H100 or B200 sold out? What to rent instead — priced

First move: check the same card somewhere else

Every rate in the verdict above is the lowest verified on-demand price for that card on our board on Nov 12, 2026 — one best offer per card, the same rule the board itself uses. Terms that change the bill (minimums, whole-node rental, VAT) are printed with the price wherever they apply. “Sold out” is almost always a sentence about one provider's shelf. The H100 SXM is listed by eight providers on our board, and the spread between them is wider than the gap between most substitutes: on Nov 12, 2026 the same card, on-demand, runs from $2.25 to $5.14 a GPU-hour before you reach the hyperscalers — TensorDock $2.25 (8-GPU minimum), RunPod Community $2.69, Thunder Compute and Hyperstack $3.20 each, RunPod Secure and Lambda $3.99 each, Verda $5.14. Microsoft Azure lists it at $12.29 per GPU — but that is an 8-GPU VM at $98.32 an hour, a whole machine, not a card you can rent one of (see “What not to do”). Before you substitute anything, price the card itself across the board — the substitute decision only starts when the card is gone everywhere you would actually rent it. The B200 tells the same story. Nine verified offers, Nov 12, 2026: RunPod Community $5.98, Hyperstack $6.00, Lambda $6.69, RunPod Secure $7.99, Nebius $8.50 (excluding VAT), CoreWeave $8.60 (an 8-GPU node — $68.80 an hour for the node), Verda $9.32, then the whole-node hyperscalers, Oracle OCI $14.00 and Google Cloud $18.53, both per-GPU figures divided out of 8-GPU machines. One warning with a date on it: Lambda has announced its B200 rises to $7.69 on Nov 16, 2026 (Lambda's own announcement of Nov 6) — if your fallback plan is “Lambda when RunPod is full,” it gets 15% more expensive on Monday.

Why these two cards go first

The H100 SXM and the B200 sit where the demand is: the H100 is the default training and serving card of the last three years, and the B200 is what the biggest jobs moved to. Our own record — every move read on providers' own pages and banked daily — shows what that demand does to prices:

Rising prices are consistent with growing demand — that is our reading of the pattern, labelled as interpretation; capacity and cost moves can push the same way. Either way, the practical fact is the same: the two most-wanted cards are the two whose shelves empty first — and whose prices move against you while you wait. A substitute plan made in advance, at checked prices, beats a scramble.

The substitutes, ranked by closeness

Same class, one step up — H200 for H100, B300 for B200. The H200 is the H100's own generation with 141GB of faster memory: anything an H100 SXM runs, an H200 runs, with headroom. Its best verified rate is RunPod Community's $3.59/GPU/hr (Nov 12, 2026) — $0.90 an hour above the cheapest single-card H100 SXM, and cheaper than four of the eight H100 SXM offers on the board. For the B200, the B300 is the same Blackwell family one step up, with 288GB: Hyperstack $7.40, RunPod Secure $8.99, Nebius $9.50 excluding VAT (all Nov 12, 2026). Stepping up costs more per hour and can cost less per job — a bigger card finishes sooner — but we publish prices, not benchmarks, so we make that claim no further than this sentence.

The same chip in a different shape — H100 PCIe and H100 NVL. The H100 PCIe is the same silicon at a lower power limit, without NVLink: for one card doing serving or fine-tuning, the difference barely shows; for cards that must work together, it is the whole ballgame. RunPod Community $1.99, Hyperstack $2.50, RunPod Secure $2.89 (Nov 12, 2026). The H100 NVL carries 94GB — the card for a model that nearly fits on 80GB — at RunPod Community $2.59 and RunPod Secure $3.19, but it is sold as linked pairs, not as a lone card, so budget for two. One step down — A100 80GB. The honest workhorse: same 80GB as the H100, an older generation, no FP8. If your stack predates FP8 or your work is 32B–70B models in INT8/INT4, it does the job at about 40% of the H100's cheapest single-card rate: Thunder Compute $1.09, RunPod Community $1.19 (PCIe) and $1.39 (SXM), TensorDock $1.50 (Nov 12, 2026). Below it, the A100 40GB (Verda $1.69, Lambda $1.99) covers smaller models and learning workloads — but 40GB fills up fast, and a substitute that runs out of memory is not a substitute. One step sideways — two smaller cards, or a different 96GB. For work that splits — batch jobs, parallel experiments, one model replica per card — two cards can beat one big one on price: two A100 80GB at Thunder Compute cost $2.18 an hour together, less than one single-card H100 SXM at RunPod Community ($2.69), and put 160GB of memory on the table across two cards. What two cards cannot do is pool that memory into one fast whole without NVLink; a single model that needs 100GB in one place still needs one big card. For exactly that job there is a sideways option this board keeps flagging: the RTX PRO 6000, 96GB on a single card at RunPod Community $1.69/GPU/hr (Nov 12, 2026) — our board's most VRAM per dollar, and enough for a 70B model on one card, quantized or in FP8. And one specialist: CoreWeave's GH200 ($6.50, 1-GPU instance, 96GB) pairs its GPU with an ARM processor on one fabric — powerful for work built for it, and a software-change risk for work that was not.

What not to do

The worked comparison: 100 hours — one GPU

Assumptions, stated plainly: 100 hours of GPU time, on-demand, card for card, at the Nov 12, 2026 rates above. Hours are not jobs — a faster card may finish your job in fewer hours, and we hold no verified timings, so we price the hours and leave the speed judgement to you. Storage and downloads are excluded; they vary by provider and are not part of the card rate. What the table prices is the decision a sold-out message forces: pay more per hour for a closer card, or less for a further one.

The worked comparison: 100 hours — eight GPUs

Read the first table's shape, not just its numbers: over 100 hours, the whole substitute decision for a single card is worth between $70 saved and $90 extra. Against that, price the wait honestly — an idle team, a missed deadline, a training run that starts next week instead of tonight. We put no figure on your deadline, because only you know it; but when the premium for the closest substitute is $0.90 an hour, “wait for the card I wanted” is usually the most expensive option on the page.