A GB200 NVL72 is not a GPU you rent by the card. It is a 72-GPU rack — 36 GB200 Superchips in one NVLink domain — and almost every price quoted for it is a division of a whole-system price, not a rate you can buy one GPU at. On the board we checked on Nov 15, 2026, exactly two providers publish a verified on-demand price for it: Oracle OCI at $16.00 per GPU per hour and Google Cloud at $18.6875. But their smallest cheques differ by a factor of twenty-one: Oracle will sell you one 4-GPU node for $64.00 an hour, while Google’s own page sets an 18-VM minimum for this shape — 72 GPUs, the entire rack — so the smallest Google commitment is $1,345.50 an hour. This guide prices the rack, not the sticker.
The short answer
If you need a full NVL72 rack and want a published price you can budget against today, there is exactly one verified route whose minimum is the rack itself: Google Cloud at $1,345.50 an hour — 18 VMs at $74.75, checked on our board Nov 15, 2026. Run that rack around the clock for a 730-hour month and the verified arithmetic is $982,215, before storage, data transfer and everything else a training programme consumes. If you do not need all 72 GPUs in one NVLink domain, the comparison changes completely. Oracle OCI’s verified node is $64.00 an hour — a buyable 4-GPU unit at $16.00 per GPU. CoreWeave states a cheaper per-GPU division, $10.50, on a $42.00 4-GPU instance (stated, read Nov 15, 2026). Microsoft’s retail list states the most expensive division, $27.04 per GPU on a $108.16 VM (stated, read Nov 15, 2026). Nebius, the third provider listed on our board for this system, publishes no price at all. So the spread that matters is not $16.00 versus $18.6875 — the two verified per-GPU figures sit within 17% of each other. It is the spread of minimum commitments: $42 to $64 an hour to start with a node, $1,345.50 an hour to start with a rack.
TL;DR — every published GB200 NVL72 price we could verify or read first-hand (Nov 15, 2026)
| Provider | Shape you actually buy | Per GPU | Minimum or system total | Checked |
|---|
| Oracle OCI | 4-GPU node | $16.00/GPU/hr | $64.00/hr for one node — the smallest verified buy | Verified Oct 27–28, 2026; board-checked Nov 15, 2026 |
| Google Cloud | 4-GPU VM, 18-VM minimum | $18.6875/GPU/hr | $1,345.50/hr — the minimum is the full 72-GPU rack | Verified Oct 27–28, 2026; board-checked Nov 15, 2026 |
| CoreWeave (stated) | 4-GPU instance (2 Superchips) | $10.50/GPU/hr (division of the stated instance price) | $42.00/hr for one instance | Stated on CoreWeave’s pricing page, read Nov 15, 2026 — not a verified board offer |
| Microsoft Azure (stated) | 4-GPU VM (ND128isr GB200 v6) | $27.04/GPU/hr (division of the stated VM price) | $108.16/hr for one VM (Linux, West US 2) | Stated in Microsoft’s retail price list, read Nov 15, 2026 — not a verified board offer |
| Nebius | Rack-scale clusters | No published price | Its own pricing page lists GB200 NVL72 with no rate; large clusters are “contact us” | Read Nov 15, 2026 |
Every per-GPU figure in this table is a division. Nobody in it sells a single GB200 GPU by the hour: the smallest unit on sale anywhere is four GPUs. The two verified rows are the only GB200 NVL72 offers in GPUJournal’s Nov 15, 2026 snapshot (87 offers). The CoreWeave and Azure rows are provider-stated figures read first-hand on Nov 15, 2026; they are printed here because a buyer will meet them, and they are labelled stated because our board has not verified them as buyable offers. No spot, term or reserved figure appears in this table: the snapshot contains none for this system.
Cheapest verified routes, by what you actually need
| What you need | The route |
|---|
| You need a few GB200 GPUs, not a rack | Oracle OCI’s verified 4-GPU node at $64.00/hr ($16.00/GPU/hr, verified Oct 27–28, 2026, board-checked Nov 15, 2026) is the smallest verified commitment in this market. CoreWeave states a lower entry — $42.00/hr for a 4-GPU instance (stated Nov 15, 2026) — worth quoting against Oracle, with the stated label kept in view. |
| You need the full 72-GPU NVLink domain | Google Cloud’s verified rack total is $1,345.50/hr (18 × $74.75). At Oracle’s verified per-GPU rate, the same 72 GPUs total $1,152.00/hr — but that is arithmetic on Oracle’s rate, not a rack product Oracle lists; whether Oracle will assemble and price a full NVL72 for you is a sales question. |
| You are negotiating, not buying self-serve | Nebius lists GB200 NVL72 with no published price and routes large clusters to its sales team (its pricing page, read Nov 15, 2026). In a negotiation, the verified figures above are your anchors: $16.00 and $18.6875 per GPU per hour are published, dated, and checkable — a quote that lands far above them needs a reason attached. |
What an hour becomes over time
Rack-scale prices are hourly the way mortgages are monthly: the unit is small, the commitment is not. The table below is simple arithmetic on the figures above — 24 hours, a 168-hour week, a 730-hour always-on month. Rows marked “rack-equivalent” multiply a 4-GPU unit price by 18 to reach 72 GPUs; they show what the rate implies at rack scale, not a listed rack product. Only Google’s row is a listed rack-scale minimum.
| Provider | Basis | Per hour (72 GPUs) | 24 hours | 168 hours | 730 hours |
|---|
| Google Cloud | Verified; the 18-VM minimum is the rack | $1,345.50 | $32,292.00 | $226,044.00 | $982,215.00 |
| Oracle OCI | Verified rate, rack-equivalent arithmetic (18 × $64.00 nodes) | $1,152.00 | $27,648.00 | $193,536.00 | $840,960.00 |
| CoreWeave (stated) | Stated rate, rack-equivalent arithmetic (18 × $42.00 instances) | $756.00 | $18,144.00 | $127,008.00 | $551,880.00 |
| Microsoft Azure (stated) | Stated retail meter, rack-equivalent arithmetic (18 × $108.16 VMs) | $1,946.88 | $46,725.12 | $327,075.84 | $1,421,222.40 |
Read the gaps, not just the rows. At rack scale for a month, the difference between the cheapest stated division (CoreWeave, $551,880) and the most expensive stated division (Azure, $1,421,222.40) is about $869,000 — for the same 72 GPUs. The difference between the two verified routes (Google’s listed rack at $982,215 and Oracle’s rate extended to rack scale at $840,960) is about $141,000. These are the numbers a reservation negotiation is actually about.
Lead times and availability: what is actually public
Here is the honest answer first: no provider publishes a delivery lead time for GB200 NVL72 capacity on its public pages. We read the public pricing and product pages of CoreWeave, Nebius, Google Cloud, Microsoft Azure and Oracle OCI for this guide (Nov 15, 2026), and not one states “order today, running in N weeks” for this system. A guide that prints a precise lead time for cloud NVL72 capacity is printing an estimate or a sales anecdote. What is public instead is the shape of each buying route — minimums, booking windows and term structures — and that is what a buyer can plan around:
| What is public | The detail |
|---|
| Minimum units are published; dates are not | Google’s 18-VM minimum (a full rack) is stated on Google’s own page and recorded on our board. Every other seller’s smallest public unit is one 4-GPU node, VM or instance. Nobody attaches a delivery date to any of them. |
| Reservation structures are where the real differences sit | Spheron’s Nov 13, 2026 comparison of reservation terms states that AWS sells GB200 only through EC2 Capacity Blocks that book at most eight weeks ahead, charge the full fee upfront and allow no cancellations; that Azure’s GB200 VM sits outside Azure’s capacity-guarantee product, leaving 1- or 3-year Reserved Instances that discount billing without guaranteeing capacity; and that CoreWeave’s new capacity is typically sold on multi-year contracts (a weighted average of about five years, per CoreWeave’s Q1 2026 earnings call as reported there). Those are Spheron’s stated findings, attributed because they describe contract terms we have not independently verified — but the questions they raise are the right ones to put to any seller. |
| Quote-only routes quote fast, on their own account | Spheron’s own GB200 listing states a quote back within one business day from an 8-GPU node minimum (Spheron’s pages, read Nov 15, 2026). Nebius publishes no equivalent promise; its page simply routes large clusters to sales. |
| The “6 to 12 months” figures are about buying hardware, not renting cloud capacity | Lead times in that range circulate in third-party analyses of purchasing NVL72 racks outright. They are not provider-published cloud delivery times, and this guide does not present them as such. |
The practical consequence: if your programme has a hard start date, the lead time is a contract term to extract in writing from whichever seller you choose — it exists nowhere public to be looked up.
Watch-outs before you commit
| Watch-out | The detail |
|---|
| Do not compare a division with a total | “$10.50/GPU/hr” (CoreWeave, stated) and “$1,345.50/hr” (Google, verified) describe the same system at different slicing. Convert everything to the full shape you will actually run, for the full hours you will actually run it, before ranking anything. |
| GB300 NVL72 is a different system | This guide prices GB200 NVL72 only. GB300 rows you will meet elsewhere — CoreWeave lists its GB300 NVL72 as “contact sales” (read Nov 15, 2026), and our board carries a GB300 label at $14.27/GPU/hr from Verda — are a newer rack generation with its own prices, not a discount on this one. |
| A retail meter is not a reservation | Azure’s stated $108.16 VM meter (West US 2, Linux, Nov 15, 2026) does not include the reservation terms Azure actually sells this capacity under; region changes the meter, too — the same VM lists higher in several other regions of Microsoft’s price list. |
| Stated is not verified | The CoreWeave and Azure figures above are printed exactly as those providers publish them, read Nov 15, 2026. Our board has not verified them as buyable at those figures, and CoreWeave’s volume GB200 business is contract-led in practice. If a stated figure anchors your budget, get it confirmed in a quote before the budget hardens. |
| The rack is not the whole bill | Storage, data transfer, networking between racks, and the staff time of a multi-week negotiation all sit outside every figure in this guide. At $1,345.50 an hour, idle time is the most expensive line: a rack left running unused for one day burns $32,292.00 at Google’s verified rate. |
What the headline rate hides
The per-GPU figure is a division, not a product. An NVL72 rack holds 72 GPUs in 18 compute trays, and every seller slices it the same way: a 4-GPU unit — two Superchips — is the atom. Oracle calls it a node, Google and Microsoft call it a VM, CoreWeave calls it an instance. When a page says “$16.00 per GPU,” it means “$64.00 for the 4-GPU unit, divided by four.” No checkout anywhere sells one GB200 GPU for one hour. A comparison that treats the divided figure as buyable understates the smallest commitment four times over before it starts. Google’s minimum is the whole rack. The second thing the sticker hides is the minimum. Google’s own page notes an 18-VM minimum for this shape (recorded on our board, checked Nov 15, 2026). Eighteen 4-GPU VMs are 72 GPUs: Google is not selling you a slice of an NVL72, it is selling you an NVL72. That makes Google’s $18.6875 the only verified per-GPU figure in this guide that is also a true rack rate — and it makes its $1,345.50 hourly floor twenty-one times Oracle’s $64.00 entry point. Neither is “cheaper” in the abstract. They are different products wearing the same per-GPU costume. A published meter is not the same as capacity you can start. Microsoft’s retail price list publishes a real hourly meter for its ND GB200 v6 VM — $108.16 in West US 2 for Linux, read Nov 15, 2026. A meter existing is not the same as on-demand capacity sitting open: our board carries no Azure GB200 offer — a published meter is not an offer we have verified as buyable self-serve, and Spheron’s Nov 13 summary describes Azure GB200 access as running through Reserved Instances. Treat the $27.04 division as a stated reference price for negotiations, not as a rate you can spin up against this afternoon.
How we check
GPUJournal’s Nov 15, 2026 snapshot contains 87 offers read from providers’ own published pages and banked by date. For GB200 NVL72 the board carries exactly three listings: Oracle OCI and Google Cloud with verified on-demand prices (first verified Oct 27–28, 2026, and still listed in the Nov 15, 2026 board check), and Nebius as quote-only with no published price. Rack totals are arithmetic on those verified figures, labelled wherever the multiplication goes beyond the shape the provider lists. The CoreWeave and Azure figures are stated: CoreWeave’s pricing page was read first-hand on Nov 15, 2026 (GB200 NVL72 4-GPU instance, $42.00/hr on-demand, spot not offered; its GB300 NVL72 is “contact sales”), and Microsoft’s retail price list was read the same day (ND128isr GB200 v6, $108.16/VM/hr for Linux in West US 2; Microsoft’s documentation specifies four Blackwell GPUs and two Grace CPUs per VM). Nebius’s pricing page, read the same day, lists GB200 NVL72 with no price in either its preemptible or on-demand column. RunPod’s pricing page (read Nov 15, 2026) lists no GB200 NVL72, and Hyperstack’s (read Nov 15, 2026 by our raw-page method, 13 rows) lists none either. Live board: https://gpujournal.ai/compare-prices. The system: https://gpujournal.ai/gpu/gb200-nvl72. What is deliberately not priced. AWS carries no GB200 NVL72 offer on our board, so no AWS price appears in this guide’s tables; the Capacity Block terms quoted in the lead-times section are Spheron’s stated summary of AWS’s structure (their guide, Nov 13, 2026), not our verified finding. Lambda’s current rates sit behind a login (checked Nov 15, 2026), so Lambda is not priced here either. Spheron’s own marketplace figures are never treated as verified prices on this site; where Spheron’s claims are mentioned, they are attributed to Spheron. No lead time is printed anywhere in this guide, because no provider publishes one — that absence is reported, not filled.
Our reading
Our reading — interpretation, not fact. The GB200 NVL72 market, as published, is thinner than it looks. Strip away the divisions and there are two verified on-demand prices, one stated instance price, one stated retail meter, and a row of quote-only doors. Spheron’s Nov 13 piece argues the real blocker is reservation fine print rather than price, and on AWS and Azure it makes that case well — but it never totals a rack, and it omits the two providers (Oracle and Google) whose published rates let a buyer total one today: $64.00 an hour to start at node scale, $1,345.50 an hour at rack scale, $982,215 for an always-on month. Price is not the whole decision at this end of the market. It is, however, a knowable number — and a buyer who walks into a quote-only negotiation knowing the verified rack total negotiates from a different position than one who knows only a per-GPU sticker.
GPU Pricing · 2026-11-14 · 13 min
On the board we checked on Nov 14, 2026, the cheapest verified on-demand offers in the table below start at $0.34 an hour for an RTX 4090 and top out at $14.27 an hour for a GB300. That range is real, but it is not the whole answer. The lowest H100 SXM rate, TensorDock's $2.25 per GPU per hour, comes with an eight-GPU minimum — an $18.00 hourly commitment, not a $2.25 rental. This guide separates the sticker price from the amount you actually commit to, then adds the costs the hourly rate leaves out.
GPU Pricing · 2026-10-18 · 10 min
$/hr is the beginning of the bill, not the bill. Spin-up, idle time, storage, egress and the spot-interruption tax — totalled honestly, with dated prices from Oct 18, 2026.
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.