56 supply bottlenecks, five sectors.

Three OEMs make nearly all heavy-duty gas turbines. Slot reservations stretch years out and set the pace for new firm generation.

Three suppliers make essentially all HBM. Each HBM bit consumes several times the wafer area of standard DRAM, crowding out conventional supply.

Interconnection queues measure in years. Energised megawatts — not empty shells — pace AI buildout, and new lines take the better part of a decade.

Large power transformers are hand-built, need grain-oriented steel and copper, and have years-long lead times. Switchgear followed.

China controls most rare-earth separation and magnet production and has used export licensing. Every motor, robot and drone depends on it.

One company produces almost all leading-edge AI logic. N3/N2 capacity is allocated years ahead and every new fab is a multi-year, tens-of-billions bet…

The scarce part of an 800G/1.6T link is often the laser — EML and CW lasers grown on indium-phosphide wafers that few fabs can make at volume.

An accelerator is useless until logic and HBM are integrated on an interposer. 2.5D/3D capacity, hybrid bonders and TC bonders gate output.

Small markets with outsized importance: gallium for GaN and InP, germanium for optics, antimony for defence. China has restricted exports.

Four US hyperscalers guided to roughly $700B+ of 2026 capex, with 2027 modelled near $1T. Their guides set demand for every supplier on this site; neo…

EUV is a single-vendor tool with a single-vendor optics chain. Mask inspection and photoresist are similarly concentrated in a handful of Japanese fir…

Electricians, pipefitters and commissioning crews are finite. Contractor backlogs are the human bottleneck behind every campus.

Cluster scale pushes past copper. 800G→1.6T pluggables ramp now; co-packaged optics shift value toward the switch and laser.

Inside the hall, every megawatt needs medium-voltage switchgear, UPS, busway and power distribution. A handful of makers supply most of it and backlog…

Western enrichment capacity is thin and Russian supply is being phased out. Conversion and enrichment, not mined uranium, are the tightest steps.

HVDC and extra-high-voltage cable comes from a handful of European and Japanese plants with order books reaching into the 2030s.

Rack densities above ~100kW make direct-to-chip liquid cooling the default. CDUs, cold plates and heat rejection scale with GPU power.

Grids, transformers, data centres and EVs all draw on copper. New mines take a decade or more; grades are falling.

Deployable AI compute tracks the minimum of logic, HBM and packaging. Merchant GPUs and hyperscaler ASICs compete for the same upstream capacity.

AI training data and inference logs need exabytes of storage. Two companies make nearly all hard drives and a few make NAND, so supply is slow to foll…

When the grid can't deliver in time, campuses build their own: recip engines, fuel cells and aeroderivatives as a bridge — at higher cost per MWh.

Bigger dies, more chiplets and HBM stacks multiply test time. SoC and memory testers become a quiet throughput gate for every accelerator shipped.

AI accelerator boards and switch trays need high-layer, low-loss PCBs and copper-clad laminate. Few makers qualify, and layer counts rise every genera…

Hyperscalers buy electrons with duration. Existing reactors, uprates and restarts are the only large firm carbon-free supply available this decade.

Each GPU needs a cold plate or vapor chamber, plus fans, manifolds and quick-disconnects. Taiwanese makers supply most of it and content per rack rise…

AI halls use many times the fiber of a conventional data centre. High-density fiber and connectors became capacity-constrained alongside optics.

Powered land, permits and local consent decide when megawatts become racks. Power access now matters more than steel and concrete.

Gate-all-around transistors, backside power and HBM stacking add process steps. More steps per wafer means more tool spend per unit of capacity.

100kW+ racks force 800V DC architectures, GaN/SiC conversion and new power shelves. Power density per rack is now a design constraint.

Large AI packages need big, many-layer ABF substrates. The build-up film itself comes from one supplier; glass-core substrates are the next transition…

Every robot joint needs a strain-wave or RV reducer. Two Japanese firms dominate; Chinese suppliers are scaling fast.

Every additional GPU in a training domain multiplies switch ports and retimers. Ethernet vs proprietary fabrics decides who captures it.

An AI server board uses many times the capacitors of a normal server, and the high-capacitance, high-temperature parts come mainly from Murata and Sam…

AI agents and AI-written attacks expand the attack surface faster than security teams can hire. Spend shifts to platforms that use AI to defend AI wor…

Autonomous systems are now central to defence budgets. Motors, batteries and sensors with non-China supply chains are scarce.

Robots and vehicles see through CMOS sensors and LiDAR. Dexterous manipulation still lacks good, cheap tactile sensing.

One provider dominates launch mass to orbit. Alternatives for medium-lift are scarce, which gates every constellation outside it.

Most new firm AI power in the US burns gas. Pipeline takeaway, not reserves, limits where it can be delivered.

Enterprises need clean data, orchestration and monitoring before AI agents do useful work. Data platforms and observability are the picks-and-shovels …

800V DC racks need silicon-carbide and gallium-nitride switches at every conversion step. Capacity is ample after the EV slowdown, so the constraint i…

Motors and drives are the muscle of automation, and rely on rare-earth magnets that come mostly from China.

Rack-scale systems are integrated by a few Taiwanese ODMs. Execution and liquid-cooling readiness matter more than raw capacity.

Driverless fleets are scaling city by city. Sensors, compute, mapping and permits gate the rollout.

Fabs need ultra-pure nitrogen, hydrogen, neon, helium and fluorine compounds on long contracts. Helium-3 for cryogenics is genuinely scarce.

Humanoids use planetary roller screws and precision bearings in quantity. Roller-screw capacity is small and specialised.

Direct-to-device, broadband and Earth observation compete for launch, spectrum and optical inter-satellite links.

Prime wafers and ultra-pure chemicals are concentrated, but capacity is ahead of non-AI demand. A latent rather than active constraint.

Running vision-language-action models on the robot needs power-efficient edge accelerators — a smaller cousin of the datacentre race.

Models now design molecules and antibodies, but clinical trials are still the gate. Value accrues to platforms with wet-lab data and pharma partnershi…

Cell capacity is abundant but concentrated in China. Tariffs and foreign-entity rules make non-China supply the scarce version.

Reshoring and labour scarcity push automation spend. Integrators and controls vendors are the existing, profitable expression of physical AI.

On-device AI drives a phone and PC upgrade cycle and pulls more memory per device — competing with data centers for the same DRAM.

Pilots and funding are real; paid serial deployments are early. Public exposure is mostly incumbents and Chinese platforms.

Small reactors promise factory-built firm power, but first-of-a-kind licensing, HALEU fuel and forging capacity push meaningful volume into the 2030s.

Qubits need dilution refrigerators, helium-3, precision lasers and control electronics — all small, specialised supply chains.

After the price bust, supply exceeds demand. Non-China graphite and processing remain the strategic gap.