GPU Rental Prices Rose 25% Then 10% More: What It Means for Your 2027 AI Budget
November 13, 2026 · 5 min read · Etzal Group Team
CoreWeave, a GPU cloud provider, reported third-quarter results on 10 November 2026. According to TechCrunch's coverage the next day, the company raised prices by about 25% in July and another 10% later in the same quarter, and CEO Michael Intrator said prices are still rising in the current quarter.
One point needs stating at the outset: this is one provider's number, not the market's. The source also does not say that per-token API prices are rising. But for IT and finance leaders drafting 2027 budgets, the news is enough to test an assumption that often goes unwritten, namely that AI always gets cheaper.
The reported figures
According to TechCrunch, CoreWeave's third quarter:
| Item | Figure |
|---|---|
| Revenue | $3.66 billion, up 168% year on year |
| Net loss | $709 million |
| Quarterly interest expense | $846 million |
| Backlog | $132.5 billion |
On pricing:
- An increase of about 25% in July, then another 10% later in the quarter.
- Intrator said pricing is still rising this quarter, which "more than offset[s] cost increases".
- Short-term capacity is selling at more than $40 million per megawatt per year.
- Pricing for new infrastructure on four and five-year contracts is moving to $20 million to $25 million per megawatt.
On funding, CoreWeave expects about $9 billion in customer prepayments and deposits over the next three months. TechCrunch also notes that yields on some of the company's bonds rose above 15% last month.
How to read those numbers
Three things can be read without going beyond the source.
Short-term capacity is priced far above long commitments. More than $40 million per megawatt per year for short-term, against $20 million to $25 million on four and five-year contracts. Buyers who want flexibility pay heavily for it.
The provider is asking customers to pay up front. Large prepayments and deposits shift part of the funding burden to customers.
Strong demand is not the same as a healthy supplier. Revenue rose 168%, yet the quarter carried a $709 million net loss and $846 million in interest expense. For buyers, that is a reminder to ask what happens to a service if the compute supplier at the bottom of the chain runs into trouble.
What cannot be read from those numbers: that your API bill will rise by some percentage. GPU rental prices do not automatically become per-token API prices, and the source draws no such link. So an increase is a risk to budget for, not a certainty.
For regional context, the e-Conomy SEA 2026 report published by Google on 21 October 2026 names affordable model running costs as one of the conditions for AI agent adoption, and puts the needed data centre capacity expansion at about three times over five years.
Three ways to buy AI capability
Most mid-sized firms do not rent GPUs directly. They buy the output. That is where the GPU prices AI budget question becomes practical: there are three routes, and each carries price risk differently.
| Route | Suits | Price risk | What you give up |
|---|---|---|---|
| Per-token API | Small to medium volume, changing needs | Follows the provider's price list, which can change | Control over long-term price |
| Capacity or spend commitment | Large, stable volume | Price locked for the contract term | Flexibility, often with a prepayment |
| Small model run yourself | Narrow high-volume tasks, sensitive data | Depends on hardware prices or GPU rental | Team time to operate and maintain it |
This table is our judgement. Many firms use APIs for most work, then move one or two high-volume tasks to a small model once the numbers are proven.
For the third route this news applies directly: if you rent GPUs to run your own model, you are a capacity buyer, and the gap between short-term and long-term pricing above applies to you.
Building the 2027 AI budget
The steps we suggest:
- Calculate cost per task, not per month. What does it cost to process one invoice, one ticket, one summary? That is the figure you can compare across routes.
- Separate fixed load from growing load. Load that grows with customers or transactions is the most price-sensitive.
- Build three scenarios. Prices flat, prices up, prices down. The size of the increase in the second is your own assumption. What matters is that the board sees a range, not one number.
- Set review points. Quarterly, for instance: if cost per task passes a set limit, the team must assess another route.
- Hold a reserve. A contingency line for compute is easier to approve now than to request mid-year.
Pricing clauses worth asking for
When renewing or signing with a model provider, a cloud provider, or an application vendor that charges for AI:
- Notice before prices change, long enough for you to react.
- A cap on increases during the term, if you are committing to volume.
- An exit right without penalty if prices rise past that cap.
- Clarity on prepayments: what happens to unused balance, and whether it is refundable.
- For application vendors: whether AI costs are passed through to you, and by what formula.
The cost of long commitments deserves a mention too. Locking a price means locking a technology choice. If a more efficient model appears mid-term, you may not be able to move. For a mid-sized firm, a one-year commitment at a conservative volume is usually safer than a long contract chasing a large discount. Contract terms depend on your own case, so confirm them with your legal or procurement adviser.
The best protection: being able to move
Clauses help, but the strongest protection is in system design:
- Model and provider names are settings, not buried in code.
- You hold your own set of test examples, so a replacement model can be assessed in days.
- Simple tasks are routed to small models, and large models are used only where needed.
- Usage is monitored per task, so a spike is visible before the invoice arrives.
If you would like cost per task calculated and the three routes compared for your workloads, Etzal Group's AI consulting team can help.
Sources
- TechCrunch, 11 November 2026: report on CoreWeave's third-quarter results
- Google, 21 October 2026: e-Conomy SEA 2026 report
Frequently asked questions
Did GPU prices worldwide rise 25%?
No. The figure comes from a single provider, CoreWeave, as reported by TechCrunch. It is not a market-wide number.
Will per-token AI API prices rise as well?
The source does not say so. Treat an increase as a risk to budget for, not a certainty.
How should we budget AI for 2027 given this uncertainty?
Calculate cost per task, prepare a price-rise scenario, ask vendors for pricing clauses, and design systems so models and providers can be swapped.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
