On this page · 11 sections
- What AWS changed in July 2026
- The numbers: what a Capacity Block costs now
- Why AWS raised prices, and why analysts disagree
- How Capacity Blocks compare to the other ways to buy GPUs
- What this does to your AI budget
- The FinOps response
- India-specific considerations
- What to watch next
- FAQ
- How eCorpIT can help
- References
Summary. AWS raised prices on EC2 Capacity Blocks for ML by roughly 20% effective July 7, 2026 (AWS first said July 1), covering the P6-B300, P6-B200, P5, P5e, P5en and P4de instance families. It was the third increase of 2026. On Saturday, January 3, AWS lifted H200-linked Capacity Block rates 15%, pushing p5e.48xlarge from $34.61 to $39.80 per hour in most regions and from $43.26 to $49.75 in US West (N. California). A 10% rise on US P5, P5en and P6-B200 rates came on April 1, and a fourth rise on October 7, 2026. P6-B300 now bills at $16.146 per accelerator hour outside GovCloud, up from $14.04 in July, and P5 capacity tied to H100 accelerators bills at $5.970 per accelerator hour in all listed regions, up from $5.191 in US regions and $4.72 elsewhere. All other EC2 prices were unchanged. Amazon now expects about $220 billion of capital expenditure for 2026. Four hikes inside ten months tell you AWS now treats constrained Nvidia supply as a pricing lever rather than a one-off correction, and that reserved GPU capacity is no longer a line item you can set and forget.
Update, 2 November 2026. AWS raised Capacity Block rates again on October 7, 2026: 15% on most families and about 26.5% on P5 and P5en outside the US. One rate now applies across regions, apart from GovCloud on P6, and AWS says the next update is due in January 2027. P5 in Mumbai has gone from $3.933 to $5.970 per accelerator hour since March, about 52%. We also corrected three things: AWS's own notice moved the July rise from July 1 to July 7, a 10% US rise on April 1 was missing, and AWS's H100 on-demand list price is $6.88, not about $3.90. Prices and arithmetic below use rates read on 2 November 2026, with rupee figures at ₹95 per dollar. For the IndiaAI route, see our guide to the IndiaAI GPU empanelment.
What AWS changed in July 2026
Capacity Blocks for ML are a reservation product. You book a defined block of GPU capacity for a fixed window, from a single day up to several weeks, and AWS guarantees the hardware is there when your job starts. That guarantee is the whole point. Spot capacity can vanish mid-run; a Capacity Block cannot. Teams use them for training runs, fine-tuning jobs and evaluation sweeps where a mid-job eviction costs more than the premium.
The July change touched only the accelerator rates on that product. According to reporting from Network World and SDxCentral, the increase applies to P6-B300, P6-B200, P5, P5e, P5en and P4de, and leaves the rest of the EC2 price list alone. That scoping matters. If your GPU workload runs on on-demand or Savings Plan capacity rather than Capacity Blocks, your bill did not move in July or October. If your training pipeline books blocks, it did.
AWS did not publish a blog post or send a customer advisory. The notice and the new rates appeared on the Capacity Blocks pricing page, and the October rise took the same route. This is the same pattern as January, when The Register noted the 15% rise landed on a Saturday with no announcement.
The numbers: what a Capacity Block costs now
Here are the rates AWS listed on 2 November 2026, in force since October 7, alongside the earlier rates they replaced.
| Instance or family | Accelerator | Rate on AWS's page, 2 November 2026 (effective October 7, 2026) |
|---|---|---|
| P6-B300 | Nvidia Blackwell B300 | $16.146 per accelerator hour outside GovCloud ($14.04 in July) |
| P5 (US regions) | Nvidia H100 | $5.970 per accelerator hour ($5.191 in July) |
| P5 (non-US regions) | Nvidia H100 | $5.970 per accelerator hour ($4.72 in July) |
| p5e.48xlarge | Nvidia H200 | $54.924 per hour in all listed regions, or $6.866 per accelerator hour ($39.80 in most regions after January) |
| p5e.48xlarge (US West, N. California) | Nvidia H200 | $54.924 per hour, the same as other regions ($49.75 after January) |
| P6-B200, P5en, P4de | Blackwell B200, H200, A100 | $14.208, $7.895 and $2.546 per accelerator hour respectively, outside GovCloud |
Two things stand out. First, the per-accelerator framing hides the cluster arithmetic. A single 8-GPU B300 block at $16.146 per accelerator hour runs about $129 per hour before storage, networking and egress. Book that for a two-week training window and you are at roughly $43,400 for the compute alone. Second, the regional spread has closed. The same p5e instance cost $9.95 per hour more in US West (N. California) than in most other regions after January, but AWS now lists one Capacity Block rate across regions, apart from GovCloud on P6. On-demand still varies: an H100 costs $6.88 per hour in N. Virginia and $8.256 in Mumbai.
Why AWS raised prices, and why analysts disagree
AWS gave a consistent explanation to reporters. Its statement, quoted by Network World: "EC2 Capacity Blocks for ML pricing vary based on supply and demand patterns... This price adjustment reflects the supply/demand patterns we expect this quarter."
Analysts split on whether to take that at face value.
Pareekh Jain, CEO at EIIRTrend and Pareekh Consulting, reads it as ordinary scarcity pricing. "The most defensible explanation is simply market-based pricing tied to supply and demand," he said, adding that as demand for H100 and H200 GPUs outstrips supply, AWS is applying a scarcity premium to guaranteed inventory and recovering higher infrastructure and capital costs from urgent capacity rather than from overall capacity.
Cloud economist Corey Quinn disagrees with the framing. "This was AWS updating the published base rates on their pricing page... That's a policy decision, not supply/demand," he said. His point is narrow and worth holding onto: a spot market clears dynamically, but a list price is a choice somebody makes. Calling a list-price edit "supply and demand" borrows the language of a market mechanism the product does not actually use.
Sanchit Vir Gogia, CEO and chief analyst at Greyhound Research, pointed at the competitive angle, noting that Google Cloud's approach differs because "they're using scheduling to compete, not price."
You do not have to resolve the argument to act on it. Both readings point the same direction for buyers. If it is scarcity, prices rise until supply loosens. If it is policy, prices rise until customers push back. Either way, the base case for 2027 planning is not a price cut.
How Capacity Blocks compare to the other ways to buy GPUs
Capacity Blocks sit in a specific slot: short commitment, guaranteed availability, a rate well above spot. They are not the cheapest way to get a GPU, and they were never meant to be. The mistake teams make is defaulting to them for workloads that do not need the guarantee.
| Buying model | Typical saving vs on-demand | Best fit |
|---|---|---|
| On-demand | Baseline | Unpredictable, short, interactive work |
| Spot or preemptible | 70–91% reduction, per nOps analysis | Fault-tolerant training with frequent checkpointing |
| Capacity Blocks | On P5, P5en, B200 and B300, level to 13% below on-demand list rates in N. Virginia | Fixed-date training runs where eviction is unacceptable |
| Reserved, 1 year | 30–40% | Steady-state inference and known baseline load |
| Reserved, 3 year | 50–60%, pulling H100 toward roughly $3.00 per GPU hour | Long-horizon platforms with confident forecasts |
| Committed hybrid | Blend of the above | Most real AI teams (see below) |
The hyperscaler comparison is worth a second table, because the per-GPU normalised rates diverge more than the headline instance prices suggest.
| Provider | Approximate H100 on-demand, per GPU hour | Note |
|---|---|---|
| Google Cloud | About $3.00 | Lowest published on-demand for H100 class |
| AWS | $6.88 | p5.48xlarge at $55.04 per hour in N. Virginia, after the 44% H100 price cut AWS made in June 2025 |
| Azure | $6.98 | Enterprise agreements reduce this substantially |
| AWS Capacity Blocks (B200) | $14.208; H100 is $5.970 | Guaranteed-capacity product, not comparable to plain on-demand |
| Spot, AWS or Google Cloud | $1.95–$2.50 | Eviction risk; requires checkpointing |
Read the two tables together and the shape of a sensible policy appears. On-demand rates favour Google Cloud. Spot favours Google Cloud and AWS. Long-term commitments roughly equalise the three. The differentiator is rarely the sticker price. It is whether your workload can tolerate eviction.
What this does to your AI budget
The July increase is small next to the waste most teams are already carrying. That is the uncomfortable part.
nOps puts GPU idle waste at 30–60% of GPU spend for most AI teams, and GPU instances are typically the largest single line item on an AI cloud bill. A 20% rise on a reservation product is painful. A 45% idle rate on the same hardware is worse, and it is inside your control in a way that the AWS pricing page is not.
The macro numbers explain why nobody should expect relief. AI infrastructure spending for 2026 is projected at about $1.5 trillion. Alphabet now guides to $195–205 billion of 2026 capital expenditure and Amazon to about $220 billion, per CNBC's reports of their October 2026 results, with Microsoft at roughly $190 billion for calendar 2026 on its earlier guidance. In Q1 2026, Google Cloud grew 63% year over year, AWS 28% and Azure 40%. Providers are spending at a rate that assumes demand holds. Nothing in that picture creates downward pressure on guaranteed GPU capacity in the next few quarters.
The demand side is not disciplined either. In LeanOps' account of FinOps X 2026, the keynote opened with the statistic that by June, many organisations had already burned through three times their entire annual AI budget. Budgets set in January were written before anyone knew what agent workloads would cost in production.
The real cost is usually the idle hour, not the price per hour.
The FinOps response
Seven moves, roughly in order of return on effort.
1. Separate the workloads that need a guarantee from the ones that do not
This is the highest-use decision and it is free. Capacity Blocks carry a premium over spot because availability is guaranteed. If your job checkpoints every ten minutes and can restart, you are paying that premium for nothing. Spot at $1.95–$2.50 per GPU hour against a Capacity Block at $5.970 for H100 or $14.208 for B200 is not a rounding difference. Audit which of your booked blocks genuinely cannot survive an eviction.
2. Adopt a hybrid commitment ratio rather than an all-or-nothing one
The pattern that FinOps practitioners converged on in 2026: commit 60–70% of capacity to cover baseline steady-state usage, and keep 30–40% on-demand or spot for the variable portion. You take the discount on the predictable base and keep flexibility for spikes. Committing 100% locks you into hardware that will be superseded; committing 0% pays retail on load you could have forecast.
3. Measure utilisation before you buy anything
If you do not know your GPU utilisation rate, you cannot tell whether a price rise or an idle cluster is the bigger problem. Most teams discover the answer is the cluster. Instrument first.
4. Revisit region choice
The p5e spread between US West (N. California) and most other regions was $9.95 per hour after January, but AWS now lists one Capacity Block rate across regions, apart from GovCloud on P6. On-demand rates still differ by region: an H100 costs $6.88 per hour in N. Virginia and $8.256 in Mumbai. Many training workloads have no residency or latency constraint and are running in an expensive region because that is where the account was set up.
5. Checkpoint aggressively so spot becomes viable
Spot delivers 70–91% reductions for fault-tolerant training. The engineering work to checkpoint reliably is a one-time cost that pays back on every subsequent run. This is the clearest case in AI infrastructure where an engineering investment beats a procurement negotiation.
6. Price the alternatives before renewing
Specialised GPU providers price well below the hyperscalers on paper, and the gap runs 40–85% across major GPU models on on-demand rates. The honest caveat: hyperscalers add egress, storage, networking and reserved capacity overhead that compound on real workloads, and moving a pipeline has migration cost. Price the alternatives anyway. A quote you never requested is a negotiation you never had.
7. Rebuild the budget on a quarterly cycle
An annual AI budget set in January is a guess about a market that repriced four times by October. Organisations burning 3x their annual AI budget by June are not overspending so much as under-forecasting. Move to quarterly re-planning.
India-specific considerations
For teams running AI workloads from India, the arithmetic differs enough to change decisions.
AWS lists an H100 in Mumbai at $8.256 per hour on demand and $5.970 on a Capacity Block, about ₹784 and ₹567 at ₹95 per dollar, near the ECB cross-rate of ₹94.96 on 30 October 2026. An A100 there is $3.293 on demand, about ₹313. Domestic GPU providers publish materially lower rates. Cyfuture AI lists H100 (80GB SXM5) at ₹219 per hour, A100 (80GB) at ₹195 per hour, L40S at ₹61 per hour and V100 from ₹39 per hour. E2E Networks, India's only publicly listed pure-play GPU cloud company, lists H100 at ₹379.05 per hour, H200 at ₹502.55 and B200 at ₹759.05 per GPU on demand, before GST, on its rate card of 30 October 2026, billed in rupees through UPI, NEFT, cards or direct debit with GST-compliant invoices.
| Route | Indicative H100 rate | Trade-off |
|---|---|---|
| AWS Mumbai | About ₹784 on demand, ₹567 on a Capacity Block, at ₹95 per dollar | Full AWS service surface; highest rate |
| Cyfuture AI | ₹219 per hour | Lower rate; smaller service ecosystem |
| E2E Networks | ₹379.05 per GPU hour on demand; ₹70 spot | Rupee billing, GST invoices, listed company |
| IndiaAI Mission | ₹153 per GPU hour on demand, ₹117 on a 12-month reservation, before GST | Subsidy capped at 40%; eligibility rules apply |
| Reserved, 3–12 months | ₹130–150 per hour effective | 30–40% off; requires a confident forecast |
The IndiaAI Mission route deserves attention. PIB said on 29 October 2026 that more than 45,000 GPUs had been onboarded. The portal lists an H100 SXM at ₹153 per GPU-hour on demand and ₹117 on a 12-month reservation, before GST and a subsidy capped at 40%. The on-demand rate is about 73% below the Mumbai Capacity Block rate before any subsidy. We could not trace the 100,000-GPU target some reports cite for the end of 2026 to an official source. For startups and research teams that qualify, no FinOps optimisation on a hyperscaler bill competes with that rate.
Data residency is now a genuine option rather than an aspiration. Credible providers host A30, A100, L40S and H100 GPUs in Indian data centres with rupee billing and DPDP-aligned infrastructure, so training data no longer has to leave the country. Under the Digital Personal Data Protection Act 2023, that removes a category of transfer questions from your architecture review. If you are working through DPDP obligations more broadly, our FinOps guide for Indian teams cutting cloud spend and our India FinOps cloud cost playbook cover the adjacent ground.
One caution on the domestic comparison. Several of the published Indian rates come from provider marketing pages rather than independent measurement, and the AWS Mumbai figures are dollar list prices converted at ₹95, so they move with the exchange rate. Treat the table as a prompt to request quotes, not as a settled benchmark.
What to watch next
July was not the last increase of 2026. Three signals will tell you which way the January 2027 update goes.
Nvidia supply. The 2026 hikes were tied to H100, H200 and Blackwell scarcity. If Blackwell supply loosens, the scarcity premium has less cover.
Google Cloud's scheduling play. Gogia's observation that Google Cloud competes on scheduling rather than price is the interesting divergence. If scheduling-based allocation wins customers, AWS has a reason to stop pricing guaranteed capacity as a premium good.
The January 2027 pricing page. AWS said the July adjustment "reflects the supply/demand patterns we expect this quarter." That sentence had a short shelf life: AWS raised rates again on October 7. Its page now says the next update is due in January 2027. Watch the page from late December and set an alert, because the July and October notices appeared there.
FAQ
How eCorpIT can help
eCorpIT is a CMMI Level 5 certified technology organisation in Gurugram, and our senior engineering teams work with AWS, Microsoft and Google platforms on cloud cost and AI infrastructure. We help teams instrument GPU utilisation before they commit, separate workloads that need guaranteed capacity from those that can run on spot, and set a commitment ratio their forecast actually supports. For Indian teams we also model AWS Mumbai against domestic and IndiaAI Mission routes with DPDP considerations included. If GPU spend is moving faster than your forecast, contact us and we will review your workload mix.
References
- AWS hikes prices for EC2 Capacity Blocks amid soaring GPU demand — Network World
- AWS raises GPU prices 15% on a Saturday — The Register
- AWS quietly increases prices for H200 EC2 instances by 15% — Data Center Dynamics
- Amazon quietly raises price tag on the AI boom — TheStreet
- Reserved vs. On-Demand GPU in 2026 — Compute Exchange
- Can AI answer the $3 trillion question? — TechCrunch
- Google Cloud, AWS, Microsoft Azure: The AI vertical integration race — Constellation Research
- GPU Cloud Pricing: India vs Global Providers — Cyfuture AI
- IndiaAI Mission: 34,000 GPUs at Rs 150/Hour for startups — Abhishek Gautam
Last updated: November 2, 2026.
