Is the cloud still the cheap option? For a decade, the answer was assumed to be yes. Budgets were built on the idea that computing gets cheaper every year. That assumption is now breaking down, and 2027 is the year the bill comes due. Microsoft has told customers that selected virtual machines and storage will cost 8% to 17% more in seven Azure regions from 1 February 2027, and about 25% more worldwide on its oldest machine series. On the same date, its reservations lose flexibility many companies rely on. AWS has already raised reserved GPU prices three times in a single year. If you are building a 2027 cloud budget on last year’s numbers, you are already behind.
This article lays out the actual figures, names the providers that matter, and explains what pushes prices up or down. It also covers the trends that will shape decisions: selective repatriation, capacity anxiety, and a possible price war that may arrive later than the pain.
The Numbers You Need Before You Budget
Real figures beat vague warnings. Here is what is documented, with dates attached, so you can see which apply to your planning window.
| Item | Figure | Applies to | Date |
|---|---|---|---|
| Azure VM and storage increase | 8% to 17% | Seven Azure regions | From 1 February 2027 |
| Azure oldest machine series | About 25% more | Worldwide | From 1 February 2027 |
| CoreWeave GPU rental increase | About 25% | GPU cloud customers | July 2026 |
| CoreWeave second increase | Another 10% | GPU cloud customers | Later in Q3 2026 |
| CoreWeave revenue | $3.66 billion, up 168% year on year | Company results | Q3 2026 |
| Big four capex guidance | About $730 billion combined | Google, Meta, Microsoft, Amazon | Full-year 2026 |
| Storage share of top-11 cloud capex | 73% of $1.0403 trillion | Eleven hyperscalers | 2027 projection |
| AWS expected ROIC | 18.9% | Provider returns | 2027 estimate |
| Google Cloud expected ROIC | 16.0% | Provider returns | 2027 estimate |
| Azure expected ROIC | 14.6% | Provider returns | 2027 estimate |
The pattern is clear. Prices are not falling across the board. Some line items are climbing fast, and the increases cluster around memory, storage, and GPUs.
What Moves Cloud Prices Up or Down

Three forces are pushing costs higher at the same time, and each one has a different timeline.
The first is memory and storage scarcity. OVHcloud and Hetzner have repriced for the memory shortage, and contract prices for memory are expected to keep rising. Storage is where the money goes: 73% of the projected $1.0403 trillion in 2027 capex from the eleven largest hyperscalers is expected to go to storage, up from $72.8 billion in 2025 to $761.3 billion. When providers spend that much on storage, the cost shows up in your bill eventually.
The second is GPU pricing power. CoreWeave raised prices by roughly 25% in July 2026 and another 10% later in the same quarter, and its CEO said prices were still rising in the current quarter. That is one provider’s number, not the whole market, and it does not mean per-token API prices are rising. But it does test the unwritten assumption that AI always gets cheaper.
The third is capex intensity. Morgan Stanley research cited in industry coverage found that capital spending as a share of sales for Amazon, Google, Meta, Microsoft, and Oracle is expected to reach 26% by 2027, close to the 32% peak of the dot-com era and above the 20% seen during the shale oil boom. Financing leases are being used to speed up data center builds, which means headline capex numbers understate the real investment. Microsoft’s capex-to-sales ratio, adjusted for finance leases, jumps from 28% to 38% in fiscal 2026; Oracle’s goes from 41% to 58%.
That spending cuts both ways. It supports capacity, but it also creates depreciation bills. By 2027, market forecasts for depreciation at just Google, Amazon, and Meta may be understated by nearly $16.4 billion. If supply growth outpaces demand, more aggressive pricing could break out as early as 2027. So the same year could bring both higher list prices and sudden discounts. Budget for the former, and stay ready to renegotiate if the latter shows up.
The Provider Picture: Who Is Actually Cheaper
Comparing providers is harder than comparing sticker prices, because returns and risk matter too. HSBC’s benchmark analysis, as covered in a cloud economics podcast episode, found Microsoft Azure and Google Cloud the most attractive on a risk-adjusted basis, while CoreWeave was seen as the most expensive.
The return figures tell you why. AWS leads expected 2027 ROIC at 18.9%, followed by Google Cloud at 16.0% and Azure at 14.6%. Oracle Cloud Infrastructure and CoreWeave sit in the single digits. On 2028 expected EV/EBIT, Google Cloud is lowest at 10.1x, then Oracle at 14.0x, Azure at 14.1x, AWS at 18.2x, and CoreWeave at 27.0x.
There is a concentration risk hiding under those numbers. OpenAI and Anthropic together account for roughly 50% of remaining performance obligations across OCI, Azure, Google Cloud, and AWS. AWS also has more short-term contract exposure: its backlog covers 2.5 times the next twelve months of revenue, versus 3.4 times for Azure and Google Cloud, while CoreWeave and OCI have coverage stretching 6.4 years and 18.8 years. Short contracts can mean higher prices but also more re-lease risk.
For buyers, the practical lesson is that the cheapest headline rate is not the cheapest outcome. A provider with strong returns can hold prices steadier. A provider with heavy GPU concentration may move faster in both directions.
Capacity Anxiety: Full or Just Tight?
Headlines claiming cloud capacity is “mostly booked through 2027” have worried many planning teams. The details differ. As one industry opinion piece argued, “capacity is tight” and “capacity is sold out” are not the same sentence. Large multi-year hyperscale GPU allocations suitable for training advanced models may be hard to get. Everyday workloads face a different market.
What actually works in practice: split your workloads into three buckets before you talk to any sales rep. Bucket one is steady, predictable usage that runs every hour. Bucket two is spiky or experimental. Bucket three is AI training that needs large GPU blocks. Only bucket three faces real scarcity. Companies that lump everything together end up over-reserving and overpaying.
A common mistake is treating a provider’s capacity warning as a reason to sign a long lock-in immediately. Another is assuming reserved instances always save money. With Azure reservations losing flexibility from 1 February 2027, a reservation that no longer matches your workload can cost more than on-demand. Check the terms before you commit.
Selective Repatriation Replaces the All-or-Nothing Exit
For years the infrastructure debate was binary: stay in the public cloud or leave it. That framing is breaking down. The companies making the best decisions are looking at which parts of their usage run every hour of every day and asking whether they are paying cloud prices for capacity they could rent or own for less.
Four forces converge in 2027 to push that question onto more agendas: a hard deadline in European law, a power market with almost nothing left to lease, a memory shortage, and a steady run of VMware renewals. The result is partial moves rather than full exits. A database that runs constantly might move to owned hardware or a colocation deal. A bursty web front end stays in the cloud.
If you are weighing that mix, it helps to understand what the terminology actually means before vendors repackage it. A plain explanation of what the cloud native computing definition really means in practice is a useful starting point, and a companion piece on what cloud native computing really means beyond the marketing helps separate architecture from sales language. The same discipline applies to adjacent spending. Teams comparing AI tool budgets for 2027 that buyers will actually pay for often find the cloud line item and the AI tool line item are describing the same workload twice.
Practical Steps to Protect a 2027 Budget
The most useful advice is unglamorous: find out this month which workloads are exposed, then act on the items with a deadline.
- Pull a usage report by SKU, not by department. Increases target specific machine series and storage tiers, so a blended average hides your exposure.
- Flag every workload on the oldest machine families. The roughly 25% worldwide increase on Azure’s oldest series is the single biggest line-item risk.
- Re-check reservation terms against the 1 February 2027 date. Flexibility you assumed may no longer exist.
- Separate GPU commitments from general compute. GPU prices are moving on their own schedule, as the CoreWeave increases show.
- Model a scenario where prices rise and a scenario where a price war cuts them. Both are plausible in the same year.
- Do not assume adjacent hardware costs stay flat. Buyers tracking handheld console prices in 2027, confirmed versus reported, are learning the same lesson about memory-driven cost pressure.
Storage deserves its own line of questioning. With 73% of top-11 hyperscaler capex expected to go to storage in 2027, ask providers how tiering, lifecycle rules, and egress affect your bill. A surprising number of teams pay for cold data sitting in hot tiers for years. For a consumer-facing comparison of how protection and pricing differ, this look at the best cloud storage for photos in 2027 and which options actually protect them shows how much the details matter.
Trends Worth Watching Beyond Price
Two quieter trends will shape 2027 as much as any price sheet.
The first is custom silicon. Hyperscalers are splitting their infrastructure between commercial GPUs and in-house chip designs. That matters to buyers because workloads that fit a provider’s custom silicon tend to be cheaper and more available than those that need scarce commercial GPUs.
The second is capital discipline. The chair of a memory maker noted that AI data center demand looks clear for 2027 and 2028, with funding and infrastructure partly arranged, while 2029 may be the key turning point if money, land, and power run short. That is a reminder that today’s capacity crunch is a financing story as much as a physics story. The same financing pressure is visible in other compute-heavy sectors, where quantum computing meets AI banking and what is coming after 2027 raises similar questions about who pays for the next wave of infrastructure.
Conclusion
The state of cloud computing in 2027 is not a story of steady, predictable decline in prices. It is a story of split outcomes. Azure list prices rise in specific regions and on old machine series from 1 February 2027. GPU rental prices have already moved sharply. Storage dominates hyperscaler spending. At the same time, returns at AWS, Google Cloud, and Azure remain strong enough to fund capacity, and a price war remains possible if supply outruns demand. The winner in this environment is not the company that picks the single cheapest provider. It is the company that knows exactly which workloads are exposed, which are flexible, and which belong somewhere other than the public cloud entirely.
FAQ
What is happening to Azure prices in 2027?
Selected virtual machines and storage will cost 8% to 17% more in seven Azure regions from 1 February 2027. The oldest machine series will cost about 25% more worldwide on the same date, and reservations lose flexibility that many companies depend on.
Are cloud prices going up or down in 2027?
Both are possible. Documented increases affect Azure, GPU rentals, and memory-driven services. At the same time, if supply growth outpaces demand, more aggressive pricing could break out as early as 2027. Budget for increases and stay ready to renegotiate.
Which cloud provider is cheapest in 2027?
On a risk-adjusted basis, Azure and Google Cloud were judged most attractive, while CoreWeave was seen as the most expensive. AWS leads expected 2027 returns on invested capital at 18.9%. The cheapest headline rate is not always the cheapest total outcome.
Is cloud capacity really sold out through 2027?
Not in the way headlines suggest. Large multi-year GPU allocations for advanced model training may be hard to secure, but everyday workloads remain available. Split steady, spiky, and GPU-heavy workloads, then plan each separately.
Why are GPU rental prices rising?
CoreWeave raised GPU rental prices about 25% in July 2026 and another 10% later that quarter, with further increases signalled. Demand for AI compute is high, supply is limited, and power is a constraint. This is one provider’s pricing, not the whole market.
What is selective cloud repatriation?
It is moving only part of your workload off the public cloud, usually steady services that run constantly and could be rented or owned for less. It replaces the old all-or-nothing exit debate and is expected to gain ground in 2027 alongside European legal deadlines and VMware renewals.


