Renting an AI chip is starting to feel like booking a hotel in a sold-out city. You pay to hold the room, and the rate keeps climbing. On AWS, it just climbed again.
Amazon Web Services has raised prices for EC2 Capacity Blocks for ML by roughly 20%, starting in July. Business Insider first reported the change, and AWS confirmed it. The service lets companies reserve Nvidia GPUs in advance, so a long training run keeps going instead of stalling halfway.
This is the second increase in six months. AWS had already lifted the same prices by about 15% in January. Stacked together, the cost of locking in this compute has jumped sharply since the new year. AWS said the prices change “periodically based on supply and demand.”
The rise is narrow, not blanket. It hits one purchasing option: the reserved blocks favoured by serious AI teams training or fine-tuning large models. Other options keep fixed prices, AWS said, and the company says it will hold them there. The increase spared Trainium, Amazon’s in-house AI chip, according to The Information.
The scope still matters, because of how much sits on top. AWS is the world’s largest cloud provider, and a sprawl of AI services runs on its servers. When the priciest tier of its compute goes up, the cost ripples out to the start-ups and enterprises renting it. AWS, for its part, framed the move as proof of how strong demand for GPUs has become.
Why the cloud could not stay cheap
The shift worth watching is where the squeeze has reached. For two years, the limit on AI was software and know-how. Now it is physical. The bottleneck is high-bandwidth memory, the chips stacked beside AI processors, and there is only so much of it to go around. The constraint on AI has moved from code to silicon, as Business Insider put it, and silicon takes years to build.
The chain is short and unforgiving. Less memory means fewer GPUs. Fewer GPUs means fewer data centres. “As there is a limit to how much memory can be produced, then there is a limit to how many GPUs can be produced, which means that there’s a limit to how many data centers can be built,” Peter Berezin, chief economist at BCA Research, wrote on X.
That scarcity hands the cloud giants a lever. When GPU capacity is tight, customers have few alternatives, so AWS, Microsoft, Google, and Oracle can pass higher costs straight through. The shortage raises their own bills, Berezin noted, yet it also keeps demand above supply, which gives them pricing power over who gets to compute at all.
A price rise that is now everywhere
Amazon is not alone, and that is the point. Apple raised prices across its Macs and iPads this week, blaming memory. Xbox did the same. Elon Musk called the jump in memory the biggest price increase he has seen in anything. Now the same memory prices are turning up in cloud bills.
The other side of the squeeze is a windfall. The shortages lifting AWS GPU prices have pushed memory makers Micron and SK Hynix to record valuations. Investors are betting the high-bandwidth memory crunch keeps the market tight, and prices high, for years.
For AWS customers, the message is plain. The cheapest AI compute is behind them, and the reserve button now costs more to press. The open question is how far this travels. If a 15% rise became another 20% in six months, the teams building the most ambitious models are left guessing what the next reservation will cost.
Stephan is the sports journalist for the Maple Grove Report.
