AM Intelligence has locked in two more firm and binding orders for 20,000 NVIDIA Rubin GPUs, pushing the Singapore-based AI infrastructure operator’s total committed fleet to roughly 29,000 Rubin chips across its data center buildout in India and Malaysia. The orders, reported October 5, 2026 by outlets including TelecomTalk, mark one of the largest single disclosed Rubin commitments outside the US hyperscaler tier, and they land at a moment when every gigawatt of AI compute capacity in Asia is being fought over by governments, chipmakers, and capital.
The new GPUs are specified as NVIDIA Vera Rubin NVL72 rack-scale systems, the liquid-cooled, 72-GPU-per-rack architecture NVIDIA has positioned as the successor to its Blackwell-generation Grace Blackwell NVL72 platform. Delivery is scheduled for Q2 2027, and the two orders together represent approximately 70 MW of additional data center capacity. Combined with AM Intelligence’s earlier commitment for 9,000 Rubin GPUs at its first AI factory in Hyderabad, India, announced August 25, 2026 (a cluster TechTimes reported will use 125 NVL72 racks and 30 MW), the company’s total Rubin exposure now sits at close to 100 MW across the two markets.
No product prices were disclosed in the reports covering this order, and a widely circulated figure putting the deal at roughly $4 billion has not been confirmed as an official transaction value in the cited coverage. What is confirmed is the scale and the timeline: AM Intelligence is betting that India and Malaysia will absorb a meaningful share of the AI compute demand that NVIDIA’s next-generation Rubin platform is built to serve, and it is doing so with binding paper, not a letter of intent.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What AM Intelligence Actually Ordered
The headline number is 20,000 NVIDIA Rubin GPUs, but the more useful detail for anyone tracking the AI hardware market is the system-level specification. These are Vera Rubin NVL72 rack-scale systems, meaning the GPUs are not being procured as standalone accelerators but as fully integrated 72-GPU racks with NVIDIA’s Vera CPU, NVLink interconnect fabric, and liquid cooling built in. That architecture choice matters because it is the same rack-scale approach NVIDIA used to push Blackwell GPUs into hyperscaler data centers at volume, and Rubin is the direct successor generation.
AM Intelligence is deploying the hardware across two countries rather than concentrating it in one site. Reports describe planned deployment in India and Malaysia, with delivery penciled in for Q2 2027. That 18-month-plus runway between order and delivery is typical for Rubin-class hardware right now, given how constrained GPU supply and the associated power infrastructure remain heading into 2027.
The 70 MW figure attached to these two orders is a useful proxy for just how much electricity this single commitment will draw once it is live. For context on how power-hungry AI data centers have become, see our coverage of Trane Technologies’ stock surge on its AI cooling bet, which traces how thermal management has become its own multibillion-dollar sub-industry riding the same GPU buildout wave.
From 9,000 to 29,000 Rubin GPUs in Ten Weeks
The timeline here is the story. On August 25, 2026, AM Intelligence announced its first AI factory commitment: 9,000 NVIDIA Rubin GPUs destined for Hyderabad, India. That deal alone would have been a notable regional win for NVIDIA in the Indian market. Later, on October 5, 2026, the company tripled its exposure with two additional firm and binding orders totaling 20,000 more Rubin GPUs, spreading the expanded footprint across India and Malaysia.
Stacking the two commitments together puts AM Intelligence’s total disclosed Rubin GPU order book at approximately 29,000 units and nearly 100 MW of planned capacity, a total AFP also reported across the company’s three orders. That pace of escalation, roughly tripling a GPU commitment inside a single fiscal quarter, is the kind of signal that regional telecom and tech trade press in South and Southeast Asia have been quick to flag as evidence that AI infrastructure investment in the region is accelerating faster than most 2026 forecasts anticipated.
| Milestone | Date | Rubin GPUs | Location | Capacity Added |
|---|---|---|---|---|
| First AI factory commitment | August 25, 2026 | 9,000 | Hyderabad, India | ~30 MW (est. share) |
| Two new firm and binding orders | October 5, 2026 | 20,000 | India and Malaysia | ~70 MW |
| Cumulative committed total | As of October 5, 2026 | ~29,000 | India and Malaysia | ~100 MW |
It is worth being precise about what “firm and binding” means in procurement language for GPU supply contracts. Unlike a reservation or an option, a firm and binding order typically locks both sides into the transaction: NVIDIA (through its supply chain partners) commits to delivering the hardware on the stated timeline, and AM Intelligence commits to paying for it, with limited ability to cancel without penalty. That framing is what separates this disclosure from the steady stream of “planned” or “in talks” AI infrastructure announcements that have populated the sector throughout 2026.
Why India and Malaysia, Not Just the US
The geographic split matters as much as the GPU count. For most of the generative AI boom since 2023, the overwhelming majority of frontier-scale NVIDIA GPU deployments have landed in the United States, with secondary clusters in the Gulf states and parts of Europe. AM Intelligence’s decision to concentrate a near-30,000-GPU commitment in India and Malaysia specifically reflects a broader shift that regional infrastructure operators and sovereign-adjacent entities have been pushing since early 2026: building AI compute capacity physically inside fast-growing digital economies rather than exporting workloads to US-based cloud regions.
India’s own government-linked push for domestic AI compute capacity, coordinated loosely through agencies tracked by India’s Ministry of Electronics and Information Technology, has made Hyderabad and other tech hubs attractive sites for exactly this kind of large rack-scale deployment. India’s broader investment-promotion apparatus, including Invest India, has spent much of 2026 courting exactly this type of large-scale digital infrastructure commitment. Malaysia, meanwhile, has spent the past two years positioning itself as a data center hub for Southeast Asia, leveraging comparatively cheap land and power alongside proximity to Singapore’s financial and cloud infrastructure; the Malaysian Investment Development Authority has actively promoted the country as a destination for exactly this class of AI data center project. Placing Rubin capacity in both countries lets AM Intelligence hedge geographically while still serving a single regional demand pool.
This dynamic echoes what has already happened in the chip supply chain itself. Our earlier report on Broadcom and Samsung’s chip pricing deal matching NVIDIA’s terms showed how price benchmarks set in one region of the AI hardware market now ripple quickly into adjacent deals elsewhere. The same gravitational pull appears to be happening with GPU allocation itself: once one operator commits at scale in a region, it becomes easier for the next large order to follow the same geography.
How the Hyderabad Buildout and Rival Indian Orders Fit Together
The Hyderabad project is the best-documented piece of the AM Intelligence buildout. The company ordered 9,000 Rubin GPUs for the site in August 2026, and TechTimes reported that the cluster will use 125 NVL72 racks and 30 MW. In the same month, a company source said AMI committed more than $8 billion for 200 MW of AI capacity. That figure describes a capacity commitment, so readers should not treat it as a confirmed price for the October GPU order, and no official AMI price for GPU rental, Compute-as-a-Service, rack systems, or facility access can be confirmed.
The business model explains why the company is building ahead of demand. AMI’s stated model is GPU-based Compute-as-a-Service, intended to give cloud providers, enterprises and AI developers access to large-scale compute near deployment locations. On October 19, 2026, AMI was reported to be exploring an equity raise of more than $1 billion, which is a reported plan and not a completed transaction. Separately, AMI’s approximately 400 MW pipeline is a development pipeline, not a confirmed order total, so it should not be added to the roughly 29,000 committed GPUs.
AMI is also not the only buyer chasing Rubin supply in India. In September 2026, Yotta announced 40,000 Rubin chips for Greater Noida’s 120 MW D4 site, according to TechTimes. For readers following the story, the practical point is that every figure here describes committed or planned capacity with delivery in Q2 2027, and the available reports do not confirm that the 29,000 GPUs are operational. Watch for AMI’s financing outcome and for any delivery update rather than treating the order book as live compute.
The Vera Rubin NVL72 Platform, Explained
NVIDIA’s Rubin platform is the generation that follows Blackwell in the company’s roughly annual data center GPU cadence, a cadence NVIDIA has publicly committed to maintaining through its architecture roadmap. The NVL72 designation refers to the rack configuration: 72 GPUs networked together via NVLink inside a single liquid-cooled rack, paired with NVIDIA’s Vera CPUs for the general-purpose compute and orchestration layer that sits alongside the accelerators. NVIDIA says Vera Rubin NVL72 is in full production, on track to ship in the second half of 2026. TechTimes reported each Rubin GPU with 288 GB of HBM4 memory, and one NVL72 rack delivering 3.6 exaFLOPS of NVFP4 inference.
That rack-scale design is deliberate. Training and serving today’s largest frontier models requires GPUs to communicate with each other far faster than traditional networking allows, which is why NVIDIA bundles the interconnect fabric directly into the rack rather than leaving it to data center operators to wire up GPUs across standard server racks. Readers who want the hardware-comparison angle on how this generation stacks up against what shipped before can see our breakdown of NVIDIA reportedly ending RTX 5090 production to feed RTX Pro GB202 output, which illustrates the same supply-reallocation logic NVIDIA applies across its product lines when a new architecture generation ramps.
Official technical detail on the Vera Rubin architecture, including the NVL72 rack specification, is published on NVIDIA’s own data center hardware pages, and the company’s newsroom has carried a steady stream of regional deployment announcements throughout 2026 as Rubin-based orders have piled up across multiple continents.
Market Impact: What 29,000 GPUs Means for Supply
NVIDIA’s data center GPU supply chain has been running at or near capacity for most of 2026, with memory pricing serving as one of the clearest signals of just how stretched the broader AI hardware ecosystem has become. A commitment of this size, even spread across an 18-month delivery window, adds to an order backlog that already includes hyperscaler-scale commitments from the largest cloud providers in the US, plus a growing list of sovereign and regional operators making similar bets.
The ripple effects show up fastest in memory and packaging. High-bandwidth memory capacity allocated to GPU makers has been one of the tightest links in the chain throughout 2026, a dynamic our coverage of the DRAM price spike pushing TV chip costs above panel costs and the AMD and Intel price hikes tied to rising TSMC costs both documented from different angles. Every additional 20,000-GPU order competing for the same constrained memory and advanced packaging capacity pushes those pressures further out across the whole semiconductor supply chain, not just NVIDIA’s own production lines.
There is also a competitive dimension worth tracking. As US-based and Gulf-region buyers continue absorbing the bulk of near-term Blackwell and early Rubin supply, operators like AM Intelligence effectively have to place orders well ahead of need just to secure a delivery slot. The Q2 2027 delivery window on this order, nearly 18 months out from the October 2026 announcement, reflects exactly that dynamic.
Competitive Landscape: Rubin vs. the Field
NVIDIA does not have the AI accelerator market to itself, even if its data center GPU share has historically dwarfed every competitor. AMD has continued pushing its Instinct accelerator line as the most credible alternative at hyperscaler scale, while a separate competitive front has emerged from Chinese chipmakers trying to build viable alternatives to NVIDIA’s CUDA software ecosystem rather than just the silicon itself. Our report on DeepSeek and Huawei shipping six Ascend tools to rival CUDA is a useful companion piece here, because it shows the competitive pressure on NVIDIA is coming from two different directions simultaneously: hardware-only competitors like AMD competing on raw accelerator specs, and software-ecosystem competitors like the Huawei-DeepSeek pairing trying to peel developers away from CUDA lock-in entirely.
AM Intelligence’s decision to commit to NVIDIA Rubin rather than hedge across multiple accelerator vendors is itself a data point. It suggests that for an operator building AI factories from scratch in India and Malaysia, CUDA’s software maturity and the broader NVIDIA ecosystem, spanning networking, system integration partners, and software tooling, still outweighs whatever price or availability advantage a diversified hardware strategy might offer. That is consistent with how most large AI infrastructure buyers have behaved throughout 2026, even as alternatives have grown more capable.
| Factor | NVIDIA Rubin NVL72 | Alternative Accelerator Paths |
|---|---|---|
| Rack-scale interconnect | NVLink fabric built into 72-GPU rack | Varies by vendor; often requires separate integration |
| Software ecosystem | CUDA, mature tooling, broad framework support | Ascend/CUDA-alternative tooling improving but earlier-stage |
| Delivery timeline (as ordered) | Q2 2027 for this order | Varies; dependent on vendor-specific supply constraints |
| Disclosed large-scale India/Malaysia commitment | ~29,000 GPUs (AM Intelligence) | No comparable disclosed order of this size in the region |
Historical Context: How We Got to Rubin
NVIDIA’s annual architecture cadence has moved fast enough that it is easy to forget how recent each generational jump actually is. Hopper GPUs powered the first wave of ChatGPT-era infrastructure buildouts in 2023. Blackwell, announced in 2024 and ramping through 2025 and into 2026, became the generation that hyperscalers used to push training clusters past the 100,000-GPU mark for the first time. Rubin is the next step in that same progression, and AM Intelligence’s order is one of the earliest large disclosed commitments to the platform outside of NVIDIA’s traditional hyperscaler customer base.
What makes the India and Malaysia angle notable in that historical arc is timing. Previous architecture generations took longer to reach large-scale regional deployments outside the US; Blackwell-class hardware largely stayed concentrated in US hyperscaler and Gulf sovereign-fund-backed data centers for much of its early production run. Rubin-based commitments reaching a near-30,000-unit scale in South and Southeast Asia while Vera Rubin NVL72 is already in full production, with AMI’s delivery scheduled for Q2 2027, suggests the lag between US hyperscaler access and regional operator access is compressing.
What Remains Unconfirmed
Several details circulating around this story have not been confirmed by the cited reporting and should be treated as open questions rather than fact. First, headlines framing this as NVIDIA “sending” GPUs to India overstate what has actually happened: the confirmed fact is that AM Intelligence placed the orders, and the cited reports describe a planned future delivery in Q2 2027, not a completed shipment. Second, claims that all 20,000 GPUs from the new orders will land exclusively in India are not supported; the reporting specifies deployment split across India and Malaysia. Third, a circulated figure putting the deal’s value at roughly $4 billion has not been established as an official or confirmed transaction value in the available coverage.
No named individual has gone on record with a direct quote about this order in the cited reporting, and no pricing details for the GPUs or the racks have been disclosed. Readers should treat any specific dollar figures attached to this deal elsewhere online with caution until NVIDIA, AM Intelligence, or a primary financial filing confirms them.
Power and Cooling: The Real Bottleneck
GPU allocation gets the headlines, but power delivery and thermal management are what actually determine how fast a commitment like this one can go live. A combined ~100 MW of Rubin capacity across India and Malaysia is roughly equivalent to the continuous draw of a mid-sized industrial facility, and liquid-cooled NVL72 racks concentrate that heat load far more densely than air-cooled server rooms of a decade ago. That is exactly the shift driving demand for specialized cooling infrastructure, a trend we covered in depth when looking at Trane Technologies’ stock climbing on its AI data center cooling bet.
India’s grid operators and Malaysia’s utility providers will both need to factor commitments like AM Intelligence’s into longer-term capacity planning, particularly since Q2 2027 delivery means the power infrastructure has to be ready well before the racks themselves arrive. Any slippage on the power side, rather than GPU supply, is the more likely source of delay for a buildout of this size.
Memory and Packaging: The Hidden Cost Driver
Every Rubin GPU shipped carries high-bandwidth memory and advanced packaging requirements that compete directly with the memory capacity other parts of the tech industry rely on. That competition has already shown up in consumer and enterprise pricing well outside the data center. Our reporting on the RX 9070 GRE price swings tied to a 200% jump in GPU memory costs and on Micron’s record $133 billion fiscal 2026 revenue both trace the same underlying pressure: memory suppliers are running flat-out to keep pace with AI accelerator demand, and every large new GPU order, including this one, adds to that queue.
Predictions: Where This Goes Next
- Expect additional Rubin order disclosures from other Asia-Pacific infrastructure operators within the next two to three quarters, as regional buyers compete to secure delivery slots ahead of 2027 capacity constraints.
- Power availability, not GPU allocation, will likely become the binding constraint on how fast AM Intelligence can actually bring its combined ~100 MW of committed capacity online in India and Malaysia.
- Pricing transparency on deals like this one will probably stay limited through at least mid-2027, continuing the pattern where GPU volume gets disclosed publicly but transaction value does not.
- Watch for India’s domestic policy apparatus to use deployments like the Hyderabad AI factory as a reference case in future sovereign AI compute announcements, given the government’s stated push for in-country AI infrastructure.
- Competitive responses from AMD-aligned or CUDA-alternative operators in the same Southeast Asia and South Asia corridor are likely, as rivals look to avoid ceding the region entirely to NVIDIA-based buildouts.
Why This Matters Beyond One Company
AM Intelligence is not a household name the way NVIDIA, Microsoft, or Google are, but that is precisely why this order is worth tracking. The largest, most visible AI infrastructure commitments tend to come from a small handful of US hyperscalers whose capital spending gets dissected every earnings season. A near-30,000-GPU, roughly 100 MW commitment from a regional operator most casual tech readers have never heard of is a signal that the AI buildout has moved well past its first wave of marquee participants and into a second tier of serious, well-capitalized regional players.
That second-tier expansion has consequences for everything downstream of the GPU order itself: power grid planning in India and Malaysia, cooling infrastructure demand of the kind tracked in our Trane Technologies coverage, memory and packaging supply chains already strained by 2026’s DRAM price surge, and the broader competitive question of whether CUDA’s software moat holds as more buyers outside NVIDIA’s traditional customer base make long-term platform bets.
Frequently Asked Questions
What did AM Intelligence actually order?
AM Intelligence placed two further firm and binding orders for 20,000 NVIDIA Rubin GPUs, specified as Vera Rubin NVL72 rack-scale systems, for deployment across India and Malaysia.
Did NVIDIA ship 20,000 GPUs to India?
Not according to the confirmed reporting. The confirmed fact is that AM Intelligence ordered the GPUs, with delivery planned for Q2 2027 and deployment split between India and Malaysia, not shipment to India alone.
How many total Rubin GPUs has AM Intelligence committed to?
Including an earlier August 25, 2026 commitment for 9,000 Rubin GPUs at its Hyderabad AI factory, AM Intelligence’s total committed order is approximately 29,000 Rubin GPUs and nearly 100 MW of capacity.
When will the new GPUs be delivered?
Delivery for the two new 20,000-GPU orders is scheduled for Q2 2027.
How much is this deal worth?
No product prices or an official transaction value were stated in the cited reports. A circulated figure of roughly $4 billion has not been confirmed.
What is a Vera Rubin NVL72 system?
It is NVIDIA’s rack-scale AI infrastructure design pairing 72 Rubin GPUs with NVIDIA Vera CPUs and NVLink interconnect in a single liquid-cooled rack, succeeding the Blackwell-generation Grace Blackwell NVL72 platform.
Why is AM Intelligence deploying in both India and Malaysia?
The split reflects a broader regional push to build AI compute capacity closer to fast-growing digital economies in South and Southeast Asia rather than relying solely on US-based cloud regions.
How does this compare to other NVIDIA GPU supply news in 2026?
It is one of several large disclosed Rubin or Blackwell-era commitments reported in 2026, alongside hyperscaler orders and other regional deployments, reflecting continued high demand for NVIDIA’s data center hardware heading into 2027.
