AWS switched on Blackwell-powered GPU instances in its Asia Pacific (Hyderabad) region on September 3, 2026, giving Indian developers direct access to Nvidia’s newest data center silicon for the first time without routing training jobs through Singapore, Tokyo, or the US. The move puts Amazon EC2 P6-B200 instances, each packing eight Nvidia Blackwell GPUs, into a region that Microsoft and Google have also been racing to build out over the past year. It also lands at effective pricing that undercuts Azure’s comparable Blackwell instance by roughly 40%, based on published rate cards.
The launch is a small line item in AWS’s release notes, but it sits at the center of a much bigger story: hyperscalers are now fighting over GPU capacity, region coverage, and per-hour pricing in India at the same time that global GPU supply remains tight. Cloud buyers comparing AWS, Azure, and Google Cloud for large language model training or inference workloads now have a fresh, direct data point to work with, and it favors AWS on cost.
What AWS Actually Shipped in Hyderabad
According to AWS’s official What’s New announcement, P6-B200 instances became available in the Asia Pacific (Hyderabad) Region starting September 3, 2026, offered in a single configuration: p6-b200.48xlarge. Each instance packs 8 Nvidia Blackwell GPUs, 1,440 GB of high-bandwidth HBM3e GPU memory, and 5th-generation Intel Xeon Scalable (“Emerald Rapids”) CPUs delivering 192 vCPUs. AWS pairs that with 2,048 GiB of system RAM, roughly 30 TB of local NVMe storage across eight 3.84 TB drives, and up to 14.4 TB/s of NVLink bandwidth stitching the eight GPUs together for fast intra-node communication, per AWS’s P6 instance family page.
Networking gets a bump too. AWS lists up to 3.2 Tbps of Elastic Fabric Adapter (EFAv4) bandwidth and 100 Gbps of dedicated EBS throughput on the instance, numbers aimed squarely at large, multi-node training jobs where inter-GPU communication speed often matters more than raw compute. AWS says the jump from its previous flagship, P5en, delivers up to 2x the performance for AI training and inference, plus a 60% increase in GPU memory bandwidth, driven by the shift to Blackwell’s HBM3e memory stack.
Hyderabad is not the first Indian region to get the hardware. AWS lit up P6-B200 in Asia Pacific (Mumbai) back in June 2026, three months earlier. Before that, the rollout ran through US West (Oregon), US East (N. Virginia and Ohio), and AWS GovCloud (US-West) starting in May 2026. By early September, the P6-B200 footprint spans at least seven regions, and a related, more powerful SKU, P6-B300 with Blackwell Ultra GPUs, has already reached Hyderabad and Seoul as well.
Why India, and Why Now
The timing is not incidental. Global cloud infrastructure revenue hit $143.4 billion in the second quarter of 2026, up 43% year over year, and the nine largest cloud providers are projected to spend roughly $830 billion on capital expenditures across 2026, a 79% jump from the prior year. India has become one of the fastest-growing pieces of that spending story, with AWS, Microsoft, and Google all racing to add regional GPU capacity so that Indian banks, telcos, SaaS vendors, and government-backed AI programs don’t have to route training jobs overseas.
Data residency is the practical driver. Financial institutions and public-sector bodies in India increasingly need to keep training data and model weights inside the country, both for regulatory reasons and to cut round-trip latency on inference workloads serving Indian users. A Hyderabad region with genuinely current-generation GPUs, not a scaled-down or delayed version, removes one of the last arguments for building AI infrastructure abroad. Cloud market share data reinforces why the region race matters commercially: Google Cloud gained roughly two points of global share year over year through Q2 2026, moving to a record 15%, while AWS slipped two points, according to Synergy Research Group data reported by CRN. Every region win, especially in a market growing as fast as India’s, has become a share-defense move as much as a customer-service one.
The Pricing Gap: AWS, Azure, and Google Cloud Compared
Published rate data shows a real spread between the three major clouds on Blackwell B200-class capacity, and it is wide enough to change purchasing decisions for teams running sustained training jobs. In US East (Ohio), AWS lists the 8-GPU p6-b200.48xlarge at an effective $98.84 per hour total, or roughly $12.36 per GPU-hour. Google Cloud’s A4 instances, also built on 8 B200 GPUs, carry an introductory list price around $110 per instance-hour, or about $13.75 per GPU-hour, for short-term reservations, with Spot and committed-use discounts available on top. Azure’s ND GB200 v6 line runs meaningfully higher: a sample ND128isr-NDR-GB200-v6 configuration priced through Azure’s calculator comes to $162.73 per instance-hour in the Australia East region, and a broader AI infrastructure pricing index puts Azure’s average around $27 per GPU-hour, more than double AWS’s rate.
| Provider / Instance | GPUs | GPU Memory | List Price (Total/Hr) | Approx. Price per GPU-Hour |
|---|---|---|---|---|
| AWS P6-B200 (p6-b200.48xlarge, US East Ohio) | 8x Blackwell B200 | 1,440 GB HBM3e | $98.84 | ~$12.36 |
| Google Cloud A4 (B200, intro list) | 8x Blackwell B200 | ~1,440 GB HBM3e | ~$110.00 | ~$13.75 |
| Azure ND128isr-NDR-GB200-v6 (Australia East) | ~8x Blackwell B200 | ~1,440 GB HBM3e | $162.73 | ~$20-21 |
| Azure ND GB200 v6 (broad average, per pricing index) | 8x Blackwell B200 | 1,440 GB HBM3e | — | ~$27.00 |
| AWS reserved GB200 UltraServer (72-GPU, rack scale) | 72x Blackwell GB200 | — | $761.90 | ~$10.58 |
The gap holds up even on a reserved basis. AWS now publishes a one-year reserved rate of roughly $10.58 per GPU-hour on its rack-scale GB200 UltraServer offering, described by GPU pricing tracker Coloprice as the first published contract figure for that SKU. That reserved rate sits well below even AWS’s own on-demand P6-B200 price, and far below Azure’s on-demand average. None of this means AWS is cheap in absolute terms. Multi-thousand-dollar-a-day training bills are still the norm for any team running sustained multi-node jobs on Blackwell-class hardware. But relative to Azure, the spread is large enough that FinOps teams evaluating multi-cloud GPU strategy now have a concrete number to plug into total-cost models rather than a vague sense that “AWS is usually cheaper.”
How This Fits the Broader GPU Cloud Shortage
None of this is happening in a market with excess supply. GB200-class capacity remains scarce across every major provider, and pricing analysts have flagged the gap between AWS’s reserved rate (about $10.58 per GPU-hour at rack scale) and Azure’s roughly $27 per GPU-hour average as evidence of a persistent supply-demand imbalance rather than simple price competition. Oracle Cloud, meanwhile, prices bare-metal Blackwell access around $16 per GPU-hour, putting it between AWS and Azure. Nvidia’s own manufacturing constraints, tight HBM3e memory supply, and continued demand from foundation-model labs all keep upward pressure on prices even as providers add more regions and capacity.
That scarcity is precisely why regional expansion matters as a competitive lever. A provider that can offer current-generation Blackwell capacity in more places, sooner, effectively captures demand that would otherwise queue for capacity elsewhere or get diverted to a competitor’s region. AWS’s decision to push P6-B200 into Mumbai in June and Hyderabad in September, ahead of a confirmed equivalent rollout from Azure in the same cities, is a bet that being first with real inventory beats waiting to match a rival’s exact specs.
Historical Context: From P4 to P6-B200
AWS’s GPU instance lineage has moved fast even by cloud standards. The P4 generation, built on Nvidia A100 GPUs, established AWS’s large-scale training instances several years ago. P5 and P5en, based on Nvidia H100 and H200 GPUs respectively, extended that line through 2024 and into 2025, with P5en adding higher memory bandwidth for inference-heavy and mixed workloads. P6-B200, built on Blackwell, is explicitly benchmarked by AWS against P5en, not against P5, underscoring how quickly the prior generation became the baseline for comparison rather than the ceiling.
| AWS GPU Generation | Nvidia GPU | GPU Memory (8-GPU instance) | Key Advance Over Prior Gen |
|---|---|---|---|
| P4d/P4de | A100 | 320-640 GB | First AWS instance built for large-scale distributed training |
| P5 | H100 | 640 GB | Major training throughput jump for transformer workloads |
| P5en | H200 | 1,128 GB | Higher memory bandwidth for inference and mixed workloads |
| P6-B200 | Blackwell B200 | 1,440 GB HBM3e | 2x P5en performance, 60% more memory bandwidth, 3.2 Tbps EFAv4 networking |
| P6-B300 | Blackwell Ultra | 2,100 GB | 6.4 Tbps EFA networking, 300 Gbps ENA, largest memory footprint yet |
What stands out in that progression is the pace: AWS has now shipped two distinct Blackwell-class SKUs, P6-B200 and P6-B300, within months of each other, rather than waiting a full generation cycle between meaningful upgrades. That cadence mirrors Nvidia’s own shift toward faster GPU refresh cycles and puts pressure on Azure and Google Cloud to match not just specs but release timing.
Competitive Positioning: AWS vs Azure vs Google Cloud on Blackwell
All three major clouds now offer Blackwell B200-class instances, but they differ on region coverage, pricing structure, and how the capacity is packaged. AWS offers a single, clearly-specced instance size (p6-b200.48xlarge) across a growing region list, paired with a rack-scale GB200 UltraServer option for customers who need beyond-single-node scale. Google Cloud’s A4 line follows a similar single-SKU approach with aggressive introductory list pricing, though the standard price is subject to change once the promotional period lapses. Azure’s ND GB200 v6 line is priced noticeably higher on both a total-instance and per-GPU basis, though Azure has offset some of that with broader regional availability in specific corporate accounts and deeper integration with its existing Azure AI Foundry tooling.
For engineering teams choosing a provider, the decision increasingly comes down to three variables: whether the region they need is actually live, not just announced, whether the on-demand or reserved price fits their training budget, and how tightly the GPU instance needs to integrate with a provider’s existing AI platform stack, such as Amazon Bedrock, Azure AI Foundry, or Google Vertex AI. AWS’s Hyderabad launch strengthens its position on the first two variables for Indian and South Asian customers specifically, while Azure and Google Cloud still hold advantages in regions or accounts where AWS has not yet deployed equivalent capacity.
Who Actually Uses These Instances
AWS positions P6-B200 for large-scale LLM training, retrieval-augmented generation systems, and high-throughput inference for generative AI applications, with particular emphasis on financial services, e-commerce, media, and SaaS companies. In practice, that means fine-tuning proprietary large language models, training vision and speech models, and running high-capacity inference for chatbots, copilots, and recommendation engines, workloads that Indian banks, telcos, and enterprise software vendors have been scaling up through 2025 and 2026. AWS has not published a list of named Hyderabad customers for this specific launch, and no vendor names should be assumed without direct confirmation.
The broader pattern across AWS’s 2026 region expansions suggests demand is coming from two distinct buyer types: large enterprises consolidating AI workloads that previously ran on-premises or through smaller regional providers, and AI-native startups that need current-generation GPU access without the multi-month lead times associated with buying and racking their own hardware. Both groups benefit disproportionately from a local region, since it removes cross-border data transfer costs and latency that previously made overseas GPU capacity a compromise rather than a first choice.
Security and Compliance Considerations for Indian AI Workloads
Running large training jobs inside a local region solves a data residency problem, but it does not automatically solve a security one. GPU instances at this scale typically sit behind the same identity and access management controls as any other AWS resource, which means misconfigured IAM roles, unrotated access keys, or overly permissive service accounts remain the most common way cloud AI infrastructure gets compromised, regardless of which region the GPUs physically sit in. Enterprises moving sensitive training data into a new region for the first time should treat that migration as a fresh opportunity to audit access controls rather than assume existing policies transfer cleanly.
There is also a cost-control angle specific to GPU workloads. Multi-node Blackwell training jobs can burn through budget quickly if instances are left running between experiments, and FinOps teams tracking GPU spend across regions need visibility into utilization, not just the sticker price per GPU-hour. As more Indian enterprises shift training workloads onto local infrastructure, the same governance discipline that applies to any high-cost cloud resource, tagging, budget alerts, automated shutdown policies, becomes more important, not less, simply because the per-hour cost of an idle Blackwell instance is far higher than an idle general-purpose VM.
Market Impact: What This Means for Cloud Spending
The immediate impact is narrow: one region, one instance family, one incremental line in AWS’s release notes. The broader impact is that it adds another data point to a trend that has been building through 2026, hyperscalers treating region-by-region GPU rollout speed as a genuine competitive weapon, not just an operational afterthought. Every week that AWS, Azure, or Google Cloud has exclusive access to current-generation GPU capacity in a fast-growing market is a week where that provider can win workloads that would otherwise be split three ways.
For enterprise buyers, the practical takeaway is that price comparisons need to be refreshed constantly. A provider that looked expensive six months ago on a given GPU generation may now be the cheapest option in a specific region, and vice versa. Teams running recurring, budget-sensitive training workloads increasingly need automated cost-tracking across providers rather than a one-time procurement decision, since the underlying rate cards and regional availability are both moving targets in 2026’s GPU market.
What to Watch Next
Five things are worth tracking as this plays out over the next two quarters.
- Azure and Google Cloud will likely announce their own Hyderabad or comparable Indian-region Blackwell capacity within the next two to three quarters, given the pattern of each provider matching the others’ regional moves within roughly six months in 2026.
- AWS’s reserved GB200 UltraServer pricing, currently the cheapest per-GPU-hour rate among the three major clouds, will likely see matching reserved offers from Azure and Google Cloud as they try to close the cost gap for long-term commitments.
- P6-B300 (Blackwell Ultra) coverage will expand faster than P6-B200 did, following AWS’s accelerated release cadence between the two SKUs this year.
- GPU pricing across all three providers should stay elevated through at least early 2027, given continued HBM3e memory tightness and sustained demand from foundation-model training.
- India-specific AI infrastructure investment, from both hyperscalers and domestic players, will keep accelerating as data residency requirements tighten across financial services and government-adjacent sectors.
The Bottom Line for Cloud Architects
AWS’s Hyderabad launch is not a headline-grabbing announcement on its own, but it is a useful proxy for how the cloud GPU market is actually behaving in late 2026: fast regional expansion, real pricing gaps between providers on comparable hardware, and persistent scarcity that keeps prices well above what raw manufacturing costs would suggest. Teams evaluating where to run their next large training run now have a clearer, numbers-backed reason to at least price out AWS’s Indian region alongside their usual defaults, particularly if data residency or latency to South Asian users is part of the calculation.
Frequently Asked Questions
What is AWS EC2 P6-B200?
P6-B200 is an AWS EC2 instance family built around 8 Nvidia Blackwell B200 GPUs, offering 1,440 GB of HBM3e GPU memory, 192 vCPUs on 5th-generation Intel Xeon processors, and up to 3.2 Tbps of EFAv4 networking bandwidth, aimed at large-scale AI training and inference.
When did AWS launch P6-B200 in Hyderabad?
AWS made P6-B200 instances available in the Asia Pacific (Hyderabad) Region starting September 3, 2026, according to its official What’s New announcement.
How does AWS P6-B200 pricing compare to Azure and Google Cloud?
AWS lists p6-b200.48xlarge at roughly $12.36 per GPU-hour in US East (Ohio). Google Cloud’s comparable A4 (B200) instance runs about $13.75 per GPU-hour on introductory list pricing, while Azure’s ND GB200 v6 line averages around $27 per GPU-hour, more than double AWS’s rate, based on published rate data.
Which other AWS regions have P6-B200 instances?
As of September 2026, P6-B200 is available in US West (Oregon), US East (N. Virginia and Ohio), AWS GovCloud (US-West), Asia Pacific (Mumbai), and Asia Pacific (Hyderabad).
What is the difference between P6-B200 and P6-B300?
P6-B300 uses Nvidia Blackwell Ultra GPUs with 2,100 GB of GPU memory and 6.4 Tbps of EFA networking, compared to P6-B200’s 1,440 GB and 3.2 Tbps. P6-B300 is available in fewer regions so far, including Hyderabad, Seoul, Oregon, N. Virginia, and São Paulo.
Why does AWS keep expanding GPU regions so quickly?
Persistent GPU scarcity and rising demand from AI training workloads mean that regional availability itself has become a competitive advantage. Being first to offer current-generation Blackwell capacity in a growing market like India lets AWS capture workloads that would otherwise wait for a rival’s matching launch or route through a distant region.
Is Blackwell GPU capacity still hard to get in 2026?
Yes. Pricing analysts point to the persistent gap between AWS’s roughly $10.58 per GPU-hour reserved rate and Azure’s roughly $27 per GPU-hour average as evidence that GB200-class capacity remains supply-constrained across the industry, not simply a matter of one provider undercutting another on cost.
