{"id":"518ea3a6070f9ce63a0f8b934e6a1b3bb516b07f7a118ccc31034f24d917574b","title":"Carmen Li argues in Silicon Data that the H200 rental premium widened","argument":"Carmen Li’s Silicon Data article was published on August 4, 2026. Li argues that the H200 rental premium over the H100 widened across the twelve weeks of daily Neo-Cloud index data, after narrowing briefly in June. The reported H200 rate rose 14.4% over the window, while the H100 rate rose 7.0%; the premium ranged from 0.4% to 16.2%. Li reads the mid-June near-parity as an anomaly and the later premium as consistent with demand for the H200’s memory advantage in memory-bound inference. That matters because, in Li’s account, the H200’s greater memory capacity and bandwidth can support larger models and longer context windows on one GPU, while compute-bound training gains less. But Li stresses that the data measures the price movement, not its cause: memory-bound demand is one possible explanation, while provider composition and supply mix could also contribute. The index has less than three months of history, and the mid-June compression could reflect an H100 demand spike, an H200 supply bulge, or panel effects. The series does not distinguish among these explanations, and the article says it is too early to establish whether July’s rise will continue.","topicIds":["investment-booms-and-busts"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 28, 2026; originally published August 4, 2026. Li says the index has under three months of history and cannot distinguish among possible explanations for the premium’s movement. The article says whether July’s rise continues cannot be established from three months of data.","speaker":{"name":"Silicon Data","roleAtSource":"Publisher"},"source":{"title":"H200 vs H100 Rental Prices: The Premium That Doubled","publisher":"Silicon Data","format":"Article","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","publishedAt":"2026-08-04","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Carmen Li reports that the H200 rental rate rose more than the H100 rate during the period covered by the index data.","excerpt":"the H200 Neo-Cloud rate rose 14.4%, the H100 rate rose 7.0% with a June round trip","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"},{"text":"Li reports that the daily premium ranged from near parity in June to a peak in July.","excerpt":"the daily premium between them ran from 5% to 0.4% to 16.2%","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"},{"text":"Li says the price data does not establish why the premium widened; memory-bound demand, provider composition and supply mix remain possible explanations.","excerpt":"A widening premium is consistent with memory-bound demand favoring the newer chip, but a young index can also widen as its provider panel matures, and quarterly context is still thin with under three months of history.","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"},{"text":"Li argues that the H200’s memory advantage is most relevant to memory-bound inference, while compute-bound training gains less.","excerpt":"In application terms that means larger models and longer context windows served from a single GPU, more room for the KV cache that high-concurrency serving consumes, and faster memory-bound decoding, while compute-bound training gains less.","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"As the index history grows, can it distinguish memory-bound demand from changes in provider composition or supply mix as causes of the widening premium?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Carmen Li argues in Silicon Data that the H200 rental premium widened","summary":"Carmen Li’s Silicon Data article was published on August 4, 2026. Li argues that the H200 rental premium over the H100 widened across the twelve weeks of daily Neo-Cloud index data, after narrowing briefly in June. The reported H200 rate rose 14.4% over the window, while the H100 rate rose 7.0%; the premium ranged from 0.4% to 16.2%. Li reads the mid-June near-parity as an anomaly and the later premium as consistent with demand for the H200’s memory advantage in memory-bound inference. That matters because, in Li’s account, the H200’s greater memory capacity and bandwidth can support larger models and longer context windows on one GPU, while compute-bound training gains less. But Li stresses that the data measures the price movement, not its cause: memory-bound demand is one possible explanation, while provider composition and supply mix could also contribute. The index has less than three months of history, and the mid-June compression could reflect an H100 demand spike, an H200 supply bulge, or panel effects. The series does not distinguish among these explanations, and the article says it is too early to establish whether July’s rise will continue.","question":"As the index history grows, can it distinguish memory-bound demand from changes in provider composition or supply mix as causes of the widening premium?","claims":[{"text":"Carmen Li reports that the H200 rental rate rose more than the H100 rate during the period covered by the index data.","excerpt":"the H200 Neo-Cloud rate rose 14.4%, the H100 rate rose 7.0% with a June round trip","support":"source-stated"},{"text":"Li reports that the daily premium ranged from near parity in June to a peak in July.","excerpt":"the daily premium between them ran from 5% to 0.4% to 16.2%","support":"source-stated"},{"text":"Li says the price data does not establish why the premium widened; memory-bound demand, provider composition and supply mix remain possible explanations.","excerpt":"A widening premium is consistent with memory-bound demand favoring the newer chip, but a young index can also widen as its provider panel matures, and quarterly context is still thin with under three months of history.","support":"source-stated"},{"text":"Li argues that the H200’s memory advantage is most relevant to memory-bound inference, while compute-bound training gains less.","excerpt":"In application terms that means larger models and longer context windows served from a single GPU, more room for the KV cache that high-concurrency serving consumes, and faster memory-bound decoding, while compute-bound training gains less.","support":"source-stated"}],"limitations":["Text as retrieved on October 28, 2026; originally published August 4, 2026.","Li says the index has under three months of history and cannot distinguish among possible explanations for the premium’s movement.","The article says whether July’s rise continues cannot be established from three months of data."],"mediaId":null,"correctionReason":null,"id":"518ea3a6070f9ce63a0f8b934e6a1b3bb516b07f7a118ccc31034f24d917574b","publicationId":"c83798fb-3077-21ae-a5bc-d5e03a6a7639","version":1,"sourceUrl":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","sourceTitle":"H200 vs H100 Rental Prices: The Premium That Doubled","publisher":"Silicon Data","publishedAt":null,"originalPublished":{"value":"2026-08-04","precision":"day","basis":"backfill:article_published"},"firstObservedAt":"2026-10-28T01:54:11.239Z","createdAt":"2026-11-03T00:07:26.032Z","historical":true,"entityIds":[],"topicIds":["investment-booms-and-busts"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"0788c92e50449849e5074ec249da971030b8fa67478a027200044c33598d1906","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-28T01:54:08.811Z"},"provenance":{"textHash":"f033e8b6bd3ad8790fe422048c2d449e84985a9494e4ee4ea730729d54ee6e0f","observationId":"595fa32ea200d8fcc5a9ee7697f0d095030a6a0e7a9960aa4202e15debc2e471","retrievedAt":"2026-10-28T01:55:53.398Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"f45b328b-d4dc-39b9-bc0d-cafe827655d1"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"7919dfa0-08c0-5cde-2ea8-20b65b498a97"}},"audit":{"status":"passed","reviewedAt":"2026-11-03T00:07:23.336Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/carmen-li-argues-in-silicon-data-that-the-h200-rental-premium-518ea3a6","formats":{"html":"/readings/carmen-li-argues-in-silicon-data-that-the-h200-rental-premium-518ea3a6","markdown":"/readings/carmen-li-argues-in-silicon-data-that-the-h200-rental-premium-518ea3a6.md","json":"/readings/carmen-li-argues-in-silicon-data-that-the-h200-rental-premium-518ea3a6.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-nvidia-h200","name":"NVIDIA H200","explanation":"An NVIDIA GPU built on the Hopper architecture with 141 gigabytes of memory at 4.8 terabytes per second: nearly double the H100's memory capacity, with 1.4 times its memory bandwidth.","referenceUrl":"https://www.nvidia.com/en-us/data-center/h200/"},{"id":"term-nvidia-h100","name":"NVIDIA H100","explanation":"An NVIDIA graphics processing unit (GPU) built on the company's Hopper architecture and designed to accelerate AI workloads, including large language models.","referenceUrl":"https://www.nvidia.com/en-us/data-center/h100/"},{"id":"term-neocloud","name":"Neocloud","explanation":"A new breed of cloud compute provider focused on renting out GPU computing power: a pure-play GPU cloud whose economics are still evolving.","referenceUrl":"https://newsletter.semianalysis.com/p/ai-neocloud-playbook-and-anatomy"},{"id":"term-context-window","name":"Context window","explanation":"How much text a model can take in and consider at one time.","referenceUrl":"https://developers.google.com/machine-learning/glossary#context-window"},{"id":"term-gpu","name":"Graphics processing unit","explanation":"A chip that performs many calculations at once; the main hardware used to train and run AI models.","referenceUrl":"https://en.wikipedia.org/wiki/Graphics_processing_unit"}],"entities":[]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"518ea3a6070f9ce63a0f8b934e6a1b3bb516b07f7a118ccc31034f24d917574b","revision":"c83798fb-3077-21ae-a5bc-d5e03a6a7639","revisionKind":"publication-id","title":"Carmen Li argues in Silicon Data that the H200 rental premium widened","url":"https://escapemeridian.com/readings/carmen-li-argues-in-silicon-data-that-the-h200-rental-premium-518ea3a6"},"source":{"type":"source","id":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","revision":"f033e8b6bd3ad8790fe422048c2d449e84985a9494e4ee4ea730729d54ee6e0f","revisionKind":"sha256-source-text","title":"H200 vs H100 Rental Prices: The Premium That Doubled","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"518ea3a6070f9ce63a0f8b934e6a1b3bb516b07f7a118ccc31034f24d917574b","title":"Carmen Li argues in Silicon Data that the H200 rental premium widened","argument":"Carmen Li’s Silicon Data article was published on August 4, 2026. Li argues that the H200 rental premium over the H100 widened across the twelve weeks of daily Neo-Cloud index data, after narrowing briefly in June. The reported H200 rate rose 14.4% over the window, while the H100 rate rose 7.0%; the premium ranged from 0.4% to 16.2%. Li reads the mid-June near-parity as an anomaly and the later premium as consistent with demand for the H200’s memory advantage in memory-bound inference. That matters because, in Li’s account, the H200’s greater memory capacity and bandwidth can support larger models and longer context windows on one GPU, while compute-bound training gains less. But Li stresses that the data measures the price movement, not its cause: memory-bound demand is one possible explanation, while provider composition and supply mix could also contribute. The index has less than three months of history, and the mid-June compression could reflect an H100 demand spike, an H200 supply bulge, or panel effects. The series does not distinguish among these explanations, and the article says it is too early to establish whether July’s rise will continue.","topicIds":["investment-booms-and-busts"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 28, 2026; originally published August 4, 2026. Li says the index has under three months of history and cannot distinguish among possible explanations for the premium’s movement. The article says whether July’s rise continues cannot be established from three months of data.","speaker":{"name":"Silicon Data","roleAtSource":"Publisher"},"source":{"title":"H200 vs H100 Rental Prices: The Premium That Doubled","publisher":"Silicon Data","format":"Article","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","publishedAt":"2026-08-04","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Carmen Li reports that the H200 rental rate rose more than the H100 rate during the period covered by the index data.","excerpt":"the H200 Neo-Cloud rate rose 14.4%, the H100 rate rose 7.0% with a June round trip","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"},{"text":"Li reports that the daily premium ranged from near parity in June to a peak in July.","excerpt":"the daily premium between them ran from 5% to 0.4% to 16.2%","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"},{"text":"Li says the price data does not establish why the premium widened; memory-bound demand, provider composition and supply mix remain possible explanations.","excerpt":"A widening premium is consistent with memory-bound demand favoring the newer chip, but a young index can also widen as its provider panel matures, and quarterly context is still thin with under three months of history.","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"},{"text":"Li argues that the H200’s memory advantage is most relevant to memory-bound inference, while compute-bound training gains less.","excerpt":"In application terms that means larger models and longer context windows served from a single GPU, more room for the KV cache that high-concurrency serving consumes, and faster memory-bound decoding, while compute-bound training gains less.","url":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"As the index history grows, can it distinguish memory-bound demand from changes in provider composition or supply mix as causes of the widening premium?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Carmen Li argues in Silicon Data that the H200 rental premium widened","summary":"Carmen Li’s Silicon Data article was published on August 4, 2026. Li argues that the H200 rental premium over the H100 widened across the twelve weeks of daily Neo-Cloud index data, after narrowing briefly in June. The reported H200 rate rose 14.4% over the window, while the H100 rate rose 7.0%; the premium ranged from 0.4% to 16.2%. Li reads the mid-June near-parity as an anomaly and the later premium as consistent with demand for the H200’s memory advantage in memory-bound inference. That matters because, in Li’s account, the H200’s greater memory capacity and bandwidth can support larger models and longer context windows on one GPU, while compute-bound training gains less. But Li stresses that the data measures the price movement, not its cause: memory-bound demand is one possible explanation, while provider composition and supply mix could also contribute. The index has less than three months of history, and the mid-June compression could reflect an H100 demand spike, an H200 supply bulge, or panel effects. The series does not distinguish among these explanations, and the article says it is too early to establish whether July’s rise will continue.","question":"As the index history grows, can it distinguish memory-bound demand from changes in provider composition or supply mix as causes of the widening premium?","claims":[{"text":"Carmen Li reports that the H200 rental rate rose more than the H100 rate during the period covered by the index data.","excerpt":"the H200 Neo-Cloud rate rose 14.4%, the H100 rate rose 7.0% with a June round trip","support":"source-stated"},{"text":"Li reports that the daily premium ranged from near parity in June to a peak in July.","excerpt":"the daily premium between them ran from 5% to 0.4% to 16.2%","support":"source-stated"},{"text":"Li says the price data does not establish why the premium widened; memory-bound demand, provider composition and supply mix remain possible explanations.","excerpt":"A widening premium is consistent with memory-bound demand favoring the newer chip, but a young index can also widen as its provider panel matures, and quarterly context is still thin with under three months of history.","support":"source-stated"},{"text":"Li argues that the H200’s memory advantage is most relevant to memory-bound inference, while compute-bound training gains less.","excerpt":"In application terms that means larger models and longer context windows served from a single GPU, more room for the KV cache that high-concurrency serving consumes, and faster memory-bound decoding, while compute-bound training gains less.","support":"source-stated"}],"limitations":["Text as retrieved on October 28, 2026; originally published August 4, 2026.","Li says the index has under three months of history and cannot distinguish among possible explanations for the premium’s movement.","The article says whether July’s rise continues cannot be established from three months of data."],"mediaId":null,"correctionReason":null,"id":"518ea3a6070f9ce63a0f8b934e6a1b3bb516b07f7a118ccc31034f24d917574b","publicationId":"c83798fb-3077-21ae-a5bc-d5e03a6a7639","version":1,"sourceUrl":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","sourceTitle":"H200 vs H100 Rental Prices: The Premium That Doubled","publisher":"Silicon Data","publishedAt":null,"originalPublished":{"value":"2026-08-04","precision":"day","basis":"backfill:article_published"},"firstObservedAt":"2026-10-28T01:54:11.239Z","createdAt":"2026-11-03T00:07:26.032Z","historical":true,"entityIds":[],"topicIds":["investment-booms-and-busts"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.silicondata.com/blog/h200-vs-h100-rental-prices-may-july-2026","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"0788c92e50449849e5074ec249da971030b8fa67478a027200044c33598d1906","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-28T01:54:08.811Z"},"provenance":{"textHash":"f033e8b6bd3ad8790fe422048c2d449e84985a9494e4ee4ea730729d54ee6e0f","observationId":"595fa32ea200d8fcc5a9ee7697f0d095030a6a0e7a9960aa4202e15debc2e471","retrievedAt":"2026-10-28T01:55:53.398Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"f45b328b-d4dc-39b9-bc0d-cafe827655d1"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"7919dfa0-08c0-5cde-2ea8-20b65b498a97"}},"audit":{"status":"passed","reviewedAt":"2026-11-03T00:07:23.336Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"c83798fb-3077-21ae-a5bc-d5e03a6a7639","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly attributes the article and its analysis to Carmen Li, founder at Silicon Data, and keeps her interpretations framed as her own.","excerpt":"Written by Carmen Li\nFounder at Silicon Data"},{"dimension":"summary","status":"pass","explanation":"The summary accurately conveys the rise in rates (14.4% vs 7.0%), the premium's range (0.4% to 16.2%), Li's reading of mid-June parity as the anomaly, the memory-advantage rationale and her caveats about mechanism. H200, H100, Neo-Cloud, GPU and context window are defined beside the reading. Memory-bound inference is explained through its stated practical consequences. The question follows directly from the source's stated open issue.","excerpt":"Read against the hardware, the near-parity of mid-June looks like the anomaly and the current spread like normalization."},{"dimension":"date","status":"pass","explanation":"The stated date of August 4, 2026 matches evidence.originalPublished and the article dateline, at day precision. The page is a publisher page, not an archive copy. Nothing is presented as happening after publication. The source's uncertainty about whether July's rise continues is reported as unresolved, not as an outcome.","excerpt":"Industry Aug 4, 2026 5 Mins Read"},{"dimension":"support","status":"pass","explanation":"Each claim is supported by the source text: the rate rises, the premium range, the stated limits on establishing a mechanism (including provider composition and supply mix), and the inference-versus-training distinction. The reading implies no benefits or costs beyond what the source states.","excerpt":"though the series is young enough that provider composition and supply mix could account for part of the move"},{"dimension":"media","status":"pass","explanation":"No media is selected, and no verified media candidates were supplied. This is appropriate.","excerpt":null}],"createdAt":"2026-11-03T00:07:23.336Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}