ByteDance Builds Massive AI Supercluster in Malaysia with NVIDIA Chips

The TikTok parent company has found a way to access NVIDIA's most advanced processors without violating U.S. export restrictions — by building one of the world's largest AI computing clusters in Southeast Asia.

The Deal

ByteDance, the Beijing-headquartered company behind TikTok and the rapidly growing AI chatbot Doubao, is constructing an enormous artificial intelligence supercluster in Malaysia that will house approximately 36,000 NVIDIA B200 Blackwell chips arranged across roughly 500 NVL72 GB200 rack-scale computing systems [1][2]. The Wall Street Journal first reported the plans, placing the cost of the hardware alone at more than $2.5 billion [3].

The arrangement works through a third-party intermediary: Aolani Cloud, a Southeast Asian cloud computing startup established in late 2023 by a group of investors that includes Singapore-based venture capital firm K3 Ventures, with a holding company registered in the Cayman Islands [4]. Aolani holds NVIDIA's Tier-1 cloud partner designation, granting it priority access to the chipmaker's latest accelerators. The servers themselves are being built by Aivres, a company specializing in NVIDIA GPU-based systems [4].

Under the arrangement, Aolani Cloud formally owns and operates the hardware in Malaysia. ByteDance leases the computing capacity — a structure that allows the Chinese company to harness the power of NVIDIA's most advanced processors without the chips ever crossing into Chinese territory [5].

Why Malaysia? The Geography of Circumvention

The choice of Malaysia is no accident. U.S. export controls on advanced AI semiconductors, first imposed in October 2022 and progressively tightened since, primarily regulate where the hardware is physically shipped — not where the resulting compute is consumed [6]. As long as the Blackwell chips remain in a country not classified as "controlled" under U.S. regulations, their use by a Chinese entity does not automatically trigger a violation.

Malaysia sits in a regulatory sweet spot. It is categorized as a Tier 2 country under the January 2026 AI Diffusion Framework finalized by the Bureau of Industry and Security, meaning it can receive advanced AI chips subject to certain aggregate caps and compliance requirements, but is not subject to the near-total embargo applied to China, Russia, Iran, and North Korea [7].

ByteDance had already been leasing servers equipped with NVIDIA's previous-generation H100 GPUs from Aolani in Malaysia since February 2025 [4]. The Blackwell deployment represents a massive escalation — from a few thousand older chips to 36,000 of NVIDIA's fastest current-generation processors.

The company is not limiting its Southeast Asian buildout to Malaysia alone. The Wall Street Journal previously reported that ByteDance was also in talks to deploy more than 7,000 B200 chips at a data center in Indonesia [3], suggesting a broader regional computing strategy spanning multiple countries.

NVIDIA's Position: Technically Compliant

NVIDIA has confirmed it has no objections to the arrangement. "All Nvidia cloud partners are evaluated and cleared by Nvidia's field operations, finance, and compliance teams before they can receive our products, directly or through an OEM," a company spokesperson said [5].

The chipmaker's legal position rests on several points. ByteDance is not listed on the U.S. Bureau of Industry and Security's Entity List or its Military End-Use list, which means transactions involving ByteDance do not automatically trigger heightened screening [5]. The hardware is being shipped to Malaysia, a non-restricted destination. And the end user — Aolani Cloud — is a NVIDIA-vetted Tier-1 partner operating in a Tier 2 country.

Yet the arrangement exposes a fundamental tension in the U.S. chip containment strategy. The export controls were designed to prevent China from acquiring the advanced computing capability needed to train frontier AI models. If a Chinese company can effectively access the same computing power through a cloud lease in a neighboring country, the policy achieves far less than intended.

NVIDIA's stock dipped 1.8% on the day the news broke [8], reflecting investor uncertainty about whether this kind of arrangement might attract tighter regulatory scrutiny.

ByteDance's $23 Billion AI Ambition

The Malaysia supercluster is one piece of a far larger puzzle. In December 2025, reports emerged that ByteDance planned to spend approximately $23 billion on AI-related capital expenditure in 2026 alone — a figure that would place it among the world's largest AI infrastructure investors alongside Microsoft, Google, Amazon, and Meta [9].

That spending breaks down into roughly $5.5 billion for AI chip acquisitions within China (where the company is limited to less advanced processors like NVIDIA's China-specific H20 and domestic alternatives) and approximately $6.8 billion on overseas infrastructure using cutting-edge NVIDIA hardware [10]. The remainder covers data center construction, networking, and operational costs.

The spending reflects a strategic pivot. Under CEO Liang Rubo, ByteDance is repositioning itself as an AI-first company, not just the parent of a social media platform facing existential regulatory threats in its largest Western market. The company's AI chatbot Doubao has surpassed 100 million daily active users since its 2023 launch [11], and ByteDance now operates five of the top 50 most popular AI applications globally by monthly users [12].

ByteDance has also been aggressively hiring AI talent in the United States, posting dozens of roles in machine learning and generative AI even as TikTok's future in the country remains uncertain [12]. Research teams operate out of San Jose, Seattle, and Singapore, among other locations.

Source: Benzinga / SiliconANGLE / WSJ Reports

Data as of Mar 13, 2026CSV

Malaysia's Data Center Gold Rush

ByteDance's investment lands in the middle of a broader technology gold rush transforming Malaysia — and particularly the southern state of Johor — into one of Asia's fastest-growing data center hubs.

The numbers are staggering. Johor has reached more than 900 megawatts of data center capacity in just three years, a milestone that took neighboring Singapore 12 to 14 years to achieve [13]. Malaysia's national data center capacity nearly doubled between 2021 and 2024, reaching 54 centers with a combined 504.9 megawatts of operational power [14]. Applications for new data centers have surged to over 11,000 megawatts — representing more than 40% of Peninsular Malaysia's entire installed power generation capacity of approximately 27,000 megawatts [15].

The investment is not coming from ByteDance alone. Google has announced a $2 billion investment for its first Malaysian data center and Google Cloud region. Microsoft committed $2.2 billion to expand cloud and AI services. YTL Power International and NVIDIA signed a $2.36 billion agreement for AI infrastructure in Johor, and Malaysia's first NVIDIA-powered AI data center became operational in October 2025 [14][16].

The Asia-Pacific Data Center Association projects that Malaysia's AI industry will create 30,900 jobs annually by 2030, with data center construction contributing $24 billion annually in economic output [14].

Source: World Bank Open Data

Data as of Feb 24, 2026CSV

The Environmental and Infrastructure Strain

The breakneck pace of data center construction has not come without costs. Malaysia's power grid is under severe strain, with data centers alone projected to require between 5 and 6 gigawatts of electricity by 2035 — roughly one-fifth of the country's current capacity [15].

Malaysia's electricity generation remains heavily dependent on fossil fuels, with more than 75% coming from coal and natural gas as of 2023 [17]. Every new AI supercluster adds to the country's carbon footprint unless paired with renewable energy commitments that have so far been slow to materialize.

Water consumption presents another challenge. Cooling a 100-megawatt data center requires as much water per day as a city of 10,000 residents [18]. In Johor, where water scarcity has historically triggered interstate tensions with Singapore, the implications are significant.

The Johor state government has begun pushing back, rejecting up to 30% of new data center applications in 2024, citing misaligned utility timelines, infrastructure disputes, and planning complications [15]. The state now requires green energy commitments from new applicants [19].

The Export Control Chess Game

ByteDance's Malaysia play is the latest move in an escalating chess game between Chinese technology companies and U.S. regulators. The pattern has become familiar: Washington imposes restrictions, and Chinese firms find creative ways to access the technology they need through third countries, cloud leases, and intermediary structures.

The January 2026 AI Diffusion Framework was supposed to close some of these gaps by creating a tiered global system for chip distribution, with aggregate caps on how many advanced processors could flow to Tier 2 countries [7]. But the framework's primary enforcement mechanism targets the point of sale — the physical shipment of chips — not the downstream allocation of computing power.

A former gaming company with Chinese government ties was previously accused of smuggling banned AI GPUs, and investigators found NVIDIA H100 chips available on various marketplaces in China despite the restrictions [20]. Over a million China-specific accelerators that technically comply with the rules — such as the NVIDIA H20 and AMD MI308X — have been legally shipped to China [20].

The ByteDance-Aolani arrangement is legally distinct from smuggling. It operates in the open, with NVIDIA's explicit approval and a compliance structure designed to satisfy the letter of U.S. law. But it raises the same fundamental question: if the goal is to prevent Chinese companies from training frontier AI models with advanced American chips, does it matter whether those chips sit in Shenzhen or Johor if the compute flows to the same researchers?

U.S. lawmakers have taken notice. Members of Congress have sought greater legislative control over export licensing, frustrated by what they see as an executive branch that has been inconsistent in its approach — tightening rules in one area while relaxing them in another [7].

What Comes Next

The ByteDance supercluster in Malaysia is expected to come online in the coming months, joining an already extensive network of AI computing resources the company operates globally. If the deployment proceeds as planned, ByteDance will have access to one of the largest Blackwell GPU clusters outside the United States — a fact that will likely intensify the debate over whether U.S. export controls are achieving their stated objectives.

For Malaysia, the investment reinforces its rapid emergence as a critical node in the global AI supply chain, but also amplifies the infrastructure, environmental, and geopolitical pressures that come with that role. For NVIDIA, the deal represents both robust commercial demand for its most advanced chips and a compliance tightrope that could attract unwanted regulatory attention.

And for the broader U.S.-China technology competition, the arrangement is a reminder that in a globally interconnected semiconductor supply chain, restricting access to technology is far harder than restricting its physical movement.

Source: GDELT Project

Data as of Mar 13, 2026CSV