Skip to main content
Scrutica
Method 2026-09-26

Compute Leverage

Pinned viewThe assessments below use Google’s training setup. Return to the baseline.

EDA tools (chip design) and EUV lithography can each delay Google’s next frontier training run by more than five years

Evidence register ↓

  1. More than five years

    1. EDA tools (chip design)Eighteen months: can delay past period. Five years: can delay past period.
    2. EUV lithographyEighteen months: can delay past period. Five years: can delay past period.

  2. More than eighteen months; five years uncertain

    1. Advanced logic foundry (≤5nm/3nm)Eighteen months: can delay past period. Five years: uncertain.
    2. Advanced packaging / CoWoSEighteen months: can delay past period. Five years: uncertain.

  3. More than eighteen months, not more than five years

    1. HBM memory (HBM3E / HBM4)Eighteen months: can delay past period. Five years: cannot delay past period.

  4. Eighteen months uncertain; not more than five years

    1. Capital ($1B+/yr runs)Eighteen months: uncertain. Five years: cannot delay past period.
    2. Cloud / datacenter capacityEighteen months: uncertain. Five years: cannot delay past period.

  5. Not more than eighteen months; five years uncertain

    1. Power / grid interconnectionHeld by several institutions acting togetherLabs it cannot delay: xAI, Google and Anthropic, at eighteen monthsEighteen months: cannot delay past period. Five years: uncertain. Cross-run baseline: can delay past period, then uncertain.

  6. Not more than eighteen months

    1. AI acceleratorLabs it cannot delay: Google and AnthropicEighteen months: cannot delay past period. Five years: cannot delay past period. Cross-run baseline: can delay past period, then uncertain.
    2. Scale-out interconnect (InfiniBand vs Ethernet)Eighteen months: cannot delay past period. Five years: cannot delay past period.
    3. Scale-up interconnect (NVLink / NVSwitch)Labs it cannot delay: Google and AnthropicEighteen months: cannot delay past period. Five years: cannot delay past period. Cross-run baseline: can delay past period, then uncertain.
Can delayUncertainEnds before five yearsRuns past five yearsCross-run baseline, where Google’s assessment differsTop edge: source of the redeployment timePrimary / researchPress · analystNo source found

Each band runs as far as its point is assessed to delay a new training run; only eighteen months and five years are assessed.

The assessments use Google’s training setup. The comparison below includes xAI, Anthropic and the baseline.

The assessment covers eleven points where a supplier or institution could prevent a lab from starting a new training run. Five are assessed as able to do so over an eighteen-month horizon; two are assessed as barriers at five years: EDA tools (chip design) and EUV lithography. These are editorial assessments of whether a lab could switch supplier or infrastructure in time; the register grades the evidence behind each. Each of the five has an identified institutional or market constraint. Export controls apply to all five and restrict whom the holder may supply. The counts concern starting a new run; interrupting a run under way is not assessed.

Reach of each check

Google’s run · Eighteen-month assessment

Can refuse supplyCan decide who is servedAt 18 monthsChip-design software — held by Synopsys, Cadence, Siemens EDA. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Chip-design softwareChip lithography — held by ASML. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Chip lithographyLeading-edge fabrication — held by TSMC, Samsung Foundry, Intel Foundry. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Leading-edge fabricationHigh-bandwidth memory — held by SK Hynix, Samsung, Micron. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.High-bandwidth memoryAdvanced packaging — held by TSMC, OSATs (ASE, Amkor, SPIL). At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Advanced packagingData-center capacity — held by AWS, Microsoft Azure, Google Cloud, CoreWeave, Oracle, neoclouds. At eighteen months: the verdict is uncertain. The check over it: a market rival has no power to compel supply.Data-center capacityCapital — held by hyperscaler balance sheets, sovereign funds (QIA, PIF, Mubadala), mega-VC. At eighteen months: the verdict is uncertain. The check over it: a market rival has no power to compel supply.CapitalAI accelerators — held by NVIDIA, Google TPU, AWS Trainium, AMD. At eighteen months: cannot delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.AI acceleratorsIn-server interconnect — held by NVIDIA. At eighteen months: cannot delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.In-server interconnectBetween-server networking — held by NVIDIA/Mellanox InfiniBand, Broadcom/Arista Ethernet. At eighteen months: cannot delay training beyond eighteen months. The check over it: a market rival has no power to compel supply.Between-server networkingGrid power — held by ISO/RTO (MISO, PJM, ERCOT, SPP), local utility, state PUC, FERC, EPA. At eighteen months: cannot delay training beyond eighteen months. The check over it: the holders are themselves public bodies running a queue, so the same process that can refuse a connection also decides whose connects first.Grid poweruncertain / nocan delay >18 monthscannot delay >18 monthsuncertainthe check over that chokepoint reaches this farCan refuse supplyCan decide who is servedAt 18 monthsChip-design software — held by Synopsys, Cadence, Siemens EDA. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Chip-design softwareChip lithography — held by ASML. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Chip lithographyLeading-edge fabrication — held by TSMC, Samsung Foundry, Intel Foundry. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Leading-edge fabricationHigh-bandwidth memory — held by SK Hynix, Samsung, Micron. At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.High-bandwidth memoryAdvanced packaging — held by TSMC, OSATs (ASE, Amkor, SPIL). At eighteen months: can delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.Advanced packagingData-center capacity — held by AWS, Microsoft Azure, Google Cloud, CoreWeave, Oracle, neoclouds. At eighteen months: the verdict is uncertain. The check over it: a market rival has no power to compel supply.Data-center capacityCapital — held by hyperscaler balance sheets, sovereign funds (QIA, PIF, Mubadala), mega-VC. At eighteen months: the verdict is uncertain. The check over it: a market rival has no power to compel supply.CapitalAI accelerators — held by NVIDIA, Google TPU, AWS Trainium, AMD. At eighteen months: cannot delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.AI acceleratorsIn-server interconnect — held by NVIDIA. At eighteen months: cannot delay training beyond eighteen months. The check over it: a licensing regime can forbid a transfer to a destination or a party; none of these regimes obliges a holder to supply any particular buyer.In-server interconnectBetween-server networking — held by NVIDIA/Mellanox InfiniBand, Broadcom/Arista Ethernet. At eighteen months: cannot delay training beyond eighteen months. The check over it: a market rival has no power to compel supply.Between-server networkingGrid power — held by ISO/RTO (MISO, PJM, ERCOT, SPP), local utility, state PUC, FERC, EPA. At eighteen months: cannot delay training beyond eighteen months. The check over it: the holders are themselves public bodies running a queue, so the same process that can refuse a connection also decides whose connects first.Grid poweruncertain / nocan delay >18 monthscannot delay >18 monthsuncertainthe check over that chokepoint reaches this far

Bands appear on the links recorded as able to delay training beyond eighteen months. Each separates the power to refuse supply from the power to decide who receives it.

Each band describes what a check of that kind can do to any lab; no cited source addresses the selected lab’s own supply.

The register

2 corrections · latest Sep 22, 2026
Redeployment time describes how long a lab would need to switch suppliers or infrastructure. “Checked by” identifies the rule, institution or competition that could constrain the holder.
Leverage point18 months5 yearsChecked byRedeployment time
EUV lithography · ASMLcan delay past periodcan delay past periodExport control — A sole-supplier monopoly over EUV scanners, governed by Dutch + US multilateral export-control law: a state has asserted jurisdiction over whom ASML may sell to, and no second supplier exists to check it commercially.No commercial alternative at any horizon for a Western lab; state-built EUV at scale assessed 5-10+ years out · estimated
Advanced logic foundry (≤5nm/3nm) · TSMC, Samsung Foundry, Intel Foundrycan delay past perioduncertainExport control — Leading-edge logic fabrication is governed by US export-control law: the Foreign Direct Product Rule reaches any chip made with US tooling, and the Jan-2025 Foundry Due-Diligence Rule mandates quarterly OSAT/foundry KYC compliance reporting to BIS.No published figure for a full TSMC→Samsung inter-foundry port of a leading-edge AI accelerator
EDA tools (chip design) · Synopsys, Cadence, Siemens EDAcan delay past periodcan delay past periodExport control — EDA for leading-edge design is governed by US export-control law: BIS exercised a standing enforcement order (Cadence $95M penalty, 28 Jul 2025, for EDA exports to Entity-List parties NUDT/Phytium under EAR §744). The broader China-sales licensing letters to Synopsys/Cadence/Siemens (imposed 23 May 2025) were rescinded 2 Jul 2025 under the US–China framework, so the enforcement order is what remains in force.No sourced timetable for a production full-flow EDA alternative at the leading edge.
HBM memory (HBM3E / HBM4) · SK Hynix, Samsung, Microncan delay past periodcannot delay past periodExport control — HBM is governed by US export-control law — the Dec-2024 rule controls HBM at/above a bandwidth-density threshold (>2 GB/s/mm²), on top of the Oct-2022/Oct-2023 advanced-computing controls.12-24 months to qualify a secondary supplier, anchored to the Samsung→NVIDIA case (~18-month qualification for a supplier with a finished product) · estimated
Advanced packaging / CoWoS · TSMC, OSATs (ASE, Amkor, SPIL)can delay past perioduncertainExport control — Since 16 Jan 2025, with compliance required from 31 Jan, a BIS regime reaches the packaging layer directly: the Foundry Due-Diligence Rule (90 FR 5298, FR Doc. 2025-00711) makes approved-OSAT attestations the mechanism by which packaged advanced ICs escape licensing presumptions, keeps the approved-OSAT list at Supplement No. 7 to Part 740, and requires quarterly OSAT/foundry compliance reporting — the same rule already cited on the advanced-foundry layer, whose named obligors include this layer’s own holders. The Dec-2024 rule separately controls advanced-packaging equipment.No published figure for how long customer-level requalification of a given accelerator takes
AI accelerator · NVIDIA, Google TPU, AWS Trainium, AMDcannot delay past periodcannot delay past periodExport control — Frontier AI accelerators are the primary object of US export-control law (ECCN 3A090/4A090; the Oct-2023 Total Processing Performance metric closed the A800/H800 work-arounds).Committed labs (CUDA-bespoke): >18 months. Diversified/vertically-integrated labs: already routed around NVIDIA at production scale (not applicable). · estimated
Scale-up interconnect (NVLink / NVSwitch) · NVIDIAcannot delay past periodcannot delay past periodExport control — Reached by the same US advanced-computing export-control architecture as the accelerator it binds: ECCN 3A090 has an I/O-bandwidth control parameter (≥150 GB/s aggregate bidirectional) and 4A090 covers electronic assemblies/components incorporating controlled accelerators (e.g. GB200 NVL72 racks).No production UALink alternative until ~late 2026 at earliest (committed labs) · estimated
Scale-out interconnect (InfiniBand vs Ethernet) · NVIDIA/Mellanox InfiniBand, Broadcom/Arista Ethernetcannot delay past periodcannot delay past periodCompetition — Ethernet (Broadcom, Arista, Ultra Ethernet Consortium) is a production-scale alternative; demonstrated at frontier scale.Routable to Ethernet now for a lab with fabric-engineering capacity · estimated
Cloud / datacenter capacity · AWS, Microsoft Azure, Google Cloud, CoreWeave, Oracle, neocloudsuncertaincannot delay past periodCompetition — Several providers sell dedicated cluster capacity and the labs examined use more than one; whether a lab could switch within eighteen months is uncertain.Switching depends on contractual commitments and capacity already arranged with alternative providers; no switching-time estimate is supplied. · estimated
Power / grid interconnection · ISO/RTO (MISO, PJM, ERCOT, SPP), local utility, state PUC, FERC, EPA · institutions acting togethercannot delay past perioduncertainPublic body — The holders are regulators (FERC, state public utility commissions, the EPA) and the system operators and utilities they oversee.New site: ~4.5yr average to commercial operation (transformer ~160wk OR interconnection study alone exceeds 18 months). On-site gas can substitute for a grid connection (the xAI Colossus route), so the new-site assessment is not applied to a lab already self-generating its power or running on capacity it has already arranged. · estimated
Capital ($1B+/yr runs) · hyperscaler balance sheets, sovereign funds (QIA, PIF, Mubadala), mega-VCuncertaincannot delay past periodCompetition — Syndicated funding rounds, including compute-for-equity arrangements, and participation by sovereign funds and infrastructure investors support the competition assessment. The effect of losing a funder within eighteen months remains uncertain.The cited funding rounds have multiple investors (Anthropic’s $30B Series G: 2 leads, 5 co-leads, ~30 named investors; OpenAI’s $122B round: 6 co-leads plus Amazon/NVIDIA/SoftBank/Microsoft). · estimated

Needs verification Three redeployment-time assessments lack a published source: EDA tools (chip design), advanced logic foundry (≤5nm/3nm), advanced packaging / CoWoS. Their assessments rest on reported capacity constraints or the absence of a qualified alternative; no switching time has been measured.

The same stack, three labs

The lab-specific counts use different assumptions about accelerator choice and existing power supply. For Google and Anthropic, the model allows alternatives to NVIDIA accelerators and NVLink/NVSwitch, which removes those two barriers. For all three labs, it assumes existing power arrangements avoid a new grid connection within eighteen months; the five-year assessment remains uncertain. The hardware descriptions below give the basis for each case.

7
xAI — 7 of 11 leverage points could delay this run beyond eighteen months, 2 beyond five years. NVIDIA H100 / H200 / GB200 NVL72, on Spectrum-X Ethernet scale-out. View assessments for this lab
5
Google — 5 of 11 leverage points could delay this run beyond eighteen months, 2 beyond five years. Google TPU v5e / v5p / v6e (Trillium); proprietary ICI interconnect; zero NVIDIA in the training path. This page uses this training setup.
5
Anthropic — 5 of 11 leverage points could delay this run beyond eighteen months, 2 beyond five years. AWS Trainium2 (>1M chips in production), alongside Google TPUs and NVIDIA GPUs. View assessments for this lab

How a verdict is made method 2026-09-26

1 · How assessments are counted

Each point has a recorded assessment of whether a lab could switch suppliers or infrastructure within eighteen months and within five years. Where a lab-specific assessment is recorded, it replaces the baseline; otherwise the baseline applies. The count sums the points marked can delay past period.

Switching-cost and redeployment-time notes explain these judgements, with their source quality shown beside them. Market share is context; the assessment turns on whether the lab could switch in time.

2 · What constrains the holder

Each layer has one accountability category, assigned in this order of priority: export control, antitrust, a public body, competition, then no identified constraint. The category stays the same across labs; the register gives the reason for each assignment.

A layer can be classified under competition and still delay a run if the lab cannot switch to the rival within eighteen months. Across the full register, 7 of 11 layers are classified under export control. Reach of each check, layer by layer.