Butterfly Effect

You are in6 Nov 2026the origin

Plate I. A butterfly woven from orbit lines. Two futures that begin a billionth apart end on different wings.

Small flaps. Big price storms.

Butterfly Effect forecasts what a GPU will rent for. Starting on 6 November 2026 its world model plays out 4,000 futures for every price it tracks and reports the range they land in, because two futures that begin a billionth apart can end on different wings.

This page is a time machine. It leaves from 6 November 2026, the last day the forecast was allowed to see, and stops wherever the forecast has something to say.

Depart for 13 Nov 2026or scroll to travel

Asking reruns the pipeline on our backend and replays the 6 Nov 2026 analysis. Your question titles it; it does not change the forecast.

You have arrived in13 Nov 20267 days after departure

One week out, most futures have barely moved

The scored forecast covers 3 dates: 7 Nov, 9 Nov and 13 Nov. For each of 37 price series it gives a low, a middle and a high. One future in ten should land below the low, and one in ten above the high.

Posted rates sit still for days at a time, so their range starts as a single line. Asking prices on the open marketplace spread out at once.

7 Nov9 Nov13 Nov
Neocloud rate cardH100 SXM
On 13 Nov, per GPU-hour
lowmiddlehigh
$3.95$3.95$3.95
7 Nov9 Nov13 Nov
Neocloud spotH100 SXM
On 13 Nov, per GPU-hour
lowmiddlehigh
$2.29$2.29$2.35
7 Nov9 Nov13 Nov
Hyperscaler spotA100 40GB
On 13 Nov, per GPU-hour
lowmiddlehigh
$1.89$1.99$2.09
7 Nov9 Nov13 Nov
Marketplace askH100
On 13 Nov, per GPU-hour
lowmiddlehigh
$1.58$1.88$2.63

The four charts share one vertical scale, from −18% to +50% around each series’ middle price on 7 Nov. Band: low to high. Orange line: the middle. Read from submission/forecasts.csv when this page was built (strategy pooled_state_v1).

You have arrived in6 Dec 202630 days after departure

If you are buying this week, lock the rate

In the world model’s futures, for a 30-day block on rate cards and hyperscaler spot, the locked discount beats paying on demand in 90 to 100% of them. The call is lock for every H100 and A100 series.

For a single week, locking is insurance and nothing more. In the backtest the policy saved +0.08% against paying on demand, somewhere between −0.50% and +0.52%. It has no skill at timing.

Lock.The 30-day lock wins in 90 to 100% of the world model’s 4,000 futures per rate-card and hyperscaler-spot series.

You have arrived inNov 2027a year after departure

A year out, the futures split four ways

Eight actors sit in the model: the US government, China, Taiwan, TSMC, NVIDIA, AMD, the hyperscalers and the neoclouds. 20 weighted relations tie them together. Six of them play a game, and its solution puts odds on four futures.

The game’s 35 payoffs are judgement calls set by hand, so every report shows how the odds move when they change.

  1. 46.1% Tighter export controls
  2. 23.0% A glut as new generations ship
  3. 17.7% Status quo
  4. 13.2% A supply squeeze
46.1%Tighter export controlsDemand moves to capacity that complies, and the H100 holds near $3.23 an hour in November 2028.Plate II: a chip at the centre of a labyrinth
23.0%A glut as new generations shipBlackwell and Vera Rubin arrive in volume. The H100 falls to $1.37 an hour and the A100 to $0.71.Plate III: a wave breaking over a heap of circuit boards
17.7%Status quoNothing breaks and nothing floods. The H100 sits at about $2.50 an hour.Plate IV: one graphics card turning inside a steady orbit
13.2%A supply squeezeChips get scarce. The H100 lands at about $2.50 an hour here too.Plate V: an hourglass tied off at the waist

The plates are illustrations made with Midjourney. They are not data. The odds and prices are.

You have arrived inNov 2028two years after departure

Two years out, the range is the forecast

Geopolitics sets the spread, not the centre. New chips pull the middle down: the same work costs about 40% less on Blackwell than on an H100. Which of the four futures arrives decides how far the old chips fall.

H100

Now

$3.85 an hour

Nov 2027

$1.62 low, $3.04 middle, $4.40 high

Nov 2028

$0.91 low, $2.56 middle, $4.59 high

A used H100 is worth about $16,700 in 2027 and $9,800 in 2028.

A100

Now

$1.79 an hour

Nov 2027

$0.83 low, $1.48 middle, $2.34 high

Nov 2028

$0.55 low, $1.30 middle, $2.68 high

A used A100 is worth about $3,200 in 2027 and $300 in 2028.

On-demand median rent, dollars per GPU-hour. “Now” is the week of 26 October 2026. Each bar runs from the low to the high: one future in ten falls below it and one in ten above. The orange dot is the middle.

You are back in15 Nov 2026the present, 9 days after departure

193 tries came back with us. Most were buried

Butterfly Effect also tries to improve itself. It has scored at least 193 forecasting configurations from 46 past start dates with Pacific’s own scorer, and a critic with ten checks judged every one. The critic’s bar for luck rises with each try, so every idea makes the next one harder to keep.

The world model’s own variants never beat the simplest forecast. A second forecaster, which pools prices across series, does: 0.204 against 0.163 on those 46 dates. The same critic rejects it on the later dates alone, so that number is a selection score, not a result.

One week, 24 to 30 October 2026, was held back and scored once. Our own pre-registered rule picked the simplest forecast. We submit the pooled forecaster instead, because it leads on the organisers’ headline measure, pooled skill (+0.125 against +0.096); on our rule’s own measure, the average of per-series skill, it is last (−0.281 against −0.047), and we say plainly that we chose it after seeing the result. The rule’s pick is kept beside it, and Sunday’s score will be reported for both.

Five of the buried ideas, all changes to the world model’s forecaster. The check they failed most often was the placebo: feed the model the same signal from 30 days earlier, when it should be useless, and see whether it helps anyway. It did.

  1. p001Model each kind of price its own way: marketplace asks move in small, frequent steps; rate cards and spot move in rare jumps.Rejected: placebo could not be run, ranges too narrow.
  2. p002Widen the ranges using the forecaster's own recent errors.Rejected: luck, placebo, did not hold across the period, ranges too narrow.
  3. p003Rate cards jump when a provider joins or leaves the panel, and those changes come in bursts.Rejected: placebo, ranges too narrow.
  4. p004Each provider keeps its recent price direction for weeks.Rejected: placebo, did not hold across the period, ranges too narrow.
  5. p005Neoclouds follow each other's repricing.Rejected: placebo, ranges too narrow.

What it cannot do yet

  • 0.159The world model does not beat the simplest forecast.“Next week equals today” scores 0.163 across 46 past start dates. The world model’s variants score 0.124 to 0.159.
  • 62–72%Its ranges are too narrow.They should catch 80% of outcomes. On 46 past start dates they catch 62 to 72%. The submitted forecaster caught 80% in the one held-out week, too little to call it fixed.
  • 0.997Export-control news does not move one-week prices.Step rates on AWS spot after an announcement, against before: a ratio of 0.997, somewhere between 0.96 and 1.04.

Still in the presentthe paper, §5.12

What we tested, and what survived

Fourteen hypotheses, each with its test and its verdict. One model was kept: pooled_state_v1, with 36 wins and 10 losses over 46 origins. Nothing proposed afterwards was kept.

DevThe incumbent, pooled_state_v1, scores 0.204 per origin [90% 0.172, 0.235], pooled 0.199, coverage 0.717: +0.041 over the best baseline, last_value_empirical (0.163). A selection score, not out of sample. Unchanged since the freeze: 0 incumbent changes in 4 weekly ledgers.

Hypotheses, results and verdicts
HypothesisResultVerdict
H1Pooling recent price changes across series of the same kind and collector beats last value36 wins, 10 losses over 46 origins; 0.163 → 0.204 per originkept; not confirmed on the purged windows
H2A candidate beats last value on the sealed holdout week (the rule’s measure)mean of per-series skill −0.281 against −0.047 (pooled +0.125 against +0.096)rule picks last value; declared post-hoc deviation
H3Export-control news raises the chance of a spot price stephazard ratio 0.997 [0.962, 1.037], placebo p 0.35rejected
H4Mixing the game’s scenarios, or an event hazard, improves forecastsBrier 0.2488 against 0.2537, unstable; signal gain −0.0024rejected
H5Single-mechanism changes to the hurdle world model help0.129 to 0.159 against 0.163; coverage 0.63 to 0.69rejected, 5 of 5
H6Language-model proposers and shared memory find better specificationsbest accepted 0.1667 in all 10 arms; memory arms re-tested 0 rejected specs, against 2 to 4null; fewer re-tests
H7A web search in every round helps the proposerbest accepted 0.1667 in both armsnull
H8Some combination of the four knobs is both calibrated and better768 configurations; best in the coverage band 0.1504 against 0.1667null
H9Weekly updates change the model0 of 16, 0 of 13 and 0 of 16 proposals keptnothing kept
H10Last value on rate cards, pooling everywhere else8 wins, 38 losses; 0.204 → 0.186rejected
H11A guard for series that do not move0 wins, 0 losses, 46 tiesnull by construction
LRA linear trend beats last valuebest variant +0.147 and +0.089 against +0.189 and +0.192not supported
GateA news classifier beats the label-frequency prior by 5 points68.9 against 66.2failed (provisional)
DecisionThe rent / lock / wait policy saves money+0.08% [−0.50, +0.52] against paying on demandnull

193 configurations in the trial ledger and 957 more pending; the critic log holds 13 reviews, 2 kept and 11 rejected, and both kept reviews are of pooled_state_v1. Read the paper.

Still in the presentthe architecture

How it is built

Two paths share one data pack. The scored path writes the CSV and calls no model, Exa or past.dev. The research loop runs on dev origins only, and nothing in it writes the submission.

Pinball loss
ρτ(y, q) = max(τ(y − q), (τ − 1)(y − q)), ℓ = ⅓ Σ ρτ over τ ∈ {0.1, 0.5, 0.9}Each forecast is three quantiles; a cell’s loss is the mean pinball loss of the three, with Pacific’s scorer unchanged.
Skill against last value
K = 1 − Σ ℓ / Σ ℓ⁰, ℓ⁰ = |y − v| / 2The baseline puts all three quantiles at the last observed price v, so it loses nothing when the price holds still.
The critic’s bar
sign test over origins: p < 0.05 / N, N = 1,150 nowN counts every configuration ever scored on dev, kept or not. pooled_state_v1 passed at N = 171 (p 7.8 × 10⁻⁵ against 2.9 × 10⁻⁴); the bar is now 4.3 × 10⁻⁵.

The bar falls with every try (193 configurations in the ledger, 957 pending), so every idea makes the next one harder to keep. Read the architecture.

Running it again next week

When a week of new prices arrives, one command scores last week’s forecast against what happened, lets the proposer try four changes and lets the unchanged critic decide whether the incumbent is replaced. A human only confirms the data pack and the cutoff; the command never writes the submission.

make weekly AS_OF=<date> DEV_DIR=<dev files>

The weekly runbook.

Still in the presentthe sponsors’ tools

Sponsor technology, exactly

What each one does in Chokepoint, how much of it ran, what it found, and where the code is.

All four run in the research loop on dev origins, not in the scored forecast: make reproduce makes no model, Exa or past.dev call, and nothing ran on Modal in the forecast path. None got a change past the critic. Memory and search did not improve forecasts.

Still in the presentthe conclusion

What we conclude

What the work supports, each point with the numbers behind it. None is stronger than its evidence, and the last one says which number to wait for.

  1. One week ahead, the simplest forecast is hard to beat.“Next week equals today” scores 0.163 across 46 past start dates. Every variant of the world model’s forecaster scores below it, 0.124 to 0.159.
  2. One method cleared the bar, by a modest margin.The pooled forecaster scores 0.204 (90% range 0.172 to 0.235) against 0.163 for the best baseline and 0.157 for drift. It was found after at least 171 tries, and nothing proposed after it was kept, so that number is a selection score.
  3. Its ranges are better than the baselines’, and still a little narrow.The low-to-high range should hold the real price 80% of the time. On the past start dates it held it 71.7% of the time, against 61.8% for drift. On the held-back week it held it 79.6% of the time.
  4. News about chokepoints did not move one-week prices. Market structure did.Three tests of events came back null: export-control announcements left the rate of price steps unchanged (hazard ratio 0.997, 0.962 to 1.037). Rate cards followed other providers’ repricing, and spot and marketplace asks fell back to a going rate. Events stay in the two-year view.
  5. Two years out, H100 rent falls about a third at the middle, and the range is wide.From $3.85 an hour now to $2.56 in November 2028, with one future in ten below $0.91 and one in ten above $4.59. Geopolitics sets the spread, not the centre.
  6. The only clean test is the week nobody chose on.Every score above was used to pick or tune something. Pacific holds back the real prices for 7 to 13 November. That score, reported for the submitted file and for our rule’s pick, is the only one nobody selected on.

The findings with every caveat are in the write-up, and the models and tests in the paper.