Understand the system
How it works
The intelligence buildout begins upstream of any chip: a chain of uncertain workload forecasts translated into expensive, long-lived commitments.
From demand to financed capacity
Select a capability to explore its role and connections.
Selected: AI workload demand
AI workload demand
Training, inference, latency, and product demand translated into capacity requirements.
Explore AI workload demand in AtlasDocumented connections
- AI workload demand → Project capital
Forecast demand unlocks financing and procurement commitments.
Selected documented dependencies. Arrows retain their Atlas direction; they do not represent quantities or a complete engineering process.
AI demand arrives as tokens, training runs, latency targets, and product forecasts, but the supply chain must answer in factories, megawatts, and years. This chapter follows the conversion from a software growth thesis into purchase commitments, project finance, construction schedules, and risk allocation—and asks who absorbs the cost when forecasts move faster than infrastructure.
Demand is a stack of forecasts
Training, inference, fine-tuning, and retrieval create different utilization and latency profiles. A credible plan separates them before converting accelerator counts into rack, site, and grid demand.
- Separate peak reservations from realized utilization
- Track training and inference as distinct load shapes
Supporting evidence · 4
- [1] Data centers used about 415 TWh of electricity worldwide in 2024, around 1.5 percent of global electricity consumption.
- [2] The IEA estimates that global data-center electricity consumption rose 17 percent in 2025, with AI-focused data centers growing faster.
- [3] The IEA's 2025 base case projects worldwide data-center electricity demand to more than double to about 945 TWh by 2030.
- [4] LBNL's 2025 Update places U.S. data-center electricity use at 649 TWh in its 2030 Reference Case, with compounded-uncertainty bounds of 521 to 843 TWh, or 9.5 to 15.3 percent of U.S. electricity use.
Capital crosses mismatched clocks
Software roadmaps turn in months, while fabs, transmission, and power plants can take years. Financing structures bridge that mismatch through reservations, take-or-pay contracts, phased campuses, and portfolio diversification. The major-platform disclosures are intentionally presented as infrastructure-adjacent capital measures—not as a normalized AI-only league table—because Microsoft reports a fiscal-year property-and-equipment measure while Alphabet, Meta, and Amazon report calendar-year capital measures with different boundaries.
- Price schedule and utilization risk explicitly
- Stage commitments around real delivery gates
Supporting evidence · 6
- [1] Global investment in data centers reached roughly half a trillion U.S. dollars in 2024, nearly double the 2022 level.
- [2] Intel characterizes a modern fab as an investment of roughly 10 billion dollars that can take three to five years to build.
- [3] Microsoft reported $64.551 billion of property and equipment additions in fiscal 2025, compared with $44.477 billion in fiscal 2024; its fiscal year ended June 30 and the measure is not AI-only.
- [4] Alphabet reported $91.4 billion of capital expenditure in 2025, primarily for technical infrastructure including servers, networking equipment, data-center land, and construction; the amount is not AI-only.
- [5] Meta reported $72.22 billion of 2025 capital expenditure including finance-lease principal, principally for servers, data centers, and network infrastructure; cash property-and-equipment purchases were $69.69 billion.
- [6] Amazon reported $128.3 billion of cash capital expenditure in 2025, primarily for technology infrastructure—mostly supporting AWS growth—and fulfillment capacity; the amount is not data-center-only or AI-only.
Procurement is capacity strategy
Long-lead purchase agreements for accelerators, HBM, optics, switchgear, transformers, and cooling equipment can determine deployment pace as much as access to land or cash. DOE's distribution-transformer evidence shows how a constrained enabling component can expand a project schedule long before racks arrive.
- Map every reservation to a physical dependency
- Avoid counting the same constrained capacity twice
Supporting evidence · 2
- [1] Semiconductor design, equipment, materials, wafer fabrication, and assembly remain distributed across highly specialized regional clusters.
- [2] The U.S. Department of Energy reports that distribution-transformer lead times expanded from roughly three to six months in 2019 to twelve to thirty months in 2023.
Economics depend on useful work
Rated compute and connected power are not outcomes. The economic denominator is useful model work, shaped by utilization, software efficiency, network stalls, maintenance, and energy price.
- Report useful throughput alongside nameplate capacity
- Treat efficiency improvements as deployable supply
Supporting evidence · 3
- [1] PyTorch documents data, fully sharded data, tensor, and pipeline parallelism as complementary approaches for distributed model training.
- [2] LBNL scenarios place U.S. data-center electricity use between 325 and 580 TWh in 2028, or 6.7 to 12 percent of national consumption.
- [3] LBNL's 2025 Update places U.S. data-center electricity use at 649 TWh in its 2030 Reference Case, with compounded-uncertainty bounds of 521 to 843 TWh, or 9.5 to 15.3 percent of U.S. electricity use.
Featured evidence
Evidence in context
2025 disclosed infrastructure-adjacent capital measures
Microsoft FY2025 property-and-equipment additions alongside Alphabet, Meta, and Amazon calendar-year 2025 capital measures, presented as separate disclosures rather than one standardized industry series.View chart values
| Category | Company disclosure |
|---|---|
| Microsoft PPE additions | $64.55 USD billions |
| Alphabet capex | $91.4 USD billions |
| Meta capex | $72.22 USD billions |
| Amazon cash capex | $128.3 USD billions |
Key indicators
Key measures & constraints
- 01
Global investment in data centers reached roughly half a trillion U.S. dollars in 2024, nearly double the 2022 level.
- Class
- estimate
- Geography
- Global
- Period
- 2024
- Confidence
- high
- 02
The IEA's 2025 base case projects worldwide data-center electricity demand to more than double to about 945 TWh by 2030.
- Class
- projection
- Geography
- Global
- Period
- 2030
- Confidence
- medium
- 03
Microsoft reported $64.551 billion of property and equipment additions in fiscal 2025, compared with $44.477 billion in fiscal 2024; its fiscal year ended June 30 and the measure is not AI-only.
- Class
- fact
- Geography
- Global
- Period
- Microsoft fiscal year 2025
- Confidence
- high
- 04
Meta reported $72.22 billion of 2025 capital expenditure including finance-lease principal, principally for servers, data centers, and network infrastructure; cash property-and-equipment purchases were $69.69 billion.
- Class
- fact
- Geography
- Global
- Period
- Calendar year 2025
- Confidence
- high
- 05
Amazon reported $128.3 billion of cash capital expenditure in 2025, primarily for technology infrastructure—mostly supporting AWS growth—and fulfillment capacity; the amount is not data-center-only or AI-only.
- Class
- fact
- Geography
- Global
- Period
- Calendar year 2025
- Confidence
- high
System map
Connections across the system
Explore inputs, outputs, and shared capabilities in the Atlas. Open the register for every documented relationship.
Enters this system 0
No published input relationships cross this boundary.
Leaves this system 4
- Leading-edge fab← Project capitalTrace relationship
- Generation & storage← AI workload demandTrace relationship
Shared capabilities can belong to more than one System. These links carry materials, energy, information, capital, or permissions; they describe dependencies, not quantities or a complete process model.
Explore 6 capabilities and operating boundaries
All documented relationships · 10
- criticaloutput
Long-duration capital funds fab construction, tools, and ramp.
Trace relationship - importantinternal
Forecast demand unlocks financing and procurement commitments.
Trace relationship - importantoutput
Workload forecasts become regional electricity requirements.
Trace relationshipSupporting evidence · 3
- [1] The IEA estimates that global data-center electricity consumption rose 17 percent in 2025, with AI-focused data centers growing faster.
- [2] The IEA's 2025 base case projects worldwide data-center electricity demand to more than double to about 945 TWh by 2030.
- [3] LBNL's 2025 Update places U.S. data-center electricity use at 649 TWh in its 2030 Reference Case, with compounded-uncertainty bounds of 521 to 843 TWh, or 9.5 to 15.3 percent of U.S. electricity use.
- importantoutput
Project finance funds site, shell, utilities, and commissioning.
Trace relationship - importantinternal
Model training demand supplies a distinct operating capability within its broader Atlas system.
Trace relationshipSupporting evidence · 2
- importantinternal
Inference demand supplies a distinct operating capability within its broader Atlas system.
Trace relationship - importantinternal
Capacity reservations supplies a distinct operating capability within its broader Atlas system.
Trace relationshipSupporting evidence · 2
- [1] Microsoft reported $64.551 billion of property and equipment additions in fiscal 2025, compared with $44.477 billion in fiscal 2024; its fiscal year ended June 30 and the measure is not AI-only.
- [2] Amazon reported $128.3 billion of cash capital expenditure in 2025, primarily for technology infrastructure—mostly supporting AWS growth—and fulfillment capacity; the amount is not data-center-only or AI-only.
- importantinternal
Project finance & risk supplies a distinct operating capability within its broader Atlas system.
Trace relationshipSupporting evidence · 2
- importantinternal
Planned training demand is translated into reserved compute, network, site, and power capacity.
Trace relationshipSupporting evidence · 2
- [1] Microsoft reported $64.551 billion of property and equipment additions in fiscal 2025, compared with $44.477 billion in fiscal 2024; its fiscal year ended June 30 and the measure is not AI-only.
- [2] Amazon reported $128.3 billion of cash capital expenditure in 2025, primarily for technology infrastructure—mostly supporting AWS growth—and fulfillment capacity; the amount is not data-center-only or AI-only.
- importantoutput
Financing terms depend on land control, permits, utility milestones, and construction risk.
Trace relationshipSupporting evidence · 2
- [1] Global investment in data centers reached roughly half a trillion U.S. dollars in 2024, nearly double the 2022 level.
- [2] CRS finds that individual data centers typically fall under state and local siting, while the federal permitting nexus varies by project; FERC, NRC, the Army Corps, and delegated air and water programs may apply depending on configuration and location.
Named companies + institutions
Organizations & their roles
- principal · public company
Alphabet
Alphabet Inc.Technology holding company whose Google businesses procure and operate global AI compute and cloud infrastructure.
Roles, locations & evidence
- Role
- financier · operator
- Products / service
- Technical infrastructure capital investment · Google Cloud and AI services
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Project capital
Alphabet's filing-backed operating portfolio includes technical infrastructure capital investment, and the attached quantitative operating or investment measure supports principal status at this Atlas boundary; no subregional operating place is asserted.
- AI workload demand
Alphabet's filing-backed operating portfolio includes google cloud and ai services; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- principal · public company
Amazon / AWS
Amazon.com, Inc.Commerce and cloud company whose AWS segment is a major buyer and operator of AI compute, network, and data-center capacity.
Roles, locations & evidence
- Role
- financier · operator · buyer
- Products / service
- AWS infrastructure capital investment · AWS cloud and AI services · AWS compute and data-center procurement
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Project capital
Amazon / AWS's filing-backed operating portfolio includes aws infrastructure capital investment, and the attached quantitative operating or investment measure supports principal status at this Atlas boundary; no subregional operating place is asserted.
- AI workload demand
Amazon / AWS's filing-backed operating portfolio includes aws cloud and ai services; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Capacity reservations
Amazon / AWS's filing-backed operating portfolio includes aws compute and data-center procurement; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- principal · public company
Meta
Meta Platforms, Inc.Digital-platform company that designs, procures, and operates large AI training and inference infrastructure.
Roles, locations & evidence
- Role
- financier · operator · buyer
- Products / service
- AI infrastructure capital investment · Production AI inference services · Large-scale model-training infrastructure
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Project capital
Meta's filing-backed operating portfolio includes ai infrastructure capital investment, and the attached quantitative operating or investment measure supports principal status at this Atlas boundary; no subregional operating place is asserted.
- Inference demand
Meta's filing-backed operating portfolio includes production ai inference services; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Model training demand
Meta's filing-backed operating portfolio includes large-scale model-training infrastructure; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- principal · public company
Microsoft
Microsoft CorporationGlobal cloud and software company that procures, finances, and operates large-scale AI infrastructure.
Roles, locations & evidence
- Role
- financier · operator · buyer
- Products / service
- Data-center and compute infrastructure investment · Azure cloud and AI services · Azure AI capacity and infrastructure procurement
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Project capital
Microsoft's filing-backed operating portfolio includes data-center and compute infrastructure investment, and the attached quantitative operating or investment measure supports principal status at this Atlas boundary; no subregional operating place is asserted.
- AI workload demand
Microsoft's filing-backed operating portfolio includes azure cloud and ai services; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Capacity reservations
Microsoft's filing-backed operating portfolio includes azure ai capacity and infrastructure procurement; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
All other organizations · 3
- material · public company
Digital Realty
Digital Realty Trust, Inc.Data-center real-estate company developing and operating colocation and hyperscale facilities across global markets.
Roles, locations & evidence
- Role
- financier
- Products / service
- Data-center real-estate investment
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Project finance & risk
Digital Realty's filing-backed operating portfolio includes data-center real-estate investment; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · public company
Oracle
Oracle CorporationEnterprise software and cloud infrastructure company that procures and operates accelerated computing capacity.
Roles, locations & evidence
- Role
- operator · buyer
- Products / service
- Cloud and database AI services · Oracle Cloud Infrastructure capacity
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- AI workload demand
Oracle's filing-backed operating portfolio includes cloud and database ai services; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Capacity reservations
Oracle's filing-backed operating portfolio includes oracle cloud infrastructure capacity; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · private company
OpenAI
OpenAI, L.L.C.AI research and deployment company that procures and uses large-scale training and inference infrastructure.
Roles, locations & evidence
- Role
- buyer
- Products / service
- contracted AI compute capacity
- Documented activity
- United States · Texas · Abilene
- Headquarters
- United States
Why this organization belongs in the System- Capacity reservations
Primary-source evidence connects this institution to contracted ai compute capacity in the specified Atlas capability. Principal status is withheld because this edition does not close a concentration, capacity, shipment, revenue, or installed-base measure for the role.
Affected policies
Relevant policies
Legal status is separated from policy objective. Every record keeps its jurisdiction, mechanism, affected nodes, and verification date.
- effective
U.S. CHIPS and Science Act incentives
Expand domestic semiconductor manufacturing, research, workforce, and supply-chain resilience.
Mechanism, scope & sources
Manufacturing grants, loans and guarantees, research programs, and an investment tax credit with program conditions; current project records distinguish proposed and final awards, with disbursement tied to milestones.
United StatesOpen full policy record - effective
European Chips Act
Strengthen European semiconductor capacity, research leadership, monitoring, and crisis response.
Mechanism, scope & sources
Coordinates the Chips for Europe initiative, first-of-a-kind facility support, supply monitoring, and crisis tools; the Commission's Chips Act 2.0 work remains a proposal rather than replacement law.
European UnionOpen full policy record
Read this System in the field guide
Continue in the Volumes
Start with three selected readings, or explore the complete reading list.
Complete reading list
Permission
4 spreadsEvidence & review
currentEvidence health23 active claims · next review Oct 13, 2026+
- Active claims
- 23
- Sources
- 28
- High volatility
- 5
- Next review
- Oct 13, 2026
- Due soon
- 0
- Overdue
- 0
- Superseded
- 0