us-west-2 NVMe Spot market review

Technical summary

All 282 configured partition/type/AZ memberships are structurally offered and have Linux/UNIX Spot price history. The actionable problem today is not missing EC2 offerings; it is price gating and weak large-instance capacity.

The strongest recommendation is to temporarily exclude 20 partition/type/AZ memberships whose 72-hour minimum Spot price never fell below DYEC's $9.99/hour global ceiling. Most are i384nvme types. Leave the other currently price-gated types defined: their prices are nearer the computed resource bid, and some have successful launches in the same 72-hour window.

The account sample contains 469 relevant Spot `RunInstances` events: 229 launched instances, 20 insufficient-capacity errors, and 255 `SpotMaxPriceTooLow` errors. Those errors are workload-correlated, but they confirm that price-cap gating is real. The i384nvme placement score is only 1-2 in every AZ, so do not depend on 384-vCPU Spot capacity for a cluster built now.

Best current AZ by partition

i96nvme — All four AZs score 9; us-west-2c has no currently price-gated type.
i128nvme — us-west-2a scores 6 (best), after excluding x2iedn.metal; us-west-2b is next at 3.
i192nvme — us-west-2b scores 9 (best) with no current price-gated type.
i192hugenvme — us-west-2b scores 7 (best) with no current price-gated type.
i384nvme — No healthy Spot choice: scores are 1, 1, 1, and 2. If unavoidable, us-west-2d is least weak; prefer the on-demand template for reliability.

Source: DYEC us-west-2 Spot market evidence snapshotFile: docs/plans/20260716T031101Z_usw2_spot_market_review_snapshot.json

Read-only EC2 offerings, Spot price history, placement scores, CloudTrail RunInstances evidence, and current working-tree template membership.

Spot price distribution across configured NVMe types

Boxes compare the configured instance-type price distributions within each partition and AZ; individual dots preserve every type.

Source: DYEC us-west-2 Spot market evidence snapshotFile: docs/plans/20260716T031101Z_usw2_spot_market_review_snapshot.json

Read-only EC2 offerings, Spot price history, placement scores, CloudTrail RunInstances evidence, and current working-tree template membership.

Partition-pool Spot placement scores by AZDirectional 1-10 score for an eight-node-equivalent target; 10 is strongest.

Source: SQLite derivation for partition placement scoresTable: partition_summary

Actual query executed against the normalized partition summary rows.

SELECT partition, availability_zone, placement_score
    FROM partition_summary
    ORDER BY CASE partition
      WHEN 'i96nvme' THEN 1 WHEN 'i128nvme' THEN 2 WHEN 'i192nvme' THEN 3
      WHEN 'i384nvme' THEN 4 WHEN 'i192hugenvme' THEN 5 END,
      availability_zone

Partition-pool Spot placement scores by AZ data
Partition	Placement score	Availability Zone
i96nvme	9	us-west-2a
i96nvme	9	us-west-2b
i96nvme	9	us-west-2c
i96nvme	9	us-west-2d
i128nvme	6	us-west-2a
i128nvme	3	us-west-2b
i128nvme	2	us-west-2c
i128nvme	1	us-west-2d
i192nvme	3	us-west-2a
i192nvme	9	us-west-2b
i192nvme	1	us-west-2c
i192nvme	5	us-west-2d
i384nvme	1	us-west-2a
i384nvme	1	us-west-2b
i384nvme	1	us-west-2c
i384nvme	2	us-west-2d
i192hugenvme	3	us-west-2a
i192hugenvme	7	us-west-2b
i192hugenvme	1	us-west-2c
i192hugenvme	1	us-west-2d

us-west-2a NVMe partition evidence

Source: SQLite derivation for us-west-2a partition summaryTable: partition_summary

Actual query executed against the in-memory normalized market snapshot.

SELECT partition, configured_types, offered_types, placement_score,
          launched_instances_72h, capacity_errors_72h, price_cap_errors_72h,
          exclude_today_count
        FROM partition_summary
        WHERE availability_zone = 'us-west-2a'
        ORDER BY CASE partition
          WHEN 'i96nvme' THEN 1 WHEN 'i128nvme' THEN 2 WHEN 'i192nvme' THEN 3
          WHEN 'i384nvme' THEN 4 WHEN 'i192hugenvme' THEN 5 END

us-west-2a NVMe partition evidence
Partition	Types	Score	Exclude today
i96nvme	31	9	0
i128nvme	13	6	1
i192nvme	14	3	1
i384nvme	10	1	5
i192hugenvme	8	3	0

us-west-2b NVMe partition evidence

Source: SQLite derivation for us-west-2b partition summaryTable: partition_summary

Actual query executed against the in-memory normalized market snapshot.

SELECT partition, configured_types, offered_types, placement_score,
          launched_instances_72h, capacity_errors_72h, price_cap_errors_72h,
          exclude_today_count
        FROM partition_summary
        WHERE availability_zone = 'us-west-2b'
        ORDER BY CASE partition
          WHEN 'i96nvme' THEN 1 WHEN 'i128nvme' THEN 2 WHEN 'i192nvme' THEN 3
          WHEN 'i384nvme' THEN 4 WHEN 'i192hugenvme' THEN 5 END

us-west-2b NVMe partition evidence
Partition	Types	Score	Exclude today
i96nvme	31	9	0
i128nvme	13	3	0
i192nvme	14	9	0
i384nvme	10	1	6
i192hugenvme	8	7	0

us-west-2c NVMe partition evidence

Source: SQLite derivation for us-west-2c partition summaryTable: partition_summary

Actual query executed against the in-memory normalized market snapshot.

SELECT partition, configured_types, offered_types, placement_score,
          launched_instances_72h, capacity_errors_72h, price_cap_errors_72h,
          exclude_today_count
        FROM partition_summary
        WHERE availability_zone = 'us-west-2c'
        ORDER BY CASE partition
          WHEN 'i96nvme' THEN 1 WHEN 'i128nvme' THEN 2 WHEN 'i192nvme' THEN 3
          WHEN 'i384nvme' THEN 4 WHEN 'i192hugenvme' THEN 5 END

us-west-2c NVMe partition evidence
Partition	Types	Score	Launched	Capacity errors	Price-cap errors	Exclude today
i96nvme	31	9	76	1	0	0
i128nvme	13	2	46	2	0	0
i192nvme	14	1	57	12	13	0
i384nvme	10	1	2	0	186	5
i192hugenvme	8	1	42	8	13	0

us-west-2d NVMe partition evidence

Source: SQLite derivation for us-west-2d partition summaryTable: partition_summary

Actual query executed against the in-memory normalized market snapshot.

SELECT partition, configured_types, offered_types, placement_score,
          launched_instances_72h, capacity_errors_72h, price_cap_errors_72h,
          exclude_today_count
        FROM partition_summary
        WHERE availability_zone = 'us-west-2d'
        ORDER BY CASE partition
          WHEN 'i96nvme' THEN 1 WHEN 'i128nvme' THEN 2 WHEN 'i192nvme' THEN 3
          WHEN 'i384nvme' THEN 4 WHEN 'i192hugenvme' THEN 5 END

us-west-2d NVMe partition evidence
Partition	Types	Score	Launched	Capacity errors	Price-cap errors	Exclude today
i96nvme	23	9	4	0	0	0
i128nvme	5	1	4	0	0	0
i192nvme	12	5	29	1	0	0
i384nvme	8	2	11	4	56	2
i192hugenvme	6	1	20	1	0	0

Exclusions for clusters built today

us-west-2a: x2iedn.metal, m8id.48xlarge, m8idb.96xlarge, m8idn.96xlarge, r8id.metal-96xl, r8idb.96xlarge, r8idn.96xlarge
us-west-2b: m8idb.96xlarge, m8idn.96xlarge, r8id.96xlarge, r8id.metal-96xl, r8idb.96xlarge, r8idn.96xlarge
us-west-2c: m8id.96xlarge, m8idb.96xlarge, m8idn.96xlarge, r8idb.96xlarge, r8idn.96xlarge
us-west-2d: m8idb.96xlarge, m8idn.96xlarge

These are temporary market exclusions, not permanent claims about the instance families. Re-run the collector before a later build.

Source: DYEC us-west-2 Spot market evidence snapshotFile: docs/plans/20260716T031101Z_usw2_spot_market_review_snapshot.json

Read-only EC2 offerings, Spot price history, placement scores, CloudTrail RunInstances evidence, and current working-tree template membership.

us-west-2a temporary exclusions

Source: SQLite derivation for us-west-2a exclusionsTable: type_audit

Actual query executed against the normalized per-type audit rows.

SELECT partition, instance_type, current_spot_usd_hour,
          min_72h_usd_hour, computed_bid_usd_hour, price_cap_failure_events
        FROM type_audit
        WHERE availability_zone = 'us-west-2a' AND recommendation = 'EXCLUDE TODAY'
        ORDER BY partition, instance_type

us-west-2a temporary exclusions
Partition	Instance type	Current	72h min	Today's bid
i128nvme	x2iedn.metal	$12.97	$12.97	$7.58
i192nvme	m8id.48xlarge	$10.64	$10.64	$8.5
i384nvme	m8idb.96xlarge	$10.91	$10.89	$9.99
i384nvme	m8idn.96xlarge	$10.79	$10.79	$9.99
i384nvme	r8id.metal-96xl	$13.14	$13.14	$9.99
i384nvme	r8idb.96xlarge	$14.64	$14.61	$9.99
i384nvme	r8idn.96xlarge	$13.84	$13.84	$9.99

us-west-2b temporary exclusions

Source: SQLite derivation for us-west-2b exclusionsTable: type_audit

Actual query executed against the normalized per-type audit rows.

SELECT partition, instance_type, current_spot_usd_hour,
          min_72h_usd_hour, computed_bid_usd_hour, price_cap_failure_events
        FROM type_audit
        WHERE availability_zone = 'us-west-2b' AND recommendation = 'EXCLUDE TODAY'
        ORDER BY partition, instance_type

us-west-2b temporary exclusions
Partition	Instance type	Current	72h min	Today's bid
i384nvme	m8idb.96xlarge	$10.42	$10.42	$9.99
i384nvme	m8idn.96xlarge	$11.02	$10.79	$9.99
i384nvme	r8id.96xlarge	$11.07	$11.07	$9.99
i384nvme	r8id.metal-96xl	$10.46	$10.46	$9.99
i384nvme	r8idb.96xlarge	$13.54	$13.44	$9.99
i384nvme	r8idn.96xlarge	$13.77	$13.66	$9.99

us-west-2c temporary exclusions

Source: SQLite derivation for us-west-2c exclusionsTable: type_audit

Actual query executed against the normalized per-type audit rows.

SELECT partition, instance_type, current_spot_usd_hour,
          min_72h_usd_hour, computed_bid_usd_hour, price_cap_failure_events
        FROM type_audit
        WHERE availability_zone = 'us-west-2c' AND recommendation = 'EXCLUDE TODAY'
        ORDER BY partition, instance_type

us-west-2c temporary exclusions
Partition	Instance type	Current	72h min	Today's bid	Cap failures
i384nvme	m8id.96xlarge	$25.06	$25.06	$9.99	0
i384nvme	m8idb.96xlarge	$11.41	$11.41	$9.99	55
i384nvme	m8idn.96xlarge	$12.09	$10.98	$9.99	26
i384nvme	r8idb.96xlarge	$14.93	$14.9	$9.99	53
i384nvme	r8idn.96xlarge	$14.11	$14.11	$9.99	48

us-west-2d temporary exclusions

Source: SQLite derivation for us-west-2d exclusionsTable: type_audit

Actual query executed against the normalized per-type audit rows.

SELECT partition, instance_type, current_spot_usd_hour,
          min_72h_usd_hour, computed_bid_usd_hour, price_cap_failure_events
        FROM type_audit
        WHERE availability_zone = 'us-west-2d' AND recommendation = 'EXCLUDE TODAY'
        ORDER BY partition, instance_type

us-west-2d temporary exclusions
Partition	Instance type	Current	72h min	Today's bid	Cap failures
i384nvme	m8idb.96xlarge	$10.96	$10.96	$9.99	15
i384nvme	m8idn.96xlarge	$10.55	$10.55	$9.99	15

Scope, data, and definitions

Window: 2026-07-13T03:11:01Z through 2026-07-16T03:11:01Z (72 hours).
Price: EC2 Linux/UNIX Spot price per instance-hour in the named AZ.
Boxplot value: one 72-hour median price per configured instance type, so volatile types do not receive extra statistical weight merely because their price changed more often.
Placement score: eight-node-equivalent target expressed in vCPU, using AWS's required capacity-optimized scoring assumption. The templates use price-capacity-optimized, so the score is directional.
Exclude today: type is offered, current price exceeds every containing compute-resource bid, and its 72-hour minimum remained above the $9.99 global cap.

Methodology

The collector deduplicated repeated instance types within each partition, queried current AZ offerings, gathered 72-hour Spot price history, requested a partition-pool placement score, and classified account-observed CloudTrail launches and failures. Today's compute-resource bid was reproduced as min(1.70 × current resource median, $9.99).

Limitations and robustness

AWS does not expose a definitive current capacity flag for each individual Spot pool. Price history is not capacity evidence. Placement scores describe the submitted pool as a whole, are not guaranteed, and are only directly applicable when a later request matches the scored configuration and uses capacity-optimized. CloudTrail is an account-specific, retry-correlated sample; zero attempts means no evidence, not unavailability. The chart uses per-type 72-hour medians rather than time-weighted prices.

Recommended next steps

For a build today, apply the listed temporary exclusions only in the matching AZ template.
Prefer us-west-2b for 192-vCPU NVMe work, us-west-2a for 128-vCPU NVMe work, and any AZ for 96-vCPU NVMe work (with us-west-2c the cleanest price-gate snapshot).
Avoid relying on i384nvme Spot. Use the Intel on-demand template when 384-vCPU reliability matters.
Re-run this collector immediately before later cluster creation; Spot state and placement scores are time-sensitive.

Further questions

Do the 255 price-cap failures materially delay Slurm scaling, or are they only EC2 Fleet override telemetry?
Would a separate, explicitly approved global-cap policy for i384nvme be cheaper than on-demand for the expected job duration?
Should this report become a pre-create advisory command that never mutates templates?

Source: DYEC us-west-2 Spot market evidence snapshotFile: docs/plans/20260716T031101Z_usw2_spot_market_review_snapshot.json

Read-only EC2 offerings, Spot price history, placement scores, CloudTrail RunInstances evidence, and current working-tree template membership.

Technical summary

Best current AZ by partition

Spot price distribution across configured NVMe types

Exclusions for clusters built today

Scope, data, and definitions

Methodology

Limitations and robustness

Recommended next steps

Further questions

Sources