Build the Taiwan AI Inference Factory

Two phone calls. A different business.

The v0.9 memo was a brokerage thesis: low-capital, services-led, 90-day path to revenue, the only player in the 5–20 MW sourcing seam. The pivot we're now scoping: build small AI inference data centers in Taiwan, sell tokens. It is not the same business. It is a heavier business — and a much larger one if the unit economics hold.

Sunday is not a yes-or-no decision. It is a scoping conversation: how we sequence the work, who leads which workstream, and which open questions the next four weeks need to close.

Tokens. The unit, the meter, the margin.

A token is ~4 characters of text. When OpenAI charges per million tokens, that's the meter. "Tokens-as-a-Service" means renting GPUs out as a metered API, OpenAI-compatible so customers can swap providers without rewriting code.

The two markets are different. Open-weight inference (Llama, DeepSeek, Qwen) has collapsed below $1/Mtok and continues falling. Frontier closed-model inference (GPT-5, Claude Opus, Gemini Ultra) holds at $3–15/Mtok because the model is the moat, not the compute. Andrew's pitch lives in the open-model market. Margins there come from utilization, energy cost, and ops efficiency — not differentiated software.

$300M to put 5,000 GPUs in a building. What it actually buys.

Andrew's "10 MW at $300M is the best unit economics" is in the right ballpark — on the lower end. The realistic capex stack:

Revenue math at 5,000 GPUs

• Theoretical peak (retail): $1.50–2.50/GPU/day → $33–55M/year
• With dedicated-contract uplift (1.5–2× retail blended): $50–110M/year gross
• Opex: ~$15–25M/year. Debt service (~$200M @ 8%): ~$16M/year
• Net pre-tax, good case: $15–60M/year. Payback: 4–7 years on equity.

Three raises, not one. Roughly two years before ground breaks.

Four sequential stages with different bars and different investors. The good news: Andrew's Macquarie relationship is the right relationship — Macquarie's institutional capital lives at stages 03–04 (project equity and DDTL-style GPU debt are exactly their wheelhouse), and a senior principal opening that door from the seed stage is an unusually short path. The seed check itself is almost certainly personal or venture-arm, not a MIRA-led commitment — and that's fine; it's the normal pattern. The work for the next four weeks is scoping what Macquarie wants to see at each stage, not whether they're in.

Reference points: CoreWeave raised ~5 years and $20B+ before IPO (DDTL-style GPU debt). Crusoe took ~7 years to scale to its 933 MW Texas project. Nscale ~$155M Series A in late 2025 after demonstrating Norway ops. None raised full project finance on a deck. Skipping stages is how DC ventures die.

The niche pitch is defensible. The global pitch is harder to defend.

The Taiwan story is real, but it's strongest when it's specific. Where it wins is with buyers who explicitly need Taiwan — APAC latency, hardware proximity, sovereignty position. Where it weakens is the "Taiwan beats everywhere else" framing — that one collapses under any serious comparison. Both panels below are arguments we want to be ready for in any pitch conversation.

One sentence. Draft form. For stress-testing Sunday.

Why this framing and not something else

• "Dedicated capacity" not "retail tokens" — avoids the race-to-zero on per-Mtok pricing.
• "APAC + sovereign" not "global" — concedes the global token market to Together / Fireworks / Groq.
• "Inference" not "training" — frontier training requires hyperscaler-grade clusters and direct Nvidia relationships we don't have.
• "Hardware-proximate" not "cheap" — TW power is not cheap; the value prop is reliability + engineering proximity.
• "10 MW" not "100 MW" — matches Andrew's unit-economics thesis and our credible capital range.

This wedge fails if (a) no anchor customer materially values Taiwan-soil inference enough to sign reserved capacity, or (b) global open-model retail prices drop so far that even contracted prices follow. Both are real. Both are what Sunday is for.

Eight statements. If any one is false, the business is wrong.

1. PartialAnchor-customer pool exists. At least 3 credible APAC / sovereign / ODM candidates would consider signing a 5+ MW reserved-capacity contract within 12 months.
2. OpenTaiwan-soil inference commands a premium ($2–4/Mtok dedicated vs $0.88 retail). If sovereignty/latency/proximity buyers pay retail, unit economics collapse.
3. Open10 MW power can be secured in central/southern Taiwan within 12–18 months. >24 months and we miss the demand window.
4. Trend againstOpen-model token prices stabilize above ~$0.30/Mtok by 2028. If they collapse to $0.10, contracted prices follow.
5. PartialThe Macquarie engagement converts to a Stage-01 check within 90 days of a complete deck — and Macquarie signals appetite for Stages 03–04 conditional on anchor + power.
6. OpenThe two of us cover the operator skill mix — Andrew on build + capital relationships, Lawrence on demand-side + customer success — without a third co-founder or expensive early hires.
7. OpenStage-02 development raise closes within 18 months of seed. >24 months and runway burns out before anchor + permits land.
8. OpenToken pricing for dedicated APAC sovereign inference can be discovered through 3–5 real buyer conversations in the next 60 days. Without price discovery, the unit-economics scenarios are speculation.

The five questions Sunday has to answer.

Most of the WMBT statements are open — not because the thesis is weak, but because we haven't done the work yet. The conversation isn't whether to do this. It's how we sequence the next four weeks so each of these moves from open to defensible.

1. Sequence — brokerage first, build second, or commit straight to build? The strongest version may be services first (Y1 revenue + validates demand + surfaces anchors + maps sites), build second. The alternative is committing to build now and using seed capital to do anchor + site discovery in parallel. Today decides which.
2. Anchor customer — three real names we can pursue this month? Without a signed reserved-capacity LOI, project finance does not happen. Japanese conglomerates, Korean sovereign-AI, Middle East programs, SEA governments, ODM-internal demand. Who do we both already know? Who needs a warm intro?
3. Power — what's our 10 MW path in central or southern Taiwan? Taipower's >5 MW restriction north of Taoyuan rules out the easy answer. Tier-2 industrial parks, retrofits, and behind-the-meter (gas / solar / fuel-cell hybrids per Exhibit 04) are all live options. Who at Taipower / IDA / park authorities do we start with?
4. Macquarie engagement — what does the next 60 days with them look like? The relationship is the right one — Macquarie's institutional capital sits at Stages 03–04 exactly. The work now: what does the principal want to see for a Stage-01 check, and what signals (anchor LOI, site option, permit motion) unlock the Stage-03 conversation? Concrete asks, dates, deliverables.
5. Operating roles + governance — who leads what, and how does Ally fit? Andrew: build, capital relationships, supply-side. Lawrence: demand-side analysis, customer development, governance. Do we form the entity now or after Stage-01? What disclosure / structure keeps Andrew's Ally CIO role clean? Decide before any joint-branded outreach goes out.