Oracle OCI bare-metal AMD Instinct MI300X / MI355X per-GPU-hour list price

PER AMD MI300X GPU-HOUR, ON-DEMAND list price (ASKING, not realised, not committed-use). USD, observed 2026-10-05 from Oracle's own 'GPU - Accelerated Compute' price table, which states the column as 'GPU price per hour' and footnotes it '\* The server price per hour is calculated by multiplying the GPU price per hour by the number of GPUs.' - i.e. the figure is per CARD and a node price is the per-GPU rate times the GPU count. price_usd = MI300X. MI355X is $8.60 per MI355X-hour. Full table read on the same page 2026-10-05: BM.GPU.MI300X.8 = 8x AMD MI300X 192GB Matrix Core, CDNA 3, 8x1x400 Gb/sec RDMA, $6.00/GPU-hr; BM.GPU.MI355X.8 = 8x AMD MI355X 288GB Matrix Core, CDNA 4, 8x1x400 Gb/sec RDMA, $8.60/GPU-hr. The same column on the same page gives NVIDIA BM.GPU.H100.8 = $10.00, BM.GPU.H200.8 = $10.00, BM.GPU.B200.8 = $14.00, BM.GPU.B300.8 = $15.00, BM.GPU.A100-v2.8 (8x A100 80GB) = $4.00, BM.GPU4.8 (8x A100 40GB) = $3.05, BM.GPU.L40S.4 = $3.50, BM.GPU.RTXPro.8 = $4.50, BM.GPUA10.4 = $2.00, V100 = $2.95, P100 = $1.275. OCI ALSO publishes a separate 'NVIDIA AI Enterprise' line with much lower figures - H100 $2.50, H200 $2.50, B200 $3.50, GB200 $4.00, A100 80 $1.00, L40S $0.88, A10 $0.88, A100 40 $0.76 - which is a DIFFERENT product line (an enterprise licence tier), NOT the BM.GPU accelerated-compute rate; do not mix the two. PRICE CLASS: PUBLISHED LIST PRICE - an asking rate a new customer is quoted, NOT a realised, discounted or negotiated transaction rate. A real bill is at or below this and typically well below it once commitments, credits and committed-use discounts apply. RE-VERIFY BY: 2027-01-04. UNIT: PER AMD MI300X GPU-HOUR READING RULE: DO NOT DIVIDE BY 8. Oracle's own footnote defines the figure as a per-GPU rate; the '.8' in BM.GPU.MI300X.8 is the GPU count in the bare-metal shape name, exactly the trap that was already caught once in this repo on Oracle's H100 record (oracle-bm-gpu-h100-h200) and on CoreWeave (coreweave-h100-h200).

lead_time_weeks is null because the quantity is structurally not applicable: this is on-demand cloud bare-metal capacity that starts on request, so there is no order-to-delivery lead time. The constraint on actually getting an MI300X hour is capacity, not lead time. Oracle's own long-standing pattern of out-of-capacity errors on GPU shapes means availability should be treated as UNVERIFIED from the price list alone - the price list says nothing about capacity, and this repo has an explicit finding on that in oracle-bm-gpu-h100-h200. Do not plan around this record without a real account.

THIS RECORD SUPPLIES THE MISSING DENOMINATOR FOR THE REPO'S LONGEST-STATED COST GAP. IT IS A CORRECTION RATHER THAN AN ADDITION. READ THE 'WHAT IT DOES NOT YET DO' SECTION BEFORE ASSUMING THE GAP IS CLOSED IN THE TOOLING. IMPORTANT LIMIT, stated first because it is the thing a reader would otherwise get wrong: the price-THROUGHPUT join in tools/cost_per_token.py is driven by a hand-maintained PRICE_QUOTES table in tools/, not by scanning data/supply. Adding this record does NOT by itself make the 13 AMD benchmarks appear in that tool's table, and `python tools/cost_per_token.py gaps` still reports them as unpriceable. The supply-side fact is established; the tooling needs one entry per accelerator added to PRICE_QUOTES, and tools/ is outside this wave's write scope, so that edit is left to a sibling. Verified on 2026-10-05: gaps output still shows 11 'no priced per-GPU-hour supply record for amd-instinct-mi355x' rows and 2 for amd-instinct-mi300x after this record was added. WHAT WAS WRONG BEFORE. docs/08-cost-per-token.md, section 'What the records cannot answer yet', item 2, states: 'Every AMD MI300X/MI325X/MI355X benchmark is unpriceable. Seven throughput records ... because no supply record prices AMD Instinct per GPU-hour. Every AMD supply record (amd-instinct-oem-systems, dell-poweredge-xe-ai, hpe-proliant-compute-xd, qct-gpgpu-servers) has price_usd: null. This is the highest-value gap in the repo.' The mechanism behind the tool's own gaps output repeats the same claim 11 times for amd-instinct-mi355x plus 2 for amd-instinct-mi300x. As of 2026-10-05 that statement is FALSE: Oracle publishes a per-GPU-hour price for both MI300X and MI355X, on a page that also prices NVIDIA H100, so the join that was described as impossible is now possible. WHY THE EARLIER NEGATIVE WAS WRONG, NOT JUST OUT OF DATE. The four AMD supply records named above are all correct as they stand: they describe OEM and reseller channels (Dell, HPE, QCT, AMD's own OEM programme) and none of those channels publishes a per-GPU-hour rate, because per-GPU-hour is not a unit an OEM system sale is quoted in. The error was one of SCOPE: the claim generalised from 'AMD hardware is unpriceable' to 'AMD hardware is unpriceable', when the correct statement was 'no SUPPLY RECORD in this repo prices AMD per GPU-hour'. The distinction matters because the fix is a supply record, and the supply channel that publishes per-GPU-hour rates for AMD is a cloud provider, not an OEM. This is the same shape of error as rent-cerebras-api-token-pricing-negative earlier in this repo: a negative asserted on one channel generalised into a negative on the whole market. WHY TWO METHODS BEFORE ASSERTING IT. Method 1: the price list fetched as text, which returns the GPU table with its price column populated. Method 2: the same page's live rendered DOM read through a browser, returning the literal innerText of each matching table row - 'BM.GPU.MI300X.8 8x AMD MI300X 192 GB Matrix Core CDNA 3 8x1x400 Gb/sec RDMA $6.00' and 'BM.GPU.MI355X.8 8x AMD MI355X 288GB Matrix Core CDNA 4 8x1x400 Gb/sec RDMA $8.60'. Both agree exactly. A third path, a plain-text reader proxy, returned the shape names and the GPU counts but DROPPED the price column entirely; had only that path been used, this record would have been filed as another negative and the repo's error would have been deepened rather than corrected. Source w5s-oracle-gpu-accelerated-compute-table. WHAT IT BUYS, AND WHAT IT STILL DOES NOT. The supply-side fact is now established: MI300X ($6.00/GPU-hr) and MI355X ($8.60/GPU-hr) have a per-GPU-hour denominator, which is what the 11 amd-instinct-mi355x and 2 amd-instinct-mi300x benchmarks need to become priceable. Two caveats a reader must not lose: (1) $6.00 is an ON-DEMAND ASKING rate on a bare-metal shape with no vCPU/NRAM bundling, so a like-for-like AMD-vs-NVIDIA comparison must use Oracle's own H100 figure of $10.00 rather than AWS p5's $6.88, which bundles 24 vCPU and 256GB of RAM into the GPU hour; (2) Oracle's availability on GPU shapes is unverified from any public page, so these figures describe price, not purchasability. WHAT IT DOES NOT YET DO. tools/cost_per_token.py reads a hand-maintained PRICE_QUOTES table rather than scanning data/supply, so until entries for amd-instinct-mi300x and amd-instinct-mi355x are added there - an edit in tools/, outside this wave's write scope - the tool's table and its gaps output are unchanged, and 13 benchmark records remain listed as unpriceable. A sibling should add two entries of the form dict(accel='amd-instinct-mi300x', usd=6.00, tier='on-demand', gpus=1, basis='BM.GPU.MI300X.8 = $6.00') and the MI355X equivalent, which would make AMD the largest joinable hardware block after NVIDIA. Until then the correct statement is that the gap is closed on the supply side and still open in the tooling. WHY THE FIGURE IS CHEAPER THAN THE NVIDIA ONE, WHICH IS THE INTERESTING PART. On the same page and the same column, MI300X ($6.00) is 0.60x the H100 ($10.00) and MI355X ($8.60) is 0.86x it. That is a hardware-price statement and it does NOT license any claim about cost per token: it says nothing about throughput, and the repo's own MLPerf records are the only thing that could bridge that gap. Pairing them is out of scope for this record and belongs in docs/.