A post making the rounds on October 15, 2026 says rental prices for Nvidia's Blackwell Ultra (B300) GPUs rose for 17 consecutive days to a new all-time high on a Compute Desk index of verified transactions from U.S. neoclouds. The claim comes from Matthew Sigel's post on X, as relayed by secondary outlets. We could not open Compute Desk's own index, so we cannot confirm the streak or the "record." What we can do is compare it with two public indexes that were readable the same day.
The short version: GPU rental prices are rising across public trackers, and B300 is up in both we checked. The specific 17-day record is unverified, and the absolute prices differ a lot between indexes.
TL;DR
| Question | Answer | Source |
|---|---|---|
| Claim | B300 rentals up 17 days in a row to a record on a Compute Desk index | Matthew Sigel post, relayed by secondary reports |
| Could we verify it? | No: Compute Desk index not accessible to us | Our check |
| Price of Compute, Oct 15 | B300 $8.26 per GPU-hour; +4.5% 7 days; +7.9% 30 days; 24-hour range $6.00-$12.19 across 4 providers | Price of Compute |
| Silicon Data, Oct 15 | B300 neo-cloud index $6.77 per GPU-hour; +6.4% over 7 days | Silicon Data |
| Wider market | Price of Compute overall index 136.3, +2.9% over 7 days, +11.9% over 30 days, 36.3% above August 10 | Price of Compute |
| Takeaway | Direction is up in both; the record claim is unconfirmed | Our reading |
What the public indexes say
Price of Compute. Its page, dated October 15, 2026, shows a composite index at 136.3, up 2.9% on the week and 11.9% on the month. It says renting a GPU costs 36.3% more than on August 10. For the B300, it lists $8.26 per GPU-hour on demand across four providers, up 4.5% over seven days and 7.9% over 30 days, with a 24-hour range of $6.00 to $12.19 (a 1.7x spread). For comparison, it shows the B200 at $7.87 (+7.7% in seven days), the H100 SXM at $4.14 and the H200 SXM at $5.87. The A100 80GB PCIe shows the largest weekly move on the page, +81.2%, which the site notes.
Silicon Data. Its B300 Rental Price Index, as of October 15, shows $6.77 per GPU-hour for the neo-cloud segment, up 6.4% over seven days. The page's FAQ text lists a different figure, $6.95, so the page itself is not fully consistent between its live reading and its static text. Its other indexes the same day: B200 $6.01 (+4.9%), H100 $2.81 (+5.6%), H200 $3.37 (+1.8%).
Neither index, in the pages we read, shows a "17 consecutive days" statistic or an all-time-high label. Silicon Data and Price of Compute do show upward moves over the past week. That is consistent with the direction of the claim, not proof of the record.
Why $6.77 and $8.26 are both "the B300 price"
The gap is about method, not error:
- Price of Compute normalizes listed on-demand prices from a small number of providers, four for B300. A thin set means one provider's change moves the number.
- Silicon Data says its index blends neo-cloud, hyperscaler, colocation and private-market observations and standardizes each for term length, cluster scale and interconnect, so it "reflects the underlying compute cost, not the headline rate on any single provider's website."
- Compute Desk, per the claim, uses verified transactions from U.S. neoclouds, which would reflect what buyers actually paid, not list prices. We have not seen its method.
So "the price" depends on whether you mean list price, realized price or a blended standardized rate, and on which providers are counted. When someone cites a record high, ask: which index, which segment, which term?
How a rental price index is built, and where it can mislead
A grid of GPU chips with one row lit, illustrating the cluster capacity behind rental price indexes
A rental price index looks like a thermometer but behaves more like a poll. Each reading depends on a sample, and a few choices decide what the sample says.
- Which prices count. List prices are what providers advertise. Transaction prices are what buyers paid after discounts. Spot prices swing with leftover capacity. Reserved prices are lower and sticky. An index that mixes these will move for reasons that have nothing to do with demand, such as a provider adding a cheaper spot tier.
- Which providers are in. Price of Compute counts four B300 providers. If one of them runs a promotion or sells out, the median moves. A transaction index from U.S. neoclouds, as Compute Desk is described, excludes hyperscalers and non-U.S. providers, which price differently.
- How consecutive days are counted. "17 days in a row" is a statement about a daily series. Daily indexes can rise on small moves, so a 17-day streak could be 17 tiny increases of a fraction of a percent. The streak says persistence, not size. The record high says the level exceeds earlier levels in that index's history, which may be short for a product as new as the B300.
- How the history is handled. A new product has a short record. An "all-time high" for a GPU generation that began renting widely this year is a weaker statement than an all-time high for the H100, which has years of data.
None of this means the claim is wrong. It means the claim should be read with its index attached. The public readings we checked point the same way, but at +4.5% to +6.4% in a week, they describe a steady climb, not a spike.
Who is affected by a rising B300 price
Different buyers feel it differently:
- Startups renting on demand feel it first and fully. A model-serving company that scaled onto B300 for long-context workloads sees margin fall if it cannot pass costs on.
- Research labs on committed contracts are insulated until renewal, then face a reset.
- Neoclouds benefit from higher rates but face the capital cost of the hardware. A rising rental price helps the payback math on a cluster and supports further buildout, which is part of the buildout story covered below.
- Hyperscalers price differently, and Silicon Data's note that the premium for the B300 over the B200 should compress as supply broadens implies that the current rise may not persist.
- Investors watch these indexes as a read on AI demand. That is why a post about a streak travels, and also why it deserves a check before it is used as evidence of anything bigger. A daily rental index covers one slice of the market, not total AI spending.
What drives B300 prices
Silicon Data's page offers its own explanation, which is an index vendor's view and worth reading as such. It says B300 rents above B200 and well above H200 because of its 288 GB of HBM3e memory and higher inference throughput, which suits frontier reasoning and long-context serving. It expects the premium to compress as Blackwell Ultra supply broadens, while frontier training and inference demand keeps a "durable floor" under rates.
The demand side shows up elsewhere in our coverage. Nebius raised GPU rental rates for the second time since May in September, as in our report on Nebius's 20% hike. Chip supply constraints and memory shortages feature in our posts on ASML's EUV orders and TSMC's record quarter. Supply scarcity also has an illegal face: Taiwan indicted nine people over 74 smuggled B300 servers. And on the supply-build side, see CoreWeave's first multi-rack Vera Rubin cluster.
What this means for what you build or pay
A cream server cabinet beside a small wind turbine, representing the data center capacity that sets GPU rental prices
- Compare more than one index. A single post citing one index is a prompt to check, not a price. Use at least two trackers and your own provider quotes.
- Ask about term length. On-demand prices are the highest and most volatile. If you have steady load, reserved or committed capacity can cost less and avoid the swings.
- Check interconnect and cluster size. A single GPU-hour price hides whether you get the networking needed for training. Silicon Data explicitly normalizes for it; list prices do not.
- Right-size the tier. B300's selling point is memory (288 GB). If your model fits in a smaller GPU, an H200 or B200 may cost less per token. Price of Compute lists H200 SXM at $5.87 and B200 at $7.87, against $8.26 for B300, on its own scale.
- Plan for volatility. Price of Compute's index moved 11.9% in 30 days; budget with a range.
Claimed versus verified
| Claim | Source | Status |
|---|---|---|
| 17 consecutive daily increases | Sigel post citing Compute Desk | Unverified: index not accessible to us |
| New all-time high | Same | Unverified; no public index we read showed an all-time-high label |
| B300 prices rising | Price of Compute (+4.5% 7d, +7.9% 30d); Silicon Data (+6.4% 7d) | Supported by two public indexes |
| Overall GPU rental prices rising | Price of Compute composite (+2.9% 7d, +11.9% 30d) | Supported by that index |
| A single "B300 price" | Indexes range from $6.77 to $8.26 | Differences due to method |
How to check it yourself
- Open Price of Compute and read the B300 row, noting the number of providers and the 24-hour range.
- Open Silicon Data's B300 index and read the 7-day trend and the methodology link.
- Look for Compute Desk's own page or the original X post and compare the methodology: transactions or list prices, which providers, what date range.
- Get two or three quotes from neoclouds for your actual workload and term, and compare them to the indexes.
Readings are from the pages as of October 15, 2026 and change daily.