Weekly Pricing Pulse

Cloud GPU pricing moves every week. Most buyers find out months later, when their renewal quote arrives or when a sharper procurement team pulls a number from a different provider. The Weekly Pricing Pulse exists to close that gap.

Every Monday, AIForge Works runs a fresh harvest across the 9 cloud providers we track — AWS, Azure, GCP, OCI, CoreWeave, Lambda, Vultr, Nebius, and Crusoe — diffs the new data against the prior week, and publishes what moved. Same numbers, same providers, same methodology, every time.

What's in each Pulse

A Pulse post has these core sections:

• What moved this week — the three biggest movers as headline claims. These are filtered to "high confidence" only, meaning at least 2 hyperscaler providers AND at least 2 neocloud providers carry the affected GPU family. If a movement is real but only one side of the market has coverage, it doesn't make this section.
• Editor's note — a short paragraph of human context. Why does this movement matter? Is it a continuation of a prior pattern or a reversal? Written by hand each week, not auto-generated.
• Top on-demand movers by provider — the full table, ranked by absolute percentage change on a per-GPU basis (instance price divided by GPU count, so a 1-GPU instance and an 8-GPU rack compare like-for-like). Both raw instance prices and per-GPU rates appear side by side.
• Provider roll-up — the weekly counts for New Silicon, Deprecated SKUs, existing comparable price changes over 1%, and likely price/SKU rolls.
• New Silicon this week — provider + region + instance/SKU observations that appeared in this week's harvest but not last week's, excluding strict same-provider, same-geo, same-instance likely SKU rolls. Useful for tracking catalog visibility, but not proof of net-new installed capacity.
• Deprecated SKUs this week — exact provider catalog SKUs that disappeared. They could be retired, region-moved, package-rolled, or temporarily absent; we surface the change without claiming which.
• Catalog continuity checks — optional notes for patterns that may reflect SKU/package relabeling, region-label normalization, repricing, or true availability movement. These are not collapsed into the New Silicon / Deprecated SKU counts unless the strict likely-roll rule matches.

The Pulse ends with a methodology footer naming the comparison window, the join logic, the per-GPU normalization, the confidence gate, and the harvest coverage status per provider.

What the Pulse is NOT

A few things worth being explicit about, because they're frequent confusions in the cloud cost-tracking space:

• Not a live quote. The numbers reflect public catalog prices at the moment of harvest, not what the provider will quote you on a phone call today. For a current quote, ask the provider.
• Not workload-specific advice. The Pulse reports market movements. It does not say "you should switch from AWS to CoreWeave for your training run." For that question, Cloud Advisor takes your workload requirements and ranks the 9-provider options against them.
• Not an audit of your own spend. The Pulse compares public prices. For an analysis of what you're actually paying versus the comparable floor, GPU Audit is the tool. It runs the same data against your scope.
• Not real-time. The harvest runs once a week. Intra-week movements wait for the next Pulse.
• Not personalized. Same Pulse for everyone. Customer-specific intelligence happens inside the platform.

How to read the confidence flags

Every row in a Pulse table carries a confidence label. There are three buckets:

• high — at least 2 hyperscaler providers and 2 neocloud providers carry the same GPU family at matched VRAM/interconnect class. This is the strongest comparison and is what feeds the headline claims.
• thin — at least 1 hyperscaler and 1 neocloud, but not both at 2 or more. Visible in tables, never used in headline claims.
• uncovered — only one side of the market (hyperscaler-only or neocloud-only) carries the SKU. Useful for tracking New Silicon appearances but never used for delta claims.

The labels appear inline in each row, with the actual provider counts (e.g., high (3H+4N) means 3 hyperscaler providers and 4 neocloud providers). The point is to make the strength of every claim auditable at a glance. We learned this discipline the hard way: extrapolating a single-source claim into a "the gap holds across all architectures" generalization is exactly the kind of confident-sounding error a weekly digest can quietly make. The confidence flag is the technical enforcement that prevents it.

The methodology in one paragraph

For every Pulse, we compare verified weekly harvest artifact CSVs before load, compute a per-GPU rate as price_ondemand / GPUCount, and surface rows where the per-GPU rate moved by more than 1% week-over-week. Existing price changes require comparable provider + region + instance/SKU observations across both harvests. New Silicon and Deprecated SKU sections compare exact public-catalog visibility, with same-provider, same-geo, same-instance pairs tagged as likely price/SKU rolls instead of being counted as New Silicon. GPU canonical naming resolves through gpu_provider_aliases.yaml (provider-aware, because every cloud labels GPUs differently). Currency is USD throughout; non-USD rows would abort the harvest. The full methodology lives at the bottom of each Pulse for cross-checking.

Cadence and what comes next

Pulses publish weekly, anchored to Monday's post-harvest run. Each one is reviewed by hand before going live — auto-publish is something we'll consider only after several weeks of clean cycles establish trust in the pipeline.

If you want to follow the Pulse feed, the RSS endpoint is at /blog/pulse/feed.xml. If you want to dig into a specific GPU family or region beyond what the Pulse summarizes, request access to the full platform — Beta opens this week.

The Weekly Pricing Pulse is one of several outputs from the AIForge Works query library. Future query-driven posts may include monthly trend analyses, per-GPU floor tracking, and provider-specific deep dives. The Pulse is the always-on weekly anchor.