Shortages

AI compute

AI compute is the chip capacity associated with the AI work in the model.

What-if model · Chosen targets, conditional projections and calculated gaps.

Conditional constraint
Calculated output blocked
$0bnin this scenario, one constraint
Range across tested cases
$0bn – $34.8T4 cases; not a probability interval

This is a conditional test of the chosen target economy. It does not predict how much GDP will be lost. Changing the target mix, supply assumptions or production method can change the result.

[A] AssumedSCENARIO (binds only in another way of producing)UNTESTED
$0Cyan · tested range   Yellow · base$34.8T

What is being compared?

The model converts AI worker-equivalents into gigawatts of chip thermal design power using an assumed ratio. This makes the result sensitive to the amount of useful work attributed to that hardware.

Quantity2035 amount
Needed for the full target146.3
Supply on course84.5
Supply at precedent12,455.7

All three use the same input unit — GW of AI chip TDP in service in 2035.

Show the shortage calculation

Input shortage = maximum of zero and requirement minus supply. Coverage = supply ÷ requirement.

Supply pathInput shortfallFull-target need met
On course61.857.8%
At precedent0.08515.3%

The physical shortfall and the blocked-output amount use different units. The model converts between them using each layer's input requirements and the allocation described below.

Source files · shortages_v4.csv

How this becomes blocked output

The model asks two questions. With the whole target attempted, how much tier-2 output must be removed at precedent input supply? With only on-course output and tier 1 attempted, how much tier-1 output must be removed at on-course supply? It adds the resulting tier-2 and tier-1 blocks for this shortage.

$0bn tier 2 + $0bn tier 1 = $0bn blocked

The allocation drops eligible output from the model's most input-intensive layers first and protects the on-course baseline. The ordering is fixed within a test. Labour-dependent requirements are recalculated as output is removed.

Understand the allocation and its limits

What changes the answer?

Each row recomputes the constraint under a different production or demand assumption. The displayed range takes the smallest and largest output blocks across all these rows.

CaseWhat changesOutput blocked
basev3 split: humans spread over all tasks; the gap split cognitive (AI) / physical (robots) by each layer's high-skill share$0bn
cogfirstAI agents take every cognitive task first; humans move to physical tasks; robots only for physical work beyond them$34.8T
humanscoghumans keep the cognitive tasks; AI adds no work; robots fill the whole gap$0bn
trendprodv3 split, with output per worker on its trend (Agent 4 2035 intensity) instead of each sector's fastest 9-year growth$0bn
Assumption behind the verdict

binds only under: AI agents take every cognitive task first; humans move to physical tasks; robots only for physical work beyond them

See its reported rank in each scenario
ScenarioRankOutput blocked
base16$0bn
AI takes cognitive tasks first3$34.8T
humans keep cognitive tasks16$0bn
labour productivity on trend17$0bn
v3 second sourced ratios15$0bn
2026 intensities / K-Y trend19$0bn
land- and approval-saving technology, task reallocation15$0bn

The published ranking takes the first matching variant for each scenario. Clinical approval and cropland each have a second S6 variant that appears in the range above and is omitted from those ranking columns.

Read every scenario

What evidence would improve this?

Epoch chip-sales update for 2027-28; TSMC CoWoS and HBM capacity plans to 2030

Growth and the historical comparison

AI chip power sold x1.87 in 2025 (Epoch) held 2028-2035; bounded in practice by K01/K14 (power and grid)

RateAnnual growth
Needed to reach the target input25.5%
Implied by precedent supply105.7%

The growth comparison uses a nine-year interval from 2026 to 2035.

Sources and exact values

Source files · shortages_v4.csv · variants_v4.csv · own_blocks_v4.csv

base · stage1-v3/requirements/requirements.csv (K18 work_ratio_2035); stage1-v3/repair/agent4_bottlenecks/work/model_memo.json (cog, gw_per_we); stage1-v3/repair/agent4_bottlenecks/work/need_by_layer.csv (K03, K04 on course); stage1-v3/requirements/model_constants.csv (MK13, MK14)

cogfirst · stage1-v3/requirements/requirements.csv (K18 work_ratio_2035); stage1-v3/repair/agent4_bottlenecks/work/model_memo.json (cog, gw_per_we); stage1-v3/repair/agent4_bottlenecks/work/need_by_layer.csv (K03, K04 on course); stage1-v3/requirements/model_constants.csv (MK13, MK14)

humanscog · stage1-v3/requirements/requirements.csv (K18 work_ratio_2035); stage1-v3/repair/agent4_bottlenecks/work/model_memo.json (cog, gw_per_we); stage1-v3/repair/agent4_bottlenecks/work/need_by_layer.csv (K03, K04 on course); stage1-v3/requirements/model_constants.csv (MK13, MK14)

trendprod · stage1-v3/requirements/requirements.csv (K18 work_ratio_2035); stage1-v3/repair/agent4_bottlenecks/work/model_memo.json (cog, gw_per_we); stage1-v3/repair/agent4_bottlenecks/work/need_by_layer.csv (K03, K04 on course); stage1-v3/requirements/model_constants.csv (MK13, MK14)

Open the input files and source registries

Copy this prompt

Copy the selected text using your keyboard or touch menu, then paste it into your assistant.