Market snapshotKalshi-impliedB200 $7.72 thin
AI COMPUTE
TRACKER

Open record· latest 2026-10-15

Corrections

Every material error we have corrected in the site’s own numbers and claims, newest first: what the site said, what the correction changed it to, where it appeared, and the evidence. Each entry is dated history; a figure corrected again later links the newer entry. Routine price moves are not corrections; quotes keep their own observation dates.

Spotted an error? Email data@aicomputetracker.com, or use “Report a wrong price” in the evidence details of any quote row. Every correction is also listed in the desk log, and its RSS feed carries the newest 10. The release named with each entry is a pull request in the site’s private repository; ask if you need its details.

Hardware specs#

The memory-cost insight rested on a B300 price no reputable source gives

The site said
The /hardware memory-cost insight put the B300's 270 GB of HBM3e at roughly $4,200, about 9% of a ~$45,000 street price; no reputable source gives a per-unit B300 price, and NVIDIA publishes none.
Corrected to
The B300 was left without a street price, as were the H200 and MI325X, for which no reputable per-unit price was found either. The insight moved to the B200, whose ~$35,000 estimate is the midpoint of the $30,000 to $40,000 range NVIDIA's CEO gave CNBC: its 180 GB of HBM3e came to roughly $2,800, about 8% of that price.
Where
/hardware memory-cost insight.
Found by
A scheduled survey.
Evidence
CNBC: Nvidia's Blackwell AI chip will cost more than $30,000, CEO says
Release
PR #136
Forward curves#

Forward curves read thin Kalshi ladders as price levels, and archived sessions were called unchangeable

The site said
The forward curves and their archive priced every strike from its bid/ask midpoint or last trade, repaired each ladder with a running minimum and smoothed the back months, so thin books published as levels. The Oct 12, 2026 session showed B200 Jan ’27 at $6.98, quality-passed, though its bids and asks bounded it only at $5.67 or more. /archive said an archived session’s values never change.
Corrected to
The 13 archived sessions from Jul 23 to Oct 14, 2026 whose raw Kalshi ladders were retained were re-derived from those ladders under estimator v2: each point carries its bid/ask bracket, only identified points draw a line or feed a move, and an unidentified point is shown as its bracket, never as a price. Oct 12’s B200 Jan ’27 became an unidentified point whose bids put it at $5.67 or more. Each re-derived permalink was changed to keep the values first published (rederived.superseded), and /archive to say a closed session changes only through a recorded re-derivation.
Where
/archive/curves/{date}.json permalinks for Jul 23–Oct 14, 2026; curve time machine; /archive; /log; homepage forward curves, terminal and chip pages.
Found by
A review of published pages and data.
Evidence
Oct 12, 2026 curve session, with the values first published · Jul 23, 2026 curve session, the first re-derived
Release
PR #126
Hardware specs#

GB200 NVL72 was priced per PFLOP-hour at the HGX B200's dense rate

The site said
The price per PFLOP-hour ranking divided GB200 NVL72's cheapest specialist on-demand price by the HGX B200's 2,250 TF dense BF16, so CoreWeave's $10.50 per GPU-hour read as $4.67 per PFLOP-hour.
Corrected to
GB200 NVL72 got a spec row of its own from the GB200 NVL72 column of NVIDIA's Blackwell datasheet, 2,500 TF dense BF16 per GPU (its 5 PF sparse figure, halved as the datasheet's footnote directs), so the same $10.50 per GPU-hour read as $4.20 per PFLOP-hour.
Where
/hardware price per PFLOP-hour ranking.
Found by
A review of published pages and data.
Evidence
NVIDIA Blackwell datasheet (individual GPU specifications, GB200 NVL72 column) · NVIDIA GB200 NVL72 specifications
Release
PR #119
GPU quotes#

The quarterly big-cloud floors misread several archived list prices

The site said
The quarterly price-history charts put the H100's big-cloud floor at $12.29 per GPU-hour from 2024Q2 to 2025Q1 and $6.88 in 2025Q2, the A100 80GB's at $4.10 in 2021Q1 and 2021Q3 and $2.74 from 2025Q3 (the A100 40GB's price), and the H200's at $10.71 in 2024Q4 and 2025Q1 and $7.91 in 2025Q2.
Corrected to
Read from archived price lists: the H100's floor was Google Cloud's $11.06 from 2024Q2 to 2025Q2, since AWS cut its price only in June 2025; the A100 80GB's was AWS's $3.43 from 2025Q3, with no archived price before Azure's record of December 2021; and the H200's was $10.60 in 2025Q1 and 2025Q2, with no archived price in 2024Q4.
Where
H100, A100 80GB and H200 chip pages: quarterly price history; homepage H100 price history.
Found by
A review of published pages and data.
Evidence
Google Cloud VM instance pricing, archived May 21, 2024 (a3-highgpu-8g) · AWS Price List, EC2 offer version of Aug 28, 2025 (p4de.24xlarge) · AWS Price List, EC2 offer version of Mar 31, 2025 (p5en.48xlarge)
Release
PR #138
Buildout#

The OpenAI-Cerebras deal carried Reuters' $10 billion instead of the value Cerebras states

The site said
The compute deals ledger listed OpenAI's January 2026 deal with Cerebras at more than $10 billion, a press-reported figure Reuters attributed to a source.
Corrected to
More than $20 billion, the value Cerebras gave the deal in its IPO prospectus (Form S-1/A, May 4, 2026), listed as a contract value stated by a party to the deal.
Where
/deals ledger table; /deals.json.
Found by
A review of published pages and data.
Evidence
Cerebras Systems registration statement (Form S-1/A), May 4, 2026
Release
PR #139
Buildout#

The $300 billion OpenAI-Oracle deal was listed as unconfirmed by either company

The site said
The compute deals ledger listed the September 2025 OpenAI-Oracle contract at about $300 billion as a press report that neither company had confirmed, with a press-reported value.
Corrected to
More than $300 billion over five years, the value OpenAI gave its July 2025 Oracle partnership on September 23, 2025, listed as a contract value stated by a party to the deal; Oracle had not stated the figure, and neither company had confirmed the reported 2027 start.
Where
/deals ledger table; /deals.json.
Found by
A review of published pages and data.
Evidence
OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites (OpenAI, September 23, 2025)
Release
PR #140
Financials#

Oracle's FY27 capex guidance has an Oracle source

The site said
/buildout cited a secondary press report for Oracle's ~$70B FY27 net cash capex guidance, marked it a secondary source, and said no Oracle release, slide deck or transcript stated the figure.
Corrected to
The guidance row cited Oracle's own Q4 FY26 earnings slides (June 10, 2026, slide 15), which state ~$70B of FY27 net cash outlay for capex, and was no longer marked secondary; the note said the ~$90–95B gross range rests on secondary reports of the June 10 call. The $70B figure did not change.
Where
/buildout capex source notes (Oracle); /financials.json Oracle guidance row.
Found by
A review of published pages and data.
Evidence
Oracle Q4 FY26 earnings presentation, slide 15
Release
PR #133
Hardware specs#

The GeForce RTX 4090 was listed as shipping after it left production

The site said
The spec table listed the GeForce RTX 4090 as shipping.
Corrected to
Legacy: press reports relaying supply-chain sources said the card went out of production in late 2024, and the RTX 5090 succeeded it in January 2025. NVIDIA has published no end-of-life notice.
Where
/hardware spec table; /accelerators.json and /accelerators.csv.
Found by
A scheduled survey.
Evidence
Digital Trends via Yahoo Tech: RTX 40-series cards going out of production
Release
PR #136
Hardware specs#

Vera Rubin's first production shipments were dated a month late

The site said
The spec table dated Vera Rubin's launch, its first production shipments, to September 2026, an estimate made before NVIDIA reported them.
Corrected to
August 2026: NVIDIA's investor presentation for its second fiscal 2027 quarter, published Aug 26, 2026, said it commenced production shipments of Vera Rubin in early August.
Where
/hardware spec table; llms-full.txt accelerator specs; /accelerators.json and /accelerators.csv.
Found by
A scheduled survey.
Evidence
NVIDIA Q2 fiscal 2027 investor presentation
Release
PR #136

Hardware specs#

B200 and Rubin bandwidth came from superseded NVIDIA pages

The site said
B200 bandwidth read 8.0 TB/s, the DGX B200 figure, and Rubin read 22 TB/s with NVLink at 3.6 TB/s from older HGX pages, beside an unverified TSMC 3nm node.
Corrected to
HGX B200: 7.7 TB/s per GPU, from the HGX B200 column of NVIDIA's Blackwell datasheet. Rubin: 19.2 TB/s and NVLink 6 at 3 TB/s, from NVIDIA's Vera Rubin NVL72 page and datasheet, with no process node because NVIDIA stated none. NVIDIA's own pages disagreed on Rubin at the time, and the chip note said so.
Where
/hardware spec table; llms-full.txt accelerator specs; /hardware HBM bandwidth chart.
Found by
A scheduled survey.
Evidence
NVIDIA Blackwell datasheet (HGX B200 column) · NVIDIA Vera Rubin NVL72 · NVIDIA Vera Rubin datasheet
Release
PR #109
Buildout#

Cluster-ledger dates, an unsourced chip count and two citations were wrong

The site said
The cluster ledger dated Stargate Saline's announcement to Jun 1, 2026 and Reliance Jamnagar's to Jan 2025, gave Stargate Abilene ~450K GB200/GB300-class accelerators with no source behind the figure, and cited the wrong sources for claims on Fairwater Wisconsin and Stargate Argentina.
Corrected to
Stargate Saline's announcement was dated Oct 30, 2025 and Jamnagar's Aug 29, 2024; Abilene was given ~403K B200/B300 projected at full buildout, labeled as a tracker estimate; and both claims were given their own sources.
Where
/buildout cluster ledger; buildout.json.
Found by
A scheduled survey.
Evidence
Cluster ledger with each claim's source
Release
PR #109
GPU quotes#

Two desk-log entries read a venue change and a partial archive as price records

The site said
The desk log said “RTX 4090 floor cut −38.5%”, but Salad's lower tier had entered the comparison while Vast.ai's price rose. It also said “B200 floor at $6.25 — lowest in the archive's 81-day span”, but the daily price ledger shows $4.19 on Aug 5, 2026.
Corrected to
The log was changed to record a changed venue set as a venue change with its like-for-like move, and to claim a floor low only when the full daily ledger holds nothing lower in its span. The RTX 4090 entry became a venue change; the B200 floor-low entry was withdrawn.
Where
/log; RSS feed.
Found by
A review of published pages and data.
Evidence
B200 daily price ledger ($4.19 on Aug 5, 2026) · RTX 4090 daily price ledger
Release
PR #102
GPU quotes#

Chip-page price ranges hid that an endpoint was an old estimate

The site said
Chip pages printed all-market ranges such as “All-market access $1.99 → $6.00” on MI300X without saying that the $1.99 endpoint is a secondary estimate observed on Jul 6, 2026.
Corrected to
Ranges were changed to print an estimated endpoint's est flag and its own observation date, so the MI300X range marked $1.99 as an estimate observed on Jul 6, 2026.
Where
All-market ranges on the chip pages.
Found by
A scheduled survey.
Evidence
MI300X chip page
Release
PR #109
Hardware specs#

Gaudi 3 throughput used Intel's superseded 2024 figure

The site said
Gaudi 3 was listed at 1,835 TF dense BF16 and FP8, the figure in Intel's 2024 Hot Chips deck, with a $15,650 list price.
Corrected to
1,678 TF BF16 and FP8, as Intel's July 2025 white paper (Rev. 3) gives at 900 W, and a $15,625 list price ($125,000 per eight-accelerator kit).
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A scheduled survey.
Evidence
Intel Gaudi 3 AI accelerator white paper (Rev. 3)
Release
PR #109
Hardware specs#

GeForce tensor throughput used the 2× FP16-accumulate rates

The site said
RTX 4090 was listed at 330 TF dense BF16 and 661 TF FP8, and RTX 5090 at 419 TF and 838 TF: NVIDIA's FP16-accumulate rates, twice the FP32-accumulate figures that data-center datasheets give.
Corrected to
RTX 4090: 165.2 TF BF16 and 330.3 TF FP8. RTX 5090: 209.5 TF BF16, 419 TF FP8 and 1.68 PF FP4. These are the dense FP32-accumulate rates in NVIDIA's architecture whitepapers, so both cards' rental price per PFLOP-hour doubled.
Where
/hardware spec table and price per PFLOP-hour ranking; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
NVIDIA Ada GPU Architecture whitepaper V2.02, Appendix A, Table 2 · NVIDIA RTX Blackwell GPU Architecture whitepaper V1.1, Appendix A, Table 3
Release
PR #102
Hardware specs#

Smaller hardware fixes: TPU notes, estimate flags, process nodes and a launch month

The site said
TPU v5p's note said it trained Gemini 1.0, TPU v5e's gave its INT8 rate as 394 TOPS, and TPU7x's said capacity required a supported reservation. The V100, A100 and H100 street prices ($10,000, $15,000 and $27,500) carried no estimate flag although only secondary sources give them. A100 and RTX 5090 process nodes read TSMC 7N and TSMC 4NP, and MI355X's launch month read Oct 2025.
Corrected to
The notes were changed to say that TPU v5p was announced alongside Gemini 1.0, which was trained on TPU v4 and v5e; that v5e INT8 is 393 TOPS; and that TPU7x had been generally available on demand since March 2026, with quota required. The three street prices were flagged as estimates, the nodes changed to TSMC N7 and TSMC 4N, and MI355X's launch month to Sep 2025.
Where
/hardware spec table; llms-full.txt accelerator specs; /hardware chip notes.
Found by
A scheduled survey.
Evidence
Google Cloud TPU v5p · Google Cloud TPU v5e · Google Cloud TPU release notes · NVIDIA A100 specifications · Vendor sources listed with each chip
Release
PR #109
GPU quotes#

Year-on-year and streak figures compared unlike series

The site said
The B200 tile read “+42% y/y”, a change computed by array position across a missing quarter and a methodology change. Streak copy called a single quarter near $3.99 a run of “straight quarters”, and each ticker arrow came from a different series than the value beside it.
Corrected to
Changes were made to follow calendar quarters within one methodology and to compare only the venues quoted at both ends; at the correction the B200 tile read “+16.9% like-for-like since Jul 23, 2026”. Streak copy was made to need at least two quarters, and each arrow to come from the series its value shows.
Where
Homepage B200 tile, lede and brief; market tape arrows; chip mastheads and The Board deltas; /gpu-cloud quote-snapshot read.
Found by
A review of published pages and data.
Evidence
Jul 23, 2026 quote snapshot · Oct 13, 2026 quote snapshot
Release
PR #102
Hardware specs#

Maia 200 throughput was approximate

The site said
Maia 200 was listed at 5,000 TF FP8 and 10,000 TF FP4, flagged as estimates.
Corrected to
5,072 TF FP8 and 10,145 TF FP4, the peak-specification table in Microsoft's Maia 200 architecture deep dive. Its BF16 peak is a vector-unit rate, so the tensor BF16 field stayed empty.
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A scheduled survey.
Evidence
Maia 200 architecture deep dive (peak-specification table) · Microsoft Maia 200 announcement
Release
PR #109
Hardware specs#

AMD's MI400-series flagship was listed at half its dense throughput

The site said
The Instinct MI400 row read AMD's figures as sparse and halved them: 10 PF FP8 and 20 PF FP4, with no BF16 rate. It also listed 19.6 TB/s, no board power, announced status and a Dec 2026 launch.
Corrected to
The row became the Instinct MI455X: 5,033 TF BF16, 20,133 TF FP8 and 40,265 TF MXFP4, which AMD publishes as dense rates, with 23.3 TB/s of HBM4 bandwidth, 2,500 W board power, ramping status and an estimated Oct 2026 launch.
Where
/hardware spec table; llms-full.txt accelerator specs; /hardware HBM bandwidth chart.
Found by
A scheduled survey.
Evidence
AMD Instinct MI455X specifications
Release
PR #109
Financials#

Microsoft's 2023 capex was understated

The site said
Microsoft's calendar-2023 capital expenditure read $41.0B.
Corrected to
$41.2B, the sum of Microsoft's four quarterly investor-relations figures, finance leases included.
Where
/buildout capex chart and table; homepage capex chart and table.
Found by
A scheduled survey.
Evidence
Capex table on /buildout
Release
PR #109
Token prices#

Four current OpenAI models were missing from the token-price table

The site said
The current token-price table, reviewed on Oct 11, 2026, left out GPT-5.6 Sol, Terra and Luna (released Jul 9) and GPT-6 Sol (released Sep 22).
Corrected to
All four were added at their Standard prices per million input and output tokens: GPT-6 Sol $2 and $10, GPT-5.6 Sol $4 and $20, GPT-5.6 Terra $2 and $12, and GPT-5.6 Luna $0.20 and $1.20.
Where
/inference current price table; llms-full.txt token prices.
Found by
A scheduled survey.
Evidence
OpenAI GPT-6 Sol model page · OpenAI GPT-5.6 Sol model page · OpenAI GPT-5.6 Terra model page · OpenAI GPT-5.6 Luna model page
Release
PR #109
GPU quotes#

Quotes last observed on Jul 23 were dated to a later survey

The site said
Vast.ai rows last observed on Jul 23, 2026 (H100 $1.73, H200 $3.93 and B300 $8.34 per GPU-hour) were cited with the date of the manual survey (Oct 11) in the H100, H200 and B300 chip-page FAQs and their structured data, and the H100 row also in the homepage lede and brief.
Corrected to
Each cited row was given its own observation date (“last observed Jul 23, 2026”), and the chip-page FAQ's cheapest-venue answer was changed to name the cheapest primary on-demand row separately from the marketplace-access floor.
Where
H100, H200 and B300 chip-page FAQ and FAQPage JSON-LD; homepage H100 lede and brief; venue-spread tile and notes; /gpu-cloud venue-spread panel.
Found by
A review of published pages and data.
Evidence
H100 price ledger (Vast.ai $1.73, observed Jul 23) · H200 price ledger (Vast.ai $3.93, observed Jul 23) · B300 price ledger (Vast.ai $8.34, observed Jul 23)
Release
PR #102
GPU quotes#

Three quotes stayed listed after their sources stopped publishing them

The site said
Google Cloud (A4) B200 on-demand at $16.11 (est) stayed in the table although Google's pricing page shows no on-demand A4 price, and so did FluidStack H200 at $2.30 (est), although FluidStack publishes no GPU prices, and SF Compute H100 at $2.04, a market price SF Compute no longer publishes. The first two set the B200 all-market ceiling and the H200 all-market floor.
Corrected to
All three were retired from the quote table. With the Oct 14, 2026 update, the B200 all-market ceiling became AWS (p6) at $14.24 and the H200 all-market floor Vast.ai at $3.93.
Where
B200, H200 and H100 quote tables on /gpu-cloud and the chip pages; B200 and H200 all-market ranges.
Found by
A scheduled survey.
Evidence
Google Cloud accelerator-optimized pricing (A4 on-demand: N/A) · FluidStack pricing page
Release
PR #109
GPU quotes#

Vultr plans that cannot be rented on demand were listed as on-demand quotes

The site said
Four Vultr bare-metal plans that Vultr's plans API marks as not deployable on demand, with no locations, were listed as On-demand rows: H100 $2.99, B200 $8.50, MI300X $3.99 and MI355X $2.59 per GPU-hour. They made MI355X the cheapest dense compute on /hardware, set its chip-page title to “from $2.59”, and were the only venue in the daily H100 and B200 specialist medians.
Corrected to
The rule became that a Vultr plan counts as on-demand only when Vultr lists it as deployable on demand in at least one location. The four rows were retired from the quote table in the Oct 14, 2026 refresh, and no headline used them after it.
Where
H100, B200, MI300X and MI355X quote tables on /gpu-cloud and the chip pages; /hardware price per PFLOP-hour leader; MI355X chip-page title; daily H100 and B200 specialist-median history.
Found by
A review of published pages and data.
Evidence
Vultr public plans API (deployability and locations per plan)
Release
PR #102 (the rule) and PR #104 (the rows retired)

Token prices#

An April 2026 token-price floor was presented as current

The site said
The GPT-4-class token-price floor, a benchmark-qualified series that ends in April 2026, was presented as the current cheapest option on the homepage Explore card, the ticker, the FAQ and /inference.
Corrected to
Each surface was changed to label it as archived April 2026 history, beside a current token-price table that compares list prices and does not rank model quality.
Where
Homepage Explore card, ticker and FAQ; /inference price-collapse chart and notes.
Found by
A review of published pages and data.
Evidence
Current token prices and the archived series
Release
PR #89 and PR #90
GPU quotes#

Three AWS Capacity Block quotes cited the On-Demand pricing page

The site said
The B300 (p6-b300), GB200 NVL72 (u-p6e) and Trainium2 (trn2) Capacity Block rows linked AWS's EC2 On-Demand pricing page instead of its Capacity Blocks pricing page. The B300 row showed $11.70 per GPU-hour.
Corrected to
They were changed to cite the EC2 Capacity Blocks pricing page, read on Oct 11, 2026: B300 $16.146 and GB200 NVL72 $10.582 per GPU-hour, and Trainium2 $2.235 per chip-hour, each with its instance configuration.
Where
B300, GB200 NVL72 and Trainium2 quote tables on /gpu-cloud and the chip pages; llms-full.txt quote table.
Found by
A scheduled survey.
Evidence
AWS EC2 Capacity Blocks pricing
Release
PR #89
Hardware specs#

B200 memory per GPU was overstated

The site said
B200 was listed with 192 GB of memory per GPU.
Corrected to
180 GB per GPU, as NVIDIA's HGX reference architecture specifies.
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
NVIDIA HGX reference architecture, components
Release
PR #89
Hardware specs#

B300 specifications did not match NVIDIA's current HGX datasheet

The site said
B300 was listed with 15 PF dense FP4, 288 GB of memory, 8.0 TB/s and 1,400 W.
Corrected to
The HGX B300 column of NVIDIA's datasheet: 13.5 PF dense FP4 (108 PF per eight-GPU system divided by eight), 270 GB, 7.7 TB/s and 1,100 W. GB300 NVL72 lists different per-GPU figures, so the chip note recorded the discrepancy.
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
NVIDIA Blackwell Ultra datasheet, page 5 · NVIDIA HGX specifications
Release
PR #89
Market structure#

CME compute futures were still shown as launching Oct 5

The site said
The homepage, the About page and the venue monitor described CME's Silicon Data H100 and B200 rental-index futures as scheduled to launch on Oct 4, 2026, with a first trade date of Oct 5, pending CFTC review.
Corrected to
The venue record, and with it the homepage, the About page and the desk log, was changed to say that CME's Sep 28, 2026 notice had postponed the launch pending regulatory review, without a replacement date.
Where
Homepage futures brief; /about market-structure section and venue monitor; price-signal notes; /log and RSS feed.
Found by
A review of published pages and data.
Evidence
CME notice, Sep 28, 2026 (launch postponed) · CME clearing notice 26-274 (effective date to be announced)
Release
PR #90
Hardware specs#

MI355X throughput used AMD's sparse figures

The site said
MI355X was listed at 5 PF dense BF16, 10.1 PF FP8 and 20.1 PF FP4: AMD's sparse throughput, twice the dense rates.
Corrected to
2.5 PF BF16, 5 PF OCP-FP8 and 10.1 PF MXFP4, AMD's explicit dense specifications, so its rental price per PFLOP-hour doubled.
Where
/hardware spec table and price per PFLOP-hour ranking; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
AMD Instinct MI355X specifications
Release
PR #89
Financials#

Oracle's FY27 funding plan was described as new debt

The site said
/buildout said Oracle ran −$23.7B of free cash flow in FY26 and was raising ~$40B more debt.
Corrected to
/buildout was changed to report Oracle's own description, each figure beside its source: a ~$40B FY27 funding plan of debt and equity, no further debt issuance expected in calendar 2026, and $20B of common-stock sales completed in FY27 Q1.
Where
/buildout financing notes.
Found by
A review of published pages and data.
Evidence
Oracle FY26 Q4 and full-year results release (funding plan) · Oracle FY27 Q1 results release
Release
PR #89
Hardware specs#

Rubin throughput was estimated from a sparse package figure

The site said
Rubin was listed with only an FP4 rate, 25 PF, estimated by halving NVIDIA's ~50 PF NVFP4 inference figure, with 20 TB/s of bandwidth and no BF16 or FP8 values.
Corrected to
The individual-GPU table on NVIDIA's HGX page at the time: 4 PF dense BF16, 17.5 PF FP8/FP6 training and 35 PF NVFP4 training, 22 TB/s, and NVLink 6 at 3.6 TB/s.
Corrected again
: B200 and Rubin bandwidth came from superseded NVIDIA pages.
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
NVIDIA HGX specifications
Release
PR #89
Hardware specs#

TPU memory was shown in GB, and published formats were missing

The site said
TPU v5p, v6e and v7 memory read 95, 32 and 192 GB although Google specifies GiB. FP8 was blank for v5p and v6e, BF16 was blank for v7, and v7 bandwidth read 7.37 TB/s.
Corrected to
95, 32 and 192 GiB; FP8 459 TF (v5p) and 918 TF (v6e); v7 BF16 2,307 TF; and v7 bandwidth 7.38 TB/s (7,380 GB/s), from Google's TPU7x architecture comparison.
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
Google Cloud TPU7x architecture comparison
Release
PR #89
Hardware specs#

Trainium2 throughput and memory unit were approximate

The site said
Trainium2 was listed at 650 TF dense BF16 and 1,300 TF FP8, flagged as estimates, with 96 GB of memory.
Corrected to
667 TF dense BF16, 1,299 TF FP8, 96 GiB and NeuronLink v3 at 1.28 TB/s, as AWS's Neuron documentation specifies.
Where
/hardware spec table and price per PFLOP-hour ranking; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
AWS Neuron Trainium2 architecture
Release
PR #89
Hardware specs#

Trainium3 dense BF16 was half its FP8 figure, not the published rate

The site said
Trainium3 was listed at 1.26 PF dense BF16, half its 2.52 PF FP8 figure, with no FP4 value and memory in GB.
Corrected to
671 TF dense BF16, 2.517 PF for both MXFP8 and MXFP4, 144 GiB of memory and NeuronLink v4 at 2.56 TB/s, as AWS's Neuron documentation specifies.
Where
/hardware spec table; llms-full.txt accelerator specs.
Found by
A review of published pages and data.
Evidence
AWS Neuron Trainium3 architecture
Release
PR #89