Ten things to settle together
All three pilot units are ready this week. This page takes the list Simon wrote after the conference, puts the research under each item, and leaves room to write the decision down.
All three pilot units are ready this week. This page takes the list Simon wrote after the conference, puts the research under each item, and leaves room to write the decision down.
Sectors shape the model. The model shapes the price, the price shapes funding and the waitlist, and all of it feeds the timeline, the team, sales and the raise. Brand sits across the top.
DTA AI Government Showcase (11 Nov) tells SMEs whether their exhibitor applications were accepted. Check whether we applied.
Queensland Gambling Community Benefit Fund round closes (up to A$35K for NFP equipment). Rounds recur several times a year.
End of Q3, the original target for closing the raise. NACCHO conference early-bird rate ends.
ACFID national conference, Canberra. The next room full of development buyers after Situation 2026.
Decision boxes on this page save in this browser only. Use "Copy all decisions" at the bottom to paste them somewhere permanent.
| Audience | Size | What's pulling now | Main catch | First ticket |
|---|---|---|---|---|
| Not-for-profits | ~2,500 in Australia | 50% name privacy or sovereignty as the main AI barrier; OVIC block ends 5 Nov | Thin, tied money; cloud AI ~A$12 a seat | A$12K–40K |
| Development | A$5.2b aid; A$2.2b Pacific | DFAT has no AI position; Fiji digital package; PNG sovereign AI framework | 6–18 month cycles, probity, free substitutes | A$40K scoping, A$150K–280K pilot |
| Government | ~900 bodies | Direct buys under A$125K from Australian SMEs; DTA registers; ADM 10 Dec | IRAP, Qwen provenance, long cycles | A$25K–45K pilot |
| SMEs | ~3,700 mid-sized | TPB client-consent rule; privilege; shadow AI | Crowded, ~A$30 seat anchor, brand drift | A$6K–20K setup |
| Enterprise | ~5,400 large | APRA AI letter; nothing local between Copilot and AI factories | Certifications, vendor-risk reviews | A$40K–70K per department |
A practical way to frame it: which one or two audiences get active selling effort for the next six months, which get a waitlist and a polite "not yet", and which get nothing? With a small team, "all five" means none get done well.
| Inclusion | Cadence (a starting proposal) | What it gives the client |
|---|---|---|
| Public data corpus downloads | Monthly | New grants for their sector, newly published research, legislation and regulator guidance, loaded into their Selma |
| Skills updates | Monthly | New use cases and personas (e.g. grant eligibility checker, funder report writer, audit prep) built by the Use Case Maker and tested before release |
| Model upgrades | Quarterly, or when a better open model is proven | Better answers without new hardware, with a documented model list (including non-Chinese options) |
| Obligation packs | Quarterly | Dated legislation and regulator guidance for their sector |
| Re-measurement | Annually | Fresh accuracy figures against their own documents, for their board and auditors |
| Support and health monitoring | Ongoing | Help desk, remote health checks where the configuration allows, patching |
| Hardware swap and refresh | On failure; refresh at year 3–4 | No dead box. The refresh is the natural renewal moment. |
| Capital bundle | Why that number |
|---|---|
| Over A$10K | Clears the new NSW Treasury capitalisation threshold (A$10K, Sept 2026) and most council thresholds, so it sits in capital budgets |
| Up to A$20K ex GST | Fits the Stronger Communities Programme cap (A$20K; "ICT hardware" named as eligible in Round 9) and the small business instant asset write-off (A$20K, permanent from July 2026) |
| A$30K–35K | Fits the Queensland Gambling Community Benefit Fund cap (A$35K) and half of a A$60K–75K matched equipment project |
Two tiers could cover most buyers: a "Community" bundle just under A$20K and a "Standard" bundle around A$30K–35K, each with an annual licence. The pricing calculator tests any combination against cost and Copilot.
The likely reason he had nothing to spend it on: under accounting rules (the 2021 IFRIC decision, and Finance's guidance RMG 109 for Commonwealth bodies), subscription AI like Copilot or ChatGPT has to be expensed. Organisations holding capital budgets or matched capital grants have had no AI they're allowed to buy with them. Selma is hardware the buyer owns, running on their own infrastructure, so it can be capitalised. Our reading
We don't know which scheme he meant. Worth asking him directly: "Was it a grant you applied for, a council or department capital budget, or the tax write-off?"
| Scheme | Buyer | Match | Max | Status (22 Sep) |
|---|---|---|---|---|
| Queensland Business Growth Fund | Qld SMEs, specialised equipment | Business pays at least 50% | A$75K | Round 7 closed; Round 8 not announced |
| Council or agency capital budgets | Councils, agencies, libraries, TAFEs | Internal; several grant programs require councils to match 1:1 | Budget | Annual; unspent capital peaks April to June |
| Instant asset write-off | Businesses under A$10M turnover | ~25% effective via tax | A$20K per asset | Permanent from 1 July 2026 |
| Scheme | For | Up to | Status | Fit |
|---|---|---|---|---|
| Qld Gambling Community Benefit Fund | Qld NFPs, equipment | A$35K | Round closes 28 Sep; recurring | High |
| Stronger Communities Programme R10 | NFPs (100%), councils (50%) | A$20K | A$22.5M budgeted; dates and guidelines TBA (Round 9 named ICT hardware as eligible). MPs nominate, so talk to local members early. | High if under A$20K |
| NSW Community Building Partnership | NSW NFPs and councils | A$100K | 2026 closed; next round 2027; IT eligibility to confirm | Medium |
| Lotterywest, SA Expand Together | WA NFPs; SA multicultural orgs | Varies / A$100K | Varies / next round | Medium |
| NIAA Indigenous Advancement Strategy | ACCOs, community-led proposals | Negotiated | Open by approach via regional offices | Medium |
| Fiji digital package (Vuvale Union) | Fiji, via DFAT's contractor | FJ$146.3M total | Announced 30 Aug 2026; probably the money behind the Fiji hub Inference | High, via procurement |
| REnew Pacific | Off-grid energy, Pacific and Timor-Leste | A$75M program | No open call | Watch, for solar Selma |
43.5% refundable offset on eligible R&D. Accuracy measurement, Guardrail and retrieval evaluation may qualify. Needs records kept as the work happens.
1:1 matching, up to A$80K a year for new markets, for Pacific and Timor-Leste marketing and travel.
A$20K–60K matched, open now. Independent validation of accuracy or off-grid energy use.
Industry Growth Program (paused May 2026), DFAT Business Partnerships Platform (closed), AI Adopt (closed). EMIIF invests through funds only.
A cheap, useful asset: a one-page "Selma for grant applications" sheet with an itemised quote (hardware, commissioning, licence split), an asset description ("on-premises AI server, 4-year effective life"), an ownership statement and community-benefit language. Most community grants reject items without a supplier quote.
One thing it fixes: our own Selma language currently says never to describe enquirers as "waiting", because an enquiry isn't a waiting list. A real list, with levels and definitions, makes the claim honest and gives investors a number they can trust.
| Level | What they do | What they get | How we report it |
|---|---|---|---|
| 1. Registered | Fill in a 10-question qualifying form | Updates and briefing invitations | Always shown beside the qualified number, never alone |
| 2. Qualified | Discovery call held; sensitive-data use case, budget path and decision-maker confirmed | Readiness assessment, draft Charter conversation | With a one-line definition in the deck |
| 3. Reserved | Sign a short letter of intent, or pay a refundable deposit | A place in a dated install batch, founding price locked for 12 months | "Z reservations totalling A$X". The number investors weight most. |
Tools: Tally or Typeform for the form, a free CRM (e.g. HubSpot) with a "level" field, Stripe payment links or an invoice for deposits. The QR on the pull-up banners (selma.au/conference) can point straight at it. Suggestions from the research, not legal advice.
One page: upfront, licence, what's included, founding discount and expiry, "quotes valid 90 days" so NFPs can put it in grant budgets.
Paper and QR, the same fields as the online waitlist form.
A 90-day pilot fee credited in full to year one. Objectives, milestones, baseline, responsibilities, and conversion terms. Unpaid pilots tend to stall.
Non-binding except confidentiality: use case, deployment, users, indicative price and budget source, start window.
Capability statement, insurances, security summary, how to buy (direct, quote or panel).
Pull the ethernet cable and it still answers. The most persuasive thing a stand can show.
| When | Event | Where | Why | Mode |
|---|---|---|---|---|
| 12–14 Oct | ACFID National Conference | Canberra | Development NGOs and DFAT; follows straight on from Situation 2026 | Attend, pre-book meetings |
| Oct (TBC) | GovAI26 | Canberra | AI in the APS; showcase free, partner slots open | Showcase or partner |
| 27–28 Oct | AI Leadership Summit (CEDA, National AI Centre) | Sydney | Senior business and government AI buyers; A$990 | Attend |
| 11 Nov | DTA AI Government Showcase | Canberra | 600+ APS decision-makers; SME exhibitor spots (applications closed 11 Sep) | Exhibit if accepted |
| 12–13 Nov | Fiji Tech Summit | Natadola, Fiji | The Pacific's largest IT event; hub-and-spoke story | Attend or sponsor |
| 2–4 Dec | NACCHO Members' Conference + Australasian Aid Conference | Canberra, same week | 1,000+ ACCHO delegates; aid researchers and DFAT. One trip, two rooms. | Attend, with care on data sovereignty |
| 9–11 Mar | Gov Data Summit | Canberra | 800+ public sector data and AI leaders | Exhibit |
| Mid-Mar | Infoxchange NFP Technology Awards (entries) | Online | Enter with a pioneer partner | Award entry |
| 24–25 Mar | Impact Investment Summit Asia Pacific | Sydney | Investors, foundations, family offices | Investor |
| May (TBC) | Infoxchange Tech for Social Justice | Melbourne | The NFP tech buyer event | Exhibit |
| June (TBC) | NDS state conferences | Several cities | Disability provider CEOs, pioneer-partner territory | Exhibit |
Also worth a look: LGNSW (22–24 Nov), Legal Innovation and Tech Fest (Apr), Accounting Business Expo (Mar), PITA forum in Fiji (Apr). SXSW Sydney is cancelled for 2026 and 2027, and Pitch@Palace looks inactive. Full calendar of ~35 events in the research notes.
The research surfaced a technical to-do list that sits squarely in "validate what's been built" and doesn't need Simon. The July investor pack named two hires in the use of funds (a developer and a project manager), so this is also the story the raise pays for.
| Workstream | Why it matters commercially | Unlocks |
|---|---|---|
| Independent code and security review; hardening guide; penetration test | Every government and enterprise buyer will ask. A pen test is ~A$15K–40K. | Government, enterprise |
| Model options and a published model list (Llama, Mistral, Gemma, Granite alongside Qwen) | Removes the Chinese-origin question for government, DFAT and banks | Government, development, enterprise |
| Re-quote the build at current memory prices; spec a panel-listed OEM option (Dell or Lenovo) | The A$3K cost may have drifted; OEM hardware answers supply-chain questions | Pricing, government |
| Productise commissioning: templated Constitutions per sector, standard connectors, remote measurement | 5–10 days per install limits scale; SMEs need 2–3 | Margin, SMEs, channels |
| Licence engine: monthly data and skills pack pipeline, update signing and delivery | Makes the annual fee real | Business model |
| Measurement: a metering plug on a unit for a fortnight; a low-power (Jetson) port and one-panel solar test | Turns estimates into claims DFAT can rely on | Development, offline |
| R&D records kept as work happens | Needed for the 43.5% R&D Tax Incentive refund | Cash |
| Security documentation pack: questionnaire answers, data flow, exit plan, CPS 230 schedule | Cuts weeks from every procurement | Government, enterprise |
Built from the dates already on the calendar. The earlier Selma plan said "product complete Q3 2026, go to market Q4 2026". This keeps that shape.
Pioneers running on premises with signed or ready-to-sign Charters.
Sectors, price, licence, waitlist, brand decided. Ask the capex contact which scheme he meant.
Founding price sheet, reservation form, paid pilot offer, letter of intent, grant-ready quote. In the bag for ACFID.
Founding partner rate offered to the three pioneers. First investor conversations with the new price in hand.
A reason to talk to Victorian community service organisations.
The single thing investors will weight most. DTA showcase and Fiji Tech Summit in the same month.
NACCHO and the Australasian Aid Conference. Qld grant results announced.
Before the summer slowdown.
First paying installs from the waitlist. Gov Data Summit and Impact Investment Summit in March.
Councils and agencies spending unspent capital before 30 June. The capex pitch has its best window.
| Tier | Who | Why |
|---|---|---|
| Anchors | The impact investor friend (A$75K–100K); 3–6 impact-minded angels from their network; Sydney Angels (six cycles a year); Scale Investors if the founding team meets their gender-equity criteria | Mission-aligned, realistic on valuation, fastest to commit |
| Stretch | Giant Leap; Tripple; Australian Impact Investments (gateway to foundation money); Rampersand (vertical AI thesis) | Bigger cheques, credibility, possible lead |
| For seed, later | Main Sequence, AirTree, Folklore, Investible, Square Peg | Send updates now; pitch with revenue |
| Non-dilutive | R&D Tax Incentive; NVIDIA Inception (hardware pricing, free); CSIRO Kick-Start; Startmate (A$120K, best after A$250K is committed so it comes in on our terms) | Runway and help without dilution |
| Development money | DFAT, ADB, UNCDF | Realistically customers, not investors. EMIIF invests through funds; the Business Partnerships Platform has closed. |
Anchor the friend, 20–30 conversations, first close mid-December. Fastest, but at a lower cap with free pilots only.
Convert pioneers and publish pricing Oct–Dec, small bridge from the friend if needed, raise Feb–Apr with revenue at a better cap.
Close A$250K from angels, then Startmate on the same terms. Programme time is the cost.
Raise off a funded development contract. Strongest proof, slowest and least certain.
This answers the biggest objection on the enterprise page: that a small model on a 16GB card gets compared with Copilot and loses. Open-weight models are now within about 4–6 points of the best cloud model on the Artificial Analysis Intelligence Index. A box that can hold one turns "small but safe" into "near-frontier and safe". Quick scan, ~20 sources
| Hardware route | Memory | Approx A$ | Note |
|---|---|---|---|
| 4x RTX Pro 6000 server | 384GB | 140K–165K hw | Best value for a 400B-class model |
| 8x RTX Pro 6000 server | 768GB | 260K–300K hw | Fits the largest open models; needs A$450K+ price |
| NVIDIA DGX Station GB300 | 252GB fast + 496GB slower | 133K–165K | Office-friendly tower, but only 252GB is fast GPU memory |
| Mac Studio M5 Ultra 512GB (from late Oct) | 512GB unified | ~16K–30K each | Quiet and cheap; weak with many users at once |
| 8x H200 / B200 server | 1.1–1.4TB | 450K–760K | Overkill at this price |
Model names and versions move fast and sources disagree; check before quoting. NVIDIA's AI Enterprise licence (~A$40K a year for four GPUs) is optional, since open-source serving software works without it. Full notes: research/12.
The model Simon has in mind: InternLM Atria-Dawn-Preview. An agentic model from the Shanghai Artificial Intelligence Laboratory built on GLM-5.2: 744B parameters, mixture-of-experts, 256K context, text only, MIT licence, with an official FP8 version. Strong on deep research (96.0 DeepSearchQA), function calling and cybersecurity tasks; behind Claude Opus 5 on coding (59.6 vs 74.7 SWE-bench Pro). At FP8 it is roughly 750GB, so it needs the 8-card 768GB box. A 4-bit quantisation would be around 380–420GB, tight on 384GB with little room for many users. Size estimates Two things to weigh: it's a preview, and it comes from a Chinese government-backed lab, which government and regulated buyers will ask about. Its strength in tool use and multi-step research suits Selma's corpus and grant-analysis use cases well.
| Option | What it looks like | For | Against |
|---|---|---|---|
| A. EthicAI product (today) | "Selma, by EthicAI". One company, one team; SAFE into EthicAI Pty Ltd. | EthicAI's ethics and governance credibility is the trust that sells Selma to NFPs and government. Simple. | Investors see a consultancy with a product. Blackbird's Giants excludes services businesses; ESIC may fail on consulting income. Selma's value is harder to separate for a later sale. |
| B. Endorsed brand, same company | Selma has its own identity, site and voice; "an EthicAI company" in small print. Separate accounts internally. | Keeps the trust link and lets Selma grow a brand of its own. Easy to spin out later. Sub-brands (Selma Practice, Selma Field) sit cleanly under it. | Still one legal entity for investors; needs discipline to keep the numbers apart. |
| C. Standalone company | Selma Pty Ltd, with EthicAI as founding shareholder and ethics partner. Investors put money into Selma. | Clean for investors, ESIC and a mission-aligned acquisition. Clear IP ownership. | Setup cost and time now; licence and services agreements between the two; the raise and SAFE paperwork change. |
| D. EthicAI becomes Selma | The consultancy folds into the product company over time. | Maximum focus. | Gives up consulting income that currently funds the team, and EthicAI's education and advisory work. |