AWS EC2 G6 / G6e (NVIDIA L4 and L40S, inference-optimized)

per GPU-hour. g6.48xlarge = 8x L4 at $13.3504/instance-hr / 8 = $1.6688. g6e.48xlarge = 8x L40S at $30.13118/8 = $3.7664. g5.48xlarge = 8x A10G at $16.288/8 = $2.036. Single-GPU g6.xlarge = 1x L4 at $0.8048. Fetched 2026-10-03 from the 2026-09-25 offer file; no regional variation found. PRICE CLASS: this is a 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 are applied PERISHABLE-FIGURE STAMP (added 2026-10-04, uniform across data/supply/ so it is greppable). 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. OBSERVED: 2026-10-03 RE-VERIFY BY: 2027-01-04 - the repo's quarterly re-verification bar. After that date this figure is presumed drifted and must be re-checked against its source before it is quoted again. A price without its observation date is noise; this stamp exists so a reader can tell a fresh figure from a stale one without reading prose. UNIT: per GPU-hour READING RULE: No further perishable flag is published for this figure.

lead_time_weeks is null because the quantity is structurally not applicable: on-demand cloud capacity starts on request, so there is no order-to-delivery lead time. The constraint is the price per GPU-hour and the availability snapshot, not an interval. The one caveat worth keeping: scarce tiers gate capacity behind a reservation or Capacity Block commitment (AWS caps the booking horizon at 8 weeks, source spl-aws-capacity-block-booking-horizon), and that is a commitment window, not a lead time.

EAR 3A001 or EAR99 (below the 3A090.a TPP / performance-density thresholds) + EAR 744.23 (Entity List) + EAR 744.6 (Unverified / MEU lists)

MECHANISM: the part falls BELOW the ECCN 3A090.a/.b TPP and performance-density thresholds, so the advanced-computing licence regime, its case-by-case China/Macau carve-out and its D:5 screening do NOT attach. What remains is the general EAR plus destination-country import rules and screening of the consignee and its ultimate parent against the Entity, Unverified and MEU lists. NO LICENCE REQUIRED for the sale itself; the screening is a party-listing check, not a performance-threshold check. The binding ECCN determination for any specific part is the exporter's own classification, not this record.

THE PRACTICAL INFERENCE FAMILIES on AWS, and the only ones here actually in stock - unlike every p5 record, no $0.00 sentinel appears for g6/g6e and both regions publish real rates. L4 at $1.67/GPU-hr with 24GB is the answer for a small-to-mid model that wants HBM-class bandwidth without Hopper pricing; L40S at $3.77 is the Ada step up with 48GB for models that do not fit. Single-GPU shapes exist and are the sane starting point. Spot applies and AWS advertises up to 90% off on-demand, but I could NOT retrieve a concrete spot rate for these SKUs: the per-instance-type spot table renders client-side and spot-price.s3.amazonaws.com returned HTTP 403. No spot number is recorded here rather than guessing - query describe-spot-price-history before relying on one. AWS's 'up to 40% savings on DL training and HPC infrastructure costs' versus the previous GPU generation is a vendor claim. EXPORT CONTROL (2026-10-04): the accelerator in this instance family sits BELOW the ECCN 3A090.a thresholds - TPP and performance density for this class are under the >= 4800 / >= 5.92 boundary - so the advanced-computing licence regime with its case-by-case and D:5 screening does not attach to the die. Recorded false for that reason, NOT because the part is unregulated: it remains under the EAR and subject to destination-country import rules and to consignee/ultimate-parent screening against the Entity, Unverified and MEU lists under 15 CFR part 744, and a complete system containing it can still be controlled. The practical effect is that L4/L40S/A10 capacity is materially easier to serve to a wide set of destinations than H100-class capacity, which is part of why its price band sits where it does. Source: spl-eccn-3a090-ccl. LEAD TIME: null, and correct as null. For a listed on-demand instance type the provisioning delay is effectively zero - capacity is allocated per request - so a week count would be meaningless; the real gate for scarce shapes is either a quota increase or a capacity reservation, not a lead time. Where a published commitment horizon exists it IS recorded: AWS documents that a Capacity Block may be reserved with a start time up to EIGHT WEEKS in the future (source spl-aws-capacity-block-booking-horizon), which is a floor on how far ahead scarce capacity must be committed, not a wait. On-demand and Spot tiers publish no lead time at all; their constraint is availability, and the repo records that as availability=in-stock or a spot/reality note rather than as a fabricated week figure. Probe log in source spl-negative-no-oem-publishes-lead-time.