LIVERESEARCH PREVIEW · REFERENCE RUN #001

The AI Compute Delta Feed

Tracks who is constrained by what, which bottlenecks are moving, which assumptions are becoming fragile, and how shocks propagate across chips, power, capital, and regulation.

Machine-generated strategic judgment: every claim is dated, falsifiable, and cryptographically sealed at birth, then adjudicated in public against open data — failures published alongside hits.

Data as of Jul 28, 2026 · 27 active claims, sealed · insight-level track record forming (n=0)

27
active claims
27
sealed artifacts
214
signpost checks registered
4 days
next window closes
4 for10 against13 timed out8 pending

Reality's checks so far: 4 for, 10 against, 13 timed out, 8 pending. We publish the against column.

Signpost checks adjudicate individual observations; the insight-level track record (below) publishes only when whole calls resolve — currently forming. Counts are unique questions (214 raw checks deduplicated).

The house view

The connective theses across 3 signals clusters — each joins multiple signals. House view sealed Jul 2, 2026 (older than the portfolio below; the theses carry their own stamp).

Emerging

AI infrastructure is experiencing a fundamental bottleneck migration from GPU allocation to power/interconnect capacity, creating systematic underpricing of energy-constrained compute.

Joins 4 signals · 4 substrate impacts
  • The assumption that GPU allocation remains the primary bottleneck for frontier training faces mounting contradictory signals
  • The assumption that interconnection queue times don't constrain datacenter siting faces mounting pressure from two contradictory signals
  • Grid capacity assumption supporting three portfolio decisions faces mounting contradictions without offsetting support
  • MISO demand baseline assumption faces pressure from 3.08σ spike to 121,514 MW
supports — Physical power and interconnect — not GPU supply — is becoming the binding constraint on AI-compute buildout
contradicts — GPU allocation remains the primary bottleneck for frontier training
contradicts — US grid can absorb datacenter load growth through 2027 without material curtailment
pressures — Cloud GPU on-demand pricing stays within +/-15% of current levels through 2026
What to watch next
Watch: PJM and ERCOT capacity auction clearing prices for 2027/28 delivery years
Compute pricing will bifurcate into power-constrained premium tiers and power-available standard tiers, with energy-secured facilities commanding 30-50% premiums
Non-obvious

NVIDIA's training ecosystem is experiencing accelerated development velocity that signals preparation for post-GPU-scarcity competition dynamics.

Joins 2 signals · 3 substrate impacts
  • NVIDIA Megatron-LM release activity shows 6.27σ deviation from baseline, pressuring capacity and pricing assumptions tied to typical activity levels
  • Federal Register AI-compute publications accelerated 178% this week, contradicting the consensus view that regulatory pace won't materially shift near-term assumptions
pressures — NVIDIA maintains >80% training-accelerator share through 2026
supports — Open-source framework dominance is contestable; the developer center of gravity is moving up-stack to agent orchestration
pressures — AI-compute export controls remain stable through Q3 2026
What to watch next
Watch: AMD MI300 and Intel Gaudi deployment announcements from hyperscalers for 2026 training workloads
NVIDIA is fortifying software lock-in ahead of hardware commoditization, suggesting they anticipate meaningful accelerator competition within 18 months
Non-obvious

Critical AI infrastructure exhibits the pre-failure signature of underfunded single points of failure across multiple dependency layers.

Joins 2 signals · 2 substrate impacts
  • reference infrastructure dependency configuration matches OpenSSL pre-Heartbleed: open standard with multiple commercial dependants relying on one underfunded maintainer team
  • The assumption that GPU allocation remains the primary bottleneck for frontier training faces mounting contradictory signals
pressures — HuggingFace transformers remains the default model-serving substrate
supports — Open-source framework dominance is contestable; the developer center of gravity is moving up-stack to agent orchestration
What to watch next
Watch: HuggingFace transformers GitHub issue velocity and maintainer commit frequency relative to dependent project growth
A critical failure in foundational AI infrastructure dependencies will trigger rapid enterprise migration to commercially-supported alternatives

Explore the claims

Each card is a distinct claim, leading with what would prove it wrong. Pick your seat to re-lens the decision implications, or use the filters below.

Grammar
Status
Conviction
0 shown

Verified track record

Every insight is pre-registered with a falsification threshold and sealed at birth; resolutions are sealed too. Public tier reflects calls resolved ≥ 7 days ago.

The scorecard is earned, not written

1Pre-register each claim with an explicit falsifier before any outcome is known.
2Seal it — SHA-256 → daily Merkle root → public anchor — so its timestamp can't be back-dated.
3Adjudicate automatically against open-data signposts as their windows close.
4Publish the scorecard at a public delay, with misses alongside hits.

Signposts adjudicated so far

27 signpost checks have adjudicated (4 for, 10 against, 13 timed out). None of them yet meets our scoring standard, so the scored record is still empty. A trip counts only when it was adjudicated mechanically against the evidence spine or by an independent judge — these were matched by an earlier keyword-overlap method we no longer treat as sound. Insight-level resolutions publish here automatically once a verdict seals under the current rule and the 7-day public delay elapses.

What these numbers do and don't say

Signpost tallies and claim counts are a snapshot generated Jul 28, 2026 — 19 days ago, and are out of date. They are read from the downloadable export bundle below, which is refreshed by commit, not continuously. The scorecard above is read live. Where the two disagree, the live figure at /api/v1/public/ai-compute/track-record is authoritative.

Withdrawn resolution criteria: 91 across 54 open calls. Between 2026-07-25 and 2026-08-15 we withdrew 91 machine-checkable resolution criteria across 54 open calls, before any of them resolved. Most were sigma-anomaly tests over hourly US grid demand — a comparison that cannot separate the claim being tested from summer air-conditioning load and the daily demand cycle. Two independent errors: the criteria had no stable operating characteristic (one pair's confirming side would have fired on 0.44% of scans measured on 2026-07-10 and 11.65% on 2026-07-25, with the falsifying side moving the opposite way, on an unchanged thesis), and a "3 sigma" criterion scanned hourly for 30 days offers 526 chances to fire, so it has roughly a 51% chance of firing on noise alone against the 0.135% its label implies. The affected calls remain open and visible; they are not scored, and they stay in the denominator. The withdrawal rule is arithmetic that was available on the day each criterion was sealed, was applied to every criterion that failed it regardless of which way it was trending, and was recorded while the resolution monitors were disabled — so the sealed record shows the withdrawals preceded any outcome.

Verify it yourself

Don't take the page's word for the portfolio. The claim, signpost, delta, and track-record files rendering this page are published below with SHA-256 digests; your browser can re-verify them right now. (The House View's synthesis file sits outside this manifest today — that's stated where it renders.)

In-browser verification needs a secure context (HTTPS). Download a file and hash it locally instead.

FileRowsSHA-256Verify
insights-active.jsonl10892bc3cd834d1b510…
insights-active.csv108853a83ab121aa34f…
insights-deltas.jsonl104162f901656bd68cd…
insights-signposts.jsonl214216bc846a1972194…
insights-resolved.jsonl56b5a4cc2f041f8bb…
track-record.json5d2703955ff62a113…
entity-crosswalk.jsonl2990f2b63c6b506a3c8…
entity-crosswalk.csv299cc3e1ee97458ad19…
insights-computation.jsonl0e3b0c44298fc1c14…

A mismatch usually means a mid-deploy cache; retry in a minute, then inspect. The offline path for professionals:

npx tsx scripts/verify-insight-pack.ts <pack.json>

Each seal chains SHA-256 → daily Merkle root → public anchor.

Data & documentation

insights-active.jsonl108 rowsinsights-active.csv108 rowsinsights-deltas.jsonl104 rowsinsights-signposts.jsonl214 rowsinsights-resolved.jsonl5 rowstrack-record.json5 rowsentity-crosswalk.jsonl299 rowsentity-crosswalk.csv299 rowsinsights-computation.jsonl0 rows