Is a Terrestrial Neocloud Possible in China? — Part 2

Part 1 — A Primer (15 August): what a neocloud is, how a number is verified, whose order bought the megawatts. This is Part 2.

China can already stand up terrestrial AI megawatts at CoreWeave and Nebius scale. What it does not have, and will not have under current GPU, customer, and capital rules, is a listed NVIDIA-native third-party GPU cloud with CoreWeave’s demand stack, allocation stack, and capital stack. People keep comparing GDS to NBIS because both have buildings and both have watts. That comparison is how you confuse a landlord with an airline. Runze, VNET, and GDS are airframes. A neocloud (a specialist that rents reserved, leading-edge GPU cluster time to third parties, financed like a fleet) is an airline. China nationalized the airline.

• The MW bar is already cleared. Runze is operating ~750 MW with ~6 GW planned. VNET wholesale in-service is 907 MW. GDS designed IT is an estimated ~1.56 GW. CoreWeave is at 1.5 GW active. Nebius is targeting 800 MW–1 GW connected by year-end 2026. China is not short of halls.
• The GPU bar is not. NVIDIA China DC Hopper shipments went from $4.6B in Q1 FY26 to $0 in Q1 FY27. The FY2026 10-K says the company is “effectively foreclosed” from China DC compute. Q2 FY27 guide assumes $0. Independents do not sit on the H100 / H200 / B200 / GB200 / GB300 allocation stack. 国产万卡 (a domestic 10,000-card cluster on a non-CUDA compiler) is a different product. 曙光8000 is a national supercomputer, not a checkout page.
• The customer bar is not. US overflow from Microsoft, OpenAI, Meta, and Anthropic created CoreWeave. In China the overflow stays inside ByteDance, Alibaba, Tencent, or 国家超算 (the national supercompute internet). ByteDance’s 2026 domestic IDC (internet data-center) demand is ~1.4–1.5 GW, and 85–90% of that is leased halls, not rented GPU hours. DeepSeek building captive DCs is the tell. Kingsoft’s Xiaomi line is a captive, not a demand stack.
• The capital bar is not. CoreWeave FY2026 capex is $35–39B. Kingsoft, the high-water mark for a listed Chinese independent, is guiding RMB 15–20B (~$2–3B) and 31% of revenue is Xiaomi. UCloud’s 定增 (a follow-on / private placement) is ≤RMB 1.5B. There is no RMB GPU-ABS (asset-backed securities) market.
• If you want “China NBIS” as a stock, it does not exist. If you want “China AI MW,” it trades as Runze 300442, GDS, and VNET. A terrestrial GPU factory in China is possible and is arguably overbuilt versus independent demand. A terrestrial neocloud in the CRWV / NBIS sense is blocked on GPUs and customers, not on land or power.

1. Define the object

A neocloud is not a colo (colocation: you rent the hall and bring your own kit) with a GPU SKU bolted on. It is not a hyperscaler renting you an A100 for an afternoon. It is not a token API wrapper sitting on someone else’s cluster. The Western neocloud debate has been about power and NVIDIA allocation. The China debate is about whether the renter exists.

Four tests, all required:

1. Gets leading-edge GPUs. Not last-gen remnants. Not a 国产 substitute on a different compiler.
2. Builds 万卡-scale IB (InfiniBand) or Ethernet fabrics.
3. Sells reserved cluster time to third parties at high ACV (annual contract value). Captive Xiaomi load is not this. A national supercomputer login is not this.
4. Finances GPUs like aircraft. Residual-value risk (what the card is worth when the lease ends) sits on the balance sheet.

Fail any one and you are something else. Fail all four and you are a landlord with a liquid-cooling brochure.

CoreWeave is the reference recipe. Q2 2026 (11 August PR): 1.5 GW active at 30 June. Contracted ~3.7 GW at quarter-end, ~4.2 GW as of 11 August. Year-end 2026 active target above 1.85 GW. 51 data centers. Q2 revenue ~$2.6B. Revenue backlog cited at $104.2B. That is a GPU airline with a hyperscaler frequent-flyer program, an NVIDIA allocation window, and a debt market that treats Blackwell the way aircraft lessors treat a 737.

Nebius is the other recipe. Active power YE 2025 ~170 MW. Connected-power target YE 2026: 800 MW–1 GW. Contracted-power target YE 2026: 5 GW. Q2 2026 group revenue $582.3m, AI cloud $574.9m. ARR (annual recurring revenue) $3.0bn at end-June. Q2 capex ~$5.7bn. Cash $8.0bn at 30 June. ACV on Q2 core deals above $20m per MW. Mäntsälä 75 MW live. NVIDIA put $2B of equity in and gave it Exemplar status on GB300 NVL72.

MW is an input. The product is reserved, leading-edge, third-party cluster time, financed like a fleet. China has the hangar. It does not have the carrier certificate.

2. China already cleared the MW bar

Stop pretending the constraint is land, power, or a liquid-cooled building in Langfang. The constraint is not the building.

Read the table left to right. CRWV and NBIS sell cluster-hours. Runze, VNET, and GDS sell halls. Kingsoft and UCloud sell some GPU time, at a scale that does not print a $5bn quarter of capex and does not produce a $104bn backlog.

These are airframes. Neoclouds are airlines. China nationalized the airline. The buildings are listed. The fleet is not for sale to third parties on a CoreWeave contract. ByteDance, Alibaba, and Tencent fly their own metal. 国家超算 flies the rest. CAICT 2025: 3.8B 智算 card-hours supplied, 1.4B used (36.8%). Some 国产 pools 70–80% idle. China overbuilt generic 智算 (intelligent compute) and underbuilt rentable CUDA 万卡. That is the opposite of the US neocloud shortage.

A 6 GW Runze plan against a UCloud that did ~RMB 685m of AI-related revenue in 2025 is not a tight market. It is a hall looking for a tenant who already has a hall.

3. Four missing ingredients

A. The GPU

This is no longer a vibe. It is a Federal Register.

BIS final rule FR 2026-00789, effective 15 January 2026: case-by-case for direct US exports only if TPP (Total Processing Performance, the BIS compute metric) <21,000 and DRAM bandwidth <6,500 GB/s, plus a 50% of US-end-use quantity cap, plus US-HQ lab testing of each shipment. Reexports from third countries remain presumption of denial. The 8 December 2025 “opening” is a trapdoor.

CoreWeave had HGX B300 GA (general availability) in March 2026 and is first-wave Vera Rubin NVL72 in 2H2026. Nebius took $2bn of NVIDIA equity and Exemplar on GB300 NVL72. That channel is jurisdictionally closed to a PRC entity.

31 May 2026 BIS: license required for advanced chips to any entity whose ultimate parent is HQ’d in China / Macau, wherever located. That killed the Singapore / Malaysia “offshore neocloud” as a Blackwell backdoor. Remote rental of overseas NVIDIA is still legal and under BIS review. Different product: workloads cross the border, cards do not.

国产万卡 is a different product. DeepSeek V4 inference was co-designed with Ascend 950DT (SemiAnalysis via Pandaily, 15 Jun 2026). V4 training still ran on NVIDIA (ChinaTalk). Huawei Cloud 昇腾云 had 2,663 customers by YE2025. China Mobile’s 2026–27 超节点集采 is 6,208 CANN cards, bids ~RMB 2.06B. That is the real China GPU cloud at training-relevant scale, and it is a stack owner plus SOE, not a merchant neocloud.

曙光8000 登峰 went live 10 July 2026 in Zhengzhou: first all-domestic 10万卡 cluster, Haiguang, national supercompute internet. Genuine industrial achievement. Not a store. 联通中卫 has >14,000 GPUs in the ground and >300 MW / >10万卡 on the plan. Watch it. Do not re-rate it as CRWV.

Without the GPU, the 200 MW Langfang building is a spa. (The HBM / Ascend pairing ceiling that sits behind 国产 silicon is the companion argument in Why Huawei’s Bet Isn’t on the Chip.)

B. The customer

CoreWeave was not created by cheap Oklahoma power. CoreWeave was created by overflow. Microsoft, OpenAI, Meta, Anthropic: the labs that could not wait for Azure and GCP spilled demand onto a specialist that already had the NVIDIA relationship and the willingness to lever GPUs like aircraft. That is why CRWV can cite a $104.2bn backlog and why NBIS can print ACV above $20m per MW. (NVIDIA writing checks into those same names is the flywheel.)

In China the overflow never leaves the building. ByteDance 2026 domestic IDC demand is ~1.4–1.5 GW, Q1 already ~1 GW tendered. Self-build is only ~10–15%. The rest is leased IDC from the landlords in the table. That is megawatt leasing, not GPU-hour ACV. Alibaba’s 2026 IDC plan is ~2 GW on the same pattern. Tencent trains on Tencent metal.

The Chinese Anthropic is the hyperscaler, or it is a lab the state would prefer inside the national supercompute internet. DeepSeek is the tell: V4 inference on Ascend, V4 training still NVIDIA, and 2026 recruiting for a Ulanqab 智算 campus. The one Chinese lab the West actually watches is not becoming a CoreWeave customer. It is becoming a landlord of its own.

The mid-tier labs that should be the TAM do rent. 智谱 2025 R&D was RMB 3.18B, more than 70% compute; compute as a share of R&D went from 17.3% (2022) to 71.8% in 2025H1. MiniMax is in the same squeeze. That demand is real, and it is inference / token shaped. It is not a Microsoft $60bn reserved-IB offtake.

Kingsoft is the listed name people want to be the exception. Q1 2026 AI gross billing ~RMB 1.0bn, 50.1% of public cloud. Xiaomi ecosystem RMB 838m, 31.0% of total revenue. Related-party cap raised to RMB 14.2bn (2025–27). Captives are real revenue. Captives are not a neocloud demand stack.

Jiemian, 2 July 2026, unverified by the companies: Alibaba on a 5-year lease for 3,000+ 8-GPU servers, Xiaomi GPU budget on Kingsoft from ~RMB 4bn to >RMB 10bn. Even if every word is true, a hyperscaler using a smaller cloud as a buffer is the opposite of the US pattern. In the US the hyperscaler is the source of overflow onto the neocloud. In this telling the hyperscaler is the tenant of the smaller cloud.

No third-party foundation-model overflow, no neocloud.

C. The capital

A neocloud is a capital-markets vehicle that turns NVIDIA allocation plus hyperscaler offtake into levered cluster-hours.

On 2025–26 evidence a Chinese independent cannot raise $5B+/year of GPU capex without being an SOE or a hyperscaler captive. There is no GPU residual market in RMB, no NVIDIA rent-back floor, no IG (investment-grade) offtake. STAR / HK 定增 is RMB 1–6B, not $10B HY GPU ABS. Policy banks fund buildings and 绿电, not $30k–$40k accelerators at CoreWeave velocity. Kingsoft at RMB 15–20B is the high-water mark, and 31% of revenue is Xiaomi. That is a captive cloud with a public listing, not a neocloud.

If the card is 昇腾, the secondary market is other domestic inference shops and the state. If the card is a legal NVIDIA SKU two generations behind what CoreWeave is racking, the Western renter will not show up. There is no China equivalent of the used-GPU bid that makes the aircraft-lease model close. Concrete is a REIT. China already has those. They are called GDS, VNET, and Runze.

D. The power product

Nebius is in Mäntsälä, not Zhangjiakou, for a reason that is not “Finland has empty land.”

Finland is a power product. A Western lab’s sustainability team can take it to a board. China’s AI MW is in 蒙西, 宁夏, 甘肃, 河北. Ningxia 2024 ~76% thermal. Gansu 2024 ~52% thermal. Zhangjiakou city generation can be ~80% renewable; DC consumption green share was about one-third in H1 2025, official. 80% 绿电, for a large share of the claimed figure, is 绿证 (green certificates — paper, not the kWh in the hall).

Domestic labs do not care. Global labs do. They will not put a multi-month pretraining run in a coal-majority grid dressed up with certificates. Sichuan is the honest exception on kWh, and it is not where the 6 GW of planned AIDC is clustering.

The Western analogy on power is also the wrong one. NDA director 刘烈宏, China Development Forum 23 March 2026: hub-node new compute facilities must hit 80% green-power share. NDRC weights; enforcement from 1 August 2026. Combined with PUE (power usage effectiveness) ≤1.2 at hubs (行动计划, July 2024) and a freeze on new large DCs in cities where existing DCs are >1 year old and utilization is <50%. A CoreWeave-speed greenfield by a nobody is a planning violation, not a construction problem. 中卫 110 kV in ~3 months is what a telco gets. Power in China is permissioning. In Texas it is a queue. Different. Both bind.

IREN Horizon 1 getting Microsoft acceptance on 13 August 2026 for 50 MW, against a 480 MW AI Cloud year-end target, is the Western version of this point. The offtaker is buying a power product and a GPU product together. China can sell the power. It cannot sell the bundled product to the offtaker that would make the equity story work.

4. What actually exists

Three products get collapsed into one conversation. They are not one conversation.

(i) The AIDC landlord. Runze, Chindata, VNET, GDS, 数据港, 奥飞. Combined they already clear CoreWeave’s live GW and are planning past it. The product is space, power, cooling. The tenant brings its own cards. GDS at a derived ~1.56 GW designed IT, 50/50 CPU/GPU, is the cleanest large-cap expression of “China has the halls.” A neocloud does not have a 50% CPU business. A neocloud’s CPU business is the login node. Buy these if you want China AI MW. Do not buy these if you want residual-value leverage to NVIDIA. There is no NVIDIA. There is a REIT with a GPU-shaped tenant mix.

(ii) GPU 零售 / 调度. AutoDL: ~30,000 cards, owned plus scheduled. The right product for a researcher who needs a handful of GPUs this weekend. The wrong product for a 万卡 reserved training run. SiliconFlow and the rest of the 调度 layer improve utilization of a fragmented fleet. They do not build the fleet and do not sit on NVIDIA’s allocation stack. CoreWeave is not a better AutoDL. CoreWeave eats foundation-model overflow.

(iii) Closest-to-neocloud. SenseCore is the closest listed AI-native stack: 40,400 PFLOPS FP16 at YE2025, GenAI revenue RMB 3.63B (72% of 2025 sales), WAIC 2026 plan of at least five domestic 万卡 clusters. Plan, not census. ACV is model plus token, not reserved IB. Kingsoft is the listed name with actual AI billing and a Xiaomi-shaped P&L. UCloud is the only name with a public card census worth using: 24,131 cards, GPU 91.29% sold, max single cluster ~16,000, Wulanchabu 61.7 MW usable. 24,131 cards is Voltage Park’s neighborhood on count (36k) and nowhere near Lambda’s 10k+ GB300 in one factory on generation. 61.7 MW is a rounding error on Runze’s 750 MW and on Nebius’s 800 MW–1 GW connected target.

The tell is 基流. 36Kr frames it against CRWV / NBIS as cluster EPC, a 包工头 (a general contractor). Revenue RMB 32m / 325m / 520m (2023–25). More than 90,000 GPUs delivered across 66 projects, two 万卡. China professionalized building 万卡. The US professionalized renting them against IG paper.

Huawei Cloud 昇腾云 is the actual China GPU cloud at training-relevant scale. It is a stack owner. It is not an independent. Allocation of 950s goes to ByteDance, Alibaba, Tencent first.

5. The only doors that open

Four doors. The MW door is already open and it does not get you a neocloud.

Door 1. 国产 chips become good enough for the median renter, not the frontier lab. If that median renter is most of the market by card-hours, a China-native GPU cloud can exist as a 国产 product. It will still not be CRWV. Watch 昇腾 rental GMV, not 万卡 ribbon-cuttings.

Door 2. A telco SOE runs a commercial 万卡 catalog at CRWV SLAs. 联通中卫 is the prototype. If a third party can reserve a 万卡 fabric at a published SLA without a ministry letter, China has a neocloud-shaped service. The equity lives inside a telco, not a NBIS ticker.

Door 3. Export controls reverse and NVIDIA needs a China channel. Not a one-quarter H200 headline. A repeating specialist allocation the way CoreWeave gets Blackwell. Until a specialist is actually receiving that contracted allocation, model zero.

Door 4. Global customers accept 绿证. They will not, for training. Sichuan hydro is the only generation mix in the hub set that looks like a Nordic PPA on kWh rather than on paper. The 6 GW of planned halls are not clustering there.

What does not open a door: another 200 MW liquid-cooled building, another 东数西算 subsidy, another 定增, another WAIC 万卡 announcement. Those are landlord prints. They confirm the MW table. They do not move the four missing ingredients.

6. Net net

A terrestrial GPU factory in China: possible, and on the MW evidence arguably overbuilt versus independent demand. A terrestrial neocloud in the CRWV / NBIS sense: blocked on GPUs and customers, not on land or power. If you want “China NBIS” as a stock, it does not exist. If you want “China AI MW,” it trades as Runze 300442, GDS, and VNET.

The airline is missing because the four ingredients are missing together. NVIDIA China DC compute is at $0 and Blackwell is presumption of denial. No US-style overflow: ByteDance / Alibaba / Tencent keep the demand, Kingsoft is a Xiaomi captive, DeepSeek is building its own halls. No GPU-as-aircraft capital stack: Kingsoft’s $2–3B is the ceiling, UCloud’s 定增 is RMB 1.5B, there is no residual-value bid for 昇腾 that a private-credit fund will lever. No power product a global lab will buy for training: 蒙西 / 宁夏 / 甘肃 generation is still thermal-majority in kWh, 80% 绿电 is 绿证, and that is why NBIS is in Mäntsälä.

Confusing landlords with neoclouds is how people compare GDS to NBIS. Stop doing it. GDS is a very good way to own China AI MW. NBIS is a way to own a Nordic-plus-global GPU airline with $3.0bn of ARR, $5.7bn of quarterly capex, and ACV above $20m per MW. Those are different cash-flow identities. They will not converge because a building in Langfang learned to spell “GB200.”

The median outcome is the one already on the ground: SOE and hyperscaler captive clusters behind landlord MW, a retail / 调度 layer for the long tail, Huawei Cloud as the stack-owner cloud, and one or two listed GPU clouds that look like Kingsoft and UCloud. Real, growing, captive-heavy, 国产-tilted, and not CoreWeave. The bull outcome is Door 1 plus Door 2: 国产 chips good enough for the median renter, and a telco SOE that actually sells 万卡 by the hour. That is a China-native 智算 utility. It is still not NVIDIA-native third-party overflow. Do not pay CRWV multiples for it.

Watch:

• Kingsoft 智算 mix. Does Xiaomi stay above 50% of 智算, or does third-party reserved ACV show up as a disclosed line with a customer count that is not two names?
• UCloud card census. Next print: count, generation mix, Wulanchabu MW, AI revenue versus the 2025 ~RMB 685m. If cards 3x and revenue does not, it is a 国产 ASP problem, not a neocloud ramp.
• H200 units actually landed after Kessler’s “very few” (14 Jul 2026), and whether any independent (not Ali / Tencent / ByteDance / JD) ever appears on a license list. One SKU, one quarter, one exception is not an allocation stack.
• 昇腾 rental GMV. Ribbon-cuttings are theater. GMV, utilization, and repeat third-party reserved hours on 国产 silicon would say the median renter has moved.
• 联通中卫 commercial catalog. The question is whether a third party can reserve a 万卡 fabric at a published SLA without being the state.
• BIS remote-rental of overseas NVIDIA. Still legal, under review. If banned, even the Singapore workaround for workloads dies.
• Firmus Batam rack date. Q1 2027 campus; 170,000 accelerators through 2027–28. Watch whether the $25–30bn offtake shows up as signed ACV, not a Bloomberg sentence.
• Yotta D2 / SoftBank GPU Cloud. Yotta 20,736 B300 targeted live August 2026. SoftBank “AI Data Center GPU Cloud” October 2026. Legal NVIDIA. Sovereign Lambda, not CoreWeave, until someone other than a telco or a state mission writes ten-billion take-or-pay to an independent.

The megawatts were never the question. The company is.

The hangar tour

Wulanchabu. Year-end 2025: 84 signed DC projects, capex about RMB 39.975bn. Signed racks: more than 4 million. Built: 500,000. In operation: 330,000. Occupancy: 66% (Cailian, 26 Jun 2026). UCloud cabinets above 95% at 30 June 2026; GPU 91.29%. Building D is 75% intent kilowatts, not signed leases. Zhongjin Bayin: dedicated 220 kV, own wind and PV, year-1 renewable replacement 38.74% (Xinhua, 18 Jul 2025). Alibaba’s GM claims 90% local green. Compare Zhongjin. DeepSeek is hiring civil engineers.

Zhongwei. Unicom buildings 1–2 storage, 3–4 training. Monthly power bill RMB 10–13 million (Qi Jun, Guangming, 25 May 2026). Dec 2025: 14k GPUs / 85% / 98 MW IT. Apr 2026 visit: >70k cards. Datang 500 MW PV is physical; 1.5 GW wind is a virtual PPA. China Mobile Ningxia bought 78% green kWh in 2025.

Qingyang. Securities Times walked off Xifeng airport past a “China Computing Valley” sign. Mobile park: Kingsoft, Ali, MiniMax, Kimi, Suiyuan; immersion tanks, almost no fan noise. Telecom phase-1 occupancy ~97%. Kingsoft 2025 capex >RMB 5bn, cluster light-up 18 May 2026. Tariff 0.398 via aggregation. Grid company: wind / solar “don’t obey,” halls want 7×24.

Zhangjiakou. Sun Jun, Hebei Daily: H1 2025 big-data use 4.071 TWh, green share about one-third (People.cn, 13 Jul 2025). Do not confuse with 82.88% green transaction share. Chindata Huailai HQ >200 MW IT. Runze Langfang 200 MW all-liquid: sold out, still commissioning as of April 2026 IR. No August site photo of live IT.

Lingang. Gu Ruoyu drives 60 km / one hour every morning. Buildings 1 and 3 have not been below 90% occupancy since Sep 2023 (People.cn, 14 Jul 2026). Customers who will not train in Qingyang. Latency is the product.

The idle is real, and it is not on this tour. CAICT: 3.8bn card-hours supplied, 1.4bn used, 36.8%. Shi Ke’s “digital unfinished buildings” is 2025 CPPCC. Hub cities do not have named idle MW-class halls in 2026 visitor copy.

Can Singapore — or Asia — have a neocloud?

Southeast Asia is not blocked on NVIDIA. Singapore island is blocked on power. China-parented vehicles in Singapore and Malaysia are blocked on BIS. Those three sentences are different businesses. Mixing them is how “SEA has so many cards” becomes a neocloud thesis.

Singapore the island cannot host a training neocloud

Live IT capacity is about 1.1 GW, with BMI putting 2026 live capacity around 1.46 GW, still the largest in Southeast Asia (TechNode Global / BMI, 29 Jul 2026). Data centres already use about 7% of national electricity, heading toward 12%. Land is 745 km². After the 2019 moratorium, new MW is rationed.

DC-CFA1 (2023, the first Data Centre Call for Applications) awarded 80 MW to four names: Equinix, GDS, Microsoft, and an AirTrunk–ByteDance consortium. DC-CFA2, launched 1 Dec 2025, closed 31 Mar 2026, allocates at least 200 MW, PUE 1.25 or better, Green Mark Platinum, ≥50% eligible green energy (Mondaq; TechNode, 31 Jul 2026). BMI notes only ~20 MW under construction against ~980 MW in the development pipeline. Jurong Island has been earmarked 20 hectares for a low-carbon park that could take up to 700 MW if built. That is a planning envelope, not a 2026 hall.

What Singapore is good for: low-latency inference, interconnect, banking and model-API front ends, a corporate HQ that can sign NVIDIA paper. What it is not good for: a 1 GW training factory. A CoreWeave-class campus is a power product. Singapore sells scarcity.

The MW is across the causeway and across the strait

Johor is the spillover. Operational capacity is often cited around 1–2 GW in 2025–26; committed and planned pipelines are talked about at 5 GW or more, with some consultants stretching to 8–12 GW by 2030. Treat the high end as pipeline, not live IT. Clusters: Sedenak, Kulai, Nusajaya / Iskandar Puteri.

Named campuses (announced, mixed live/pipeline):

• YTL + NVIDIA, YTL Green Data Center Park, Kulai: up to ~500 MW phased; first NVIDIA-powered hall about 20 MW, live Oct 2025.
• Vantage JHB1, Sedenak: 300 MW campus potential, Phase 1 go-live Jan 2026, liquid-cooled, 100 kW/rack class.
• Princeton Digital Group JH1, Sedenak: ~150 MW AI-ready.
• AirTrunk Johor ~150 MW early phase; Yondr 200–300 MW class; GDS Nusajaya / Kulai phased toward ~280 MW.
• ByteDance / TikTok Johor: reported 1–2 GW pipeline, RM10 bn. Pipeline.
• AWS, Microsoft, Google Malaysia regions: hyperscaler cloud, not a specialist GPU cloud.
• Equinix JH1/JH2: carrier-neutral, tens of MW.

Batam is the other lung. Firmus (Australia, Coatue round at $5.5bn, NVIDIA-backed) + Singapore-headquartered DayOne: 360 MW NVIDIA DSX AI Factory, up to 170,000 Grace-Blackwell / Vera-Rubin / Vera accelerators through 2027–28, campus targeted Q1 2027 (Firmus; DCD). NVIDIA takes product revenue plus a share of cloud revenue. Firmus has talked $25–30bn offtake over six years. That is a target, not a backlog print like CoreWeave’s $104bn. DayOne is already building; it has a 450 MW PPA at Kabil Tech Park. Runze’s planned Batam ~360 MW is a Chinese landlord going overseas. The tenant still brings the cards.

This is the actual “Singapore neocloud” geography: company and network in Singapore, factory in Johor or Batam. Same latency band, different power politics.

“SEA has so many cards” is three inventories mixed together

Inventory A: hyperscaler and landlord MW. AWS / Microsoft / Google / ByteDance halls, GDS / AirTrunk / Vantage shells. Cards belong to the hyperscaler. Not a rental catalog.

Inventory B: the China cutout, now being unwound. Megaspeed (Singapore, spun out of a Chinese gaming company in 2023) bought nearly $2bn of advanced NVIDIA product through a Malaysian subsidiary, into Malaysia and Indonesia DCs that appeared to remotely serve Chinese clients (CNBC / NYT, 10 Oct 2025). Bridge Data Centres (Bain) had allocated 68.4 MW to Megaspeed in Malaysia; by Feb 2026 that capacity was given to Zenlayer (Bloomberg via Edge, 8 Apr 2026). NVIDIA did spot checks. Megaspeed denies illegal transfers. BIS, 31 May 2026: a license is required for advanced chips to any entity whose ultimate parent is headquartered in China or Macau, wherever located. Reexports stay presumption of denial. NVIDIA also tightened SG / MY / JP approved-buyer lists. Malaysia started NVIDIA export permits in 2025.

That inventory was real cards. It was not a local neocloud TAM. It was China demand sitting in Johor and Batam. The parent-HQ rule is designed to kill it. Remote rental of remaining SEA Blackwell by PRC labs is still legally grey and under BIS review.

Inventory C: genuine non-China GPU cloud, mostly not live yet. Firmus Batam (Q1 2027). YTL Kulai first 20 MW (live, small). Yotta in India (below). SoftBank in Japan (Oct 2026). These can buy Blackwell because the parent is Australian, Malaysian-listed, Indian, or Japanese.

If you add A+B+C and call it “SEA neocloud capacity,” you have counted shells, a closing loophole, and press releases. Only C is the object. C is still small, and a lot of it is 2027.

Japan: sovereign neocloud, Lambda-scale, legal NVIDIA

SoftBank will launch “AI Data Center GPU Cloud” on GB200 NVL72 plus its Infrinia OS in October 2026, and uses the word neocloud in the press release (SoftBank, 25 May 2026). Sakura Internet: Koukaryoku expanding toward ~10,800 GPUs, including HGX B200 at Ishikari, targeting 100% renewable by 2027. GMO GPU Cloud on H200; KDDI planning GB200 NVL72; Highreso Kagawa; Rutilea >1,000 Hopper. METI-aligned. Cards are legal. Customers are domestic enterprises, robotics, telco, healthcare. Power is expensive; nuclear restart helps. This is a sovereign Lambda, not CoreWeave. There is no Microsoft $60bn overflow onto an independent Japanese specialist. SoftBank is the telco-conglomerate doing it itself.

India: the other legal NVIDIA pool

Yotta: 20,736 liquid-cooled Blackwell Ultra (HGX B300) at the 60 MW D2 hall in Greater Noida, targeted live August 2026, >$2bn, Quantum-X800 InfiniBand (Yotta PR, 18 Feb 2026). More than 10,000 of those B300s committed to the IndiaAI Mission. Separate four-year, >$1bn NVIDIA DGX Cloud engagement, billed as one of APAC’s largest. Yotta says >10,000 NVIDIA GPUs already live, another 8,000 near-term, roadmap >80,000 by FY27. Navi Mumbai campus scalable to 2 GW. Treat the million-GPU long-term line as marketing.

This is closer to a national AI factory with a commercial window than to CoreWeave. A large slice is sovereign. The commercial slice is real NVIDIA, legal, liquid-cooled, IB. Capital is Indian + NVIDIA engagement, not GPU-ABS against Meta / Microsoft paper.

Korea and the rest

Korea: SK Telecom in the telco-neocloud club (with NTT, Deutsche Telekom, Orange). National AI Computing Center broken ground 2026, target 15,000 AI semiconductors by 2028. Captive and sovereign. Taiwan: TSMC already fights for power; a specialist GPU cloud is politically and electrically crowded. Australia: Firmus’s home, CDC, plenty of MW on paper, far from Asian users. UAE / Saudi (G42, Humain) are the non-US sovereign GPU clouds that actually look like a state-backed CoreWeave. They are not Southeast Asia, but they are the template “Asia-Pacific” bulls keep reaching for.

Mapping onto the three Western products

So can Singapore have one? Can Asia?

Singapore-the-island: no, not at training scale. 80+200 MW of new approvals against a 1 GW+ live base, 7–12% of national power, is a boutique. Inference, yes. Training factory, no.

Singapore-the-company: yes, if the parent is not China. Firmus+DayOne is the working drawing. HQ, contracting, NVIDIA relationship in Singapore; 360 MW and 170k GPUs in Batam. That can be a neocloud in the Lambda-to-small-CRWV sense if the $25–30bn offtake shows up as signed ACV, not a Bloomberg sentence, and if the GPUs actually rack in 2027.

Southeast Asia: not blocked on NVIDIA for non-China parents. Blocked on confusing three inventories. Johor’s GW pipeline is mostly colo and hyperscaler. The China-linked racks are a shrinking stock. The new legal GPU clouds are 20 MW live (YTL), 170k GPUs promised (Firmus, 2027), plus whatever honest local clouds remain after whitelist cuts.

Asia as a whole: Japan and India can and are building legal NVIDIA GPU clouds. They look like sovereign Lambdas and national factories. They do not look like CoreWeave until someone other than a telco or a state mission writes a ten-billion-dollar take-or-pay to an independent. Firmus is the only name in the Singapore orbit even claiming that offtake. Watch the 2027 rack date and whether those contracts are real.

The China question does not get a Singapore escape hatch. BIS parent-HQ closed it. A Wulanchabu training run cannot legally sit on Batam Blackwell through a Singapore shell anymore. Remote rental is the leftover grey. That is a workload path, not a campus path, and it is on the BIS watch list.

Cards in Southeast Asia were never the scarce object. Clean title to those cards was.

References

• Cailian — Wulanchabu DC census (26 Jun 2026)
• Xinhua — Zhongjin Bayin 220 kV / 38.74% year-1 renewable (18 Jul 2025)
• Guangming — Unicom Zhongwei, Qi Jun (25 May 2026)
• People.cn / Hebei Daily — Zhangjiakou H1 2025 big-data kWh (13 Jul 2025)
• People.cn — SenseCore Lingang occupancy (14 Jul 2026)
• TechNode Global / BMI — Singapore DC capacity (29 Jul 2026)
• Mondaq — Singapore DC-CFA2 (pilot to power play)
• TechNode — Singapore tightens DC growth vs Johor / Bangkok / Jakarta (31 Jul 2026)
• Firmus — 170,000-GPU Batam campus with NVIDIA
• DCD — Firmus 170,000-GPU cluster in Batam
• CNBC / NYT — Megaspeed NVIDIA export-control probe (10 Oct 2025)
• The Edge — Bridge reallocates 68.4 MW from Megaspeed to Zenlayer (8 Apr 2026)
• SoftBank — AI Data Center GPU Cloud (25 May 2026)
• Yotta — 20,000 NVIDIA Blackwell Ultra GPUs (18 Feb 2026)
• Part 1 — A Primer (15 August)
• Companion — Why Huawei's Bet Isn't on the Chip
• Companion — NVIDIA Is Buying Its Own Demand