Already serving workloads, researchers, organizations, or users.
What Is AI · Cross-Layer Snapshot
Major AI Investment Projects — July 2026
A layer-by-layer map of the factories, data centers, model laboratories, agent deployments, applications, and scientific programs building the AI economy.
The figures on this page are not directly comparable. Some are construction budgets, some are capital-expenditure plans, some are financing rounds, and others are contractual or milestone-based commitments. Many projects span more than one layer; they are placed where their main industrial contribution is easiest to understand.
Physical construction, manufacturing, or deployment has started.
A formal financing, contract, partnership, or investment program has been announced.
An early-stage plan, letter of intent, or engineering program whose full execution is uncertain.
1 ApplicationsAI products directly used by people, organizations, and governments
Layer 1 budgets are difficult to isolate because application spending is usually embedded inside broader infrastructure, cloud, manufacturing, and corporate programs.
Google AI Search
Google is rebuilding Search around AI Overviews, AI Mode, Gemini models, research tools, shopping, and agent-like task completion.
The application does not have a separately disclosed project budget. Its scale comes from global distribution and the infrastructure needed to serve very large numbers of searches.
Primary sources: Google: A new era for AI Search · Alphabet Q2 2026 remarks
Apple Intelligence and Private Cloud Compute
Apple’s Houston operation manufactures advanced Apple-designed servers that support Apple Intelligence and Private Cloud Compute. Apple reported in February 2026 that advanced AI servers were already shipping from Houston.
The broader $600 billion commitment includes far more than AI, so it should not be treated as the budget of Apple Intelligence alone.
Primary sources: Apple’s $600 billion U.S. commitment · Apple Houston manufacturing update
Alexa+ international deployment
Amazon rebuilt Alexa as a generative and personalized assistant designed to complete tasks across voice, web, mobile, shopping, smart-home devices, vehicles, and partner products.
By mid-2026, Alexa+ had expanded beyond the United States to multiple countries and languages.
Primary sources: Introducing Alexa+ · Alexa+ international rollout
Layer-level observation: The largest application projects are distinguished less by a disclosed standalone budget than by distribution through search, devices, operating systems, enterprise software, retail, and government channels.
2 AgentsModels combined with tools, memory, state, permissions, and execution
Agent investment is moving from chatbot experiments toward systems authorized to perform real work inside software, data, and organizational processes.
SoftBank–OpenAI Cristal Intelligence
SoftBank committed to spend approximately $3 billion annually deploying OpenAI technology across SoftBank Group companies. The partners also created SB OpenAI Japan.
The program is designed to connect models to enterprise documents, data, software, and workflows so agents can perform multistep business work.
Primary sources: SoftBank and OpenAI announcement
Salesforce Agentforce and the Agentic Enterprise
Salesforce announced a $15 billion, five-year San Francisco investment covering AI innovation, customer adoption, workforce development, and its Agentforce strategy.
Agentforce connects models with Salesforce data, permissions, CRM actions, sales, service, marketing, commerce, and internal operations.
Primary sources: Salesforce $15 billion investment · Agentforce 360
Gemini Enterprise agent platform
Google is connecting Gemini models to Workspace, Cloud, company data, developer systems, security controls, and business applications.
The visible agent platform sits on top of Google’s much larger TPU, cloud, data-center, networking, model-training, and inference programs.
Primary sources: Google I/O 2026
Layer-level observation: The difficult part is not model intelligence alone. Reliable agents require permissions, identity, tool integration, persistent state, evaluation, cybersecurity, human oversight, and responsibility for failed actions.
3 Domain ModelsSpecialized AI for transportation, medicine, government, finance, science, and other fields
Domain systems combine general models with specialized data, simulation, sensors, rules, experts, workflow integration, and real-world validation.
Waymo Driver and robotaxi expansion
Waymo raised $16 billion in February 2026 to expand autonomous ride-hailing, fleet operations, manufacturing, geographic coverage, and the Waymo Driver.
The project combines perception, prediction, simulation, mapping, planning, safety validation, vehicle control, and continuous fleet learning.
Primary sources: Waymo $16 billion financing · Waymo manufacturing expansion
Lilly TuneLab
Lilly created TuneLab to give biotechnology companies access to drug-discovery models trained on Lilly’s proprietary experimental data.
The project illustrates how domain advantage can come from validated data, scientific expertise, laboratory processes, and experimental feedback rather than only model size.
Primary sources: Lilly TuneLab announcement · TuneLab
OpenAI for Government
OpenAI’s government program includes work on administrative processes, healthcare access, acquisition analysis, and proactive cyber defense.
Government-domain systems require specialized security, authorization, procurement, data handling, auditability, and policy controls.
Primary sources: Introducing OpenAI for Government
Layer-level observation: The most defensible domain projects may be those that control unique data and can test outputs against physical, clinical, financial, legal, or operational reality.
4 Model ArchitectureFrontier research, pretraining, post-training, reasoning, multimodality, efficiency, evaluation, and safety
Financing rounds provide a rough measure of model-laboratory scale, but the money also supports infrastructure, staff, products, operations, safety, and distribution.
OpenAI frontier-model program
OpenAI announced in March 2026 that its latest round closed with $122 billion in committed capital.
A financing round is not the same thing as a single model-training budget; the capital spans research, infrastructure, products, and global deployment.
Primary sources: OpenAI financing announcement
Anthropic frontier research and Claude
Anthropic announced a $65 billion Series H financing round in May 2026.
The program supports frontier research, model development, safety and interpretability, products, enterprise deployment, and large multi-cloud compute commitments.
Primary sources: Anthropic Series H
SpaceXAI and Grok
xAI completed a $20 billion Series E before SpaceX acquired the company in February 2026.
The combined organization connects model research with Colossus, Grok distribution, SpaceX engineering, Starlink, and the proposed STARMIND orbital-compute program.
Primary sources: xAI Series E · xAI joins SpaceX
Mistral AI
Mistral AI raised €1.7 billion in September 2025 in a round led by ASML.
The financing supports scientific research, frontier and open models, enterprise systems, agents, training tools, and European-controlled deployment.
Primary sources: Mistral AI financing
Layer-level observation: Capital is concentrating around a small number of model laboratories, but model efficiency, open-weight releases, specialized systems, and new training methods can still alter the competitive map.
5 InfrastructureData centers, energy, cooling, networking, storage, cloud, training clusters, and inference systems
Layer 5 is receiving the largest visible capital flows because it requires land, buildings, electrical generation, grid connections, cooling, networking, storage, servers, and long-lived physical assets.
How to read the metrics:
“GPU / accelerator count” includes non-GPU AI chips such as AWS Trainium and Google TPUs.
A blank disclosure is shown as not disclosed rather than estimated. GW figures may
describe compute capacity, electrical supply, or contractual capacity, depending on the source.
The capital figures are not automatically additive.
Layer 5 comparison
| Project | GPU / accelerator count | GW or power capacity | Capital expenditure or commitment |
|---|---|---|---|
| Stargate | More than 2 million chips were specified for the first more-than-5-GW phase; the count for the now-secured capacity above 10 GW is not disclosed. | More than 10 GW secured as of April 2026. | $500 billion original four-year U.S. program. |
| SpaceXAI Colossus | More than 220,000 NVIDIA GPUs operating; 1 million GPUs stated as the Memphis objective. | More than 300 MW (0.3 GW) available through Colossus 1. | Not disclosed. |
| SpaceX STARMIND | Processor count not disclosed; AI1 uses an interchangeable compute payload. | AI1 concept: 120 kW average and 150 kW peak per satellite; total constellation capacity not committed. | Not disclosed. |
| Anthropic–AWS / Project Rainier | More than 1 million Trainium2 chips in use. | Nearly 1 GW by end-2026; agreement scales up to 5 GW. | More than $100 billion AWS technology commitment over ten years. |
| Google global infrastructure | TPU and third-party accelerator counts not disclosed. | Global GW capacity not disclosed. | $195–205 billion Alphabet 2026 CapEx guidance; not all of it is AI-only. |
| Meta Hyperion | GPU / accelerator count not disclosed. | 5 GW compute-capacity expansion in Richland Parish. | More than $50 billion regional investment. |
| Microsoft Pecos | GPU / accelerator count not disclosed. | Approximately 2 GW. | Multibillion-dollar campus; exact figure not disclosed. |
| Alibaba cloud and AI | GPU / accelerator count not disclosed. | GW capacity not disclosed. | At least RMB 380 billion (about $53 billion) over three years. |
| European AI Gigafactories | More than 100,000 advanced AI processors per Gigafactory; up to five facilities planned. | GW figure not disclosed. | €20 billion Gigafactory facility within a €200 billion InvestAI mobilization target. |
Project details
Stargate
GPU / accelerator count
More than 2 million chips for the first >5-GW phase; current >10-GW total not disclosed
Power capacity
More than 10 GW secured
Capital / commitment
$500 billion original U.S. program
OpenAI reported in April 2026 that Stargate had already surpassed its original 10-GW U.S. capacity objective. An earlier Oracle expansion specified more than two million chips across over five gigawatts under development.
The $500 billion figure is the original multi-year investment program. It should not be interpreted as money already spent, and the accelerator count for all capacity above 10 GW has not been published.
Primary sources: OpenAI: compute infrastructure for the Intelligence Age · OpenAI–Oracle 4.5-GW expansion · Original Stargate announcement
SpaceXAI Colossus — Memphis
GPU / accelerator count
More than 220,000 NVIDIA GPUs; 1 million-GPU objective
Power capacity
More than 300 MW (0.3 GW)
Capital / commitment
Not disclosed
Colossus 1 supports training, fine-tuning, inference, and high-performance computing. SpaceXAI reports more than 220,000 H100, H200, and GB200 GPUs.
Anthropic described access to more than 300 MW of Colossus 1 capacity. SpaceXAI’s Memphis page states an objective of equipping the facility with one million GPUs.
Primary sources: SpaceXAI compute partnership · SpaceXAI Memphis · Colossus
SpaceX STARMIND orbital AI compute
GPU / accelerator count
Not disclosed; interchangeable processor payload
Power capacity
AI1: 120 kW average / 150 kW peak per satellite; total not committed
Capital / commitment
Not disclosed
STARMIND proposes solar-powered orbital compute satellites connected through SpaceX communications systems. The AI1 concept describes one satellite-sized compute module rather than a terrestrial multi-gigawatt campus.
No total processor count, constellation power commitment, or capital budget has been publicly committed. It should therefore remain labeled proposed.
Primary sources: SpaceX STARMIND · SpaceXAI–Anthropic compute announcement
Anthropic–AWS Project Rainier and 5-GW expansion
GPU / accelerator count
More than 1 million Trainium2 chips currently in use
Power capacity
Nearly 1 GW by end-2026; up to 5 GW under agreement
Capital / commitment
More than $100 billion over ten years
Anthropic and Amazon expanded their compute partnership in April 2026. Anthropic stated that it was already using more than one million Trainium2 chips to train and serve Claude.
The agreement secures up to five gigawatts of AWS capacity and commits more than $100 billion to AWS technologies over ten years. These figures cover Project Rainier and later generations of AWS infrastructure, not one building.
Primary sources: Anthropic–Amazon 5-GW agreement · AWS Trainium and Project Rainier
Google global AI infrastructure
GPU / accelerator count
TPU and third-party accelerator counts not disclosed
Power capacity
Global GW capacity not disclosed
Capital / commitment
$195–205 billion 2026 Alphabet CapEx guidance
Alphabet raised its full-year 2026 capital-expenditure guidance to $195–205 billion after spending $44.9 billion in the second quarter.
Alphabet said the vast majority of Q2 capital spending supported technical infrastructure for AI. Approximately 60% of technical-infrastructure spending went to servers and 40% to data centers and networking. The company does not publish a consolidated accelerator count or global GW figure.
Primary sources: Alphabet Q2 2026 earnings call
Meta Hyperion — Richland Parish, Louisiana
GPU / accelerator count
GPU / accelerator count not disclosed
Power capacity
5 GW compute-capacity expansion
Capital / commitment
More than $50 billion regional investment
Meta announced in July 2026 that the Richland Parish project would expand to five gigawatts of compute capacity. The site is intended to house Hyperion, Meta’s largest AI training cluster.
Meta described the expanded regional investment as more than $50 billion. The company has not disclosed a GPU or accelerator count.
Primary sources: Meta Hyperion 5-GW expansion · Hyperion joint venture
Microsoft Pecos, Texas data-center campus
GPU / accelerator count
GPU / accelerator count not disclosed
Power capacity
Approximately 2 GW
Capital / commitment
Multibillion-dollar investment; exact amount not disclosed
Microsoft announced the Pecos campus in June 2026 as one of the largest single capacity additions in its history.
The company plans approximately two gigawatts of new capacity over five to seven years and says it will fund the dedicated generation and infrastructure needed for the campus.
Primary sources: Microsoft Pecos announcement
Alibaba cloud and AI infrastructure
GPU / accelerator count
GPU / accelerator count not disclosed
Power capacity
GW capacity not disclosed
Capital / commitment
At least RMB 380 billion (about $53 billion) over three years
Alibaba’s program supports cloud computing, model training, inference, storage, networking, the Qwen ecosystem, and enterprise AI services.
Alibaba has disclosed the capital commitment but not a consolidated accelerator count or electrical-capacity target.
Primary sources: Alibaba RMB 380 billion program
European InvestAI and AI Gigafactories
GPU / accelerator count
More than 100,000 advanced AI processors per Gigafactory; up to five planned
Power capacity
GW capacity not disclosed
Capital / commitment
€20 billion facility within €200 billion targeted mobilization
The European Union plans up to five AI Gigafactories for training and serving next-generation models. Each Gigafactory is described as containing more than 100,000 advanced AI processors.
The Commission has not published a common GW figure. The €20 billion Gigafactory facility is part of the broader €200 billion InvestAI mobilization target and should not be read as fully spent public money.
Primary sources: European AI Factories and Gigafactories · InvestAI announcement
Layer-level observation:
Training and inference are becoming partially separate industries. Training creates or improves
model weights; inference runs those models whenever a person, application, robot, or agent
requests intelligence. The most useful comparison therefore requires all three dimensions:
accelerator count, power capacity, and capital.
6 HardwareSemiconductor equipment, chips, foundries, memory, packaging, servers, power equipment, and cooling
Layer 6 investments overlap heavily with infrastructure. A chip purchase may appear inside a data-center budget, and a foundry program supports many model companies simultaneously.
SK Group–NVIDIA AI factories and memory initiative
SK Group and NVIDIA announced letters of intent covering a two-gigawatt SK Telecom AI factory and a long-term NVIDIA–SK hynix partnership for next-generation AI memory.
Because the agreement is based on letters of intent, the headline figure is an intended long-term partnership rather than money already spent.
Primary sources: NVIDIA–SK announcement
TSMC Arizona
TSMC’s Arizona program has expanded from its earlier $165 billion plan to a current stated $265 billion program.
Current plans include six semiconductor logic wafer fabs, two advanced-packaging facilities, and an R&D center. The first fab entered high-volume production in late 2024.
Primary sources: TSMC Arizona
Micron U.S. memory expansion
Micron is expanding memory manufacturing and research in New York, Idaho, and Virginia, including advanced DRAM and high-bandwidth memory capabilities.
Memory is a central AI bottleneck because models require enormous movement of weights, activations, training data, and checkpoints.
Primary sources: Micron U.S. expansion · Micron July 2026 milestone
OpenAI–NVIDIA systems partnership
OpenAI and NVIDIA announced a strategic program to deploy at least 10 gigawatts of NVIDIA systems, representing millions of GPUs.
NVIDIA stated that it could invest up to $100 billion progressively as gigawatts are deployed. The first phase was targeted for the second half of 2026.
Primary sources: OpenAI–NVIDIA partnership
OpenAI-designed Broadcom accelerators
OpenAI and Broadcom are co-developing custom accelerators, Ethernet networking, racks, and systems targeted for deployment from the second half of 2026 through 2029.
The project shows a model developer moving deeper into hardware and system architecture.
Primary sources: OpenAI–Broadcom collaboration
OpenAI–AMD deployment
OpenAI and AMD signed a multi-year agreement covering several generations of AMD Instinct accelerators, beginning with a planned one-gigawatt deployment in the second half of 2026.
Primary sources: OpenAI–AMD partnership
Layer-level observation: The bottleneck is moving from GPUs alone to complete systems: lithography, logic, HBM, packaging, networking, power delivery, cooling, racks, firmware, compilers, and reliable manufacturing capacity.
7 Scientific AIModels connected to simulation, laboratories, instruments, experiments, and physical validation
Scientific AI is not merely text generation. It joins models with scientific data, supercomputers, domain knowledge, laboratory automation, experiments, and repeated validation.
U.S. Department of Energy Genesis Mission
The Genesis Mission connects Department of Energy supercomputers, AI systems, scientific datasets, laboratories, instruments, and research teams.
Its goal is to accelerate work in energy, materials, national security, engineering, and fundamental science.
Primary sources: DOE Genesis Mission · Genesis demonstration
NVIDIA–Lilly Co-Innovation AI Lab
NVIDIA and Eli Lilly committed up to $1 billion for researchers, computing infrastructure, models, laboratory systems, and data.
The program links biomedical models, scientific agents, autonomous laboratories, robotics, drug discovery, and pharmaceutical manufacturing.
Primary sources: NVIDIA–Lilly AI lab
Isomorphic Labs
Isomorphic Labs raised $600 million to expand its AI drug-design platform and advance therapeutic programs.
Its pharmaceutical collaborations combine model development, molecular design, laboratory work, drug-development milestones, and potential royalties.
Primary sources: Isomorphic Labs financing · Pharmaceutical collaborations
DOE–Oracle–NVIDIA Solstice
Solstice is a planned scientific AI supercomputer for Department of Energy research in energy, materials, simulation, engineering, and national laboratories.
A dedicated public budget was not disclosed in the cited announcement.
Primary sources: NVIDIA scientific AI infrastructure
Layer-level observation: Scientific AI receives less visible capital than general infrastructure today, but successful discoveries in medicine, energy, materials, manufacturing, agriculture, and engineering could produce unusually large long-term economic value.
Projects that connect all seven layers
OpenAI ecosystem
- Layer 6: NVIDIA, AMD, and Broadcom systems.
- Layer 5: Stargate and multi-cloud infrastructure.
- Layer 4: frontier-model research and post-training.
- Layer 3: government, health, finance, science, and enterprise adaptations.
- Layer 2: agents, tools, memory, and workflows.
- Layer 1: ChatGPT and embedded applications.
- Layer 7: scientific computing and health initiatives.
SpaceXAI ecosystem
- Layer 6: GPU systems and proposed satellite hardware.
- Layer 5: Colossus and proposed STARMIND orbital compute.
- Layer 4: Grok model research.
- Layer 3: coding, engineering, government, and scientific systems.
- Layer 2: tool-using and multi-agent products.
- Layer 1: Grok, X integration, APIs, and enterprise products.
- Layer 7: scientific simulation and discovery ambitions.
Lilly–NVIDIA ecosystem
- Layer 6: accelerator systems.
- Layer 5: biomedical compute infrastructure.
- Layer 4: scientific foundation models.
- Layer 3: drug-discovery models and TuneLab.
- Layer 2: scientific agents and laboratory workflows.
- Layer 1: tools for researchers and biotechnology companies.
- Layer 7: medicine discovery and physical validation.
What these projects reveal
Infrastructure absorbs the most visible capital
Data centers, power, chips, memory, networking, and cooling require physical construction and generate the largest public numbers.
Complete systems matter more than chips alone
Accelerators are useful only when combined with memory, packaging, power, cooling, networking, storage, software, and skilled operation.
Training and inference are diverging
Training creates or improves model weights. Inference runs those models every time a user, application, robot, or agent requests intelligence.
Model companies are moving into hardware
Custom accelerators, networking, racks, and vertically integrated systems are becoming strategic capabilities for leading model developers.
Governments view compute as strategic infrastructure
National and regional programs increasingly treat AI capacity like energy, telecommunications, transportation, research, or defense infrastructure.
Scientific AI may create the deepest long-term value
Discoveries in medicine, energy, materials, agriculture, and engineering could ultimately outweigh the value of conversational applications alone.
Editorial and source policy
This is a dated snapshot, not a permanent ranking. Project amounts use the currencies and units stated by the organizations. Financing rounds, capital expenditure, contract ceilings, processor counts, gigawatts, and letters of intent are deliberately kept distinct.
Primary corporate or government sources are linked under each project. Proposed projects should be revisited regularly because financing, construction, capacity, ownership, and timelines can change quickly.