Published 07:13 AEST · Before US Open · Wednesday, 9 September 2026Pipeline 04:30 AESTVol. I  No. 121
The Grid
Independent Price Reporting for the Compute Market
thegridco.ai
EST. 2026 · SYDNEY
Derived Instruments

Instruments

The assessed series of record, printed daily with the methodology public — for the analyst marking a book or the buyer sizing a commitment. Three derived reads on the same physical market: what the market pays to lock capacity (the term structure), whether the equity tape agrees with the physical layer (the divergence view), and what a unit of inference costs by route (the inference spread). Every series has a specification page; where the rules withhold or suspend a print, the state renders in place of the number.

· The Term Structure

GAP-H100 spot — daily print · reserved-1Y and GTS-1Y suspended from 2026-08-22 (notice 2026-08-14)
GAP-H100 spot · print of 2026-09-09
WITHHELD
min-obs rule
panel below two contributing providers — methodology change notice 2026-08-16. Printing resumes when the panel supports it. The previous print stands on its own date and is not carried forward.
Observations: 3 bookable asks across 1 provider(s) (vast.ai 3d unchanged) · excluded from panel: hyperstack BENCHED 15d unchanged, lambda BENCHED 15d unchanged, runpod BENCHED 112d unchanged · transactions: 0 · judgment: none (arithmetic) · tier-2 bookable asks only
Final curve · 2026-08-21 · suspended
SPOT1Y RSV$2.69$6.46
GTS-1Y · final print 2026-08-21
+$3.77
suspended
GTS-1YSUSPENDEDFINAL PRINT 21 AUG 2026since 22 AUG 2026 · notice 2026-08-14full disclosure below ↓
GTS-1Y ($/GPU-hr)series begins 2026-07-24 · final print 2026-08-21suspended from 2026-08-22
$3.50$4.00+$3.77MOONSHOT AI · 27 JULMSFT · 29 JULMETA · 29 JUL24 Jul29 Jul3 Aug9 Aug14 Aug21 Aug
Regime · history to the final print 2026-08-21
CONTANGO — MARKET PAYS TO LOCK
Table view
DateGTS-1Y $/hr (final print 2026-08-21)Regime
2026-08-21+$3.77CONTANGO
2026-08-20+$3.26CONTANGO
2026-08-19+$3.77CONTANGO
2026-08-18+$3.98CONTANGO
2026-08-17+$3.77CONTANGO
2026-08-16+$3.82CONTANGO
2026-08-15+$3.91CONTANGO
2026-08-14+$3.90CONTANGO
2026-08-13+$4.02CONTANGO
2026-08-12+$3.77CONTANGO
2026-08-11+$3.88CONTANGO
2026-08-10+$4.23CONTANGO
2026-08-09+$3.99CONTANGO
2026-08-08+$4.02CONTANGO
2026-08-07+$3.98CONTANGO
2026-08-05+$3.77CONTANGO
2026-08-04+$3.92CONTANGO
2026-08-03+$3.77CONTANGO
2026-08-02+$3.77CONTANGO
2026-08-01+$4.00CONTANGO
2026-07-31+$4.02CONTANGO
2026-07-30+$4.02CONTANGO
2026-07-29+$3.90CONTANGO
2026-07-28+$4.02CONTANGO
2026-07-27+$4.02CONTANGO
2026-07-26+$3.98CONTANGO
2026-07-25+$4.02CONTANGO
2026-07-24+$4.08CONTANGO
GAP-H100 — printed history, range, realised movement · series begins 2026-07-24 (v3 panel)
$2.50$3.00$3.2024 Jul30 Jul5 Aug12 Aug18 Aug7 Sept
RANGE OF PRINTS $2.24$3.20MEDIAN $2.54n=31 PRINTS · 24 JUL 20267 SEPT 202610 WITHHELD DAYS IN THE STORED RECORD
Largest print-to-print moves — the dated record
From (print)To (print)MoveSpan
2026-08-23 · $2.482026-09-07 · $3.20+$0.72across 15 days (no print between)
2026-08-19 · $2.692026-08-20 · $3.20+$0.51next print day
2026-08-20 · $3.202026-08-21 · $2.69$0.51next print day
2026-08-10 · $2.242026-08-11 · $2.58+$0.34next print day
2026-08-12 · $2.692026-08-13 · $2.45$0.24next print day

Ranges and moves are display arithmetic over this series’ own printed history — computed at render, never stored, never printed as a series. The 2026-07-24 re-base is the series’ inception; nothing here crosses a methodology boundary. Panel-composition context for any move is in the print of record’s provenance stamp.

GAP-H100 (spot)Reclassified 2026-08-14 from assessment to indicator: the panel holds two nominal contributors, one static across its entire 93-day history, and 50.3% of the printed series' realised variance is panel composition, not price. Printing continues unbroken. Returns to assessment class when the panel carries at least two independently moving contributors (methodology change notice 2026-08-14). Minimum-observation rule adopted 2026-08-16, binding from the 2026-08-22 print: the daily print is withheld — disclosed as WITHHELD with the provenance stamp, never a silent gap — when the panel has fewer than two contributing providers or every contributor is unchanged for 7 or more days (methodology change notice 2026-08-16).
GAP-H100-R1YDefect disclosed 2026-08-14: every print since inception (2026-07-24) is the identical value — zero variance. The reserved panel (near-static Azure retail plus a Lambda fallback) does not move, so the series has carried no market information. Suspended from 2026-08-22 (methodology change notice 2026-08-14); published history stands as-published.
GTS-1YDefect disclosed 2026-08-14: the reserved leg has zero variance, so GTS-1Y reduces to a constant minus GAP-H100 spot — correlation −1.0 with the spot leg across every print. It carries no independent term-structure information. Suspended from 2026-08-22 (methodology change notice 2026-08-14); published history stands as-published.

· Physical vs Tape — the Divergence View

Physical composite · GCCX (29-ticker) · rebased 100 = 2026-06-09
PHYSICAL COMPOSITEGCCX (EQUITY TAPE)DIVERGENCE GAP
80.0100.0120.0140.0PHYSICAL 138.1GCCX 98.3COREWEAVE, INC. · 18 JUNNEBIUS GROUP N.V. · 17 JULNEBIUS GROUP N.V. · 17 JUL9 Jun26 Jun13 Jul29 Jul16 Aug9 Sept
Divergence oscillator  (GCCX 30d% − physical 30d%, pts)
−40.0−20.00.0CROWDEDTAPE BELOW PHYSICAL−3.89 Jul21 Jul3 Aug13 Aug22 Aug9 Sept
Above zero — crowded. The tape has run ahead of the physical layer: positioning risk, trim/hedge signal for the exposed reader.
Below zero — tape below physical. Equities broke while the physical layer held: a technical correction, not a thesis break. July 2026 is the live case.
Table view
DatePhysicalGCCXDivergence 30d pts
2026-09-0998.3−3.8
2026-09-0898.2−5.0
2026-09-07138.1
2026-09-0495.9−6.0
2026-09-0393.9−9.1
2026-09-0293.3−6.2
2026-09-0193.0−3.6
2026-08-3194.5−1.7
2026-08-2493.6−7.1
2026-08-23120.7
2026-08-22125.094.7−25.2
2026-08-2194.3−10.7
2026-08-20137.594.1−11.2
2026-08-1994.3−7.4
2026-08-18119.195.8−15.6
2026-08-17100.3−0.3
2026-08-16122.9
2026-08-15120.999.5−17.3
2026-08-14121.099.0−20.3
2026-08-13118.399.7−16.6
2026-08-1298.9−8.6
2026-08-11120.896.7−26.6
2026-08-10112.296.8−13.2
2026-08-09117.8
2026-08-08117.097.8−3.4
2026-08-07117.796.6−20.8
2026-08-0596.6−14.1
2026-08-04118.997.8−12.4
2026-08-0393.8−16.3
2026-08-01116.7

· The Inference Bill & Spread

RW-S / RW-L · today's cost by route · $/day · index in the footnote
RW-S · small workload · 1M tokens/day· today’s bill by route
Frontier API$2.25/day9 SEPT
Open-weight APICHEAPEST$0.09/day9 SEPT
Self-host · spotWITHHELD · PANEL <2 · print of 9 SEPT · min-obs rule · inherited from GAP-H100 spot
Self-host · reservedCEASED · FINAL 21 AUG · since 22 AUG · notice 2026-08-14 §6 · input leg suspended
STAYING ON A FRONTIER API COSTS $2.16/DAY MORE THAN THE CHEAPEST ROUTE (9 SEPT)
RWI-S 50.00 (9 SEPT) · 100 = 2026-05-07 · cheapest path: open-weight API
RW-L · large workload · 100M tokens/day· today’s bill by route
Frontier API$200.00/day9 SEPT
Open-weight APICHEAPEST$8.40/day9 SEPT
Self-host · spotWITHHELD · PANEL <2 · print of 9 SEPT · min-obs rule · inherited from GAP-H100 spot
Self-host · reservedCEASED · FINAL 21 AUG · since 22 AUG · notice 2026-08-14 §6 · input leg suspended
STAYING ON A FRONTIER API COSTS $191.60/DAY MORE THAN THE CHEAPEST ROUTE (9 SEPT)
RWI-L 50.00 (9 SEPT) · 100 = 2026-05-07 · cheapest path: open-weight API
FRONTIER APICHEAPEST OPEN PATHTHE PREMIUM
$0.00$2.00$4.00FRONTIER $2.25OPEN $0.09MOONSHOT AI · 27 JULMSFT · 29 JULMETA · 29 JUL24 Jul1 Aug10 Aug18 Aug1 Sept9 Sept
v1.0 is token-denominated; models differ in tokens-per-task; efficiency-adjusted v1.1 in development.
Table view
DateFrontier $/dOpen $/dSpread $/d
2026-09-092.250.092.16
2026-09-082.250.092.16
2026-09-072.250.092.16
2026-09-062.250.062.19
2026-09-052.250.092.16
2026-09-042.250.092.16
2026-09-032.250.092.16
2026-09-022.250.092.16
2026-09-012.250.092.16
2026-08-312.250.092.16
2026-08-242.250.072.18
2026-08-232.250.062.19
2026-08-222.250.082.17
2026-08-212.250.102.15
2026-08-202.250.102.15
2026-08-192.250.102.15
2026-08-182.250.102.15
2026-08-172.250.102.15
2026-08-162.250.082.17
2026-08-152.250.062.19
2026-08-142.250.062.19
2026-08-132.250.062.19
2026-08-122.250.102.15
2026-08-112.250.102.15
2026-08-102.250.102.15
2026-08-092.250.112.14
2026-08-082.250.112.14
2026-08-072.250.112.14
2026-08-054.000.113.89
2026-08-044.000.113.89
Honest display constraints

An index chart would look nearly flat today and that is the honest picture: 41 prints since inception, 15 of them at 64.29. The route repricings so far have moved the spread, not the cheapest path. The display leads with the bill by route (ruling of 2026-08-25), not a drama-less index chart — the index chart earns its place as history accrues (the page’s standing awaiting-first-print discipline). A missing print renders as a gap, never an interpolation.

Self-host · reserved (RW-S / RW-L)Ceased 2026-08-22 with the suspension of its input series GAP-H100-R1Y (methodology change notice 2026-08-14 §6): a self-host cost priced on a suspended leg would restate the defect daily. Published history stands as-published. RWI-S / RWI-L cheapest-path indices are unaffected — the cheapest path was the open-API leg on every print of both.
v1.0 is token-denominated; models differ in tokens-per-task; efficiency-adjusted v1.1 in development.

· The Ledger

forward claims — recorded when made, checked when due · 40 on record
The honesty frame

The Grid does not forecast. It records who did, and checks. Every entry is a dated, attributed, third-party forward claim, stored verbatim the day it was captured. When a claim’s date arrives, the register’s print of record scores it — hit, miss, or partial — and the score is permanent. The ledger never argues with a claim in flight; an OPEN row is a record, not an endorsement or a rebuttal.

Forecasting stands suspended (identity decision of record, 2026-07-24). This surface is the inverse of forecasting: the register’s dated memory of other people’s numbers, held against the prints that eventually exist.

Status walk: OPEN → DATE ARRIVED → SCORED (hit / miss / partial, per the register print)
Who · where · when madeThe claim (stored verbatim)Due / triggerScored byStatus
PETER DIAMANDISmoonshots · 2026-08-13Grok 4.5 came out two weeks ago, 4.6 just this morning, and 4.7 is rumored to be coming out in two weeks.commentator_signals 69032026-08-27dated record — model release calendardated record — model release calendarDATE ARRIVED
DAVE BLUNDINmoonshots · 2026-08-13Grok 5 was supposed to be out in May, I think it was, and it's now August.commentator_signals 69052026-08-31dated record — model release timeline slippagedated record — model release timeline slippageDATE ARRIVED
JASON CALACANISallin · 2026-08-14According to the Financial Times, this would obviously break the record-setting $1.75 trillion IPO by SpaceX. FT reported Wednesday night that investors expect the company to go public in October, which is like six to eight weeks away.commentator_signals 68902026-10-31Anthropic IPO filing/completion recordAnthropic IPO filing/completion recordOPEN
EMAD MOSTAQUEmoonshots · 2026-08-13You could shoot a movie just like that one probably for $100,000 of compute and then $10,000 of compute probably by the new year.commentator_signals 69102026-12-31video gen compute cost trajectory by year-end 2026video gen compute cost trajectory by year-end 2026OPEN
JASON CALACANISallin · 2026-08-14Annualized run rate, we talked about this sacks a bunch, will end the year between $100 and $120 billion.commentator_signals 68912026-12-31Anthropic quarterly earnings or public filingsAnthropic quarterly earnings or public filingsOPEN
CHAMATH PALIHAPITIYAallin · 2026-07-31There are some incredible efficiencies about to be demonstrated that will cut token consumption by 50-75% for the same task.commentator_signals 58112027-01-27STANDING WATCH: unstated horizon ('about to be') — 180-day review window from episode date, closes 2027-01-27token_pricingDaily token-pricing prints + RWI route prints. Stored speaker 'Chamath' (pre-canonicalization row; stored rows not rewritten).OPEN
DIGITIMESdigitimes · 2026-07-12TSMC CoWoS output reportedly to reach at least 200K wafers in 2027news_signals 54502027-12-31s4_cowos[gap] trade-press capacity prints (digitimes cadence already ingested); no scoring series yet — arrives with the capacity register.OPEN
DWARKESH PATELdwarkesh · 2026-07-30AI will go from 60% of N3 to 86% wafer allocation by end of 2027, when wafer allocation will hit a wall.commentator_signals 60242027-12-31s2_1_wafer_fab[proxy] monthly trade-press wafer-start prints; instrument gate (013 §5.1) applies before any series.OPEN
Showing 8 of 40 unscored claims — earliest due first.
Undated claims. A claim with no stated date is legal only as LEVEL-TRIGGERED (scores if the named level prints; no expiry) or STANDING WATCH (a stated review window or event trigger) — the trigger is printed in the row so the convention is auditable.
No scoring series yet. A claim mapping to an uninstrumented node stays OPEN with that tag — it scores against the dated public record when its date arrives, with a one-line evidence citation, never mechanically. Scores set by founder review; none are automatic.
THE GRID PUBLISHES CONDITIONS AND INPUTS — MARKET OBSERVATION, NEVER INVESTMENT OR PROCUREMENT ADVICE. · CLAIMS ARE THIRD-PARTY STATEMENTS RECORDED FOR SCORING, NOT GRID VIEWS.
Derived series from the Grid registry (methodology versions as shown). GTS-1Y: panel composition differs across legs — the delta over time is the signal, not the level. GCCX: derived value only, no raw quote republication. Event pins from the Grid forward calendar, day-precision entries only. Ledger: third-party claims stored verbatim from the capture row; status flips by the daily date-arrival pass; scores are founder-set and permanent.