Data Center Market Update: Week of August 3, 2026

This week’s data center news centers on Southeast Asia and South Korea: Nvidia and SK Group signed a $500 billion AI partnership, CoreWeave made its first Asia-Pacific move with a 360MW Indonesia build, AirTrunk closed Malaysia’s largest-ever data center loan, and Singapore’s capacity caps kept pushing growth to its neighbors. GPU rental prices held steady, with B200 still commanding a 75-150% premium over H200.

Key takeaways

• Nvidia and SK Group’s $500 billion partnership bundles a 2GW AI factory (online 2027) with a long-term HBM supply deal — SK Telecom aims for 15GW of AI data center capacity nationally by the early 2030s.
• CoreWeave enters Asia-Pacific for the first time: three Indonesian data centers totaling 360MW, targeting a 2028 launch.
• A four-country Southeast Asia colocation pact (SB Telecom, SC ZEUS, Robust HPC) targets Malaysia, Thailand, Indonesia and Vietnam, effective 1 August 2026.
• AirTrunk secured US$2.325 billion in green financing for its 270MW+ JHB2 campus in Johor Bahru — Malaysia’s largest single-asset data center loan.
• Singapore’s live capacity (1.1GW+) is effectively capped by its DC-CFA licensing regime, while Johor (1GW+, 0.7% vacancy), Jakarta (344MW) and Bangkok (122MW, growing to 402MW by 2027) expand unconstrained.
• US colocation demand persists alongside the AI campus wave: Flexential (108MW, Dallas-Fort Worth) and Rowan ($780M, 300MW, Temple, Texas) both announced Texas expansions.
• B200 GPU rentals run $7-10/GPU-hour, a 75-150% premium over H200’s $3.99/hour; H100 spans $1-7.50/hour depending on channel.

Nvidia and SK Group’s $500 billion AI partnership

Announced 25 July 2026, the deal centers on a 2GW AI factory that SK Telecom will build on Nvidia’s Vera Rubin DSX platform in South Korea’s Gyeongsang region, targeted for 2027. SK hynix locks in as Nvidia’s HBM memory supplier for the life of the agreement. The $500 billion figure spans the full multi-year commercial relationship — chip purchases, supercomputer acquisitions and AI factory infrastructure — rather than a single upfront capex line, per SK hynix’s newsroom disclosure. SK Telecom’s stated ambition is 15GW of national AI data center capacity, with 5GW opening in stages from 2029.

CoreWeave’s first Asia-Pacific move: 360MW in Indonesia

CoreWeave will build and operate three data centers in Indonesia totaling 360MW of IT power, its first facilities outside the US and Europe, targeting a 2028 launch. The company reported more than 1GW of active power and 3.5GW of contracted power globally as of 31 March 2026. The Indonesia move follows the same logic driving multiple operators to the archipelago this year: proximity to Singapore’s demand pool without Singapore’s capacity ceiling.

A four-country Southeast Asia colocation pact

SB Telecom Singapore (a SoftBank subsidiary), Singapore’s SC ZEUS Data Centers and Malaysia’s Robust HPC Group signed a memorandum effective 1 August 2026 to jointly develop AI data center infrastructure across Malaysia, Thailand, Indonesia and Vietnam, with Malaysia first. The division of labor is explicit: SC ZEUS handles land, design and operations; Robust HPC handles GPU cluster procurement; SB Telecom handles data hall fit-out and connectivity. Facilities will use liquid cooling and 800 VDC architecture — see our liquid cooling guide for what that implies for rack density.

AirTrunk’s $2.3 billion Johor Bahru loan

AirTrunk secured US$2.325 billion in green project financing for its JHB2 campus in Johor Bahru — Malaysia’s largest single-asset data center financing, backed by 30 lenders including BNP Paribas, DBS, HSBC and UOB. JHB2 adds 270MW+ on a 520-hectare site that already hosts the 150MW JHB1 campus, targeting a 1.37 design PUE. It is the clearest signal yet that institutional lenders now treat Johor hyperscale campuses as investment-grade infrastructure, not speculative build.

Singapore’s ceiling keeps pushing growth to its neighbors

Singapore hosts over 1.1GW of live IT capacity across 60+ facilities — the region’s largest single market — but new capacity has been rationed since 2019. Its DC-CFA rounds awarded just 80MW in 2023 and roughly 200MW+ from December 2024, with new entrants required to hit a 1.25 PUE and source at least 50% renewable electricity. The consequence is visible in the numbers: Johor already exceeds 1GW live with 0.7% colocation vacancy against Singapore’s 4.9%, Jakarta sits at 344MW and is projected past 1GW by 2027, and Bangkok is forecast to grow from 122MW at end-2025 to 402MW by 2027. See our Bangkok vs Johor vs Singapore comparison for the full market breakdown.

Market	Live capacity (2026)	Growth constraint	2027 outlook
Singapore	1.1GW+	DC-CFA licensing cap, 1.25 PUE / 50% renewables required	Incremental only
Johor, Malaysia	1GW+ (0.7% vacancy)	Land and power available	AirTrunk JHB2 +270MW live
Jakarta, Indonesia	344MW	Grid buildout pace	1GW+
Bangkok, Thailand	122MW (end-2025)	Grid connection queue	402MW

US colocation still expanding despite the AI headline capacity

Two Texas deals landed in the same week as the Asia-Pacific news: Flexential acquired 110 acres in Kaufman County for a fourth Dallas-Fort Worth facility (36MW initial phase, 108MW at full buildout, 375,000 sq ft), and Rowan committed $780 million to a 300MW, 500,000 sq ft facility in Temple, Texas. Neither is AI-hyperscale scale, but both confirm that enterprise and cloud colocation demand hasn’t been crowded out by the AI campus wave — see our colocation vs cloud TCO guide for how buyers weigh the two paths.

GPU rental prices: B200 premium holds

B200 on-demand rates ran $7-10/GPU-hour across specialist clouds in early August 2026 — Nebius at $7.15, Lambda at $6.99, RunPod at $8.64 and up — a 75-150% premium over H200’s $3.99/hour. H100 pricing remains fragmented, spanning roughly $1-7.50/GPU-hour depending on channel, with hyperscaler rates running several times marketplace medians for identical silicon. Full comparison in our H100 rental price tracker and GPU cloud comparison.

GPU	Rate range (Aug 2026)	Notes
H100	$1.00-7.50/GPU-hour	Wide spread by channel; hyperscaler vs marketplace gap persists
H200	~$3.99/GPU-hour	Reference point for B200 premium
B200	$7.00-10.00+/GPU-hour	Nebius $7.15, Lambda $6.99, RunPod $8.64+

What this means for buyers

Southeast Asia expansion decisions should default to Johor or Jakarta/Batam rather than Singapore: the capital’s DC-CFA ceiling is a policy constraint, not a temporary shortage, and lender appetite (AirTrunk’s $2.3B) confirms Johor campuses now clear institutional due diligence. For GPU procurement, treat $7-10/hour as the current B200 benchmark and $1-7.50/hour as the H100 range — any quote materially above the upper end of these bands warrants a second quote via our live quote tool. Track live colocation benchmarks in the index, browse the underlying facility data in the data center catalog, and see aggregate pricing trends in stats.