MI325X sits in an odd spot: AMD's own second-tier accelerator, sandwiched between MI300X below it and MI355X above it, and the one nobody's written a dedicated pricing rundown for. That's a gap, because MI325X is also one of only two chips named directly in this year's US export control overhaul, which changes the supply math in a way that MI300X coverage doesn't capture. This post prices it out across the providers currently renting it, explains why availability reads tighter than MI300X's, and works through when its extra memory is worth paying for over a cheaper H100.
For the sibling chips on either side of MI325X, see our AMD MI300X and MI355X pricing breakdown, and for the underlying export policy driving this post's availability numbers, read our full rundown on GPU export controls and cloud pricing.
MI325X Specs: 256GB HBM3e Between MI300X and MI355X
MI325X launched October 10, 2024 at AMD's Advancing AI event, roughly four months after MI300X shipped. It's built on the same CDNA 3 chiplet architecture as MI300X: 8 Accelerator Complex Dies on TSMC's N5 process paired with 4 I/O dies on N6, 304 compute units, 1,216 Matrix Cores. What changed is the memory package. MI325X carries 256GB of HBM3e running at 6TB/s, up from MI300X's 192GB of HBM3 at 5.3TB/s, and that jump pushes TDP from 750W to 1,000W.
FP8 dense compute lands at 2,614.9 TFLOPS (2.61 PFLOPS), rising to 5.22 PFLOPS with structured sparsity, essentially matching MI300X's compute since the silicon is identical. The step up is memory capacity and bandwidth, full stop.
Same CDNA3 Silicon as MI300X, More Memory
Because MI325X and MI300X share the same compute die, the decision between them almost never comes down to raw throughput. It comes down to whether 192GB is enough for your model at your target precision, or whether you need the extra 64GB of headroom MI325X provides. MI355X, the CDNA 4 chip above both of them, pushes further still: 288GB of HBM3e at 8TB/s, plus native FP4 and FP6 support that neither MI300X nor MI325X has. If you're weighing whether that jump is worth it, our MI300X vs NVIDIA H200 comparison covers the CDNA 3 family's positioning against NVIDIA's memory-tier competitor in more depth.
MI325X On-Demand and Reserved Pricing Across Providers (2026)
Direct answer: on-demand MI325X pricing runs roughly $2.25-$3.31/GPU-hr in mid-2026, and that range has widened as reserved capacity has tightened. Smaller neoclouds running on-demand billing, like TensorWave, sit at the floor of that on-demand range; the median has climbed sharply as demand has outpaced new supply coming online. DigitalOcean's contract-based rate is cheaper still, starting at $1.69/GPU-hr, but that's a committed-term price, not on-demand.
DigitalOcean, TensorWave, and the Rest: Real Numbers
DigitalOcean announced MI325X GPU Droplets on July 17, 2025, a month after its MI300X launch, starting at $1.69/GPU/hr through a contract, with single and 8-GPU configurations available. That's the lowest publicly quoted rate we could verify directly. TensorWave, an AMD-first neocloud running bare-metal MI300X, MI325X, and MI355X clusters, has been tracked at $2.25/GPU-hr on-demand, near the low end of the on-demand (non-contract) tier.
| Provider | Billing | Approx. $/GPU-hr | Notes |
|---|---|---|---|
| DigitalOcean | Contract | $1.69 | Single and 8-GPU configurations |
| TensorWave | On-demand | ~$2.25 | AMD-first neocloud, bare metal |
| Market median (mid-2026) | On-demand | ~$3.24 | Up from ~$2.25 in July 2025 |
The gap between DigitalOcean's contract floor and the on-demand median is the same pattern MI300X shows: committing to a term buys a meaningfully lower rate, and paying by the hour costs a real premium for the flexibility to walk away.
Pricing fluctuates based on GPU availability. The prices above are based on 28 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
Reserved and 12-Month Rates vs On-Demand
The wider market's on-demand MI325X range currently spans roughly $2.25-$3.31/GPU-hr depending on provider, instance type, and billing model, with TensorWave's on-demand rate anchoring the low end and premium on-demand neocloud listings anchoring the high end. Committed-term pricing runs well below that on-demand floor: 12-month reserved commitments start around $1.57/GPU-hr, and DigitalOcean's contract-based offering starts at $1.69/GPU-hr, neither of which is on-demand. What stands out is the trend, not just the range: the median on-demand price has risen about 44% since July 2025, climbing from around $2.25/hr to around $3.24/hr per GPU as reserved capacity has tightened. That's the opposite direction MI300X pricing has moved over the same window, where a more mature provider footprint has kept the low end competitive. MI325X's price is going up because supply is getting harder to source, not because demand cooled.
Why MI325X Availability Is Tighter Than MI300X (Export Controls)
The short version: MI325X is one of exactly two chips named by product name in a January 2026 US export control rule, and that rule is actively reshaping who can buy it and at what volume.
On January 15, 2026, the Bureau of Industry and Security published a final rule shifting export license review for NVIDIA H200 and AMD MI325X, specifically, from a presumption of denial to case-by-case review for China-bound exports. The rule applies to chips under a 21,000 Total Processing Performance and 6,500 GB/s DRAM bandwidth threshold, and approval requires exporters to demonstrate US supply protection, buyer due diligence, independent US-based third-party testing, and no diversion of production capacity away from US customers. Volume is also capped at half of what a given exporter ships to its US customers. A 25% tariff on qualifying chips took effect the same month, layered on top of an August 2025 arrangement where NVIDIA and AMD already hand over 15% of China AI chip revenue to the US government.
That case-by-case pathway sounds like an opening, and technically it is one, but the demand side dwarfs it. Tom's Hardware reports that Chinese buyers had orders on file for more than 2 million H200 units against a volume cap that lands close to half that, meaning roughly half of what's on order can actually clear. When a chip sits inside an export-controlled bracket with capped allowable volume, providers serving global customers have less headroom to build out capacity than they do for a chip with no such ceiling, and MI325X's provider list staying thinner than MI300X's is a direct symptom of that. Our export controls and cloud pricing breakdown covers the full rule and what it means for GPU cloud buyers outside China who are competing for the same constrained supply pool.
When 256GB of HBM3e Beats a Cheaper H100 or H200 for Inference
The rule of thumb: MI325X's 256GB matters when your model doesn't fit on a single H100's 80GB at your target precision, or when you'd otherwise need to shard across multiple H100s to hit that footprint. At FP16, roughly 2GB of VRAM per billion parameters is a reasonable estimate for inference, which puts a single MI325X's usable headroom at somewhere around 120B parameters before you'd need to think about multi-GPU serving. Our GPU memory requirements guide walks through that math in more detail, including how context length and KV cache multiply the baseline.
Below that threshold, the comparison usually tilts back toward NVIDIA. H200's 141GB covers a wider range of mid-size models than H100's 80GB without needing AMD's extra capacity at all, and H100's deeper CUDA tooling and TensorRT-LLM support still give it an edge on latency-sensitive, low-batch serving. MI325X's case is strongest specifically for models in the gap between what one H200 can hold and what two GPUs of either vendor would otherwise require, where consolidating onto a single card removes NVLink and tensor-parallelism overhead entirely.
MI325X vs Spheron's NVIDIA Rates: What You'd Pay Instead
Spheron's own catalog doesn't carry AMD Instinct GPUs as of this post's publish date; it's NVIDIA-only. Live rates on the pricing page put H100 on-demand from $2.01/hr, H200 SXM5 from $4.84/hr, and B200 from $9.36/hr.
Set against MI325X's roughly $2.25-$3.31/GPU-hr on-demand range (DigitalOcean's $1.69/GPU-hr contract rate is cheaper but isn't on-demand), H100 comes in close to or below AMD's on-demand low end, but H100's 80GB is nowhere near MI325X's 256GB, so the two aren't a fair memory-for-memory swap. H200's 141GB is the more honest comparison point: at $4.84/hr it costs meaningfully more per GPU-hour than MI325X's ~$2.25-3.24/hr on-demand band, but it also carries NVIDIA's software maturity, TensorRT-LLM support, and a supply chain that isn't subject to the same export-license bracket. If you're running a model that fits in 141GB, H200 avoids the availability question entirely. If you specifically need more than that and don't want to shard, MI325X is the cheaper way to get there, provided you can find a provider with capacity. Our H100 SXM5 rentals on Spheron and H200 pricing pages carry the current live rates if you're pricing out either path.
How to Decide: MI325X, MI300X, or MI355X
Run through these checks in order:
- Does your model fit in 192GB at your target precision? If yes, MI300X is the safer default. It's the more mature chip in terms of provider coverage, and its pricing has stayed comparatively stable rather than climbing the way MI325X's has.
- Do you specifically need the extra 64GB MI325X offers over MI300X? That's the entire reason to pick MI325X over its cheaper sibling. If your model fits in 192GB, there's no cost or performance reason to pay MI325X's premium.
- Can you tolerate export-control-driven availability risk? MI325X's provider list is thinner than MI300X's, and that's tied directly to its inclusion in the January 2026 BIS rule alongside H200. If your deployment timeline can't absorb a provider running out of capacity, budget extra lead time or keep an H200 fallback in your plan.
- Would MI355X's 288GB and FP4/FP6 support justify the jump instead? If you're already accepting AMD's software tradeoffs and shopping the high-memory tier, it's worth pricing MI355X directly rather than assuming MI325X is the ceiling. Our MI300X and MI355X pricing post has the current range for that chip, and our ROCm training guide covers what deploying on either AMD tier actually looks like in practice.
Docs on running production inference on rented GPU infrastructure, including deployment patterns that apply regardless of which chip you land on, are at docs.spheron.ai.
If your model's memory footprint fits inside 141GB, H200 sidesteps the export-control availability questions MI325X carries entirely, and Spheron runs it on-demand with per-minute billing.
Frequently Asked Questions
On-demand MI325X rates run roughly $2.25-$3.31/GPU-hr depending on provider, with TensorWave's on-demand rate at the low end and premium neocloud listings toward the high end. DigitalOcean's $1.69/GPU-hr figure is cheaper, but that's contract pricing, not on-demand. The median on-demand price has climbed about 44% since July 2025 as reserved capacity has tightened.
Yes. MI325X is one of only two chips, alongside NVIDIA's H200, named directly in the January 2026 US export control rule governing China-bound AI accelerators. That puts it under tighter licensing scrutiny and a 25% tariff on qualifying exports, which pulls global supply toward US and allied buyers and keeps neocloud capacity thinner than MI300X's more established provider footprint.
When your model, at your target precision, doesn't fit on a single H100's 80GB without sharding across multiple GPUs. A 256GB MI325X can hold models up to roughly 120B parameters at FP16 on one card, cutting NVLink and tensor-parallelism overhead that a multi-H100 setup carries. Below that size, H100's lower per-GPU rate and deeper CUDA tooling usually win on cost per token.
Not currently. Spheron's catalog is NVIDIA-only as of publication, with H100 on-demand from $2.01/hr, H200 from $4.84/hr, and B200 from $9.36/hr, all live-tracked on the pricing page. If your workload needs MI325X's memory tier specifically, compare it to H200 rather than H100, since H200's 141GB is the closer match.