Forward

What the index is expected to average, and what locking the same compute costs today.

For the next one, three, six and twelve months: the expected mean of the TCI EU H100 print, with its range, and the cheapest price an EU buyer can secure the same unit at today. A research output beside the index, estimated from the stored record as it stood on the day, and not a reference price for any financial instrument.

Estimated 15 Sep 2026method 1.0.0+4b8640d50d57latest.jsonhistory.csv

Expected index over the window

USD per GPU-hour. The mean of the published print over the window, as expected on the estimate date, with the 10th and 90th percentiles of that mean.

No estimate on 15 Sep 2026. The reasons, by window, are below. Nothing is carried forward from an earlier day.

Table view

Expected window mean, its percentiles, the term-implied figure and the cheapest lock, USD per GPU-hour
Window	Expected mean	10th	Median	90th	Term-implied	Cheapest lock	Lock minus expected
1 month	—	—	—	—	—	111.97	—
3 months	—	—	—	—	—	36.91	—
6 months	—	—	—	—	—	18.46	—
12 months	—	—	—	—	—	9.20	—

1 month: 4 pro-forma prints under 0.5.0, fewer than the 8 the fit needs
3 months: 4 pro-forma prints under 0.5.0, fewer than the 8 the fit needs
6 months: 4 pro-forma prints under 0.5.0, fewer than the 8 the fit needs
12 months: 4 pro-forma prints under 0.5.0, fewer than the 8 the fit needs

Cost to lock today

USD per GPU-hour actually used. The cheapest way an EU buyer could secure H100 SXM compute under the index's unit today: EU/EEA, at least two GPUs. A contract longer than the window is charged in full. It is a price to act on. It neither forecasts nor bounds the index, because a median of offers cannot be reproduced by renting one of them.

Cheapest lockable cost by window
Window	Per used GPU-hour	How	Seller	Covers, days	Offers counted
1 month	111.97	rate card, whole contract charged	azure	365	33
3 months	36.91	rate card, whole contract charged	azure	365	33
6 months	18.46	rate card, whole contract charged	azure	365	33
12 months	9.20	rate card, whole contract charged	azure	365	33

Term-implied, beside the estimate

The expected level scaled by the median discount of at least three independent EU sellers in the index population at exactly that tenor. Shown for comparison and never used to form the estimate. A schedule that prices every chip alike is not counted.

1 month: no martingale estimate to scale. Sellers at this tenor: none.
3 months: no martingale estimate to scale. Sellers at this tenor: seeweb (administered untested).
6 months: no martingale estimate to scale. Sellers at this tenor: seeweb (administered untested).
12 months: no martingale estimate to scale. Sellers at this tenor: seeweb (administered untested).

Track record

Estimates are stored as they are made and never revised in place. A window is scored once it has closed, only if at least half its days printed, and only windows that do not overlap count, so a year-long horizon adds one independent observation a year. Rows marked backfilled were computed after their date for 41 earlier sessions, from data stored on each date. They are shown apart and are not a track record.

Scored windows by horizon
Window	Estimates	Independent	Bias	RMSE	Inside 10th to 90th
1 month	as made	0 of 20	—	—	—
3 months	as made	0 of 20	—	—	—
6 months	as made	0 of 20	—	—	—
12 months	as made	0 of 20	—	—	—
1 month	backfilled	0 of 20	—	—	—
3 months	backfilled	0 of 20	—	—	—
6 months	backfilled	0 of 20	—	—	—
12 months	backfilled	0 of 20	—	—	—

How the estimate is made

A GPU-hour cannot be stored, so today's price does not fix a forward price the way carry fixes one for a metal. The forward of a good that cannot be stored is an expectation, and that is what is estimated here: the price level is treated as a martingale, so with no drift the record can measure, the expected level over any window is today's.

The print is a weighted median of a handful of prices and moves in steps. It is modelled as a level plus measurement noise, fitted by maximum likelihood on each estimate date, so a one-day spike moves the estimate a little rather than all the way.

Where a window runs past an announced change of methodology, the days after the change are estimated from the index recomputed under the new rules from the same stored prices. A version counts only once its notice was published.

No estimate is published on a day the recomputed index did not print, and no earlier value is carried forward.

Prepaid prices are restated as paid monthly at the government curve of their currency before they are compared with anything.