inference/research
verifiedhigh · 0.85updated 2026-10-04

NEGATIVE RECORD: inference pricing that genuinely is not public (fabric, cluster rates, hyperscaler GPU list prices)

global

Fields

id
supply/rent-pricing-not-public-negatives-2026-10
Kind
unknown
Region
global
Channels
contact-salesunknown
Price basis
NO price_usd, and that is the finding. Each negative below names what is missing and what the provider publishes instead. (1) HYPERSCALER INTERNAL FABRIC AND CLUSTER PRICING: not published by anyone. AWS publishes 3200 Gbps EFA and 900 GB/s NVSwitch for p5.48xlarge as SPECIFICATIONS with no price attached to any of it, and does not price EFA or NVLink as a line item. Google does not publish a price for a3 interconnect. Azure does not price Quantum-2 CX7 InfiniBand per VM. The pre-existing nic-cloud-fabric-bandwidth-not-public gotcha establishes the bandwidth half of this; the PRICING half is separate and equally absent. (2) COLOCATION AND POWER OPERATORS: quote-only. The pre-existing colo-vantage-power-density-tiers and colo-vantage-liquid-to-rack-not-to-chip records record exactly this - no per-rack-month, no per-kW-month, no published figure at all - and ai-datacentre-power-capacity-run-cost records that per-kW-month is site-specific. (3) PROVIDERS THAT SELL A GPU API BUT NOT A GPU HOUR: OpenAI, Anthropic, Google Gemini, Cohere, Mistral, xAI, Groq and Cerebras all publish per-token API prices and NONE publishes a per-GPU-hour rate or names the accelerator serving its API. Every record in this slice for those providers therefore carries an empty accelerator_ids list, and that emptiness is data, not an omission. (4) SALES-GATED WITHIN AN OTHERWISE-PUBLIC PRICE LIST: Groq's llama-3.1-8b-instant and llama-3.3-70b-versatile both read literally 'ContactSales' in the price column of an otherwise public per-token table, and OpenAI's Enterprise plan is 'Custom pricing - Contact our sales team to discuss enterprise pricing' with only credit-based and token-based rate cards behind a help-centre link rather than on the pricing page.
Availability
unknown
Lead time source
structural-na
Lead time basis
lead_time_weeks is null because NO NUMBER APPLIES: this record prices a non-accelerator input (electricity, egress, PUE/uptime) and lists accelerators only as the load that consumes it. There is no order-to-delivery step, so a lead time is the wrong quantity rather than a missing one.
Export control regime
no EAR citation: no controlled item is transferred by this transaction; not an EAR citation: provider service terms and end-use policy (the AI Diffusion Rule briefly tiered service access before its 2025-05-13 rescission)
Export control note
MECHANISM: a token-priced API, so no accelerator crosses a border and the 3A090 hardware licence regime does not attach to the transaction. Export control becomes an END-USER AND TERMS question instead: the provider applies geographic and end-use restrictions to service access, and a customer can still be refused under sanctions or end-use policy wherever the compute sits. A reader targeting a restricted destination should read the provider's terms, not the EAR.
notes
THE POINT OF THIS RECORD IS THAT 'I COULD NOT FIND IT' AND 'IT IS NOT PUBLISHED' ARE DIFFERENT CLAIMS, AND THIS ONE IS THE LATTER. Three consequences for a reader deciding where to run. First, ANY cost model of a hyperscaler at cluster scale is built on a negotiated rate nobody can verify, so a capacity plan built on list prices has a known-unknown in it and should carry an explicit contingency rather than false precision. Second, the absence of published fabric pricing is the biggest single gap in GPU cloud economics: the fabric determines whether TP=8 or TP=64 works at all, and no provider will sell you that number at any price. Third, the honest ceiling on price research differs sharply by order of magnitude: a 1-GPU-hour rental is fully published and fully comparable, and a 1000-GPU cluster has nothing published and everything negotiated. Most existing per-part records in this repo sit on the published side of that line, which is exactly why they are trustworthy and exactly why they cannot answer a cluster question. Do not fill these gaps with remembered figures, third-party aggregator 'cloud price indexes', or vendor blogs - the repo contract forbids citing aggregator spec-comparison sites, and a fabricated internal rate is worse than an open question. CROSS-REFERENCE the other negatives in this slice so a reader finds them together: ppl-sup-aws-egress-negative (AWS tiered outbound table is JS-rendered and not statically fetchable), ppl-sup-uptime-pue-negative (Uptime Institute 2026 PUE data behind a download wall), aws-ec2-spot-gpu-general (no per-SKU spot rate published in one place), vast-ai-marketplace (Live GPU Prices grid renders client-side from an API), rent-mistral-api-token-pricing (pricing STRUCTURE published, per-model rate card not). WITHDRAWN FROM THIS LIST 2026-10-04: rent-cerebras-api-token-pricing-negative was previously cross-referenced here as a negative and was WRONG - a rendered-DOM fetch of www.cerebras.ai/pricing returns Cerebras's Developer Tier rates in full (source pd-cerebras-pricing-dom-rendered), and the record is now supply/rent-cerebras-api-token-pricing carrying a real price. Only Cerebras's ENTERPRISE tier is genuinely quote-only, and only that narrower claim survives. THE LESSON, recorded because this list is where an unverified negative is most likely to hide: a negative asserted on a SINGLE extraction method is indistinguishable from a bug in that method. Require two independent failures before recording one. See gotcha/rent-vendor-pricing-tables-are-js-but-their-apis-are-not. THE PATTERN ACROSS ALL OF THEM IS IDENTICAL: providers publish their marketing pages as static HTML but their actual rate tables as JavaScript-rendered grids or JSON APIs. If you want current numbers the APIs are the reliable surface - which is how rent-aws-ec2-gpu-hourly-2026-10-04 was produced, and it is why the AWS price list is the one GPU source in this slice that is fully reproducible. EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator is transferred across a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded null rather than false precisely because a token API is where export control instead becomes an END-USER/TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded on 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. Anyone reading this record for a restricted destination should look at the provider's terms, not at the EAR. Sources for the regime: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance. EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator crosses a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded NULL rather than false precisely because a token API is where export control instead becomes an END-USER and TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded and unenforced from 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. For a restricted destination, read the provider's terms, not the EAR. Sources: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance. LEAD TIME: null - correct and not a gap. A token-priced serverless API is provisioned on request, so there is no lead time to publish; the binding constraint when a model is at capacity is provider-side queuing and any per-account concurrency limit, which this repo records in notes where a vendor publishes one (see the DeepSeek concurrency figures already here). Deliberately NOT recorded: any 'model capacity lead time' estimate, because vendors publish none and the SEO 'GPU lead time' numbers that circulate do not apply to a rented endpoint. Probe log in source spl-negative-no-oem-publishes-lead-time. ACCELERATOR LINKS ADDED 2026-10-04: this negative record is about pricing that is genuinely not published (fabric, cluster rates, hyperscaler GPU list prices), and the parts whose rates stay private are the datacenter nodes and racks - H100/H200 SXM, B200 and GB200 NVL72. Linking them makes the negative queryable against the parts a reader actually wants a number for. SELF-DESCRIBING RULE ADDED 2026-10-04, because this record is the one most likely to be trusted wrongly: EVERY CLAIM IN IT IS A CLAIM ABOUT ABSENCE, AND ABSENCE IS THE HARDEST THING IN THIS REPO TO KEEP TRUE. Each negative here is dated and names the probe that established it, so a future reader can re-run the probe rather than inherit the conclusion. The cerebras withdrawal above is the proof that this is necessary rather than decorative: one entry in this list was false, and it was false because a single failed probe was recorded as a fact about the vendor. Treat every item on this list as 'true as of its stated date, re-test before relying on it', and never as a standing property of the vendor. NEGATIVE RE-CONFIRMED 2026-10-04, so that this record is as self-describing as a priced one. These are ABSENCE claims and absence claims rot silently, so each is recorded with the date it was tested and the shape of the probe. (1) NO HYPERSCALER PUBLISHES INTERNAL FABRIC OR CLUSTER PRICING - still true. AWS publishes 3200 Gbps EFA and 900 GB/s NVSwitch for p5.48xlarge as SPECIFICATIONS and attaches no price to any of it, and does not price EFA or NVLink as a line item at all. Google publishes no price for a3 interconnect. Azure does not price Quantum-2 CX7 InfiniBand per VM. What fabric IS published is bandwidth and topology, never money: the number that determines whether TP=8 or TP=64 is physically possible is free, while the number that determines what it costs is withheld. That asymmetry is the whole finding. (2) NO TOKEN-API PROVIDER PUBLISHES A GPU-HOUR RATE OR NAMES ITS SERVING SILICON - still true, and now checkable from the data rather than from prose, which is the better way to hold a claim like this. Every token-API record in this directory that RENTS third-party GPUs has a genuinely EMPTY accelerator_ids list, re-verified 2026-10-04: rent-openai-api-token-pricing, rent-anthropic-api-token-pricing, rent-gemini-api-token-pricing, rent-cohere-api-token-pricing, rent-mistral-api-token-pricing, rent-xai-api-token-pricing. The emptiness is data, not an oversight - you can query it. The two records that DO carry an accelerator id are exactly the two providers who sell their OWN silicon rather than renting GPUs: rent-groq-api-token-pricing carries groq-lpu and rent-cerebras-api-token-pricing carries cerebras-wse-3. That is the cleanest available test of the claim, and it passes: named silicon implies vertical integration, and every provider without vertical integration says nothing at all about what it runs on. WHAT THIS MEANS FOR A COST MODEL, stated so it cannot be misread: a token rate from any of the six unnamed providers is a BLACK BOX with respect to hardware. Its cost-per-token cannot be decomposed into joules, cannot be checked against a GPU-hour from this repo, and cannot be reasoned about when a new accelerator generation lands - because nothing links the two. Anyone tempted to divide a token bill by a GPU-hour rate for the same provider is inventing the link this record says does not exist. THE COROLLARY THAT CUTS THE OTHER WAY, and is worth stating: the same opacity that makes these rates unverifiable is also why they can be CHEAP. A provider renting commodity GPUs has no hardware advantage to price in, so its margin must come from utilisation, batching and scheduling - which is exactly where DeepSeek's off-peak tiering and Groq's published throughput come from. Opacity is a weak signal about whether a price is good; it is a strong signal that you cannot verify it.

Notes

THE POINT OF THIS RECORD IS THAT 'I COULD NOT FIND IT' AND 'IT IS NOT PUBLISHED' ARE DIFFERENT CLAIMS, AND THIS ONE IS THE LATTER. Three consequences for a reader deciding where to run. First, ANY cost model of a hyperscaler at cluster scale is built on a negotiated rate nobody can verify, so a capacity plan built on list prices has a known-unknown in it and should carry an explicit contingency rather than false precision. Second, the absence of published fabric pricing is the biggest single gap in GPU cloud economics: the fabric determines whether TP=8 or TP=64 works at all, and no provider will sell you that number at any price. Third, the honest ceiling on price research differs sharply by order of magnitude: a 1-GPU-hour rental is fully published and fully comparable, and a 1000-GPU cluster has nothing published and everything negotiated. Most existing per-part records in this repo sit on the published side of that line, which is exactly why they are trustworthy and exactly why they cannot answer a cluster question. Do not fill these gaps with remembered figures, third-party aggregator 'cloud price indexes', or vendor blogs - the repo contract forbids citing aggregator spec-comparison sites, and a fabricated internal rate is worse than an open question. CROSS-REFERENCE the other negatives in this slice so a reader finds them together: ppl-sup-aws-egress-negative (AWS tiered outbound table is JS-rendered and not statically fetchable), ppl-sup-uptime-pue-negative (Uptime Institute 2026 PUE data behind a download wall), aws-ec2-spot-gpu-general (no per-SKU spot rate published in one place), vast-ai-marketplace (Live GPU Prices grid renders client-side from an API), rent-mistral-api-token-pricing (pricing STRUCTURE published, per-model rate card not). WITHDRAWN FROM THIS LIST 2026-10-04: rent-cerebras-api-token-pricing-negative was previously cross-referenced here as a negative and was WRONG - a rendered-DOM fetch of www.cerebras.ai/pricing returns Cerebras's Developer Tier rates in full (source pd-cerebras-pricing-dom-rendered), and the record is now supply/rent-cerebras-api-token-pricing carrying a real price. Only Cerebras's ENTERPRISE tier is genuinely quote-only, and only that narrower claim survives. THE LESSON, recorded because this list is where an unverified negative is most likely to hide: a negative asserted on a SINGLE extraction method is indistinguishable from a bug in that method. Require two independent failures before recording one. See gotcha/rent-vendor-pricing-tables-are-js-but-their-apis-are-not. THE PATTERN ACROSS ALL OF THEM IS IDENTICAL: providers publish their marketing pages as static HTML but their actual rate tables as JavaScript-rendered grids or JSON APIs. If you want current numbers the APIs are the reliable surface - which is how rent-aws-ec2-gpu-hourly-2026-10-04 was produced, and it is why the AWS price list is the one GPU source in this slice that is fully reproducible.

EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator is transferred across a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded null rather than false precisely because a token API is where export control instead becomes an END-USER/TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded on 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. Anyone reading this record for a restricted destination should look at the provider's terms, not at the EAR. Sources for the regime: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance.

EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator crosses a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded NULL rather than false precisely because a token API is where export control instead becomes an END-USER and TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded and unenforced from 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. For a restricted destination, read the provider's terms, not the EAR. Sources: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance.

LEAD TIME: null - correct and not a gap. A token-priced serverless API is provisioned on request, so there is no lead time to publish; the binding constraint when a model is at capacity is provider-side queuing and any per-account concurrency limit, which this repo records in notes where a vendor publishes one (see the DeepSeek concurrency figures already here). Deliberately NOT recorded: any 'model capacity lead time' estimate, because vendors publish none and the SEO 'GPU lead time' numbers that circulate do not apply to a rented endpoint. Probe log in source spl-negative-no-oem-publishes-lead-time.

ACCELERATOR LINKS ADDED 2026-10-04: this negative record is about pricing that is genuinely not published (fabric, cluster rates, hyperscaler GPU list prices), and the parts whose rates stay private are the datacenter nodes and racks - H100/H200 SXM, B200 and GB200 NVL72. Linking them makes the negative queryable against the parts a reader actually wants a number for.

SELF-DESCRIBING RULE ADDED 2026-10-04, because this record is the one most likely to be trusted wrongly: EVERY CLAIM IN IT IS A CLAIM ABOUT ABSENCE, AND ABSENCE IS THE HARDEST THING IN THIS REPO TO KEEP TRUE. Each negative here is dated and names the probe that established it, so a future reader can re-run the probe rather than inherit the conclusion. The cerebras withdrawal above is the proof that this is necessary rather than decorative: one entry in this list was false, and it was false because a single failed probe was recorded as a fact about the vendor. Treat every item on this list as 'true as of its stated date, re-test before relying on it', and never as a standing property of the vendor.

NEGATIVE RE-CONFIRMED 2026-10-04, so that this record is as self-describing as a priced one. These are ABSENCE claims and absence claims rot silently, so each is recorded with the date it was tested and the shape of the probe. (1) NO HYPERSCALER PUBLISHES INTERNAL FABRIC OR CLUSTER PRICING - still true. AWS publishes 3200 Gbps EFA and 900 GB/s NVSwitch for p5.48xlarge as SPECIFICATIONS and attaches no price to any of it, and does not price EFA or NVLink as a line item at all. Google publishes no price for a3 interconnect. Azure does not price Quantum-2 CX7 InfiniBand per VM. What fabric IS published is bandwidth and topology, never money: the number that determines whether TP=8 or TP=64 is physically possible is free, while the number that determines what it costs is withheld. That asymmetry is the whole finding. (2) NO TOKEN-API PROVIDER PUBLISHES A GPU-HOUR RATE OR NAMES ITS SERVING SILICON - still true, and now checkable from the data rather than from prose, which is the better way to hold a claim like this. Every token-API record in this directory that RENTS third-party GPUs has a genuinely EMPTY accelerator_ids list, re-verified 2026-10-04: rent-openai-api-token-pricing, rent-anthropic-api-token-pricing, rent-gemini-api-token-pricing, rent-cohere-api-token-pricing, rent-mistral-api-token-pricing, rent-xai-api-token-pricing. The emptiness is data, not an oversight - you can query it. The two records that DO carry an accelerator id are exactly the two providers who sell their OWN silicon rather than renting GPUs: rent-groq-api-token-pricing carries groq-lpu and rent-cerebras-api-token-pricing carries cerebras-wse-3. That is the cleanest available test of the claim, and it passes: named silicon implies vertical integration, and every provider without vertical integration says nothing at all about what it runs on. WHAT THIS MEANS FOR A COST MODEL, stated so it cannot be misread: a token rate from any of the six unnamed providers is a BLACK BOX with respect to hardware. Its cost-per-token cannot be decomposed into joules, cannot be checked against a GPU-hour from this repo, and cannot be reasoned about when a new accelerator generation lands - because nothing links the two. Anyone tempted to divide a token bill by a GPU-hour rate for the same provider is inventing the link this record says does not exist. THE COROLLARY THAT CUTS THE OTHER WAY, and is worth stating: the same opacity that makes these rates unverifiable is also why they can be CHEAP. A provider renting commodity GPUs has no hardware advantage to price in, so its margin must come from utilisation, batching and scheduling - which is exactly where DeepSeek's off-peak tiering and Groq's published throughput come from. Opacity is a weak signal about whether a price is good; it is a strong signal that you cannot verify it.

Cross-references

Accelerators (4)