The Taiwan
AI Infrastructure Wedge

The macro is no longer in dispute. The choke point has moved downstream.

Eight citable datapoints, drawn from filings, Reuters, DIGITIMES, Taipower commentary, and Mordor Intelligence. Every number below has a source in §Cites.

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

Everything below in this note either reinforces or refutes these. Read the rest of the document as evidence for or against each line. The open question for any market entrant is which of these cannot yet be defended.

1. Buyers need Taiwan-routed capacity, not generic compute. If a buyer is indifferent to where the rack lives, we have no edge.
2. The 5–20 MW buyer band is real. Too small for hyperscaler-grade direct procurement, large enough to pay structured fees. If this band collapses into "neoclouds get hyperscaler treatment" or "everyone routes through Foxconn direct," we have no customer.
3. Taiwan-side supply is fragmented enough that brokerage adds value. If the top three colos and Foxconn already cover the buyable surface, we are a thin layer that gets disintermediated within a year.
4. Power-constrained sites can still be found outside Tier-1 zones. Central and southern Taiwan, retrofits, second-tier industrial parks. If Taipower's 5MW ceiling has effectively closed the market everywhere worth being, no amount of brokerage helps.
5. Supply-side parties will pay success fees. If property owners and ODMs refuse to compensate the introducer, the business is buyer-side only — a much thinner P&L.
6. A neutral, conflict-free sourcing structure can be maintained. If the intermediary has any ownership or compensation stake in specific supply-side inventory, that has to be disclosed and neutrally arbitrated, or the model breaks down.
7. A first paying buyer closes within 90 days. If the cycle is 9–12 months, we are not a services business — we are a project, and the unit economics don't work.

The bottleneck has moved. From silicon, downstream — into power, real estate, and integration.

Three citable shifts define the current Taiwan picture. None requires assumptions about lead times we cannot verify.

Shift 1 — Power has become the gating constraint, not chips

Taipower has restricted new data center applications above 5 MW north of Taoyuan unless new generating capacity is built.⁷ Taipower chair Wen-sheng Tseng has publicly warned that semiconductor and AI infrastructure will add 5.3–5.4 GW of demand by 2030, with each new fab requiring ~200 MW.⁸ Within that envelope, AI facilities alone are projected to reach 1 GW by 2030 — meaning a fifth of the new load sits in workloads that did not meaningfully exist five years ago.²⁰ The Taipower System Planning Division has flagged transformers and gas turbines as being in global shortage, with northern Taiwan supply already running below demand.¹⁰

Shift 2 — Hyperscalers are bypassing the colocation layer entirely

AWS, Microsoft, and Google are now purchasing land and power blocks directly from Taipower, bypassing colocation intermediaries to enable proprietary liquid-cooling designs and renewable-energy procurement.¹¹ AWS invested over $5B to launch its three-zone Taipei region in January 2025 — the largest cloud deployment in Taiwan history.¹¹ This bypass dynamic creates a structural opening: the next tier of buyers — neoclouds (GPU-native cloud providers like CoreWeave, Lambda, Nebius that emerged to serve AI workloads, distinct from AWS / Azure / GCP), sovereign AI programs, and mid-market AI companies — cannot replicate this direct procurement, and need a partner.

Shift 3 — Foxconn itself is becoming an operator, not just a manufacturer

Foxconn's Visionbay.ai unit is building a 27 MW $1.4B GB300 cluster — Taiwan's largest advanced GPU cluster and Asia's first GB300 site — going live H1 2026.^3,12 Chairman Young Liu has committed $2–3B annually to AI infrastructure and named AI as the primary 2026 growth driver.³ Each 200-rack facility represents roughly $10M in liquid cooling hardware alone.¹³ Foxconn is now simultaneously the dominant supplier (1,000 racks/week, ~40% global AI server share) and a major end-customer of AI infrastructure capacity. That dual role creates demand for capacity it does not internally own.¹⁴ Zoom out one level: CIER estimates Taiwan as a whole supplies 80–90% of the world's AI servers — a concentration ratio that has no parallel in any other strategic category.²¹

The frictions stack — three more constraints below the surface

Clean megawatts, not just megawatts. Hyperscalers and sovereign-AI buyers procure under renewable mandates, and Taiwan's clean-power supply is binding. Google signed its first Asia-Pacific offshore wind PPA in Taiwan in 2025 to back its cloud and data center loads on the island.²⁴ TSMC has expanded its own offshore-wind PPAs to absorb AI-driven fab demand, tightening what's left for everyone else.²⁵ CommonWealth's May 2026 reporting confirms tiered industrial electricity pricing now penalizes inefficient data centers, making PUE and liquid cooling a procurement gating factor, not a nice-to-have.²³ The right way to score a Taiwan site is no longer megawatts — it is certifiable clean megawatts available before 2027.

Fiber and interconnect, not just power. A 50 kW+ AI rack carries roughly an order of magnitude more fiber than a traditional server rack; cable lead times have stretched as AI deployment compresses cycles globally.²⁶ Taiwan's compensating asset: Chunghwa Telecom operates a 27-cable global network including 12 submarine cables landing on the island, giving carrier diversity and resilience that most adjacent geographies cannot replicate.²⁷ Fiber adjacency belongs in the site-scoring rubric alongside power.

Water and heat rejection. Liquid cooling needs water rights, recycled-water access, and heat-rejection headroom. Many industrial-park retrofits look feasible on paper and fail at the heat-rejection step — cooling-tower siting, dry-cooler footprint, and ambient design conditions all gate what density a building can actually host. The next layer of diligence past "is there power?" is "can the building reject the heat?" — and most Tier-1 brokers don't ask.

Where the bottleneck moved — and where we play.

A 5-second visual of the argument above. Layers are ordered top-down from the revenue-bearing customer workload to the foundation layers that now gate everything. The accent layer in the middle is our wedge.

Five years ago, the bottleneck in this stack sat at Layer 04 (silicon). Today the bottleneck has migrated to Layers 07–09 — cooling, the wedge, and grid. Layers 01–03 have capital; Layer 04 has Nvidia; Layer 06 has Foxconn. Layer 08 has nobody serving the 5–20 MW band. That is the entire business in one diagram.

"Why Taiwan, despite the constraints?" The honest comparison.

A buyer's natural objection: "Just build in Malaysia." Directional comparison across the seven jurisdictions a 5–20 MW AI buyer actually considers. Scoring is qualitative (●●●●● strong → ○○○○○ weak), drawn from public reporting and industry benchmarks. Treat as conversation scaffolding, not audited diligence.

The honest answer to "why Taiwan?" is not that Taiwan is best on every axis — it isn't. It is that Taiwan is the only jurisdiction in the world where AI server manufacturing, advanced packaging, and rack integration physically converge. Every other geography on this table is a deployment site; Taiwan is also the supply chain. For buyers whose deployment timing is gated by rack delivery, that proximity is not a preference — it is the unlock.

This is the wedge to a "Malaysia question": "You can absolutely build there. But you'll still be sourcing the racks from Taiwan, and the lead times will be longer, not shorter."

A market seam exists between demand and supply. Here's what's on each side.

This category rewards whoever occupies a defensible position between Western AI demand and Taiwan-side capacity. The market map below shows the actors on both sides and the intermediary layer where a sourcing partner plugs in.

The category is occupied at the edges, open in the middle.

There are well-funded incumbents at the high end (consulting majors, research houses) and a wide unserved layer at the bottom (relationship brokers, no software). The middle — software-enabled supply-demand matching at deal-size $1–10M — is empty.

Where we're strong, weak, where the wind is at our back, and where it can kill us.

Why a buyer pays $25–75K for sourcing help. The arithmetic, not the assertion.

The macro proves scarcity. This section converts scarcity into the buyer's P&L. All numbers below are illustrative ranges anchored to public benchmarks; treat as the shape of the conversation, not audited diligence.

Reference deployment: a 10 MW GPU cluster

Anchor case for a mid-market AI buyer or sovereign program: a 10 MW deployment housing roughly 4,500–5,500 GB200-class GPUs (~2 kW per GPU plus networking and cooling overhead). Capex range: $400M–$700M (GPUs $300M–$500M; racks, cooling, power, fit-out $100M–$200M). Revenue range, at neocloud-equivalent rental rates of $3–6 per GPU-hour and 65–75% utilization: $80M–$220M per year.

The "shorten sourcing by 60 days" trade

Take the conservative end of the range: a 10 MW cluster earning $80M/year at 70% utilization. Sixty days of earlier deployment = ~$13M of revenue pulled forward. Even discounted for ramp and customer onboarding, the present value of 60 days saved is in the $5–10M band. Against that, a $25–75K sourcing engagement plus a 1–3% success fee on a $5–10M land/power placement is a rounding error. The buyer's decision is not "can I afford this?" — it is "is the sourcing partner credible enough to actually compress the timeline?"

The hidden urgency — leveraged capex, not VC runway

The above frames the cost of delay as foregone revenue. The sharper frame, and the one that matters for closing a deal, is that delay = real cash burn on already-financed capex. Neoclouds are not VC-backed startups extending their runway — they are leveraged infrastructure vehicles with covenants tied to deployment.

CoreWeave's Q1 2026 balance sheet, disclosed 19 May 2026, makes the point explicitly:²⁸

Translate to the math: a 60-day site delay is not just $13M of foregone revenue — it is ~$350M of accumulated interest expense burning against capex already deployed and contracts not yet earning. A $75K sourcing engagement isn't even a rounding error; it is the cost of one rounding error.

Translation into the pitch

The argument we make to a buyer in plain language: "You are not buying our Taiwan knowledge. You are buying 60–120 days of compressed timeline, $5–15M of avoided wrong-site cost, and certainty on the rack slot. Our fee is a fraction of one day of interest expense on idle capex."

One sentence. This is what we sell.

All the strategy work above resolves to a single positioning statement we test with three real buyers in two weeks.

The wedge is not a generic supply chain dashboard, AI infrastructure intelligence platform, or sourcing concierge. Those phrasings keep the door open too wide. This phrasing closes it down to a specific buyer (mid-market AI infra), a specific need (land + power + integration), a specific market (Taiwan), and a specific deal-size band ($1–10M). Every product, pricing, and GTM decision flows from this sentence.

Services-led at the start. Software emerges from the data we collect.

We do not build a SaaS product on day one. We build a high-touch service that produces proprietary supply-side data, and the software emerges as the natural artifact of doing the service well.

Four forces. Each one would be enough; together they are deterministic.

Why a CoreWeave-tier buyer is on a plane to Taoyuan this year, not next. The structural pressures putting our target customers in our addressable market.

1. Engineering proximity — A 50 kW+ liquid-cooled GB200 NVL72 rack is roughly 1.3 tons of pre-filled coolant loops and delicate fiber interconnects. When a manifold leaks at a US site, the troubleshooting cycle is days of trans-Pacific RMA. When it leaks in Taiwan, the same engineers who built the rack at Foxconn/Quanta are an hour away. Nvidia's VP Alexis Bjorlin made the implicit case at Hon Hai Tech Day: "building individual facilities may no longer make economic sense. Renting compute resources may offer a far better return on investment."³ Foxconn's Visionbay.ai 27 MW GB300 cluster is the literal proof — Nvidia and Foxconn co-engineering at sovereign scale, on the island.^3,12
2. The US grid escape hatch — US interconnection queues are now measured in years, not months. The Lawrence Berkeley National Lab Queued Up series documents ~2,600 GW stuck in transmission queues nationwide — multi-year waits for new large loads in PJM, ERCOT, and CAISO.³¹ For a neocloud whose customer contract starts paying on energization, year-three power means year-three revenue. Taiwan is constrained, but constrained in a way a sourcing partner can navigate — central and southern allocations, retrofits, second-tier parks — whereas US transmission queues simply do not move.
3. Nvidia's gravitational pull — Nvidia is not just a supplier to the neoclouds; it is the orchestrator. The capital flow runs through them (Nvidia's $2B equity stake in CoreWeave; equity stakes in Lambda, Nebius, et al), the chip allocation runs through them, and the public reference architecture runs through Taiwan via the Visionbay.ai validation site. When Nvidia anchors its sovereign-AI factory narrative on a Taiwan deployment, every Nvidia-backed neocloud is incentivized to follow — not because they were told to, but because chip allocation favors those who validate the architecture where Nvidia has chosen to validate it.^3,12
4. The death of asset-light — The old neocloud thesis ("just lease a room and plug in chips") collapsed over the last twelve months. AI hardware is too heavy, too hot, and too power-hungry for that model. Nebius and Bloom Energy announced 328 MW of behind-the-meter solid-oxide fuel cells on 20 May 2026, explicitly to bypass utility transmission queues.²⁹ Crusoe Energy's regulatory filings disclose a 933 MW off-grid natural-gas plant (the "Goodnight" campus) in the Texas Panhandle.³⁰ Neoclouds are not looking for landlords anymore — they are looking for land, substations, and proprietary power pipelines. That is exactly what the right industrial real-estate partner's inventory and utility relationships are positioned to provide. The "asset-light" customer is now the heavy-asset partner.

The first paying logo determines the next ten.

Our pioneer is not the easiest sale — it's the one whose success makes the next nine sales trivial. We need a buyer that is well-known enough to be a reference, hungry enough to move fast, and small enough that we are not overshadowed.

Frontier Training vs Inference — which sub-market we're actually in

The AI infrastructure market is bifurcating, and the bifurcation determines which targets fit our band. Frontier training (CoreWeave anchoring OpenAI / Meta multi-billion-dollar training runs; xAI's Colossus; Anthropic's clusters) requires gigantic 50–500 MW deployments and is effectively hyperscaler-or-equivalent territory. Inference and enterprise AI — serving trained models, vertical AI, sovereign deployments, edge — runs in 5–20 MW deployments distributed regionally for latency, sovereignty, and cost. That is our band, almost exactly. So the shortlist below should be read accordingly: rows 01–05 (existing neoclouds) are best approached on their inference build-out, not their frontier-training core; rows 06–10 (sovereign and mid-market) are inference-native and align directly.

Pioneer Shortlist · Eleven Neoclouds, Ranked by Taiwan-Help Need

The ranking below is based on observed operational footprint, not on size or fame. The logic: a neocloud with global cloud regions but no Taiwan operating presence is the highest-leverage target, because Taiwan-routed capacity is a real gap their customers will surface. Companies already deep in Taiwan (GMI, Firmus, SoftBank, Yotta) are not pioneers — they are competitors, partners, or too high up the stack. Companies too big for a small sourcing broker (CoreWeave) are useful benchmarks, bad pioneers.

What the ranking actually tells us: only the top three (Lambda, Nebius, Crusoe) are unambiguously good Phase-00 calls. Nscale is a worth-one-call edge case. Rows 05–11 are not Phase-00 targets — they are either too big, too local, or structurally non-buying. The wedge is narrower than the original 10-name list suggested. That is good news, not bad — it tells us where to spend the next fourteen days.

Adjacent Targets · Sovereign AI & Mid-Market (Not Neoclouds)

Sovereign-AI programs and mid-market AI labs have very different buying patterns from neoclouds — slower procurement but higher willingness-to-pay, often larger deals. Keep these on a parallel track.

First-ask discipline: a single question on the first call. "If we could compress your Taiwan capacity sourcing by 60 days and guarantee the rack-delivery slot, what would that be worth to you in your next deployment?" Listen for whether the answer is in five-figure, six-figure, or seven-figure territory. That sorts the list.

Tactical Note · Avoiding Procurement Hell (and using COMPUTEX as the accelerant)

The honest risk: a publicly-traded neocloud has a procurement function geared for nine-figure ODM contracts. A $50K advisory engagement could sit in MSA negotiation for ninety days, which kills the unit economics. We have to design around this from day one, not discover it after the first verbal yes.

Five tactics that prevent it:

• Engagement letter, not MSA. Standard boutique-advisory pattern — 1–3 page document signed by an executive sponsor. Avoids the procurement queue entirely.
• Price under the threshold. First engagements at $25–50K typically clear "expedited" approval at most enterprises; $75K+ triggers full procurement review. Sell the wedge first, scale the dollars second.
• Sell to the function that owns deployment. CTO, COO, or VP Infrastructure — not procurement. They have personal incentive to compress the timeline (their bonus depends on it). Procurement is informed after the engagement letter is signed.
• Shift the bigger dollars to the supply-side success fee. Our 1–3% brokerage take is paid by the property owner / supplier, not the buyer. The buyer's procurement only sees the small engagement fee; the venture's economics come from the supply side that has no procurement function.
• Use COMPUTEX as the accelerant. COMPUTEX Taipei (2–5 June 2026) physically converges every name on this shortlist plus Nvidia, Foxconn, Quanta, Wiwynn, and the major sovereign-AI buying delegations in one city for four days. Decisions that take ninety days through email take ninety minutes in a hotel bar with the right exec. At least two of our five Phase-00 calls should happen in-person at COMPUTEX, not over Zoom — and the handshake there becomes the engagement letter signed the following week.

Three revenue streams. Sequential, not simultaneous.

Year 1 is α and β only. Year 2 introduces γ once we have the data to make it real.

Year 1 financial picture — illustrative

Eight engagements at average $40K = $320K engagement revenue. Two closed deals at average $5M deal value × 2% blended take = $200K brokerage revenue. Total Year 1: ~$520K. Two-person team, no payroll, expenses primarily travel and tooling. Cash-flow positive by month 4–6 if first engagement closes on time. Year 2 adds platform revenue from a pre-validated buyer set. These are illustrative targets, not forecasts — none are validated until we have three buyer calls done.

Twelve months, four phases. Each phase has a kill criterion.

No phase moves forward without a hard signal from the previous one. If the signal fails, we stop and reassess. The discipline is the strategy.

Internal product spine. The proprietary asset, in outline form.

This is the operational backbone the venture eventually charges for. Each row is a step in the >5 MW Taipower approval pathway that we will document, time, and price as we run real engagements. Treat this as the skeleton — flesh comes from primary research and real deployment engagements.

The cells marked [TBD] are where the venture's proprietary value accrues. Every engagement we run fills in one more cell with a real number from a real deal — turning this table from outline into the most accurate >5 MW Taipower approval map outside Taipower itself. This is the product, in latent form.

The Taiwan industrial-site AI-suitability audit template. One row per site.

This is the diagnostic template a Taiwan industrial real-estate portfolio would need to run, site by site, to answer one question: is there any AI-suitable inventory today, or is the supply-side thesis fantasy?

If a portfolio audit clears zero sites against five-plus thresholds, the supply-side wedge for that portfolio is fantasy. If three or more sites clear, there's a real seed inventory and a credible buyer pitch. Either answer is decisive.