Amazon Web Services raised prices on its EC2 Capacity Blocks for ML on July 1, 2026, and the increase landed hard. Rates for reserved GPU clusters climbed by roughly 20% across six instance families, from the newest Blackwell-class hardware down to older Ampere-generation instances. AWS confirmed the change through its official pricing page and told reporters the update reflects supply and demand.
This is not AWS’s first move this year. The company already raised Capacity Block pricing by about 15% in January 2026. Two hikes in six months mean any company reserving Nvidia GPU clusters for AI training now pays meaningfully more than it did at the start of the year, and the pattern raises a bigger question for the broader cloud computing market: how much more will GPU access cost before the AI training boom cools off?
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 Changed on July 1, 2026
AWS did not send out a press release for this one. The change appeared on the EC2 Capacity Blocks pricing page, and it applies only to a specific set of GPU-backed instance families used for machine learning training and other accelerated workloads. Everything else in the EC2 catalog stayed the same, according to AWS’s own documentation.
Six instance families saw new hourly rates take effect: P6-B300, P6-B200, P5, P5e, P5en, and P4de. Each one reserves a cluster of Nvidia accelerators for a defined future window, which is the whole point of a Capacity Block. Instead of hoping an on-demand GPU instance is free when a training run needs to start, a customer locks in the hardware ahead of time. That certainty now costs more.
The New EC2 Capacity Block Rates, Instance by Instance
AWS lists the updated per-accelerator hourly rates on its pricing page, and the numbers vary by instance family and region. Here is what customers pay starting July 1, 2026:
| Instance Family | GPU Class | Region | New Rate (per accelerator-hour) |
|---|---|---|---|
| P6-B300 | Blackwell-class | All regions except GovCloud | $14.04 |
| P6-B200 | Blackwell-class | All regions except GovCloud | $12.355 |
| P5 | H100-class | US regions | $5.191 |
| P5 | H100-class | Non-US regions | $4.72 |
| P5e | H200-class | All available regions | $5.97 |
| P5en | H200-class | US regions | $6.865 |
| P5en | H200-class | Non-US regions | $6.241 |
| P4de | A100-class | US regions | $2.214 |
The steepest rates belong to the newest hardware. P6-B300, AWS’s Blackwell-class offering, now costs $14.04 per accelerator-hour in every available region except GovCloud. P6-B200 follows close behind at $12.355. Older families cost less in absolute terms but climbed by a similar percentage, which means the increase hit AWS’s entire GPU reservation lineup rather than a single product line.
This Is AWS’s Second GPU Price Increase of 2026
AWS raised Capacity Block pricing once before this year. In January 2026, rates went up by about 15%. InfoQ documented a concrete example at the time: a p5e.48xlarge reservation, which bundles eight H200 accelerators into a single instance, rose from $34.61 an hour to $39.80 an hour in most regions. In US West, the same instance rose to $39.80 an hour — the same uniform rate InfoQ reported for every region, with no separate US West price.
Two increases inside of six months turn what looked like a one-off adjustment into a pattern. If AWS keeps the same rhythm, customers should expect another pricing review around January 2027, which would mark a full year of roughly semiannual hikes for reserved GPU capacity.
Why AWS Says It’s Raising Prices Now
AWS has kept its public explanation short. Reporting from Morningstar and Dow Jones on June 26, 2026, quoted the company describing the move as a periodic pricing update tied to supply and demand, with all other EC2 prices remaining unchanged. That is the entire official rationale AWS has offered so far. No executive statement, no blog post, no forecast for future pricing.
Business Insider took the analysis further in its June 2026 coverage, tying the price increase to soaring memory chip costs across the industry. The outlet framed the move plainly: AWS just made one of its key AI cloud services more expensive at a moment when the hardware behind that service is getting scarcer and pricier itself.
The GPU Scarcity Behind the Sticker Shock
Blackwell Capacity Stays Tight
The two families that top the new price list, P6-B300 and P6-B200, run on Nvidia’s newest Blackwell-class silicon. Demand for that hardware has outpaced supply for most of 2026, and AWS is not alone in feeling the squeeze. Nvidia and AMD both raised prices on GPU hardware kits earlier this year, with Chinese retail prices climbing 22% as manufacturers passed component costs down the chain.
Cloud providers sit downstream of that same shortage. When accelerator supply tightens, hyperscalers either ration capacity, raise prices, or both. AWS chose price. The approach lets AWS keep serving customers willing to pay more while easing pressure on a resource it cannot simply manufacture more of on short notice.
Amazon’s Spending Bet Behind the Price Hike
Higher Capacity Block prices arrive alongside record spending inside Amazon itself. The company reported $44.2 billion in capital expenditures for the first quarter of 2026, up 77% year over year, driven largely by AI infrastructure. That number climbed again in the second quarter. Amazon’s cash capital expenditures hit $53.1 billion, and the company said the spending primarily relates to AWS and generative AI, according to its July 30, 2026 earnings report.
Amazon has guided toward roughly $220 billion in total capital expenditures for 2026, most of it aimed at data centers, chips, and AI capacity. Raising prices on the GPU reservations customers want most is one of the more direct ways AWS can offset that spending while demand for AI training capacity keeps climbing.
Timeline: A Year of Rising AI Infrastructure Costs
Line up the pricing changes against Amazon’s own spending disclosures and the picture gets clearer. Each price increase landed within weeks of a capex figure that beat the one before it.
| Date | Event | Key Figure | Source |
|---|---|---|---|
| January 2026 | First 2026 Capacity Block price increase | About 15% average hike | InfoQ |
| April 2026 | Amazon reports Q1 2026 capex | $44.2 billion, up 77% year over year | Business Insider |
| June 26, 2026 | AWS confirms second 2026 price increase | Effective July 1, about 20% average | Morningstar / Dow Jones |
| July 1, 2026 | New Capacity Block rates take effect | Up to $14.04 per accelerator-hour | AWS pricing page |
| July 30, 2026 | Amazon reports Q2 2026 capex | $53.1 billion cash capex | CNBC |
| Full-year 2026 (guidance) | Amazon’s total capex outlook | Roughly $220 billion (raised from $200B on Jul 30) | CNBC |
The pattern suggests AWS is managing GPU capacity as a constrained resource rather than a commodity it can scale freely. Every dollar spent building new data center capacity takes months to turn into usable Blackwell or Hopper clusters, and pricing is one of the few levers AWS can pull immediately.
How Rivals Compare: Azure and Google Cloud
What’s Confirmed, and What Isn’t
Here is the honest answer: there is no public record of Microsoft or Google matching AWS’s move with their own reserved GPU price increases in 2026. Azure’s pricing pages document current VM series and availability, including newer GPUs in its ND-series lineup, but neither Microsoft nor Google has announced a comparable hike to reserved AI capacity pricing this year.
That does not mean GPU access is cheap elsewhere. Independent cloud-pricing trackers, including Spheron Network and CloudZero, put on-demand H100-class instances at roughly $88 an hour on Google Cloud’s A3 family and near $98 an hour on Azure’s ND H100 v5 series. Those are third-party, on-demand estimates rather than official reserved rates, so they are not directly comparable to AWS’s Capacity Block increase. Still, the figures show GPU capacity runs expensive across every major cloud, not just AWS.
Why the Silence Might Not Last
AWS moving twice in six months while its two biggest rivals hold public pricing steady could mean one of two things. Either Azure and Google Cloud have more comfortable GPU supply lines heading into the back half of 2026, or they are absorbing the same cost pressure quietly rather than passing it to customers. Given how tight Blackwell-class supply has been industry-wide, the second explanation looks more likely to hold through the rest of the year.
What EC2 Capacity Blocks Actually Are
Capacity Blocks exist because on-demand GPU availability became unreliable once AI training demand took off. Instead of requesting a GPU instance and hoping AWS has one free, customers reserve a cluster for a defined future window and know it will be there when the training job starts. That certainty is the entire value proposition, and it explains why AWS can charge a premium over standard on-demand pricing for the same hardware.
The families affected by the July increase span AWS’s full Nvidia-backed lineup. P4de instances use older A100 accelerators. P5, P5e, and P5en step up through H100 and H200 generations. P6-B200 and P6-B300 represent AWS’s newest Blackwell-class capacity. Raising prices across that entire range, rather than just the newest hardware, tells customers that scarcity is not limited to the latest chips.
Who Feels the Increase Most
Not every AWS customer notices a price change buried in a GPU reservation product. The customers who do are running the most compute-hungry workloads on the planet: AI labs training frontier models, startups fine-tuning open-source models at scale, and enterprises running large simulation or rendering pipelines that need guaranteed accelerator access.
These customers do not have many alternatives. Switching cloud providers mid-training run is disruptive and expensive in its own right, and building private GPU infrastructure takes years and capital most companies do not have. For now, AWS customers who need reserved Blackwell or Hopper capacity have little choice but to pay the new rate or scale back their reservations.
Market Reaction and Early Fallout
Public reaction has been muted so far. Reporting on the July increase, including coverage from ITPro, has not surfaced a named customer complaint or a rival cloud provider using the moment to poach AWS accounts. That silence stands out. Price increases on infrastructure this central to AI development usually draw some public pushback, and this one has not, at least not yet.
One reason may be timing. AWS gave customers roughly a week of public notice between the confirmed announcement in late June and the July 1 effective date, leaving little room to renegotiate contracts or shift workloads elsewhere. Customers with active reservations locked in before July 1 likely avoided the increase entirely, which may explain why the loudest objections have not materialized yet.
What AWS and Industry Coverage Are Saying
Public statements on this increase are sparse, and the ones available come from official documentation and financial reporting rather than named executives. Here is what is on the record.
AWS’s own pricing page states the exact numbers behind the change: “Effective July 1, 2026, hourly rates per accelerator will be: P6-B300 at $14.04 … P5 at $5.191 … P4de at $2.214” (AWS, official documentation). The same page adds a short but important line: “All other prices remain unchanged” (AWS, official documentation).
Financial wire coverage confirms the timing and rationale. “Amazon Web Services said it would prices for certain Amazon EC2 Capacity Blocks for machine learning beginning July 1 as part of a periodic pricing update based on supply and demand”, according to a company statement reported by Morningstar and Dow Jones on June 26, 2026.
Investing.com’s coverage put the scale of the change in plain terms: “AWS will raise prices on its EC2 Capacity Block reservations for machine-learning GPU instances by approximately 20% effective July 1, 2026, citing supply and demand dynamics in a posting to its official documentation page”.
Network World summed up the mechanics of what actually happened: “Amazon Web Services (AWS) has updated the pricing structure for some of its Elastic Compute Cloud (Amazon EC2) Capacity Blocks for ML offerings”.
How Businesses Can Respond to Rising Cloud GPU Costs
Teams that depend on reserved GPU capacity have a few real options, even if none of them fully offset a 20% increase. The first is timing. Booking Capacity Blocks further in advance, before the next pricing review, locks in current rates for the length of the reservation.
The second is architecture. Not every training or fine-tuning job needs a guaranteed Blackwell-class cluster. Workloads that can tolerate interruption may run cheaper on spot capacity or standard Savings Plans instead of a Capacity Block reservation, trading certainty for a lower bill.
The third is visibility. Engineering teams can check current Capacity Block availability directly through the AWS CLI before committing budget to a reservation:
aws ec2 describe-capacity-block-offerings \
--instance-type p5e.48xlarge \
--instance-count 1 \
--capacity-duration-hours 24 \
--region us-east-1Finally, FinOps discipline matters more than ever. Cloud waste already eats up close to 29% of AI-related cloud spend at some organizations, and a GPU price increase makes that waste far more expensive. Tools built specifically to track AI infrastructure costs, including AWS’s own free FinOps agent, give teams a faster way to catch runaway reservations before the next invoice arrives.
5 Predictions for Cloud GPU Pricing Through 2027
Where does this trend go from here? Based on the pattern AWS has already set this year, here are five reasonable predictions.
- A third increase is likely around early 2027. Two hikes in six months point to a roughly semiannual review cycle, and nothing in AWS’s public statements suggests that cadence will slow down.
- Newest-generation hardware will keep carrying the steepest premiums. P6-B300 and P6-B200 already command the highest rates on the list, and any future Blackwell Ultra or next-generation capacity will likely enter at similarly high price points.
- Azure and Google Cloud will face pressure to clarify their own pricing. Staying quiet works only as long as customers do not start comparing notes across providers, and AWS’s public rate card makes that comparison easier every time it changes.
- Spot and flexible capacity will grow faster than reserved Capacity Blocks. As reservation costs climb, more customers will shift interruption-tolerant workloads toward cheaper, less guaranteed capacity.
- FinOps spending on AI infrastructure will keep expanding. Rising per-hour costs turn small inefficiencies into large invoices, giving every engineering organization a stronger reason to invest in cost visibility tools built for GPU workloads.
Frequently Asked Questions
What exactly did AWS change on July 1, 2026? AWS raised hourly rates on EC2 Capacity Blocks for ML, its reserved GPU cluster product, by roughly 20% across six instance families: P6-B300, P6-B200, P5, P5e, P5en, and P4de. All other EC2 pricing stayed the same.
Is this the same as a general EC2 price increase? No. The change applies only to Capacity Blocks for ML, the product customers use to reserve GPU clusters in advance. On-demand and other reserved EC2 pricing was not affected, according to AWS’s own pricing page.
How much did prices actually go up? New rates run from $2.214 per accelerator-hour for P4de up to $14.04 for P6-B300. AWS and multiple outlets, including Investing.com, put the average increase at approximately 20%.
Did AWS raise these prices before in 2026? Yes. AWS raised Capacity Block pricing by about 15% in January 2026, according to InfoQ. The July change is the second increase inside of six months.
Have Azure or Google Cloud made similar changes? There is no public record of either company raising reserved GPU capacity pricing in 2026 to match AWS’s move, based on available reporting as of this writing.
Why is AWS raising prices now? AWS attributes the change to supply and demand. Business Insider’s June 2026 reporting linked the timing to rising memory chip costs across the AI hardware supply chain.
What can businesses do to avoid the higher cost? Booking Capacity Blocks further ahead locks in current rates for the reservation term. Workloads that can tolerate interruption may also run more cheaply on spot capacity or standard Savings Plans instead.
Which AWS customers are most affected? Companies running large AI training jobs, fine-tuning workloads, or simulation pipelines that depend on guaranteed GPU cluster access feel this the most, since they are the primary buyers of Capacity Blocks.
Related Coverage
- AWS Reserved vs Savings Plans vs Spot: 90% Off [2026]
- AWS Fargate vs Lambda: 70% Spot Savings, 15-Min Cap [2026]
- Cloud Waste Hits 29% as AI Spend Breaks Budgets [2026]
- AWS FinOps Agent Launches Free in $16.5B Market [2026]
- Data Center Leases Hit $850B, Meta, Microsoft Lead [2026]
- AMD MI350P vs Nvidia H200: 40% Faster, $30K Price [2026]
- Nvidia, AMD Hike GPU Kit Prices: China Retail Up 22% [2026]
