The Singapore Shortage Is The Margin

On 8 May 2026 the Vast.ai marketplace listed exactly one usable part in Singapore: an RTX 4000Ada with 20 GB of memory at 0.04 to 0.08 dollars an hour. Nothing above 20 GB. No A100. No H100. No L40S. RunPod listed no Singapore region at all. Its regions were the United States, Europe and Canada. Lambda Labs ran the United States only. CoreWeave ran the United States and Europe. Three companies, and not one of them had a rack on the island. The global reference rates were published the same day. An A100 PCIe 80 GB rented from Italy at 0.13 dollars an hour. An H100 SXM 80 GB rented from France at 1.47 dollars an hour. Both are real listings. Both are thousands of kilometres from the buyer. So the buyer reads a table with two columns and one hidden third: the country. Singapore had the demand and did not have the parts.

The published requirement for a large open-weight model in production service is a 250 to 300 GB working set at 8-bit quantisation. At 4-bit the same weights fit inside about 150 GB. Two numbers, and both are memory numbers. Compute is not the constraint. Memory is the constraint. That single line rules out every consumer card in the country and every 20 GB listing on the marketplace. So the requirement converts into three shapes. A 320 GB box is four 80 GB accelerators. A 160 GB box is two. A 320 GB hybrid is two of each. Every provider row that follows is one of those three shapes, priced by the month, in United States dollars. The arithmetic is small. The decision is not.

Leaseweb Singapore runs at Digital Realty SIN10 and SIN11, with a private VLAN included in the base price. Two A100 80 GB cards at 160 GB of total memory price at 4,200 to 5,500 dollars a month. Four A100 cards at 320 GB price at 7,800 to 9,500. Two H100 cards at 160 GB price at 6,500 to 8,000. Four H100 cards at 320 GB price at 12,500 to 15,000. Each configuration ships with 64 or more CPU cores, 512 GB to 1 TB of RAM, two NVMe drives, a 10 Gbps port and 10 to 20 TB of monthly traffic. The setup fee runs 500 to 1,000 dollars. The minimum term is one month. One month minimum. No lock-in. That is the whole reason this row sits at the top of the sheet for a first cluster.

OVHcloud runs an Asia-Pacific bare metal line in Singapore on its Advance range, with vRack private networking included. Two A100 80 GB cards at 160 GB price at roughly 3,500 to 4,500 dollars a month. Four A100 cards at 320 GB price at 6,500 to 8,000. Two H100 cards at 160 GB price at 5,500 to 7,000. Four H100 cards at 320 GB price at 10,500 to 13,000. The included package is a layer 2 VLAN across the racks, DDoS protection, 500 Mbps to 1 Gbps of port speed and 10 to 30 TB of traffic. The minimum term is one month. vRack is included. For a private cluster that is the line that matters. A private network across two racks is a different product from two rented servers behind a firewall rule, and only one of them survives an audit.

A card that is not in the country is not capacity, whatever the hourly rate says.

Tencent Cloud Singapore publishes two shared configurations and one dedicated host. Two A100 cards at 160 GB list near 2,400 dollars a month shared. Four A100 cards at 320 GB list near 4,800 dollars shared. Four A100 cards on a dedicated host list at 6,500 to 8,500 dollars. Reserved terms of one year take 30 to 40 per cent off. Alibaba Cloud Singapore publishes a wider ladder. Two A100 cards at 160 GB list near 2,700 dollars a month. Four A100 cards at 320 GB list near 5,400. Four A100 cards on a dedicated host list at 7,000 to 9,000. Two H100 cards at 160 GB list at 5,500 to 6,500. Four H100 cards at 320 GB list at 10,500 to 12,500. A one-year reservation takes about 30 per cent off. Both cloud rows charge for traffic. Outbound runs 0.08 dollars per GB, a 100 Mbps committed plan near 300 dollars a month and a 1 Gbps plan near 2,500 dollars a month. Shared is cheaper per month and shared is still not dedicated. The neighbour decides your latency at peak hour, and you do not get a vote.

Traffic beyond the included allocation adds 500 to 2,000 dollars a month to any of these rows. A 320 GB box moving model weights and checkpoints between two sites will find that ceiling early. Supply is the other bill. GPU supply in Singapore is constrained, with lead times of two to eight weeks commonly quoted. A twelve-month commitment typically yields 10 to 20 per cent against the monthly rate. The setup fee, the private VLAN and the committed bandwidth are the three lines that get quoted last and hurt first. The probability that a quoted lead time holds is a function of the queue in front of it, not of the contract. Ask for the queue position in writing. Market pricing is an estimate. A binding number only exists on a signed quote.

Three cloud providers quote the same two-A100 box without a Singapore presence. RunPod lists 2,760 to 3,340 dollars a month on secure cloud. Lambda lists 1,606. CoreWeave lists 1,869. All three are cheaper than the local bare metal row, and none of them can put the machine in the country. Leaseweb quotes 4,200 to 5,500 for the same two cards. That is 1.5 to 1.6 times the nearest offshore equivalent. The premium is not for the silicon. The premium is for the postcode. For a 320 GB 8-bit box the cheapest published route is four A100 cards at 6,500 to 9,500 dollars a month from OVHcloud or Leaseweb, against four H100 cards at 10,500 to 15,000. A hybrid of two A100 and two H100 cards sits at 8,000 to 11,000. For a 150 GB 4-bit box, two A100 cards run 4,200 to 5,500 and two H100 cards run 5,500 to 8,000. The cheapest hour in the table is the most expensive month in the table. Read the last column, then sign the shortest term that fits the work.