem

No local AI hardware blocks skill development and the ability to prototype sellable AI employee solutions.

Purpose

Build a local AI platform to develop skills and sell privacy-compliant AI agents to Australian professional services firms.

🧭 Get RTX 5090 quotes by Sep 30. Validate business demand by Oct 15. Decision gate Oct 15.Stage validationHours 20Budget $12,000Next gate 2026-10-15Open actions 5

idea›concept›validation›build›market›revenue

Get RTX 5090 quotes by Sep 30. Validate business demand by Oct 15. Decision gate Oct 15.Decision point: Hardware Purchase Decisionin 15d (2026-10-15)

Recommend RTX 5090 desktop build or M5 Max 128GB. Centre Com R9700 quote rejected: CPU-motherboard mismatch, no Windows support, 2.8x slower than RTX 5090.

✅ Proceed when

Business demand validated (2+ potential customers) AND cloud prototype works AND hardware available under $10K

⏸ Park when

Business demand unclear but personal interest remains strong — defer to Q1 2027 for price drops or used market

⛔ Kill when

No business demand found AND cloud APIs serve personal needs AND $12K better deployed elsewhere

💡 Advisor: 2026-09-18: Don't buy the Centre Com build as quoted. Three critical issues: (1) CPU-motherboard mismatch (9960X is TRX50, not WRX90), (2) R9700 has no Windows support — Linux only, (3) R9700 is 2.8x slower than RTX 5090. Instead: get RTX 5090 desktop quotes from AU builders (~$11-15K) or consider M5 Max 128GB (~$9.5K) for immediate 70B capability. Validate business demand before committing capital. Decision gate: Oct 15.

validation

Stage

12d in stage

$12,000

Budget

Personal capital

Hours invested

founder time

Open actions

owner-owed

High/critical risks

of 5 total

Unvalidated assumptions

of 5

🗺️ Engagement Flow — Stages, Gate & Exits

DoneCurrentPlannedGatePark (dashed)Kill (dashed)

⚠️ What could sink it — top risks

RSK-02Business demand not validated — no willing customers foundhigh 15

RSK-01Technology obsolescence — next-gen models require >128GB VRAMhigh 12

RSK-05Cloud API providers undercut local AI value propositionhigh 12

✅ What to do next

ACT-001Contact Centre Com: confirm CPU-motherboard compatibility (9960X is TRX50, not WRX90) and ask for RTX 5090 swap quoteoverdue 5d (2026-09-25)

ACT-002Get RTX 5090 desktop build quotes from PLE, Scorptec, or other AU builders with future GPU expansion pathin 0d (2026-09-30)

ACT-003Validate business demand: interview 3-5 potential customers in healthcare/legal/financein 2d (2026-10-02)

Gate criteria — Hardware Purchase Decision

Business demand validated with 2+ potential customers expressing willingness to pay
Cloud API prototype demonstrates viable agent capability
M5 Max 128GB confirmed available at <$10,000 AUD
Personal commitment: 10+ hrs/week for AI skill development confirmed

🔍 Dig deeper

⚡ Execution 🗺️ Plan & Actions 🧠 Intelligence 💰 Value & Money 🎯 Strategy 🧰 Resources 📜 Charter

📜Charter

Without local hardware, we can't develop skills, prototype sellable AI solutions, or reduce cloud dependency. · Build a local AI platform to sell privacy-compliant AI employee solutions to Australian firms that can't use cloud APIs — healthcare, legal, and finance.

🗣 Problem Statement

No local AI hardware means no hands-on skill development, no platform to demo AI solutions to clients, and ongoing cloud API costs with nothing to show for it.

Pain points (severity)

Can't run large reasoning models locally — limited to small models on current laptop

4/5

No platform to demo AI employee solutions to potential clients

4/5

Cloud API costs accumulate with no ownership or skill development

3/5

Evidence & sources

AI agents market: $52.62B by 2030 (MarketsandMarkets)
Australian Privacy Principles require data sovereignty — local AI is the only option for healthcare/legal/finance
Current laptop (MSI Vector, RTX 5070 8GB) limited to 7B-8B models
70B models need 42.5GB+ VRAM — requires dedicated hardware

🎯 Purpose & Objectives

Build a local AI platform to sell privacy-compliant AI employee solutions to Australian firms that can't use cloud APIs — healthcare, legal, and finance.

Objectives

OBJ-1

Run 70B+ reasoning models locally at >10 tokens/second

Measure: Benchmark at setup

planned

OBJ-2

Develop 5+ local LLM projects within 6 months

Measure: Project log

planned

OBJ-3

Prototype 1+ sellable AI employee solution within 12 months

Measure: Prototype delivery

planned

OBJ-4

Achieve 2+ paying customers within 12 months

Measure: Revenue log

planned

OBJ-5

Break even on hardware investment within 12 months

Measure: Financial tracking

planned

📐 Scope

Hardware acquisition for local LLM inference, skill development, and AI employee business prototyping

✅ In scope

Hardware options analysis and purchase decision
LLM model requirements and VRAM calculations
Cost-benefit analysis vs cloud API alternatives
Australian market pricing and availability
Power, cooling, and space requirements
Skill development pathway
Business model feasibility for AI employee solutions

⛔ Out of scope

Actual hardware procurement and setup
LLM fine-tuning or training infrastructure
Cloud GPU server alternatives (vast.ai, Lambda)
Legal/business registration for selling AI solutions
Detailed software configuration guides

Constraints

$12,000 AUD hard budget ceilingLocated in Australia (hardware availability, 240V power)Single operator — no dedicated IT supportLaptop-only current setup — no existing desktop to leverage

💰Value & Money

A $12K hardware investment enables selling AI employee solutions at $400-600/mo per customer to Australian firms that need data sovereignty — break-even in 3-6 months with 2-3 customers.

💎 Value Proposition

A $12K hardware investment enables selling AI employee solutions at $400-600/mo per customer to Australian firms that need data sovereignty — break-even in 3-6 months with 2-3 customers.

Value drivers

Business revenue

$48K-120K/yr

5-10 customers at $400-600/mo plus $5-15K setup fees

Cloud cost savings

$600-6,000/yr

No per-token costs for development and experimentation

Skill value

Career enabler

Local LLM expertise is rare and high-demand

Hardware resale

~$6,500 at year 2

Apple laptops retain 65-70% value

Revenue model: Consulting/setup ($5-15K per engagement) + monthly license ($400-600/mo per agent) targeting Australian healthcare, legal, and finance firms

⚖️ Cost–Benefit Analysis

Net cost of ~$3K over 2 years after resale, with break-even at 3-6 months if business model validates

Payback3-6 months with 2 paying customers at $400-500/mo

Costs vs benefits (scaled)

M5 Max 128GB BTO cash upfront

9500

Power (8hr/day avg) cash annual

Software (Ollama, llama.cpp) cash annual

Research time (20hrs) time completed

Cloud API cost elimination yr 1

1320

Business revenue (conservative, 2 customers) yr 1

9600

Hardware resale value yr 2

6500

Method & assumptions

Business revenue is unvalidated. Conservative scenario assumes 2 customers; moderate assumes 5-8. Hardware resale based on Apple laptop value retention of 65-70% after 2 years.

🎯Strategy

Centre Com R9700 build rejected — CPU-motherboard mismatch, no Windows support, 2.8x slower than RTX 5090. Two viable paths: (A) RTX 5090 desktop build from AU builder ($11-15K) — best CUDA ecosystem, expandable, but over budget; (B) Apple M5 Max 128GB ($9.5K) — only device that runs 70B today, but macOS and no upgrade. Validate business demand before committing capital.

🧭 High-Level Strategy

📍 Positioning

Local AI solutions for privacy-conscious Australian healthcare, legal, and finance firms — a niche cloud APIs cannot serve due to data sovereignty requirements.

🏰 Moat / defensibility

Local LLM expertise + understanding of Australian privacy regulations + first-mover in local AI services

Strategic options

RTX 5090 desktop build ($11-15K) adopted

Best CUDA ecosystem, expandable to multi-GPU, mainstream support. Over $12K budget but correct long-term platform.

Apple M5 Max 128GB (~$9.5K) watching

Only device that runs 70B Q8_0 today, portable, high resale value. macOS (not Windows), no upgrade path.

MSI + XG Mobile RTX 5090 (~$4,249) watching

Cheapest meaningful upgrade, works with existing laptop, 24GB VRAM. Not expandable, eGPU bandwidth limitations.

Used RTX 3090 desktop (~$3,500) watching

Cheapest entry, validates business model. Limited VRAM (24GB), used hardware risk.

Cloud APIs only watching

Zero upfront cost. No local capability, ongoing costs, no data sovereignty.

🚦 Decision Rules (pre-committed)

✅ Proceed when

Business demand validated (2+ potential customers) AND cloud prototype works AND hardware available under $10K

⏸ Park when

Business demand unclear but personal interest remains strong — defer to Q1 2027 for price drops or used market

⛔ Kill when

No business demand found AND cloud APIs serve personal needs AND $12K better deployed elsewhere

💡 Advisor note

2026-09-18: Don't buy the Centre Com build as quoted. Three critical issues: (1) CPU-motherboard mismatch (9960X is TRX50, not WRX90), (2) R9700 has no Windows support — Linux only, (3) R9700 is 2.8x slower than RTX 5090. Instead: get RTX 5090 desktop quotes from AU builders (~$11-15K) or consider M5 Max 128GB (~$9.5K) for immediate 70B capability. Validate business demand before committing capital. Decision gate: Oct 15.

🗺️Plan & Actualisation7

7 actualisation steps · 1 done · 0 active · next: Build cloud API prototype

🧭 Dependency Map — Steps to Actualisation

Arrows show what each step needs before it can start; the longest left-to-right chain is the critical path.

DoneCurrentPlannedGate

📅 Stages & Time

Sep 2026Oct 2026Nov 2026Dec 2026Jan 2027Feb 2027Mar 2027Apr 2027May 2027Jun 2027

Research & Analysis

2026-09

Business Validation

2026-09 → 2026-10

Hardware Acquisition

2026-10 → 2026-11

Skill Development

2026-11 → 2027-04

Business Launch

2027-01 → 2027-06

🪜 Steps to Actualisation

Click any step to expand its sub-steps, owner and dependencies.

Complete hardware research done

Phase: ResearchOwner: ZCode

Define charter
DC decomposition
Academic deep research

Build cloud API prototype planned

Phase: ValidationOwner: ZaidDepends on: S1

Set up Ollama with cloud API
Build basic AI agent
Test with real use case

Interview potential customers planned

Phase: ValidationOwner: ZaidDepends on: S1

Identify 5 targets in healthcare/legal/finance
Conduct discovery interviews
Assess willingness to pay

Decision gate planned

Phase: ValidationOwner: ZaidDepends on: S2, S3

Review interview results
Review prototype results
Proceed/park/kill decision

Purchase M5 Max 128GB planned

Phase: AcquisitionOwner: ZaidDepends on: S4

Confirm pricing with Apple AU
Order BTO configuration
Set up development environment

Local LLM projects planned

Phase: DevelopmentOwner: ZaidDepends on: S5

Run 70B models locally
Build 5+ projects
Document learnings

Launch AI employee offering planned

Phase: BusinessOwner: ZaidDepends on: S6

Build sellable agent solution
Onboard first customer
Establish recurring revenue

🔁 Development Process

Phased validation: research → business demand validation → hardware purchase → skill development → business launch

Research & Analysis›Business Validation›Hardware Acquisition›Skill Development›Business Launch

Research & Analysis

Hardware requirements, pricing, business model research

Define charter
Comparison matrix
TCO analysis

Business Validation

Test demand with cloud APIs, interview potential customers

Customer interviews
Cloud API prototype
Go/no-go decision

Hardware Acquisition

Purchase and set up M5 Max 128GB

Working local LLM setup
Benchmark results

Skill Development

Build local LLM projects, learn inference optimization

5+ projects completed
Portfolio of work

Business Launch

Build and sell AI employee solutions

First paying customer
Recurring revenue

🧰Resources

Single operator with laptop, $12K capital, 10+ hrs/week commitment

💵 Budget

Total budget: $12,000 · Source: Personal capital

Budget is hard cap. Cloud API prototype costs during validation phase estimated at $50-100.

🧾 Investment Ledger

Cash committed ($)

total committed

Cash spent ($)

to date

Hours invested

founder time

Capital at risk: Up to $12,000 AUD for hardware; primary risk is technology obsolescence and unvalidated business demand

👥 People & Tools

People

Name	Role	Commitment	h/week
Zaid	Owner/Operator	Primary	10
ZCode	AI Advisor	As needed	0

Tools

Tool	Purpose	Cost
Ollama	Local LLM inference	Free
llama.cpp	Low-level inference engine	Free
OpenAI/Anthropic API	Cloud prototype during validation	$50-100

🤝 Stakeholders

Keep informed · high interest / low influence

ZCode supportive

Manage closely · high / high

Zaid champion

Monitor · low / low

none mapped

Keep satisfied · low interest / high influence

none mapped

Full stakeholder register (2)

Name	Role	Interest	Influence	Needs / notes
Zaid	Owner	high	high	Clear recommendation, validation framework
ZCode	AI Advisor	high	med	Research data, decision criteria

🧠Intelligence

5 risks (3 high/critical) · 4 upside opportunities · 5 assumptions · 0 open issues

⚠️ Risk Register

RSK-05

RSK-01

RSK-02

RSK-03

RSK-04

I1I2I3I4I5

L = likelihood, I = impact (1–5). Tint = severity band: low / medium / high (L×I ≥ 12) / critical (≥ 20).

Full risk register (5)

ID	Risk	Category	L/I	Severity	Mitigation	Owner	Status
`RSK-01`	Technology obsolescence — next-gen models require >128GB VRAM	technical	3/4	high	Choose upgradeable platform; Apple retains resale value; cloud APIs as fallback	Zaid	open
`RSK-02`	Business demand not validated — no willing customers found	market	3/5	high	Validate before buying hardware (ACT-001, ACT-003); hardware still valuable for personal skill development	Zaid	open
`RSK-03`	M5 Max 128GB BTO exceeds $10K AUD	financial	2/3	med	Confirm pricing before ordering; consider M5 Pro 48GB as fallback ($5.7K)	Zaid	open
`RSK-04`	Local 8B-24B models insufficient for business agent quality	technical	2/4	med	Test with cloud APIs first; hybrid approach (local + cloud fallback) for complex tasks	Zaid	open
`RSK-05`	Cloud API providers undercut local AI value proposition	market	4/3	high	Focus on data sovereignty niche — cloud APIs cannot address privacy requirements	Zaid	open

🌱 Upside Opportunities

good

Australian privacy-compliant AI market

Healthcare, legal, and finance firms need local AI for data sovereignty under APPs

Upside: high · Effort: med

prime

AI skills as competitive advantage

Local LLM expertise is rare and high-demand; opens consulting and employment opportunities

Upside: high · Effort: low

later

Hardware-as-a-service model

Pre-configured AI workstations for businesses — higher upfront revenue, ongoing support contracts

Upside: med · Effort: high

later

Employee skill development platform

Use hardware to train employees on AI tools, then sell training as a service

Upside: med · Effort: med

🧩 Assumptions

ASM-01

Australian professional services firms will pay $400-600/mo for local AI agents

Validation: Customer interviews (ACT-001)

unvalidated

ASM-02

8B-24B local models are sufficient for business agent tasks

Validation: Cloud API prototype testing (ACT-003)

unvalidated

ASM-03

M5 Max 128GB can run 70B models at >10 tokens/second

Validation: Benchmark after purchase

unvalidated

ASM-04

Apple laptops retain 65-70% value after 2 years in Australia

Validation: Check used M3/M4 Max pricing on eBay AU

unvalidated

ASM-05

10+ hours/week available for AI skill development

Validation: Self-assessment after 2 weeks

unvalidated

🔴 Issue Log (live)

◇ No live issues recorded — add them as they emerge.

⚡Execution

Stage KPIs · next gate: Hardware Purchase Decision (due 2026-10-15) · 5 open owner actions · full progress log.

📊 Stage KPIs

KPI	Current	Target
Budget headroom	TBD	>$2,000 remaining after hardware
70B model speed	TBD	>10 t/s usable
Break-even	3-6 months (2-3 customers)	<12 months
Business validation	Unvalidated	2+ paying customers by month 12
Centre Com quote	Needs clarification	Confirmed compatible build

🚪 Next Gate — Hardware Purchase Decision

Due: in 15d (2026-10-15)

Business demand validated with 2+ potential customers expressing willingness to pay
Cloud API prototype demonstrates viable agent capability
M5 Max 128GB confirmed available at <$10,000 AUD
Personal commitment: 10+ hrs/week for AI skill development confirmed

✅ Next Steps & Actions

ID	Action	Due	Status
`ACT-001`	Contact Centre Com: confirm CPU-motherboard compatibility (9960X is TRX50, not WRX90) and ask for RTX 5090 swap quote	overdue 5d (2026-09-25)	open
`ACT-002`	Get RTX 5090 desktop build quotes from PLE, Scorptec, or other AU builders with future GPU expansion path	in 0d (2026-09-30)	open
`ACT-003`	Validate business demand: interview 3-5 potential customers in healthcare/legal/finance	in 2d (2026-10-02)	open
`ACT-004`	Test cloud API workflow: build a basic AI agent prototype using cloud APIs to validate the concept before buying hardware	2026-10-09	open
`ACT-005`	Decision gate: choose between RTX 5090 desktop build vs M5 Max 128GB vs XG Mobile eGPU	2026-10-15	open

🕘 Progress Log (newest first)

2026-09-24 validation

Full CBA completed 2026-09-24 (workflow dwfrun-8f889beb, verifier 21/21): XG Mobile eGPU NPV-best in all scenarios (+$51.9k S2@6%, +$4.4k stress); used RTX 3090 close second (flips ahead if eGPU resale haircut); M5 Max only option passing 70B gate; RTX 5090 build BREACHES $12k cap ($13.3k); cloud-only worst (-$7.7k stress). CAVEAT: case hinges on avoided-cloud-spend est $1,873/yr and revenue UNVALIDATED (ACT-003 open, due 2026-10-02). Gate 2026-10-15 unchanged.

2026-09-18 validation

Phase 2: Analyzed Centre Com WRX90 workstation quote ($10,960 inc GST). Threadripper 9960X + WRX90E-SAGE SE (7x PCIe 5.0 x16) + R9700 32GB. This is the true multi-GPU platform the Perplexity discussion identified as ideal but assumed was over budget. Two paths: accept R9700 (AMD ROCm risk) or ask Centre Com to quote RTX 5090 swap (~$13-14K). Also reviewed Perplexity discussion confirming MSI Vector laptop (RTX 5070 8GB) and XG Mobile eGPU option.

2026-09-18 validation

Phase 1: Completed Define charter, DC decomposition, academic deep research. Key findings: RTX 4090 discontinued in AU, RTX 5090 at $7-7.7K, M5 Max 128GB BTO at $9.2-9.6K. Business model viable with 2-3 customers at $400-600/mo.

🔗 Links

Define Charter Research Report Full cost-benefit study (2026-09-24): XG Mobile NPV-best; revenue unvalidated

‹ Tax & Household Finance Archive ▦ All 9 Lakes Entrance Apartment (40/203 Esplanade) ›