MI355X vs B200 Rentals: $7.70 vs $9.65 an Hour, and 17% Less per Token

Crusoe charges $7.70 per GPU-hour for an AMD MI355X endpoint and $9.65 for an Nvidia B200 endpoint, a 20% gap on one vendor's price page. MI355X rental price gaps at 5 vendors run from 4% to 39%. On MLPerf throughput the Crusoe gap narrows to 17% per token, and it only survives if your software stack matches the benchmark.

Is the MI355X cheaper to rent than the B200? Yes, at every vendor we found publishing both. The gap is 4% to 39% per GPU-hour. After adjusting for MLPerf Server throughput, the same gap is 1% to 37% per token. Few of those rates are plain on-demand, so check the billing type first.

MI355X vs B200 rental prices on dated rate cards

This table holds every per-GPU-hour MI355X and B200 rate we could read from a vendor's own page or public API. Each row carries its billing type, because a preemptible rate and a pay-as-you-go rate are different products. Sorting them into one ranking would mislead you.

Per-GPU rates at Vultr and CoreWeave are node prices divided by 8. Host CPU, memory and network differ by vendor and are not normalised here. "Not listed" means the vendor's readable price page has no such row, which does not rule out a private quote.

Two things stand out before any arithmetic. CoreWeave and Lambda, the names most buyers start with, list no MI355X at all. And the rates are moving up: Crusoe will raise the MI355X by 6.5% and the B200 by 5.7% on 1 December 2026.

The broader market is thinner on the AMD side. The tracker GetDeploying counted 42 providers for the B200 on 6 November 2026, and its 24 priced on-demand configurations had a median of $8.11, up 25% in 90 days. Its MI355X page shows 7 providers and too few rows for a median. Treat those counts as directional, because a tracker is not a vendor. We found one of its rows contradicted by the vendor's own page.

Only 1 paired MI355X and B200 rate is plain on-demand

Of the 5 vendors publishing both chips, only Oracle sells both as ordinary pay-as-you-go instances. That matters more than the headline discount, because the other 4 pairs each carry a condition that changes what you are buying.

Crusoe is a managed endpoint, not a server

The $7.70 and $9.65 figures sit in Crusoe's Self-Serve Deployments table. Crusoe describes these as dedicated endpoints for open and fine-tuned models. If you want a raw instance, its MI355X product page says contact sales. I still lead with this pair because it is the cleanest test available: one vendor, one table, one product and one billing unit.

Preemptible rates can disappear mid-job

Vultr's plans API flags both chips as preemptible only, with on-demand deployment switched off. Daytona's page lists both under a preemptible heading. DigitalOcean's docs say its Spot MI355X at $4.50 can be reclaimed with at least 2 hours of notice. Those rates suit batch jobs that can restart, and they suit almost nothing with a latency promise.

"Starting at" is not a rate card

TensorWave's $2.95 looks like the bargain of the table. Its MI355X page does not state a term or a commitment, and the homepage routes buyers to sales. Cirrascale's "as low as" rates carry the same warning, since its page says hourly pricing is for example only. I would not model either figure until a written quote names the term.

MI355X and B200 throughput in MLPerf v6.0 is close to a tie

The MI355X matched the B200 on MLPerf Inference v6.0 Llama 2 70B, within 3.2% on the two main scenarios. Both rows in the table below are closed-division, 1 node with 8 GPUs, and they come from the same round, so this is as close to like for like as public data gets.

Those shares reconcile with AMD's public claim that it tied on Offline, delivered 97% on Server and reached 119% on Interactive. AMD's marketing numbers therefore hold against the results file. The more useful finding is how small the gap is against normal variation between submissions.

The Server gap is about the size of submission noise

In the prior v5.1 round, 9 submissions of the same 8-GPU B200 configuration ranged from 98,608 to 102,909 tokens per second on Offline. That is a 4.4% spread for one chip across different builders. A 3.2% gap between two chips is not something I would base a purchase on.

Llama 2 70B is a small, old model by current standards. AMD's blog also reports gpt-oss-120b results of 111% and 115% against B200 systems, using partner submissions as the reference. I could not confirm a same-round B200 row for that model, so it stays out of the cost table below. The direction favours AMD, but the size of the lead is AMD's claim and not our measurement.

Cost per token for MI355X and B200 rentals

At full load, the MI355X costs less per token than the B200 at all 5 vendors, by 1.0% to 36.5%. The formula is simple enough to rerun with your own quote. Divide the GPU-hour rate by tokens per GPU-hour, where tokens per GPU-hour is the MLPerf system throughput times 3,600, divided by 8 GPUs.

On MLPerf Server, that gives 45.13 million tokens per GPU-hour for the MI355X and 46.64 million for the B200. So the MI355X breaks even when it is priced 3.2% below the B200, and any larger discount is a real saving.

Zan Digital calculation. Assumes the rented machine reaches the benchmark throughput, which needs the same quantisation and a tuned stack. Preemptible rows exclude the cost of restarted work.

Utilisation matters more than the chip

Every figure above assumes a saturated machine, and no production fleet runs that way. Cost per token rises in direct proportion as utilisation falls. At Crusoe's rates, a B200 at 61% utilisation costs the same per token as an MI355X at 50%. Put another way, a team that gets 21% more useful work out of its B200s cancels the whole discount.

That is why I would not switch chips to save 17% until I had measured my own utilisation. If you are still deciding whether to rent at all, the model of where owning a B200 server beats renting one uses the same inputs from the other direction.

Renting only wins if you beat the API price. Our comparison of inference provider rate cards found gpt-oss-120b listed at $0.60 per 1M output tokens at Together, Fireworks and Groq. On AMD's Server result for that model, a $7.70 MI355X produces tokens at about $0.208 per 1M at full load. Break-even is roughly 35% utilisation, before you count input token revenue, engineers or idle capacity.

GPU memory price: the MI355X costs half as much per GB

The MI355X carries 288 GB of memory to the B200's 180 GB, a 1.6x advantage that MLPerf Llama 2 70B does not reward. Crusoe's page states both capacities. Per GB-hour, the MI355X undercuts the B200 by 40% to 62% across the 5 paired vendors. Memory bandwidth does not separate them: AMD quotes 8 TB/s per GPU, and Nvidia's DGX B200 page quotes 64 TB/s across 8 GPUs, which is the same rate.

• Oracle: $0.0299 vs $0.0778 per GB-hour, 61.6% lower.
• Crusoe: $0.0267 vs $0.0536 per GB-hour, 50.1% lower.
• Vultr: $0.0090 vs $0.0178 per GB-hour, 49.4% lower.
• Cirrascale: $0.0149 vs $0.0266 per GB-hour, 44.0% lower.
• Daytona: $0.0119 vs $0.0199 per GB-hour, 40.1% lower.

This is the strongest argument for the MI355X, and I think it beats the throughput argument. If a model fits on fewer GPUs, you pay for fewer GPUs, and large context and long KV caches eat memory first. The saving is only real if your model or batch size actually needs the room. A small model that fits on either chip gets nothing from it. Memory is also the input with the sharpest price pressure right now, as the analysis of who absorbs the DRAM premium sets out.

AMD's 97% growth is not an MI355X demand signal

AMD's data centre revenue rose 97% to a record $8.5 billion in Q3 2026, but the company did not credit the MI355X for it. The 3 November earnings release attributes the growth to EPYC processors and the launch of MI450 Series GPUs. The results text we read does not name the MI355X at all.

Coverage that links the 97% to MI355X capacity on rental price pages joins 2 things the filing does not join. I would not read neocloud availability from that number. A rate card tells you what a vendor will rent you this week, and a segment total mixes CPUs with next-generation GPUs.

The MI450 launch does raise a question for anyone signing long terms. The MI355X is no longer AMD's newest part, and the same is true of the B200, as the B300 and later Nvidia systems arrive. Crusoe will list the B300 endpoint at $13.80 from 1 December. On Nvidia's own v6.0 Server result of 107,318 tokens per second, that works out to $0.286 per 1M tokens against $0.182 for the MI355X. AMD's commercial commitments to its largest buyers are covered in the breakdown of the OpenAI warrant.

Where this MI355X and B200 comparison is weakest

The argument has 4 real holes, and the first could reverse the conclusion for your workload.

Benchmarks were tuned by the people selling the result

AMD's submission used ROCm 7.0.0 with vLLM and tuned quantisation. HPE's B200 entry used TensorRT. Neither was run by an independent lab, and MLPerf audits rule compliance rather than your experience on a Tuesday. AMD published a reproduction guide, which is to its credit. Getting those figures on your own model is an engineering cost this post does not price, and a slower port can consume a 17% saving quickly.

Newer results exist that we could not use

MLCommons published v6.1 on 16 September 2026. Nvidia's v6.1 results folder has no B200 entry, and AMD's blog on that round would not load for us, so this post uses v6.0. Treat the throughput here as a floor for both stacks, since software keeps improving.

One Nvidia figure sits 10 times away from ours

Nvidia's DGX B200 page cites about $0.02 per 1M tokens on gpt-oss-120b, attributed to SemiAnalysis InferenceX in early 2026. At $9.65 an hour, that would need about 134,000 tokens per second per GPU, against 12,954 per GPU in HPE's Llama 2 Server result. The model, test and cost basis all differ, so neither figure disproves the other. We could not confirm what hourly cost sits behind the $0.02, so it stays out of our table.

Rate cards are not capacity

A listed price does not mean you can get the GPUs. GetDeploying showed Vultr's 8-GPU plans for both chips out of stock on 6 November. Almost no vendor here publishes a reserved rate with a stated term, and reserved is how serious inference capacity is usually bought. Crusoe's own gap between contracted and running capacity is a reminder that supply and announcement are separate facts. Your negotiated number may look nothing like this table.

Frequently asked questions

How to test the MI355X discount this week

Skip the procurement debate and rent one machine of each chip for a day. Run your real model on both, using your real prompt lengths, at the latency you promise customers. Divide each hourly bill by the tokens you actually served, and compare that number with the 3.2% break-even above.

If the MI355X result lands within 5% of the B200, get a written quote that states the term, the host configuration and the reclaim policy. If it lands further behind, your finding is the engineering work needed to close the gap, and that estimate belongs in the price comparison.

Related reading