OVHcloud reprices most of its catalog as its own RAM index hits 6x, with 9x due in September

Start with the pass-through. The steepest increase in the catalog, 87 percent on 2026-edition gaming servers ^[6], is a price multiple of 1.87. Memory, on OVHcloud's own purchasing index, sits at 604 against June 2025, or 6.04x ^[3]. The first is about 31 percent of the second ^[1]. Memory is one line on a bill that also carries chassis, CPUs, drives and network cards, so this is not a clean cost-recovery sum, and the 15 to 20 percent rises on CPUs, motherboards and network cards that Klaba says he has been told to expect are not in the June index at all ^[5]. Even so, the worst-treated customer in the catalog is being asked for less than a third of the input move, and the untouched ranges, Kimsufi and Rise among them, are being asked for nothing ^[7].

The steps matter more than the level. Going from 6x to 9x between June and September is a further 50 percent in three months ^[2]; 9x to 12x by early 2027 is another 33 percent on top of that ^[3]. September's prices are being set against a figure Klaba says he already knows ^[4], and the early-2027 figure is in no published price list yet. That is what the Reddit complaint was actually about. The commenter bammitscamm objected to the timing, not the arithmetic, because April had only just happened ^[17], and two increases inside six months ^[18] make the next one a scheduling question rather than a surprise.

Amazon buys memory in the same market and has repriced one product. EC2 Capacity Blocks for ML rose roughly 20 percent in July after about 15 percent in January, which compounds to around 38 percent across the year ^[4], while the rest of the catalog, Trainium instances included, did not move, and the change appeared on the pricing page without announcement ^[13]. The gap is procurement, not engineering. Hyperscalers contract years ahead in volumes that earn priority allocation, and several design their own accelerators ^[14], whereas OVHcloud, by Klaba's own description, orders month by month over 12 months with no guaranteed purchase price and no firm read on customer demand ^[12]. His stated worry is supply rather than margin: running out of parts and being unable to deliver ^[11].

The saving plan change is where a customer has to decide something. The one-month, six-month and 24-month plans go away, leaving 12 and 36 months, each locking price for its term ^[9]. Klaba puts the end of the exceptional period at 2028 ^[10]. A 12-month lock signed this autumn expires in autumn 2027, inside that window; a 36-month lock runs to autumn 2029, past the far end of his own forecast ^[5].

The Gen3 unbundling on October 1, with storage at EUR 0.000146 per GB per hour and IP addresses at EUR 0.0027 per hour ^[8], is a separate move and reads as one: costs that used to sit inside the instance price now track usage. The planning consequence of all of it is not a percentage. Unit costs for ordinary compute are being restated in multiples, which means a 2027 budget built on 2025 memory prices is wrong by a factor, not by a few points.