Who is building it, layer by layer.
198 companies and institutions across the buildout, from the power plant to the proof checker. Point at a name, or tap it, for the company's most recent dated move since March 11, 2026, with its source.
Power · 16
Generation, interconnection and the grid that feeds the campuses.
Largest gas-turbine supplier; turbine slots are the near-term bottleneck for behind-the-meter AI campuses.
Largest US nuclear operator and counterparty to Microsoft and Meta nuclear PPAs.
Largest competitive generator with nuclear and gas fleet contracted to data centers.
Largest US renewables and nuclear developer (Duane Arnold restart for Google).
Susquehanna nuclear supply to Amazon's Pennsylvania campus.
Gas turbines, grid gear and transformers for data-center power.
Fuel-cell on-site power for Oracle/Brookfield AI campuses.
Backup and prime-power generator sets for data centers.
Fast-reactor developer with data-center offtake LOIs.
Natrium reactor developer backed by Gates and NVIDIA-linked capital.
Molten-salt SMR developer under Google's 500MW agreement.
Amazon-backed HTGR developer.
Only NRC-certified SMR design; ENTRA1/TVA pipeline.
Enhanced geothermal for Google and other data-center offtakers.
Best-funded fusion developer; Google offtake for ARC.
Fusion developer with Microsoft offtake (Orion).
Data centers · 15
Land, shells, cooling and the campuses themselves.
Leading thermal and power-distribution supplier for liquid-cooled GPU racks.
Largest colocation operator; xScale JV builds hyperscale AI capacity.
Largest data-center REIT by hyperscale leasing.
Power and cooling systems supplier for AI data centers (incl. Motivair liquid cooling).
Electrical distribution and switchgear for hyperscale campuses.
Multi-gigawatt hyperscale developer (Frontier campus, Texas).
Blackstone's hyperscale developer with the largest US land/power bank.
Hyperscale developer acquired by the BlackRock/MGX AI Infrastructure Partnership.
Blue Owl/IPI-backed hyperscale developer across US, Europe, APAC.
Polaris Forge campuses leased to CoreWeave and hyperscalers.
Bitcoin miner turned AI cloud with Microsoft capacity contracts (Sweetwater, Childress).
CoreWeave's largest third-party HPC landlord.
Fluidstack/Google-backed HPC leases at Lake Mariner.
AWS and Fluidstack HPC leases converting mining sites to AI.
Fluidstack/Anthropic-backed River Bend campus.
Chips and systems · 30
Accelerators, memory, fabrication and the racks they ship in.
Supplies the accelerators, networking and rack systems that nearly every frontier training run and inference fleet depends on.
Sole-source foundry for every leading AI accelerator (NVIDIA, AMD, Google, Broadcom, Amazon designs) and the CoWoS packaging that gates their supply.
Monopoly supplier of EUV lithography; its tool capacity bounds how fast leading-edge logic and HBM supply can grow.
Designs the custom accelerators (Google TPU, Meta, OpenAI) and Ethernet fabric that give hyperscalers an alternative to NVIDIA.
Leading HBM supplier (~half of HBM4 share); memory bandwidth is the binding constraint on accelerator throughput.
The only merchant GPU alternative at rack scale, with multi-gigawatt commitments from OpenAI, Meta and Anthropic.
Second HBM supplier, a 2nm foundry alternative to TSMC, and a strategic investor across the stack (Mistral, Rebellions, Tenstorrent).
The only US HBM maker; third leg of the memory supply that bounds accelerator output.
si-infra analysis: Micron Books a $54 Billion Quarter and Sells Most of Its 2027 HBM in Advance
China's primary domestic accelerator supplier; its SuperPoD roadmap determines how far Chinese labs can scale without NVIDIA.
Largest wafer-fab equipment maker; deposition/etch/inspection capacity gates leading-edge logic and HBM output.
Etch and deposition leader whose tools are critical for HBM stacking and gate-all-around logic.
Second-largest equipment maker, dominant in coater/developers used with every EUV scanner.
Process-control monopoly-in-practice; yield at 2nm and in HBM depends on its inspection tools.
The only potential second leading-edge foundry in the West; 18A/14A success would de-risk the TSMC single point of failure.
Second custom-silicon house for hyperscalers; now attached across Google's TPU ecosystem.
Wafer-scale alternative to GPUs with major OpenAI and G42 inference contracts.
CPU architecture under Grace, Graviton and Axion; now selling its own data-center silicon.
Dataflow inference chips; survived stalled Intel acquisition to raise at $11B.
LPU low-latency inference architecture, licensed to NVIDIA in a $20B 'reverse acquihire' (Dec 2025) and now shipping as NVIDIA Groq 3.
Transformer-only ASIC (Sohu) with $1B+ signed contracts and first racks due 2026.
RISC-V AI accelerator company and Qualcomm/Intel acquisition target.
China's leading listed AI-chip startup, targeting 500k accelerators in 2026 despite SMIC yield limits.
China's only advanced foundry; its 7/5nm-class capacity caps Huawei and Cambricon output.
Entering rack-scale inference (AI200/AI250) with Humain as first customer and circling Tenstorrent.
Connectivity (PCIe/CXL/scale-up switching) inside GPU racks.
Korea's leading AI-inference chip startup (Rebel100), backed by Samsung and SK hynix.
One of China's 'four GPU dragons'; HK-listed Jan 2026.
STAR-listed Chinese GPU maker.
Korean inference-chip startup (RNGD) that rejected a Meta acquisition.
NAND supplier for AI storage tiers.
Networks and optics · 15
The fabric that lets many sites act as one machine.
Supplies the optical fiber and connectivity every AI campus is wired with; US capacity is the binding constraint.
Largest networking vendor; its hyperscaler AI orders are a direct read on cluster buildout pace.
Lasers and 800G/1.6T transceivers for scale-out fabrics; its ramp tracks optical bandwidth per GPU.
Optical line systems and DCI that stitch multi-site clusters together (scale-across).
Transceivers, optical circuit switches and CPO for AI clusters; Nvidia-backed.
PCIe/CXL/NVLink-Fusion scale-up connectivity inside the rack.
Active electrical cables and optical DSPs linking GPUs within rows.
Leading Ethernet switch vendor for AI back-end (scale-out) networks at hyperscalers.
Optical I/O chiplets backed by Nvidia and AMD for scale-up beyond the rack.
Co-packaged optics inside Nvidia's NVLink Fusion ecosystem.
Largest independent US long-haul fiber builder for inter-campus AI traffic.
Private Connectivity Fabric supplier to the labs and hyperscalers (about $13B of contracts, including Anthropic).
Nvidia-backed telecom vendor putting GPUs into the radio access network.
Photonic Fabric scale-up interconnect, absorbed by Marvell in February 2026.
Dedicated-fiber network for blockchains carrying a majority of Solana stake; marginal to compute but relevant to agent-payment rails.
Hyperscalers · 5
The clouds that rent the compute at the largest scale.
Largest AI cloud by revenue, OpenAI's primary partner, and now a frontier-model builder in its own right via Microsoft AI.
Runs the TPU fleet and Google Cloud, the only hyperscaler with a vertically integrated chip-to-frontier-model stack.
Largest cloud, Anthropic's primary training partner (Project Rainier) and builder of the Trainium accelerator line.
Primary Stargate landlord for OpenAI with the largest contracted AI backlog of any cloud.
Stargate co-founder and OpenAI's largest investor.
Neoclouds · 9
GPU-first clouds built for training and inference.
Largest independent GPU cloud and a primary capacity source for OpenAI, Microsoft and Meta.
Developer-operator of the Abilene Stargate campus and other gigawatt-class AI factories with self-supplied power.
Fast-scaling GPU cloud with multi-year Microsoft and Meta commitments and an early Vera Rubin deployment.
Builder of Stargate UAE (1GW) and the UAE-US AI campus (5GW target).
PIF-backed national AI company with NVIDIA, AMD and Qualcomm supply deals and a 1.9GW pipeline.
Builder of Anthropic's $50B US data-center program and Google's TPU sites.
NVIDIA-backed GPU cloud heading to public markets.
Europe's largest sovereign-AI neocloud (Narvik, UK) with Microsoft as anchor tenant.
Leading open-model inference cloud; a barometer of open-weight adoption.
Frontier labs · 25
The builders of the most capable models.
Frontier lab with the largest consumer footprint and the Stargate compute program; its model cadence sets the pace for the field.
si-infra analysis: Rob Bonta Subpoenas OpenAI Over Cyber Risks From Testing to Release
Frontier lab whose Mythos-class models set the current capability ceiling and whose safety gating decisions shape how frontier capability is released.
Alphabet's frontier lab, with in-house TPUs and a near-monthly Gemini release cadence spanning reasoning, robotics and science.
Meta Superintelligence Labs plus the Prometheus/Hyperion multi-gigawatt clusters make it one of the four largest compute buyers and a frontier contender.
Operates Colossus 2, the largest single training cluster disclosed, and is scaling Grok toward a multi-trillion-parameter Grok 5.
Sets the efficiency frontier for open-weight models and forces price and architecture shifts across the field.
si-infra analysis: DeepSeek Open-Sources Huawei Ascend Kernels and Claims 99.8% of Chip Peak
China's largest cloud and most-used open-weight model family (Qwen).
Released the largest open-weight model to date, narrowing the gap to closed frontier models.
China's largest AI spender with the most-used assistant (Doubao, 345M MAU) and in-house chips co-designed with TSMC.
Microsoft's in-house superintelligence effort, freed from OpenAI exclusivity to train frontier models from scratch.
The one lab pursuing superintelligence as a 'straight shot' with no products; its compute access is the tell for progress.
Europe's frontier lab and the anchor of the EU sovereign-compute effort.
First publicly listed Chinese frontier lab; GLM-5 family is a top open-weight coding/agent model.
Open-weight lab combining frontier coding with 1M context and native video input.
Hunyuan models plus the largest Chinese consumer distribution; capex nearly tripled in Q2.
ERNIE models and the Kunlun chip line, one of China's few full-stack AI players.
Well-funded ex-OpenAI research lab with billions in NVIDIA/Google compute commitments.
The US-backed open-weights frontier bet, now with multi-billion compute deals but no released model.
Enterprise/sovereign model lab; consolidating Europe via Aleph Alpha.
Korea's sovereign-model builder (HyperCLOVA X) scaling to gigawatt-class NVIDIA infrastructure.
EXAONE open-weight models and AI-for-science (materials discovery).
Japan's most valuable AI lab, Google- and NVIDIA-backed, focused on efficient post-training methods.
India's government-selected sovereign model builder.
Shanghai multimodal frontier lab in the second tier of Chinese model makers.
Korean sovereign-LLM contender (Solar).
Agent runtime · 21
The software that lets models run long, unsupervised work.
Enterprise data and agent platform where much of the world's proprietary training and tool-use data lives.
One of the largest independent inference clouds for open models (40T tokens/day), setting the price floor for agents.
Inference platform whose run-rate tripled in a quarter; a proxy for agentic inference demand.
Serverless GPU runtime widely used for agent sandboxes, evals and RL environments.
Payments rail that lets agents transact; co-author of the Machine Payments Protocol.
x402 is becoming the default machine-to-machine payment standard, with settlement on Base.
Edge network and Workers runtime that agents traverse and run on; co-founder of the x402 Foundation; owner of Replicate.
Governed enterprise data plane now shipping Claude-powered agents (CoWork, CoCo) to 20,000+ customers.
Positioning Solana as the settlement chain for agent payments (15M agent payments claimed).
AI gateway acting as a control plane for agent traffic, acquired to secure agents at scale.
Open-source vector database for agent retrieval.
MemGPT lineage; stateful-agent runtime and memory research.
Most-downloaded agent framework (90M monthly with LangGraph); LangSmith traces 12x YoY.
Managed vector database; reportedly explored a sale in 2025.
Open-source embedding database with 15M monthly downloads.
Memory layer API for agents (186M calls/quarter by late 2025).
Temporal knowledge-graph memory (Graphiti) used by AWS and Samsung.
AI SDK (3M weekly downloads), AI Gateway and Sandbox; x402 Foundation member.
Model-deployment marketplace folded into Cloudflare Workers AI in December 2025.
Stripe-linked payments chain co-authoring the Machine Payments Protocol.
USDC is the settlement asset in x402, Pay.sh and Bedrock AgentCore Payments; x402 Foundation core member.
Verification · 14
Proof of what ran, on what, with what result.
Its time-horizon metric is the most cited external gauge of autonomous capability growth and it now audits internal agent use at the labs.
si-infra analysis: METR's Chris Painter Tells Senators Claude Leads 26% of Anthropic's AI Research
Principal external evaluator of scheming and deception in frontier checkpoints; now a public-benefit corporation.
Largest security vendor's AI unit, built on Protect AI and Portkey, now selling gated frontier models for defense.
Runtime security for agents and coding harnesses; largest pure-play AI-security round of the window.
Red-teaming arena and guardrails used by every major frontier lab.
Frontier cyber-capability evaluations used by OpenAI, Anthropic and DeepMind.
Prompt-injection defense now Check Point's AI-security R&D center.
Blind-compute network (MPC/FHE/TEE) for private AI inference and RAG; ERC-8004 agent verification.
TEE-based verifiable execution research (BuilderNet, attestation) that transfers to verifiable AI inference.
GPU-TEE confidential inference cloud with real-time attestation.
General-purpose zkVM and proof market; the substrate most zkML efforts target.
Leading open-source zkML prover (ONNX to EVM proofs).
On-chain proof systems for AI inference (models to ~18M parameters).
Verifiable on-chain agents for capital allocation; activity fell sharply mid-2026.
Evaluation and data · 9
The measurers and the data suppliers.
The default open-weights repository and community; its owner now shapes how open models are distributed and on what silicon.
Keeps a human-vs-machine general-intelligence yardstick that frontier models have not yet closed.
Once the default frontier data vendor; now 49% Meta-owned, consolidated around defense and Meta.
Expert-labor marketplace supplying RL and evaluation data to the frontier labs.
Turned a college-recruiting network into a ~$1B gross-revenue expert data business for the labs.
Largest profitable RLHF/data vendor to the labs (>$1B revenue).
Human-intelligence supplier to the labs (~$300M revenue, $2.2B valuation).
Reference source for compute, chip-stock and training-run trends (3.4x/yr AI chip compute).
Crowd-preference leaderboard that labs optimize toward; $1.7B valuation in January.
Robotics · 9
Models acting in the physical world.
Highest-volume humanoid/quadruped maker; the public-market benchmark for China's embodied-AI push.
Only automaker converting a car line to humanoid production with a stated million-unit target.
Highest-valued humanoid startup, now buying frontier-scale compute for its own robot foundation models.
Leading robot foundation-model lab (pi-series VLA models) licensing brains to hardware makers.
First company shipping home humanoids at scale from a vertically integrated US factory.
Most-deployed US warehouse humanoid (Digit), heading to public markets.
Hyundai-owned Atlas program with a 25,000-unit factory plan and a DeepMind model partnership.
SoftBank/Nvidia-backed universal robot brain at $14B+.
Boston Dynamics' owner and first mass customer (25,000 Atlas units planned) with Mobis supplying actuators.
Capital · 21
The money: closed, committed or announced.
Abu Dhabi's AI-only investor is an equity funder of Stargate and a co-lead in OpenAI, Anthropic and xAI rounds.
Financial lead of Stargate and the largest single outside funder of OpenAI and of robotics (Skild).
Largest private-credit financier of hyperscale AI campuses (Meta Hyperion, OpenAI/Oracle Jupiter, IREN).
Runs the $100B AI infrastructure program (BAIIF) spanning land, power, data centers and compute.
Largest real-asset owner in data centers; now financing non-Nvidia (TPU) compute capacity directly.
Record infrastructure fund plus a purpose-built AI campus developer with secured power.
Backer of OpenAI, Anthropic and xAI; its capital allocation signals where frontier-adjacent money flows.
OpenAI's most concentrated venture backer; now operating an OpenAI-linked holding company.
Concentrated late-stage backer of Anthropic, OpenAI, Anduril and Physical Intelligence.
Sovereign investor in both OpenAI and Anthropic, now tripling AI allocation.
Largest venture franchise by AUM; raised $15B in January and a $2.2B crypto fund in-window.
Crossover investor that led Databricks' $190B round and launched a data-center land venture.
Early backer of Anthropic, Mistral, Skild and Fireworks with $9B of fresh 2025 funds.
Private-credit lender to compute (xAI/Valor, Broadcom XPV) with $40B+ deployed in next-gen infrastructure.
Crossover fund active in late-stage AI rounds.
Sovereign full-stack AI company with a 6 GW data-center plan and Aramco as minority holder.
Foundational shareholder of MGX and G42; Abu Dhabi's diversified AI investor.
Anchor LP in both Brookfield's BAIIF and KKR's Helix, making it a quiet funder of gigawatt-scale campuses.
Growth investor concentrating on agent runtime (co-led Modal's Series C).
Led the largest inference-platform round of the window and backs Thrive Holdings.
OpenAI-partnered holding company deploying frontier models into acquired services firms.
Government · 9
The rules and the public buyers.
The US government's pre-deployment tester of frontier models and convener of agent-interoperability standards.
Best-resourced state evaluator; its findings on jailbreaks and cyber capability set the public record.
Controls federal sites, national-lab supercomputers and grid interconnection policy for AI campuses.
Enforces GPAI obligations under the AI Act for every frontier model sold in Europe.
Direct China's computing-power buildout, domestic-chip mandates and data-center siting.
Host of Stargate UAE with approvals for tens of thousands of Blackwell GB300s.
State sponsor of HUMAIN and chip-import approvals for a 6 GW sovereign buildout.
Funds Rapidus, GENIAC compute and sovereign-model programs.
Subsidized national GPU pool and sovereign-model program for the largest developer population.