GPU PRICING → · 2026-11-08 · 12 min read
GPU rental prices have been rising for six weeks, and the usual explanation — not enough chips — no longer fits. The chips exist. What is scarce is somewhere powered to plug them in: an energised building, on a grid connection, under a signed power contract. This page walks through the verified rises, shows why they point at power rather than silicon, and tells you what to ask before you rent capacity that has not been switched on yet.
A note on the verdict above: the dated rises and the figures in it are our verified record — every rise was read on the provider’s own published pricing page on both sides of the change. What the pattern points at is our reading; the argument fills the rest of this page, and you can judge it against the dated facts. Our companion guide, “GPU rental prices are rising in 2026 — the dated evidence”, carries every rise in full and is linked at the foot of this page. This page needs only the shape of it. Between Oct 1 and Nov 6, 2026, five providers raised prices we could verify: Nebius across the board on Oct 1; RunPod’s Secure Cloud on Oct 14; Verda ten times between Oct 23 and Nov 6, roughly 1% at a time, pausing only on Nov 5 and flat since; SaladCloud’s priority tiers on Nov 1; and RunPod’s Serverless B200 rate, up about 25% between our Oct 26 and Nov 2 reads. Lambda’s rise of about 15% is announced for Nov 16 and has not happened yet — we list it because a dated announcement from the provider itself is a fact, but no Lambda price on our board changes until the new rates are in effect and re-read. One thing stands out about who moved. The rises came first and most often from providers selling capacity out of large, power-hungry facilities — the operators closest to the cost of energised space. Nobody in that list launched a new GPU in those weeks, and no chip shortage was announced. The prices moved anyway. That is the observation the rest of this guide explains.
A rented GPU hour looks like a chip product. It is actually the end of a chain, and the price is set at the chain’s slowest link:
| The link | What has to happen |
|---|---|
| The chip | A GPU has to exist and be bought. This was the binding constraint in the shortage years — and it is the link that has eased. Cards ship, launches land on schedule, and the rises above happened with no silicon news attached. |
| The building | The chip earns nothing on a loading dock. It has to sit in a rack, in a hall, with cooling that can carry it — modern AI halls draw power at a density ordinary data centres were never built for. |
| The power | The hall has to be energised: a grid connection approved and built, substations and transformers installed — equipment with lead times measured in years — and a power purchase agreement that fixes what the electricity costs. |
| The queue | Behind each step is a line of other buyers waiting for the same connection, the same transformers, the same contracted megawatts. |
When the slow link was the chip, rental prices tracked chip supply. When the slow link is the energised building, rental prices track powered land — and powered land cannot be manufactured faster by anyone, at any price.
The clearest single example landed this month. On Nov 6, 2026, Crusoe confirmed — in its own announcement, carried by GlobeNewswire — that it has sold its unfinished Wyoming campus to Google. Google had first come to that campus as a customer. It ended up buying the site itself: the land, the build-out and, above all, the power position that comes with it. One caution our news desk applied to this story, and this guide repeats: the sale price was not disclosed, and Google’s separate $17bn Wyoming commitment, announced on blog.google on Nov 5, is not the purchase price — the two figures sit side by side in coverage of this deal, and only one of them is a price for anything. What the sale does show, without any number attached, is where the value in the chain now sits: a hyperscaler that can buy chips in any quantity chose to buy an unfinished campus, because the campus is the part it could not simply order. The quarter’s spending, from the same companies’ own results, shows the scale of the chase for capacity:
| Company | September-quarter capex (reported Oct 28, 2026) |
|---|---|
| Microsoft | $45.2B — including finance leases |
| Alphabet | $51.4B |
| The two together | ≈ $96.6B in one quarter |
The spending figures point the same way. Microsoft and Alphabet reported their September-quarter results on Oct 28, 2026: capital spending of $45.2B and $51.4B — about $96.6B between two companies in one quarter, much of it aimed at AI capacity. And in the same results, Microsoft’s chief financial officer pointed to delayed third-party GPU capacity deliveries while guiding the next quarter’s capex above $50B. Read that slowly: the constraint named by one of the biggest buyers on earth was not chips it could not buy, but capacity — other people’s powered buildings — arriving late. Two further reports belong in this picture, labelled for what they are. IREN’s Nov 5, 2026 results reported H100s being recontracted at double their original price — that is IREN’s own reporting of its own contracts, not a figure we can verify from a pricing page. And Bloomberg has reported that renewed capacity on older GPUs is reselling at a premium of roughly 20%, with Microsoft reportedly walking away from some of it; Bloomberg’s account rests on anonymous sources, so treat it as a lead, not a fact. Both reports are consistent with the same mechanism — when a contract on energised capacity comes up for renewal, the new price is set by the scarcity of powered space, not by the age of the cards inside it — but consistency is not proof, and we print them as reported, not verified.
Everything above would mislead if it left you thinking every GPU costs more than it did in October. It does not. The cheapest verified offers on our board did not move between Oct 19 and Nov 8, 2026:
| GPU | Best verified offer | Provider | Oct 19 → Nov 8 |
|---|---|---|---|
| H100 SXM | $2.25/hr | TensorDock (8-GPU minimum) | — unchanged |
| H100 SXM, single card | $2.69/hr | RunPod Community | — unchanged |
| H200 | $3.59/hr | RunPod Community | — unchanged |
| B200 | $5.98/hr | RunPod Community | — unchanged |
So the two halves of the market are genuinely telling different stories, and both are true. List prices at the big providers — the prices attached to contracted, energised, large-scale capacity — have risen steadily for six weeks. The floor of the open market, where a renter who compares can still buy a single card by the hour, has not moved. If you rent at spot or community rates today, today’s checkout is not where this shortage reaches you. It reaches you at renewal, when a contract reprices against a tighter market for powered space — and it reaches you at scale, when the quantity you need only exists inside the big facilities whose prices are the ones rising. Knowing which situation you are in is worth more than any single rate.
Scarcity of powered space creates a specific sales pattern: capacity sold today against a building that will be energised later. Some of those promises will be kept. Before you build plans on one, five questions separate a phase that exists from one that is hoped for. 1 — Is this capacity energised today, or is it a future phase? Ask which it is, in those words. “Available Q3” can mean racks live in Q3 or a grid connection applied for in Q3, and the difference is your launch date. 2 — Whose power contract stands behind it? A provider with a signed power purchase agreement and an approved connection has something to sell. A provider “in discussions” with a utility has a plan. Ask which document exists. 3 — What happens to my price at renewal? The rises in this guide landed on list prices and renewals, not on contracts mid-term. If the powered-space market is tighter when your term ends, your renewal is where the rise will find you. Ask what, if anything, caps it. 4 — What do I get if the phase slips? Microsoft, with all its leverage, publicly absorbed delayed third-party capacity in its Oct 28 results. A smaller renter should know in advance whether a slipped phase means a refund, a credit, substitute capacity somewhere else — or nothing. 5 — Can I see the same GPU, live, somewhere today? A provider already running the card you want, in a building already energised, has proven the part of the chain that matters. Future capacity from one with no live equivalent is a different risk class.
GPUJournal reads providers’ own published pricing pages and banks the figures daily; a price change is reported only when the before and after figures were each seen on the provider’s own page, with their dates. The rises used here: Nebius’s across-the-board rise of Oct 1, 2026 (post-rise H100 $4.50/hr, carried on our board since our records began Oct 19); RunPod Secure Cloud, Oct 14, 2026 (RTX 4090 $0.74 → $0.89, H200 $4.59 → $5.29); Verda, ten rises between Oct 23 and Nov 6, 2026 (H100 SXM $4.34 read Oct 22 → $4.94 read Nov 6; GB300 $11.82 → $13.44 over the same span; flat Nov 5, 7 and 8); SaladCloud priority tiers in effect Nov 1, 2026 (RTX 4090 High $0.33 → $0.44, RTX 5090 High $0.50 → $0.60, from SaladCloud’s own announcement and pricing page); RunPod Serverless B200 ($8.64 read Oct 26 → $10.79 read Nov 2); Nebius’s removal of its spot-price ceiling, Nov 2, 2026. Lambda’s figures (B200 $6.69 → $7.69, H100 $3.99 → $4.59, A100 80GB $2.79 → $3.19, A100 40GB unchanged at $1.99) are Lambda’s own announcement of Nov 6, 2026, effective Nov 16 — announced, not yet in effect, and not applied to our board. Best offers are from our board check of Nov 8, 2026, unchanged since Oct 19: TensorDock H100 SXM $2.25/GPU/hr at an 8-GPU minimum; RunPod Community H100 SXM $2.69, H200 $3.59, B200 $5.98. The Crusoe sale is Crusoe’s announcement of Nov 6, 2026 (GlobeNewswire); the price was not disclosed, and Google’s $17bn Wyoming commitment (blog.google, Nov 5, 2026) is a separate commitment, not the sale price. Microsoft and Alphabet figures are the companies’ reported September-quarter results of Oct 28, 2026 (capex $45.2B including finance leases, and $51.4B), including Microsoft’s guidance of following-quarter capex above $50B and its reference to delayed third-party GPU capacity deliveries. Two items are reported, not verified, and are labelled as such wherever they appear: IREN’s Nov 5, 2026 results describing H100s recontracted at double their original price (IREN’s own reporting), and Bloomberg’s report of renewed older-GPU capacity reselling at roughly a 20% premium, with Microsoft reportedly walking away from some of it (Bloomberg, anonymous sources). Our records began Oct 19, 2026; nothing in this guide claims a price from before that date except where a provider’s own dated announcement is the source.
Our reading — interpretation, not fact. Everything in this box is GPUJournal’s judgement on the verified record above it; the record itself is in the tables and the method note. If the binding constraint is powered land, list prices will keep drifting up for as long as energised capacity is scarcer than the demand for it — because a grid connection, a substation and a power contract cannot be sped up the way a chip production line can. Expect the pattern of the last six weeks to repeat: rises landing first at the providers closest to the constraint, renewals repricing harder than new spot offers, and the cheapest end of the market moving last, if it moves. The things that would change our reading are concrete and checkable: a run of verified list-price cuts at the big providers, hyperscalers reporting capacity arriving ahead of schedule rather than delayed, or contract renewals reported at flat rates. Until one of those shows up in the record, plan as if capacity you will need in 2027 costs more than the same capacity costs today — and lock the price of anything you cannot afford to reprice.
GPU Pricing · 2026-10-26 · 7 min
If you rent GPUs at a provider's own list rates, you are probably paying more than you did a few weeks ago. This page shows the rises we have verified on providers' own pages — each one dated — and the half of the market that has not moved yet.
GPU Pricing · 2026-11-01 · 10 min
Search for an H100 rental price and you will meet several different numbers, every one of them presented as current. In late October 2026 we counted five different on-demand prices for the same H100 on a single provider’s own pages — on the same day. This guide puts the stated figures next to the verified ones, names the five mechanisms that pile prices up around one GPU, and ends with the three-step check that separates a price from a claim.
GPU Pricing · 2026-10-29 · 8 min
One provider raised its GPU prices on Oct 23, Oct 24, Oct 26, Oct 27, Oct 28 and again on Oct 29 — six rises in seven days, across almost its whole range, each one small and each one in the same direction. We know the dates and the figures because we check and bank these prices every day. Here is the full record, what else moved in October, and the half of the market that still has not.