← Cockpit
INF_004predictionGeopoliticsChina-ghost-DCs

China has massive latent 'ghost data center' capacity — fully powered facilities sitting empty awaiting compute — such that if export controls relax or are bypassed, China has near-immediate physical capacity to train frontier models equivalent to Anth...

Predictor: Jensen Huang

Prior probability
68.0%
Current probability
38.9%
evolves via intake + LBP
Conviction
5/5
Signal quality
A
Resolution
partial
Window
2026-01-01 – 2026-11-30
Edges in / out
3 / 1
Tickers exposed
34

Prediction text

China has massive latent 'ghost data center' capacity — fully powered facilities sitting empty awaiting compute — such that if export controls relax or are bypassed, China has near-immediate physical capacity to train frontier models equivalent to Anthropic's Claude Mythos. | They have ghost cities, they have ghost datacenters too | Any relaxation/bypass of US export controls; domestic Chinese chip ramp

Key catalyst: Any relaxation/bypass of US export controls; domestic Chinese chip ramp

Watch events: China advanced-node wafer output; export-control exceptions or loopholes; DeepSeek/Qwen/Kimi frontier benchmarks

Verbatim quote

From episode "The Architecture of Intelligence: AI Infrastructure, Energy & Networking Predictions (2023-2026)"
They have ghost cities, they have ghost datacenters too

Resolution evidence

Status: partial

WSJ Nov 2025: INF Tech accessed advanced NVIDIA chips via Indosat Jakarta DC despite export controls. DeepSeek V3/R1 demonstrated frontier-class training with constrained hardware.

Predictor: Jensen Huang

κ + Brier as of 2026-05-22
κ (discount)
0.808
Brier
0.0128
excellent
Hits / Misses
6 / 0
of 8 resolved
Hit rate
75.0%
Calibration plot (stated vs observed)

Evidence about this node from Jensen Huang is multiplied by κ in /api/intake. Lower κ = less weight; floors at 0.10 (effectively silenced) and caps at 1.00 (full weight).

Reference class: regulatory_freeze_window

Linked via embedding similarity 0.583

Major-country regulatory pause/moratorium on AI capability research lasting >6 months

Base rate
5.0%
0/4 historical
Inside weight
Outside weight
no pull
inside 38.9% → blend 38.9% 0.0pp)

Tetlock-style outside view: at TRF=1 (just predicted), outside view dominates (w_in=0.3). At TRF=0 (deadline), inside view dominates (w_in=1.0). The blend regularizes overconfident inside views toward the historical base rate.

Probability over time

10 prob_history rows
0%25%50%75%100%prior 68%2026-04-302026-05-032026-05-24
intake v2milestone miss sweeplbp propagationreference class assignedlegacy v1prior_prob (analyst seed)current = 38.9%

Milestone chain

Pre-event signals (upstream prereqs + window checkpoints) → resolution event → downstream cascades. Status/dates update from linked nodes; re-derive nightly via scripts/ops/derive_milestones.py.
Leading chain: 3 overdue ⏱
  1. 2026-01-31overdueQ1 window check-in (25%)
  2. 2026-03-02overdueQ2 window check-in (50%)
  3. 2026-04-01overdueQ3 window check-in (75%)
  4. 2026-03-01 → 2026-09-30pendingChina data center construction capacity disclosure (MIIT or equivalent)
    How: China's Ministry of Industry and Information Technology (MIIT) or major Chinese cloud (Alibaba, Tencent, Baidu) publicly discloses idle/under-utilized data center capacity
    Source: MIIT.gov.cn, Reuters China coverage, SCMP Tech, Caixin Globalconf 40%
  5. 2026-04-01 → 2026-11-30pendingChinese frontier model trained on domestic chips reaches Claude/GPT-class capability
    How: DeepSeek, Alibaba Qwen, ByteDance Doubao, or peer releases a frontier model trained primarily on domestic Huawei Ascend / Cambricon / SMIC silicon, with benchmarks ≥85% of Claude Sonnet / GPT-4o levels on MMLU + SWE-Bench
    Source: DeepSeek-AI blog (https://www.deepseek.com), Alibaba Cloud announcements, HuggingFace leaderboardsconf 60%
    Notes: DeepSeek V3/V4 already demonstrated competitive results on H800-class chips; further frontier-class with fully domestic stack is the bar.
  6. 2026-04-01 → 2026-11-30pendingU.S. export-control modification or enforcement gap announced
    How: Bureau of Industry and Security (BIS) announces changes to chip export controls, OR Reuters/NYT reports gap in enforcement allowing >100,000 H100/H200/B200 equivalent to reach China
    Source: BIS.doc.gov, Reuters, NYT investigative reporting, Center for Strategic and International Studiesconf 55%
  7. 2026-09-01 → 2027-06-30pendingChinese AI lab announces frontier-scale training run on domestic infrastructure
    How: Public announcement of frontier model training on ≥10K-equivalent-H100 cluster located in China
    Source: Lab blog posts, Reuters, SCMP Techconf 50%

What if this resolves?

Clamp this prediction TRUE or FALSE and run a counterfactual Gibbs sample. Surfaces the predictions whose marginals shift most under that assumption.
(live posterior: 39%)

Click a button to clamp this prediction and run a Gibbs sample. Returns the predictions whose marginals shift most. ~30s per run; ideal for stress-testing "if X resolves, what else moves?"

Evidence chain

Every probability update with full Bayesian provenance — chronological, latest first
LBP2026-05-24T02:00:02Z38.9%+13.6pp
Network propagation: 25.3% → 38.9%
4-iter LBP, residual 0.01000 · damping 0.5, w_intrinsic 0.5 · method lbp_v3 · run 806b02f8
intake_event_update2026-05-21T23:15:16Z25.3%-29.1pp
intake:7afeeb9a-f217-4dd2-b910-24ff14bdfc39 bayesian_v2 inside=0.544 blend=0.253 LLR=0.000 κ=0.81 w_in=0.60 regulatory_freeze_window
Raw metadata
{
  "trf": 0.5766698637873985,
  "kappa": 0.8077,
  "base_rate": 0.05,
  "predictor": "Jensen Huang",
  "total_llr": 0,
  "bayesian_v2": true,
  "prior_logit": 0.1784432190618108,
  "bayes_factor": "1:1 (no signal)",
  "blend_reason": "blend 60% inside / 40% outside (TRF=0.577, base_rate=0.050 from regulatory_freeze_window)",
  "inside_prior": 0.5444928057782582,
  "kappa_source": "predictor_table",
  "blend_applied": true,
  "contributions": [
    {
      "llr": 0,
      "kappa": 0.8077,
      "label": "If China is blocking US chips and pushing domestic, the 'ghost data center' thesis is in tension.",
      "adjusted_llr": 0
    }
  ],
  "evidence_kind": "intake_event_update",
  "inside_source": "history_v2",
  "inside_weight": 0.596331095348821,
  "outside_weight": 0.403668904651179,
  "posterior_prob": 0.2530961100525352,
  "evidence_origin": "daily_intake",
  "llm_suggestions": [
    {
      "polarity": "mentions",
      "status_change": "partial",
      "evidence_strength": "weak",
      "delta_prob_suggestion": -0.03
    }
  ],
  "posterior_logit": 0.1784432190618108,
  "predictor_brier": 0.01276,
  "evidence_doc_ids": [],
  "inside_posterior": 0.5444928057782582,
  "blended_posterior": 0.2530961100525352,
  "reference_class_id": "regulatory_freeze_window",
  "total_adjusted_llr": 0,
  "predictor_n_resolved": 8
}
LBP2026-05-17T02:00:01Z54.4%+2.0pp
Network propagation: 52.4% → 54.4%
5-iter LBP, residual 0.00689 · damping 0.5, w_intrinsic 0.5 · method lbp_v3 · run e607fa96
LBP2026-05-10T02:00:02Z52.4%+4.0pp
Network propagation: 48.4% → 52.4%
6-iter LBP, residual 0.00584 · damping 0.5, w_intrinsic 0.5 · method lbp_v3 · run e5c18d29
LBP2026-05-03T02:00:01Z48.4%+7.8pp
Network propagation: 40.6% → 48.4%
6-iter LBP, residual 0.00677 · damping 0.5, w_intrinsic 0.5 · method lbp_v3 · run 1a683ac9
resolution_terminal2026-05-01T00:00:00Z50.0%+9.4pp
resolution_terminal partial outcome=0.5 pre_resolution=0.406
Raw metadata
{
  "source": "backfill_resolution_history.py",
  "status": "partial",
  "bayesian_v2": false,
  "outcome_prob": 0.5,
  "evidence_kind": "resolution_terminal",
  "posterior_prob": 0.5,
  "delta_to_outcome": 0.09372999999999998,
  "inside_posterior": 0.40627,
  "validation_notes": "WSJ Nov 2025: INF Tech accessed advanced NVIDIA chips via Indosat Jakarta DC despite export controls. DeepSeek V3/R1 demonstrated frontier-class training with constrained hardware.",
  "validation_status": "hit",
  "pre_resolution_prob": 0.40627,
  "resolution_evidence": "WSJ Nov 2025: INF Tech accessed advanced NVIDIA chips via Indosat Jakarta DC despite export controls. DeepSeek V3/R1 demonstrated frontier-class training with constrained hardware.",
  "does_not_update_current_prob": true
}
LBP2026-04-30T16:39:51Z40.6%+13.4pp
Network propagation: 27.2% → 40.6%
5-iter LBP, residual 0.00825 · damping 0.5, w_intrinsic 0.5 · method lbp_v2 · run 0c8a4ea3
legacy v12026-04-30T16:13:50Z27.2%-13.4pp
reference_class_assigned bayesian_v2 inside=0.680 blend=0.272 w_in=0.53 regulatory_freeze_window
LBP2026-04-30T02:18:57Z40.6%+13.4pp
Network propagation: 27.1% → 40.6%
5-iter LBP, residual 0.00825 · damping 0.5, w_intrinsic 0.5 · method lbp_v1 · run 592311ef
legacy v12026-04-30T01:56:50Z27.1%-40.9pp
reference_class_assigned bayesian_v2 inside=0.680 blend=0.271 w_in=0.53 regulatory_freeze_window

Network propagation neighbors

Top edges sorted by latest LBP cross-impact
All propagation →

Top incoming (parents)

Edges that influence THIS node's belief

KindNodeTheir probP(c|s=T)P(c|s=F)Δ implied
killerTK06
China-Taiwan Military Conflict
8.0%0.0500.680+0.241
killerTK02
AI Compute Supply Shock (TSMC/Taiwan Disruption)
12.0%0.0500.680+0.215
killerTK09
Energy Grid Cap (Data Center Power Wall)
35.0%0.0500.680+0.070

Top outgoing (children)

Predictions THIS node influences

No outgoing edges.

Ticker exposure

34 ticker(s) linked

Beneficiaries (24)

EQIXSOUNFLNCFSLRWULFAIAPLDARGANNVDACRWVNBISIRENAAPLAMTAMZNETNGOOGLHUBBIRMMETAMSFTORCLSFTBYSTX

Adverse (4)

AAPLLRCXAMATTSLA

Prerequisites (3)

Predictions that must hit first
TypePredTitleDomainLag
killerTK09Energy Grid Cap (Data Center Power Wall)
killerTK02AI Compute Supply Shock (TSMC/Taiwan Disruption)
killerTK06China-Taiwan Military Conflict

Dependents (1)

Predictions enabled by this
TypePredTitleDomainLag
correlateCOD_TECH_001A16/N2-class TSMC process availability materially supports 2027 AI accelerator rampsSemis

Expected milestones (1)

From Sheet 17 Monitoring Triggers
Expected byDescriptionStatus
2026-12-31[Geopolitics 2026-12] [INF_004] China advanced-node wafer output; export-control exceptions or loopholes; DeepSeek/Qwen/Kimi frontie [SEM_021] DeepSeek V5 release; Qwen 3 scale; MoE ternary-weight model benchmark releases [235_007]pending

Validations (2)

Resolution events
Observed atStatusByNotes
2026-05-21partialintake:7afeeb9a-f217-4dd2-b910-24ff14bdfc39If China is blocking US chips and pushing domestic, the 'ghost data center' thesis is in tension.
2026-04-29hitthesis_timeline_v1.0_importWSJ Nov 2025: INF Tech accessed advanced NVIDIA chips via Indosat Jakarta DC despite export controls. DeepSeek V3/R1 demonstrated frontier-class training with constrained hardware.

Raw metadata

From Thesis_Timeline_v1.0_FINAL workbook
{
  "nia": true,
  "mode": "FORECAST",
  "role": "Guest-CEO",
  "context": "On the Dwarkesh Patel podcast (Apr 16 2026), Huang described China's compute situation as the inverse of the US: US has chips but lacks power and land; China has power and shells but lacks chips. Bottleneck asymmetry.",
  "to_year": 2027,
  "verbatim": "They have ghost cities, they have ghost datacenters too",
  "conv_cues": "CEO FIRST_PERSON; direct statement; explicit capability",
  "direction": "HAPPEN",
  "from_year": 2026,
  "timeframe": "2026 (ongoing)",
  "conv_level": "HIGH",
  "milestones": [
    {
      "kind": "quartile_checkpoint",
      "label": "Q1 window check-in (25%)",
      "status": "overdue",
      "weight": 0.05,
      "ordinal": -3,
      "source_id": null,
      "expected_date": "2026-01-31",
      "observed_date": null,
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep"
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q2 window check-in (50%)",
      "status": "overdue",
      "weight": 0.05,
      "ordinal": -2,
      "source_id": null,
      "expected_date": "2026-03-02",
      "observed_date": null,
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep"
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q3 window check-in (75%)",
      "status": "overdue",
      "weight": 0.05,
      "ordinal": -1,
      "source_id": null,
      "expected_date": "2026-04-01",
      "observed_date": null,
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep"
    },
    {
      "kind": "event",
      "label": "China has massive latent 'ghost data center' capacity — fully powered facilities sitting empty awaiting compute — such that if export contro",
      "status": "partial",
      "weight": 1,
      "ordinal": 0,
      "source_id": "INF_004",
      "expected_date": "2026-05-01",
      "observed_date": "2026-05-01"
    },
    {
      "kind": "llm_pre_event",
      "label": "China data center construction capacity disclosure (MIIT or equivalent)",
      "source": "MIIT.gov.cn, Reuters China coverage, SCMP Tech, Caixin Global",
      "status": "pending",
      "weight": 0.4,
      "ordinal": 1,
      "source_id": null,
      "confidence": 0.4,
      "expected_date": "2026-06-15",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2026-09-30",
        "from": "2026-03-01"
      },
      "measurement_criterion": "China's Ministry of Industry and Information Technology (MIIT) or major Chinese cloud (Alibaba, Tencent, Baidu) publicly discloses idle/under-utilized data center capacity"
    },
    {
      "kind": "llm_pre_event",
      "label": "Chinese frontier model trained on domestic chips reaches Claude/GPT-class capability",
      "notes": "DeepSeek V3/V4 already demonstrated competitive results on H800-class chips; further frontier-class with fully domestic stack is the bar.",
      "source": "DeepSeek-AI blog (https://www.deepseek.com), Alibaba Cloud announcements, HuggingFace leaderboards",
      "status": "pending",
      "weight": 0.4,
      "ordinal": 2,
      "source_id": null,
      "confidence": 0.6,
      "expected_date": "2026-07-31",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2026-11-30",
        "from": "2026-04-01"
      },
      "measurement_criterion": "DeepSeek, Alibaba Qwen, ByteDance Doubao, or peer releases a frontier model trained primarily on domestic Huawei Ascend / Cambricon / SMIC silicon, with benchmarks ≥85% of Claude Sonnet / GPT-4o levels on MMLU + SWE-Bench"
    },
    {
      "kind": "llm_pre_event",
      "label": "U.S. export-control modification or enforcement gap announced",
      "source": "BIS.doc.gov, Reuters, NYT investigative reporting, Center for Strategic and International Studies",
      "status": "pending",
      "weight": 0
... (truncated)