Build the
Taiwan AI Inference Factory

From sourcing to building. A different business.

The prior note (twstrategy.pgintel.dev) laid out a brokerage thesis: low-capital, services-led, ninety-day path to revenue, positioned in the 5–20 MW sourcing seam. The thesis under evaluation here is a pivot from brokerage to build — instead of sourcing capacity for others, build small (10 MW) AI inference data centers directly in Taiwan and sell tokens. It is not the same business. It is a much heavier one. This note exists to evaluate whether it is also a better one.

What the build thesis holds

• Build, don't broker. The thesis calls for operating data centers directly, not introducing supply to demand.
• Small unit, $300M each, 10 MW. The unit-economics thesis is that 10 MW is the sweet spot — small enough to deploy quickly and serve regional markets, large enough to justify full AI networking. Multiple of them, not one big one.
• Sell tokens-as-a-service. The business model is inference-as-a-service: host open-weight models, expose an OpenAI-compatible API, meter customers by tokens consumed.
• A named institutional relationship at Macquarie is cited as a candidate source of seed capital — "a few million bucks" of development capital, not the full $300M project financing.
• Demand evidence: "Salesforce is spending $300M on tokens." Used as proxy for the size of the token-buying market.
• "Taiwan can provide this to the world that other countries can't." The geographic differentiation argument.

What we are actually selling. The unit, the meter, the margin.

A token is the metering unit for large-language-model input and output. Roughly four characters of text, or three-quarters of a word. When OpenAI charges per million tokens, this is the meter. "Tokens-as-a-Service" means renting your GPUs out as a metered API.

The mechanics, in plain terms

1. The buyer (a developer, app, or enterprise) calls an API. They send a prompt, expect a completion. Both directions are measured in tokens.
2. The provider hosts a model on GPUs. Open-weight (Llama, Mistral, DeepSeek, Qwen) or licensed proprietary. The model runs on the GPUs.
3. The provider exposes a metered API. Usually OpenAI-compatible so customers can swap providers without rewriting code.
4. Billing is per-token. Typically separate prices for input and output tokens, sometimes higher for output. Charged per million tokens (Mtok).
5. Margin is the gap between what the buyer pays per Mtok and what the GPU cluster costs to run per Mtok (power + cooling + depreciation + debt service + ops).

What tokens cost in 2026 — and why the chart is collapsing

Token pricing has fallen roughly an order of magnitude in eighteen months. For a comparable open-model API (Llama-3 70B class):

The two markets are different. Open-weight inference (Llama, Mistral, DeepSeek) has collapsed to under $1 per Mtok and continues falling. Frontier closed-model inference (GPT-5, Claude Opus 4, Gemini Ultra) holds at $3–15 per Mtok because the model itself is the moat, not the compute. This thesis is in the open-model market. That is the commoditizing market. Margins there come from utilization, energy cost, and operational efficiency — not from differentiated software.

How a token provider actually makes money

Reverse-engineering the margin: a Llama-3 70B request might generate ~50 output tokens per second on a single H100 GPU. At ~$0.88 per Mtok, one H100 running continuously at full utilization produces roughly:

• 50 tok/sec × 86,400 sec/day × $0.88 / 1,000,000 = $3.80/day per H100 at theoretical peak
• Realistic utilization (with batch, ramp, idle): $1.50–2.50/day per H100 blended
• H100 amortized capital cost (24K capex / 4-year depreciation / 365 days): $16/day per H100
• Power + cooling + ops at ~$5–8/day per H100
• Net result: open-model inference at retail pricing barely breaks even on hardware cost alone, before debt service.

This is why open-model inference providers either (a) optimize batch and quantization aggressively to push effective throughput 3–5× above naive estimates, (b) sell to enterprises at higher contract prices than retail rates, or (c) bundle higher-margin services (fine-tuning, dedicated capacity, support). The retail API price is a customer-acquisition cost, not the actual margin model.

Bigger than training. Commoditizing faster than training.

In 2024 the AI infrastructure conversation was about training clusters. In 2026 the larger and faster-growing market is inference. Inference is what enterprises actually pay for in production. The economics are completely different from training.

Market size and shape

Estimates of the global AI inference market vary, but the directional consensus across Bain, McKinsey, IDC and a16z reports in late 2025 / early 2026 puts inference compute spend at $50–80 billion in 2026, projected to $250–400 billion by 2030.³⁴ Inference is now plausibly larger than training in dollar terms because every enterprise deployment generates inference perpetually, while training is episodic.

Two distinct inference markets

The DeepSeek effect and the commoditization of open inference

DeepSeek-V3 (released January 2025) and DeepSeek-R1 (released January 2025) demonstrated frontier-comparable reasoning at roughly one-tenth the inference cost of closed-frontier alternatives.³⁵ Llama-4, Qwen 2.5, and Mistral Large 2 followed similar curves through 2025. The result by mid-2026: open-weight models are good enough for ~70% of production enterprise use cases, and the differentiator is no longer model quality but inference cost.

That is excellent news for inference providers in one sense — the addressable market is growing fast as enterprises shift workloads to open models — and brutal in another sense — every provider serving open models is selling the same thing, and competing on price.

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

The claim that "10 MW at $300M is the best unit economics" is in the right ballpark but on the lower end. The capex stack below is the realistic breakdown. Numbers are illustrative ranges based on industry-standard data-center cost benchmarks circa 2026.

Capex stack — what $300M actually buys

Revenue and margin math

At 5,000 GPUs producing inference at retail pricing equivalent:

• Theoretical maximum: 5,000 GPUs × $1.50–2.50/day = $7,500–12,500/day = $2.7–4.6M/month = $33–55M/year
• With dedicated-contract uplift (1.5–2× retail blended): $50–110M/year gross revenue
• Opex (power, cooling, ops, software, support): ~$15–25M/year for a 10 MW facility
• Annual debt service (if ~$200M debt at 8% blended cost): ~$16M/year
• Net pre-tax (good case): $15–60M/year
• Payback period: 4–7 years on equity, longer if token prices keep falling

The two scenarios that change everything

Anchor-customer scenario. If a single enterprise or sovereign-AI customer signs a multi-year reserved-capacity contract for 50–80% of the cluster at premium pricing ($2–4 per Mtok dedicated, not $0.88 retail), the math improves dramatically. Revenue could approach $100M+ in year one, and payback compresses to 3–4 years. This is the only realistic path to attractive returns. Without an anchor customer, the math depends entirely on retail-token sales that compete with Together / Fireworks.

Token-price-collapse scenario. If open-model token prices continue falling 50%+ per year (which the 2024–2026 curve suggests), retail revenue could drop below opex within 24 months. The hedge is anchor contracts that lock in pricing. Without contracts, the project is unfundable.

Seed, development, project finance, build. Three raises, not one.

"Find a rich person to give us a few million bucks" sounds like one round. It is actually the first of three sequential raises with very different bars, very different investors, and roughly two years between the first dollar and the first ground-broken.

The four stages

Macquarie's institutional capital sits primarily at stages 03 and 04. A seed-stage relationship with a Macquarie principal is almost certainly a personal check or a venture-arm bet, not a MIRA-led commitment. That's a normal pattern — but any such relationship needs to be named, dated, and confirmed as actually committed, not informal.

Reference points: how comparable companies actually raised

• CoreWeave — raised through a layered structure totaling $20B+ before IPO; current $17.3B debt is GPU-collateralized via DDTL facilities.²⁸ Sequence took ~5 years from founding to IPO.
• Crusoe Energy — vertically integrated power + DC; raised across stages and is now building a 933 MW Texas facility via permits filed April 2026.³⁰ Took ~7 years to scale to this size.
• Nscale — UK/Norway/Portugal/US capacity; raised ~$155M Series A in late 2025 after demonstrating Norway operations.
• Macquarie/Aligned — Macquarie acquired Aligned Data Centers for ~$40B in 2023; that is the model for stage 03–04 capital, on a fully developed platform.

The pattern: none of these companies raised the full project finance on a deck. They built credibility incrementally — seed for the team, development for the site, then institutional capital after the anchor was signed. Skipping stages is how DC ventures die.

The niche pitch is defensible. The global pitch is not.

The framing that "Taiwan can provide this to the world that other countries can't" works at a high level and falls apart in the specifics. The honest version is narrower and stronger: Taiwan is the right place to build inference capacity for specific buyers who specifically need Taiwan, not for the global token market.

What is actually true about Taiwan for inference

• Hardware proximity. Taiwan supplies 80–90% of global AI servers via Foxconn, Quanta, Wiwynn, Inventec.²¹ A failed liquid-cooling manifold gets diagnosed in hours, not weeks. Equipment lead times can be compressed if relationships are real.
• APAC latency. Tokens served from Taipei reach Tokyo, Seoul, Singapore, Manila 50–100ms faster than tokens served from US-East. Real differentiator for inference workloads serving APAC users.
• Sovereignty alignment. A growing pool of buyers (Japanese enterprises, Korean conglomerates, Middle Eastern sovereign-AI programs, Southeast Asian governments) want AI infrastructure that is not US-soil and not mainland-China-soil. Taiwan fits both criteria. Few other places do.
• Carrier diversity. Chunghwa Telecom operates 27 submarine cables globally with 12 landing in Taiwan.²⁷ Better connectivity foundation than most adjacent geographies.

What is overstated

• Power cost. Taiwan industrial power is not cheap relative to Texas, Malaysia, or even Japan. Tiered industrial rates introduced in early 2026 add up to 20% surcharges for inefficient DCs.²³ Power-cost arbitrage is against Taiwan, not for it.
• Power availability. Taipower restricts new >5 MW applications north of Taoyuan.⁷ Central and southern Taiwan are accessible but slower. Securing 10 MW is not trivial — it is one of the gating risks of the entire project.
• Geopolitical risk. Cross-strait tension is priced into every Taiwan infrastructure investment. Risk-averse Western enterprises (especially financial services, healthcare, government) actively prefer non-Taiwan deployments.
• "Taiwan can do this others can't." Taiwan is not the only place that can host inference. It is one of a small number of jurisdictions with a specific combination of advantages — and disadvantages.

The token market is crowded, capitalized, and racing toward zero. Where can a new entrant play?

Open-model inference has roughly a dozen serious providers, most with $50M–$500M+ in funding, all serving the global market. Pricing pressure is intense. The question for any new entrant is not "can we sell tokens" — it is "what can we sell that they cannot."

Who is in this market

What this competitive picture means

A new entrant cannot win on retail price for generic Llama-class inference. Together, Fireworks, and Groq are all selling the same tokens for $0.59–0.90 per Mtok. A new Taiwan-based provider entering at the same price has no advantage and no economic moat — the same model, the same hardware, just a different building.

A new entrant can win on specific dimensions where the incumbents cannot.

• Sovereignty. A Japanese / Korean / Middle Eastern enterprise that legally cannot send inference traffic through US-soil providers.
• APAC latency. An AI app whose users are concentrated in Tokyo/Seoul/Singapore and whose latency budget cannot accommodate a 200ms round-trip to US-West.
• Data residency. A financial services firm under Singapore MAS or Japan FSA rules requiring local data processing.
• Dedicated capacity contracts. Not retail tokens — reserved-cluster deals at $2–4/Mtok where the customer is buying assurance and locality, not lowest price.
• Sovereign-AI build-out. Government-tied programs where Taiwan-as-build-site is a strategic positive (Taiwan/US alliance, hardware proximity).

Salesforce is not our customer. Here is who actually is.

The reference point that "Salesforce spends $300M on tokens" is a directionally correct illustration of token market size but is the wrong customer profile. Salesforce buys from OpenAI, Anthropic, and Cohere — frontier-model APIs — not from Taiwan-based open-model providers. The real customer set is narrower and more specific.

The buyer segmentation

The anchor-customer question, sharpened

From §IV, the project finance only works with an anchor customer. From the segmentation above, the realistic anchor candidates are:

1. A Japanese or Korean megacorp wanting 5–10 MW of dedicated Taiwan-soil inference capacity for AI products serving APAC users (Rakuten, LINE, Samsung, SK Telecom, KDDI, NTT).
2. A Middle Eastern sovereign-AI program diversifying compute geography off Gulf or US soil (G42, HUMAIN). G42 has Cerebras anchor relationship; HUMAIN is newer and less locked-in.
3. The Taiwan government itself via a sovereign-AI program — an open-model deployment for Taiwan public-sector use and ASEAN export.
4. A Taiwan ODM (Foxconn / Quanta / Wiwynn) building inference capacity for their own products and ecosystem (Foxconn's Visionbay.ai precedent^3,12).
5. A specialty AI app with APAC users committing to 1–2 MW of dedicated capacity at premium pricing.

Without one of these signed in development stage (02), the project is unfundable. The seed (~$2–5M) can be raised on the team and thesis. Stage 02 cannot.

One sentence, draft form, for stress-testing against reality.

Same exercise as v0.9 §VI — collapse the strategy into a single positioning statement that can be tested in real conversations.

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 and competes in the specific segment where Taiwan has structural advantage.
• "Inference" not "training" — frontier training requires hyperscaler-grade clusters and direct Nvidia relationships that a new entrant typically doesn't have. Inference plays to the cost-and-locality strengths.
• "Hardware-proximate" not "cheap" — Taiwan power is not cheap; the value proposition is reliability and engineering proximity, not unit cost.
• "10 MW" not "100 MW" — matches the unit-economics thesis and a realistic capital range for a new entrant.

This wedge fails if (a) no anchor customer materially values Taiwan-soil inference enough to sign reserved capacity, or (b) the global open-model retail price drops so far that even contracted prices follow it down. Both are real risks that further diligence needs to resolve.

Eight statements. Same discipline as v0.9, retargeted at the new thesis.

If any one of these is false, this business is wrong. The open question is which of these can be defended and which cannot.

1. Anchor-customer pool exists. At least three credible APAC / sovereign / ODM candidates would consider signing a 5+ MW reserved-capacity contract within 12 months. If no one is willing to anchor, the project is unfundable.
2. Taiwan-soil inference commands a premium ($2–4/Mtok dedicated vs $0.88 retail). If sovereignty / latency / proximity buyers pay retail, the unit economics collapse.
3. Power can be secured for 10 MW in central or southern Taiwan within 12–18 months. If Taipower allocation is more than ~24 months, the project misses the demand window.
4. The operating team has verifiable data-center build experience. Not just relationships — actual prior projects, scale, role, references. This is foundational and, for any new entrant, needs to be established up front.
5. Open-model token prices stabilize above ~$0.30/Mtok by 2028. If they collapse to $0.10 or below, even contracted prices follow, and margins evaporate.
6. A named institutional relationship (e.g. at a firm like Macquarie) is real and committed to writing a seed check, not just willing to take a call. The difference is six to twelve months of fundraising.
7. The founding team can operate a build-and-operate venture — not just a brokerage. The skill mix is different (capital, construction, ops, customer success), and someone on the team needs to be credible at each.
8. The first stage 02 raise closes within 18 months of seed. If development capital takes 24+ months, runway burns and the project stalls.

Two businesses. Pick one, or sequence them.

The brokerage thesis from twstrategy.pgintel.dev and the build thesis from this document are not the same venture. They have different capital, different risk, different timelines, different team requirements. The choice is real.

The hybrid argument

The strongest version of the strategy might be to do brokerage first, build second. Year 1 brokerage generates revenue, builds relationships, surfaces the actual anchor customers, and identifies the right sites. Year 2 the venture pivots into operating, with proprietary information advantage and validated demand. Skipping brokerage skips the most efficient possible validation of the build thesis.

The counter-argument: brokerage occupies time and attention that the build thesis needs urgently. The window for Taiwan inference is closing (every quarter, sovereign-AI buyers commit to other geographies). Brokerage could be a 12-month distraction that loses the build window.

Five thresholds. If we cannot clear them, this is not yet a real venture.

Same discipline as v0.9 Phase 00. If by 5 June 2026 — including the COMPUTEX week — the five questions below cannot be answered with named, dated specifics, the project is not yet ready for serious commitment.

1. Capital-relationship verification. Any named seed-capital relationship (e.g. an institutional contact) is confirmed as a real, responsive commitment — not just a warm intro.
2. The operator's track record. Verifiable prior data-center builds the operating team was actually involved in — scale, year, role, references. Without this, the operator thesis lacks the proof point that justifies a seed raise.
3. One named credible anchor candidate. Not "we'll find one" — a specific Japanese / Korean / sovereign / ODM target with a warm path and a hypothesis for why they would sign reserved capacity. Without this, stage 02 financing is impossible.
4. One identified candidate site. A specific industrial parcel in central or southern Taiwan with a credible >5 MW power story. Ally inventory is the obvious starting point. Without a real site, the seed has nothing to develop.
5. Operating roles defined. A clear, indispensable function for each member of the founding team — commercial lead, capital lead, technical lead, or another structural role. If the answer after diligence is "we'll figure it out," the team lacks the structure to fundraise.

If 3+ of 5 thresholds are not cleared by 5 June 2026: the venture is not yet investable, the seed conversation with Macquarie should be deferred, and the brokerage thesis (v0.9) becomes the right path to revalidate. The brokerage work would itself surface the anchor customer and site for a future pivot to build.

If 4–5 of 5 thresholds are cleared: the venture is investable in seed terms, and the conversation transitions from "is this real?" to "how do we structure it?"