AGI Bottlenecks
The AGI Bottlenecks table maps the supply chain of frontier capability progress: what the next generation of capability gains requires more of, how constrained each input is today, and who controls it. Rows are inputs (compute, fabs, energy, data, talent, capital, regulation); columns are the dimensions along which an input can bind: present tightness, direction of change, time-to-relieve, cost-to-expand, controllers, and geographic concentration.
The table is intended as a strategic map. Bottlenecks define where governance interventions have leverage — a binding constraint controlled by a single firm in a single country (HBM, EUV lithography) is a different policy target from a moderately tight constraint distributed across many actors (deployment infrastructure, capital). The companion Actor Power Scorecard scores who exerts power over the AI development trajectory; this table scores the inputs they have power over.
Columns:|
How constrained is this input right now? | Is the constraint tightening or loosening? | How fast can supply be expanded? | Capex required to relieve the constraint | Who controls | How geographically concentrated is supply? | Strategic notes | |
|---|---|---|---|---|---|---|---|
Fab capacity Leading-edge semiconductor fabrication (3nm and below) — the upstream constraint behind nearly all frontier AI accelerators. | Binding TSMC sub-3nm sold out 12+ months ahead; one credible supplier for frontier nodes | Stable TSMC Arizona, Samsung Texas, Intel Foundry expanding; multi-year ramps | Very Long Fab buildout 3-5 yr; equipment lead times 1-2 yr | Very High $20-40B per leading-edge fab; only 3 firms can plausibly attempt | TSMCSamsung FoundryIntel FoundryASML (EUV supply) | Single Point TSMC Taiwan dominant; ASML EUV monopoly. Geopolitical concentration risk. | The upstream chokepoint. US CHIPS Act and Taiwan tensions make this the bottleneck with the highest strategic salience. |
HBM / memory supply High-bandwidth memory (HBM3/HBM3E/HBM4) — paired with logic dies in every frontier accelerator. Memory bandwidth, not FLOPS, often gates throughput. | Binding SK Hynix HBM3E sold out into 2026; memory bandwidth gates accelerator throughput | Tightening Each chip generation increases HBM stack count; demand outpaces DRAM fab additions | Medium DRAM fab lead times shorter than logic; HBM packaging (CoWoS) is the bottleneck within the bottleneck | High Multi-billion DRAM and packaging capex; CoWoS capacity at TSMC is gating | SK HynixSamsungMicronTSMC (CoWoS packaging) | Concentrated South Korea dominant for HBM; Taiwan for advanced packaging | Often overlooked. The reason H100 and B200 supply is constrained is as much HBM and CoWoS as logic-die wafers. |
Compute (training) Frontier training compute — H100/B200-class clusters at 100K+ accelerator scale used for foundation-model training runs. | Tight Frontier training runs use 100K+ H100-equivalents; allocations contested across hyperscalers | Tightening Frontier training compute growing ~4-10x/yr; supply ramp lags demand | Long New fab capacity 2-4 yr; new GPU generations on 1-2 yr cycle | Very High Single frontier cluster $5-50B in capex | NVIDIATSMCHyperscalers (MSFT/AMZN/GOOG)Frontier labs | Concentrated Most frontier training compute in US data centers; UAE/Singapore growing | The headline bottleneck. NVIDIA-led GPU supply, hyperscaler cluster build-out, and TSMC fab capacity are the three sub-constraints. |
Compute (inference) Serving capacity for deployed models — accelerator-hours per token across cloud and edge deployment. | Tight Major releases regularly hit capacity limits; reasoning models multiply inference cost 10-100x | Tightening Demand growth (reasoning chains, agents) outpacing efficiency gains and capacity adds | Medium Inference is more parallelizable; custom silicon (Trainium, TPU, Groq) helping | High Substantial but lower per-FLOP than training; commoditizing | Hyperscalers (AWS/Azure/GCP)NVIDIACustom silicon (Trainium, TPU, Groq, Cerebras) | Distributed Many regions, but US/EU dominant; latency forces geographic distribution | Less single-point-of-failure than training. Inference is where most users feel capacity constraints (rate limits, queues). |
Energy / power Grid interconnect capacity, generation, and transmission — increasingly the rate-limiter for new datacenter siting. | Tight Multi-GW interconnects waitlisted 5-10 yr in major US grids; UK, Ireland, Netherlands at moratorium | Tightening Gigawatt-scale clusters proposed; grid build-out 5-10x slower than datacenter demand | Very Long Transmission line permitting 5-10 yr; new generation (nuclear, gas) similar | High Interconnect upgrades $100M-$1B per site; new generation in $B | UtilitiesFERC / state PUCsIndependent power producersLocal zoning authorities | Distributed Many viable regions globally, but per-site permitting is the constraint | Replacing compute as the headline bottleneck for next-generation clusters. Stargate, Microsoft-Constellation, and Meta nuclear deals all reflect this. |
High-quality training data Pretraining corpora — text, code, multimodal data of sufficient quality to drive frontier model improvements. | Tight Public web text approaching exhaustion at frontier scale; quality filtering aggressive | Mixed Tightening for novel human-generated data; loosening via synthetic data and licensing deals | Medium Synthetic generation pipelines, real-world telemetry, and licensed corpora maturing | Medium Licensing deals (NYT, Reddit, Shutterstock) in $10s-$100s of millions; labeling at scale | Frontier labs (proprietary pipelines)Major publishers (NYT, Reddit, Stack Exchange)Data brokers and labeling firmsSynthetic-data toolchains | Global Data sources are global; legal regimes (GDPR, copyright) regional | Synthetic-data viability is the open question. If synthetic scales, this bottleneck loosens; if not, real-data licensing becomes a moat. |
Evaluation data Benchmarks, held-out test sets, and dangerous-capability evals used to measure frontier-model progress and safety. | Tight Public benchmarks saturating; contamination widespread; few credible held-out evals | Mixed Public benchmarks tightening (saturation); private holdout sets from AISIs and labs emerging | Medium Eval suite development 6-18 months; held-out test creation iterative | Low Eval data cheap relative to training data; expert annotation is the cost driver | METRUS/UK AISILabs (private holdouts)Academic groups (BIG-Bench, HELM) | Global Eval development globally distributed; AISIs concentrated in US/UK/EU | A governance-critical bottleneck. Bad evals → labs can't credibly self-report risk → safety commitments lose teeth. |
Talent supply Frontier ML researchers, alignment researchers, and senior engineering staff capable of operating at the capability frontier. | Tight Top frontier-ML researchers ~1000s globally; alignment researchers ~100s; compensation reflects scarcity | Loosening PhD pipelines growing, but slowly; lab hiring frenzy bids up the senior tier | Very Long PhD pipelines 5-10 yr; senior engineering experience harder to manufacture | High Senior researcher comp packages reach $1M-$10M; aggressive cross-lab poaching | Top frontier labs (compete for talent)Universities and PhD programsGovernment immigration policy | Concentrated US Bay Area, UK, China dominate; visa policy is a material lever | A slow-moving bottleneck. Less acute than compute today, but the long pipeline makes shocks (e.g. visa restrictions) hard to recover from. |
Cooling / siting Water rights, land, climate suitability, and zoning for hyperscale datacenters. | Moderate Water-cooled designs constrained in arid regions; direct-liquid cooling reduces water demand | Stable Air and direct-liquid cooling retrofits expanding; sites still available outside drought regions | Medium Datacenter construction 1-2 yr once site and power are secured | Medium Site acquisition and construction in $100M-$1B range | HyperscalersLocal zoning boardsWater authorities | Distributed Many viable global sites; concentration follows power and fiber availability | Less acute than power but interacts with it — water-cooled efficiency is what makes some 100MW+ sites viable. |
Algorithmic insights Research breakthroughs in architectures, training procedures, post-training methods, and reasoning techniques. | Moderate Key insights diffuse fast via papers, code, and lab transitions; some methods stay private 6-18 months | Loosening Open-weight releases (DeepSeek, Llama, Mistral) and academic publications keep diffusion fast | Short Papers and code propagate in weeks once published | Low Research is talent-intensive, not capex-intensive | Frontier labs (closed research)Academic communityOpen-weight releasers (Meta, DeepSeek, Mistral, AI2) | Global Research community is global; lab concentration in US/UK/China | The least supply-constrained bottleneck. The question is whether closed labs can sustain a multi-month research lead. |
Capital Investment capital for frontier training runs, infrastructure buildout, and operating losses during scaling. | Moderate Frontier labs have access to $10s-$100s of billions; trailing labs more constrained | Loosening Sovereign wealth (UAE, Saudi), hyperscaler commitments, Stargate-scale infrastructure deals | Short Capital can be deployed in months when investor appetite exists | Low Capital is fungible; the question is access, not creation | Hyperscalers (MSFT, AMZN, GOOG)Sovereign wealth (UAE, Saudi, Singapore)Mega-cap VC and strategic investors | Concentrated US capital markets dominant; sovereign-wealth tier is the swing player | Capital is the least constrained input for top-3 labs. For everyone else it is the binding constraint. |
Deployment infrastructure APIs, enterprise integrations, agent frameworks, and tooling needed to put model capabilities in front of users. | Moderate Major API providers mature; enterprise integration and agent tooling still nascent | Loosening API ecosystems, agent frameworks (MCP, agent SDKs), and enterprise pipelines maturing | Short Engineering effort scales with capital; not capacity-bound | Low Engineering cost, not capex; standard SaaS economics | Major API providers (OpenAI, Anthropic, Google)Cloud platforms (AWS Bedrock, Azure OpenAI, Vertex)Enterprise integrators | Distributed API access global where not legally restricted | Less a supply bottleneck than a maturity bottleneck. Affects deployment timelines more than capability frontier. |
Regulatory friction Legal, regulatory, and compliance constraints — EU AI Act, US export controls, sectoral rules, liability frameworks. | Moderate EU AI Act and US export controls binding for some uses; not capability-gating for frontier training in the US | Tightening EU enforcement ramping; US export controls expanding; California SB-53 and sectoral rules emerging | Long Legislative cycles slow; rule-making takes years | N/A This is a constraint, not a resource to expand | EU CommissionUS BIS / White HouseChina State Council / CACCalifornia, NY, and state legislatures | Regional Brussels Effect plus US-China bifurcation creating regional regulatory regimes | The bottleneck most amenable to deliberate governance intervention. Tightening it (more rules) reduces risk but slows deployment; loosening it does the opposite. |
13 bottlenecks across 5 categories. Snapshot as of 2026-05; bottleneck dynamics change on quarter-to-year timescales.
What the table is not
This is a snapshot, not a forecast. The cells reflect mid-2026 conditions. Several of the rows have plausible scenarios in which their tightness flips within twelve months — HBM packaging capacity, training-compute interconnect availability, and US export-control scope are all moving fast. The footer in the table notes the snapshot date and update cadence.
The table also flattens several distinctions worth noting:
- Compute is treated as a single bottleneck, but the H100/H200/B200 generations are simultaneously binding (current-gen supply) and not (next-gen reflects scaling needs not yet at frontier). Cells reflect the present-generation gating constraint.
- Algorithmic insights are scored as relatively unconstrained because public-research diffusion is fast, but specific frontier-lab insights (reasoning-model training recipes, post-training procedures) stay private for 6–18 months and confer real capability advantage during that window.
- Regulatory friction is the only row where "cost to expand" does not apply — regulation is a constraint to bind, not a resource to expand. The cell is marked N/A.
- Geographic concentration is scored by where supply is produced, not where it is deployed. Compute deployment is global; compute production is concentrated.
Methodology
Tightness, trajectory, time-to-relieve, cost-to-expand, and geographic-concentration cells are subjective expert judgments synthesized from public reporting through mid-2026. They are not derived from a single quantitative model. For each cell, the inline note captures the specific facts driving the rating; readers should weight the note more heavily than the badge.
The notes column is intended for strategic framing — the single most important thing to know about each bottleneck that the structured columns do not capture.
Related
- Actor Power Scorecard — scores the actors exerting power over AI development; this table scores the inputs they have power over.
- AI-Driven Concentration of Power — analysis of how control over bottleneck inputs concentrates strategic leverage.
- Concentration of Power Systems Model — the causal model behind why bottleneck control matters.
Changelog
| Date | Change |
|---|---|
| 2026-05-11 | Initial table created; 13 bottlenecks scored across 5 categories |