GPU prices just spiked again: here's what that actually means for your AI budget

The RTX 5090, a gaming card, with some listings reported an increase again. It was nearly $8,500 which was $1,999 when it was launched. 

It is a scary number. Although it is not what lands on your invoice at the end of month. But the scarier part is, GPU’s hourly/monthly rate is not the only number that lands on your invoice when you run AI workloads and rent your compute.

What lands on your invoice is smaller, slower and still worth planning for. Here is what has actually moved, why it moved, and what I would do about it.

Why have GPU prices spiked again?

AI data centres are soaking up DRAM, HBM and GDDR supply, and everything else is being priced around that shortage.

Here is what happened in the first three weeks of October 2026:

• Micron said on its 1 October earnings call that DRAM contract prices would rise another 10 to 15 percent this quarter, and that demand would outrun supply through 2027 and into 2028

• AMD raised the price of the GDDR6 memory it supplies to board partners from 1 October, its fourth memory-driven adjustment in ten months

• Nvidia has reportedly raised some card prices a third time this year, by 20 to 30 percent on certain models

• One price tracker had the RTX 5090's lowest new asking price up 64 percent between early August and mid-October, from about $3,000 to about $4,900

Is this the same as the 2021 GPU shortage?

Not really. In 2021, mining demand evaporated once mining stopped paying. This time the demand comes from data centre spending commitments that run for years.

That is my read, not a forecast, but I would not build a budget around a quick drop.

Are cloud GPU rental prices rising too, or only the cards you buy?

They are rising too, but mostly by less than retail cards. Here is what has been reported recently.

What

Reported move

When

RTX 5090, lowest new asking price

About +64% ($3,000 to $4,900)

Early Aug to mid-Oct

Nebius H100, per GPU-hour

$3.85 to $4.50 (about +17%)

From 1 Oct

Nebius H200, per GPU-hour

$4.50 to $5.40 (about +20%)

From 1 Oct

B200 cloud rental, market index

Up 79% to about $8.01 per GPU-hour

Three months to late Sept

AWS H200 instances

+15%

January 2026

These come from different trackers and providers, so treat them as a pattern, not a single market price. The pattern is clear enough: H100 and H200 rentals are rising in the high teens to around 20 percent where providers have announced increases, and the newest cards are rising much faster.

Why do memory-heavy cards move the most?

Because memory is the scarce part. One market index puts H200 rental at roughly 1.76 times the H100 rate, which is almost exactly the ratio of their memory, 141GB against 80GB.

In other words, you are increasingly paying for gigabytes, not just for compute. That changes how you should choose a card, and I will come back to it below.

How will the spike affect AI teams in India?

Mostly through rental rates, and more slowly than the headlines suggest. For gpu prices india teams actually pay, watch rental listings, not retail cards, since most Indian AI teams rent rather than buy.

A few things shape how it reaches you:

• Rupee rates follow global moves with a lag. CloudPe has forecast that Indian cloud GPU rental prices could rise 15 to 20 percent over the next 12 to 18 months, which is a conditional forecast, not a promise

• Dollar-billed providers add currency movement on top. If your invoice is in USD, a weaker rupee raises your cost even when the hourly rate stays flat

• Subsidised compute is priced differently. The IndiaAI Mission reported more than 45,000 GPUs onboarded by June, at roughly ₹67 to ₹92 per GPU-hour. Those rates are set through the programme, not the open market, so I would expect them to move least. Access is approval-gated, though

How do you work out your own exposure in ten minutes?

Take last month's GPU invoice, strip out storage and bandwidth, and look only at the compute line. Multiply it by 1.15 and by 1.20. That gives you a rough range for what the same usage could cost after the increases.

Here is what that looks like in rupees. If you spend ₹1 lakh a month on GPU time, a 20 percent rise adds ₹20,000 a month, or ₹2.4 lakh over a year. Now check how many of those hours were idle. If a third of them were, cutting that waste saves roughly ₹33,000 a month on the same bill, which is more than the rise costs you.

It is a rough sum, but it turns a vague worry into a number you can take to whoever approves the budget.

What should you actually do with your AI budget now?

Five things, in the order I would do them.

1. Add a buffer: Budget 15 to 20 percent above today's hourly rates for the next year. That is a planning assumption, not a prediction.

2. Fix idle time first: If a GPU sits idle for a third of the month, switching it off recovers more than a 20 percent rate rise costs. It is the only saving here you control completely.

3. Check whether you need the memory: Memory-heavy cards carry the steepest increases. If your model fits on a smaller card, do not rent an H200 or B200 out of habit.

4. Ask for a fixed-rate term quote if your usage is steady. Reserved rates have been climbing too, with one index showing one-year H100 contracts up about 40 percent between October 2025 and March 2026. Locking in costs you flexibility, so only do it for loads you are sure about.

5. Do not buy hardware unless you must: Retail cards are at spike prices. Buying now means paying a premium you may not get back.

Conclusion

A price spike on a gaming card is a headline. A 15-20% rise in your rental rate is a budget line. Treat the second one as likely, plan a buffer, and spend your first effort on the things you control: idle hours, card size and contract terms.

All the info is as of October 2026, and several of these numbers are moving week by week. Check your own provider's current rate card before you lock anything in.

Frequently asked questions

Will GPU prices come back down soon?

Not on the evidence so far. Micron's own guidance points to tight supply through 2027 and into 2028. I would plan as if today's prices are the floor for a while, and treat any relief as a bonus.

Should I wait for prices to drop before starting my AI project? 

Usually no. Waiting means your project sits still while the rental market stays short on capacity. A project that starts now with a 20 percent buffer is better off than one that starts in six months at an unknown price.

Is renting still cheaper than buying after this spike? 

For most teams, yes. Retail prices have risen faster than rental rates, which tilts the sums further toward renting. Buying only makes sense if you will keep a card busy almost every day for a year or more.