AI Compute Infrastructure Buildout

The thesis

The thesis is that AI compute deployment is governed by a physical, interdependent buildout whose value is realized only when all deployment gates are cleared.

The economic premise is not merely accelerator demand. Cloud and infrastructure operators must secure sites and power, acquire and install electrical and thermal systems, connect clusters with high-speed networking, and source packaged and tested semiconductors. A failure or delay at any gate can defer utilization at the others. That relationship makes the buildout broader than a single equipment category and gives enabling layers economic relevance when their products or services are necessary to turn hardware into operating capacity. The current evidence supports an active capital cycle. Oracle's latest reported capital expenditure increased sequentially, while cloud-capacity operators IREN and Nebius have maintained or raised operating capacity and revenue objectives. In parallel, Vertiv, Amkor, Coherent, Lumentum, Onto Innovation, FormFactor, Quanta, and others provide evidence of expanding revenue, capacity investment, backlog, or forward expectations within their respective layers. This is stronger evidence for continued buildout than the prior research state, particularly at the cloud-capacity, power-and-thermal, optical, packaging, and electrical-construction layers. The thesis is not uniformly proven at every company. Several anchors carry weaker profitability, capital-discipline, balance-sheet, or growth-durability signals, and some have not reaffirmed guidance. The appropriate current state is therefore a supported but uneven infrastructure buildout: broad spending evidence exists, but transmission into company-level outcomes remains differentiated.

• Compute capacity depends on a chain of semiconductor, facility, networking, power, cooling, and construction inputs.
• Capital formation is visible in both operators of compute capacity and suppliers of deployment-critical systems.
• The fundamental picture is stronger than a single-theme narrative, but company execution is uneven.

Where the spending comes from

Spending originates with hyperscaler, cloud, enterprise, utility, colocation, semiconductor-manufacturing, and data-center-development budgets, with current evidence showing capital flowing across the full deployment chain.

Cloud and AI-infrastructure operators are the most direct owners of capacity budgets in the record. Oracle's reported capital expenditure rose sequentially in its latest quarter. IREN raised its year-end AI Cloud annualized-run-rate objective and disclosed contracted support for much of that objective, while Nebius maintained its revenue and annualized-run-rate objectives and raised its contracted-power target. These disclosures indicate that capacity providers are committing capital and planning around power availability, rather than treating compute as a purely asset-light service. That spending transmits outward. Hyperscaler, cloud, and colocation construction budgets purchase electrical distribution, backup power, cooling, networking, and installation services. Foundry, integrated-device-manufacturer, and outsourced-assembly budgets fund packaging, inspection, metrology, probe-card, and fabrication equipment. Utility transmission and data-center developer budgets fund substations, high-voltage interconnection, and electrical construction. The direction is not identical across every anchor. Recent spending rose at Oracle, Vertiv, Coherent, MYR Group, Powell, Modine, Gates, and Quanta, while recent quarterly capital expenditure fell at Arista, FormFactor, EnerSys, and Atkore. This mixed pattern does not negate the broader cycle, but it limits any claim that investment is accelerating uniformly throughout the chain.

• Cloud-capacity operators provide the clearest direct evidence that compute capacity is being funded.
• Facility, networking, semiconductor, and construction layers receive spending through distinct buyer budgets.
• Recent company investment trends are broadly supportive but not uniform.

How the money flows through the stack

Economic capture is distributed across the layers that convert semiconductor content and secured power into installed, connected, operating compute capacity.

The chain begins with semiconductor production and advanced assembly. Advanced logic, memory integration, yield management, inspection, metrology, and test are required before systems can be deployed. The next layer is the facility: powered land, interconnection, data-center space, electrical distribution, backup power, and liquid-cooling infrastructure. Those physical assets then require high-speed switching, optics, and connectivity components to make a cluster function as a coordinated fabric. Finally, transmission, substation, and mission-critical electrical contractors connect the facility to the grid and execute installation. Pricing power and scarcity should be understood as an inference from necessity and execution dependence, not as a universal claim about margins. The clearest potential scarcity points in the record are advanced packaging and test capacity, powered sites and interconnection, high-density power and thermal deployment, optical interconnect capability, and skilled electrical-construction execution. Each is a gate that cannot simply be bypassed by buying more accelerators. The chain also explains why financial outcomes vary. A supplier may participate in the same buildout but have distinct customer concentration, legacy end markets, capital needs, balance-sheet exposure, or execution requirements. Consequently, evidence of system-wide spending supports the chain, not identical operating outcomes for every participant.

• Advanced assembly and test make semiconductor content deployable.
• Power, cooling, and construction make a facility operable.
• Networking and optics determine whether installed compute operates as a usable cluster.

What is constrained

The binding constraints identified in the record span deployable power, thermal and electrical infrastructure, grid connection and construction execution, advanced packaging and test, and high-speed interconnect.

Powered land and interconnection are constraints because a compute cluster requires an energized site before it can operate. The cloud-capacity theme specifically identifies scarcity of ready-to-energize capacity and large-load interconnection as strategically important. IREN and Nebius frame their operating objectives around capacity and connected or contracted power, reinforcing that power access is central to their deployment plans. High-density electrical and thermal systems remain a second deployment gate. The power-and-thermal theme identifies rack power density and heat rejection as direct deployment constraints, while Vertiv's company context links its demand driver to hyperscaler and colocation spending on high-density electrical and liquid-cooling infrastructure. Grid connection and electrical construction are a further execution constraint because substations and high-voltage connections must be built and installed, not simply ordered. The relevant theme identifies skilled labor and execution capability as central to that work. Advanced packaging, inspection, metrology, and test remain semiconductor-side chokepoints because heterogeneous systems and high-bandwidth memory require assembly quality and yield management. The record supports the importance of these gates, but it does not establish that every constraint is worsening. Capacity investment at Amkor, Coherent, Vertiv, and others can eventually relieve particular bottlenecks; the current evidence supports persistence and importance, not a precise direction of scarcity at every layer.

• Power availability and interconnection constrain when a facility can be energized.
• Thermal and electrical systems constrain how densely compute can be deployed.
• Packaging, inspection, and test constrain the conversion of advanced semiconductor content into usable systems.
• The record supports bottlenecks, but not a universal claim that each is worsening.

How the companies actually participate

The anchors participate through identifiable products, services, capacity platforms, and construction activities rather than through a generic association with AI.

Advanced packaging and test: Amkor provides outsourced semiconductor packaging and test services, making it a direct participant in assembly and final-test requirements. Onto Innovation supplies process-control tools for defect inspection and metrology on advanced-packaging lines. FormFactor supplies probe cards and testing solutions relevant to advanced packaging and high-bandwidth-memory devices. Axcelis supplies ion-implantation equipment used by semiconductor chip manufacturers. Cloud capacity and powered land: IREN owns and operates power-connected data centers and AI cloud infrastructure, so its transmission is the development and monetization of energized compute capacity. Oracle funds cloud infrastructure through its capital expenditure and sells enterprise cloud services. Nebius builds and operates GPU clusters, cloud platforms, and developer tools, linking its transmission to vertically integrated AI cloud capacity. Power and thermal: Vertiv supplies critical digital-infrastructure technologies and services for data centers, including the electrical and thermal systems required by high-density facilities. Modine supplies thermal-management products to industrial and commercial HVAC markets. Atkore sells electrical and mechanical infrastructure products used by contractors and infrastructure builders. EnerSys supplies stored-energy systems across network, infrastructure, and precision-power applications. Gates supplies engineered power-transmission and fluid-power solutions, while Powell supplies custom-engineered electrical equipment and systems for power infrastructure. Networking and optical interconnect: Arista sells networking hardware and software for AI and data-center environments. Astera Labs sells semiconductor connectivity solutions to cloud and AI infrastructure customers. Applied Optoelectronics sells fiber-optic modules and lasers to internet data-center operators. Coherent supplies lasers, optical components, and systems to data-center and communications markets. Lumentum supplies optical and photonic products serving cloud and data-center operators. Grid connection and electrical construction: MYR Group performs transmission-and-distribution and commercial-and-industrial electrical construction. Quanta designs, constructs, upgrades, and maintains transmission, distribution, substations, and related utility infrastructure.

• Each anchor is linked to a disclosed operating mechanism in the deployment chain.
• Transmission spans semiconductor tools and services, cloud platforms, facility systems, optics, and construction.
• Exposure remains company-specific even where companies occupy the same theme.

What the market currently believes

Market behavior recognizes the narrative as highly important, but anchor-level price structure is split rather than uniformly confirming the fundamental buildout.

Trading Compass Narrative Heat classifies the overall narrative as hot, and the available theme readings also classify advanced packaging and test, cloud capacity and powered land, networking and optical interconnect, and grid connection and electrical construction as hot. These readings describe current market attention and perceived importance; they do not establish demand, capacity, backlog, or spending. Trading Compass Price Action shows a fragmented anchor distribution. Constructive pullbacks are present among Axcelis, Astera Labs, Coherent, FormFactor, Gates, Nebius, and Onto Innovation. Arista is in an early trend, while Lumentum and Quanta show strong advances. Against that, damaged structure is present at Applied Optoelectronics, Amkor, EnerSys, IREN, Modine, Oracle, and Powell; MYR Group and Vertiv are classified as technical failures. The market therefore appears to distinguish sharply among implementation paths and company-specific risk rather than validate the entire chain uniformly. This divergence matters analytically. The fundamental record can support ongoing infrastructure formation while market behavior remains selective or adverse for individual participants. Neither form of evidence substitutes for the other.

This section reads market behaviour, not company fundamentals. Price and attention confirm or contradict a thesis; they are not evidence about demand or spending.

• The narrative and the available theme heat readings remain elevated.
• Anchor price behavior is bifurcated, with constructive and damaged structures both well represented.
• Market behavior indicates belief and positioning, not physical demand or capital spending.

Evidence against this thesis

The main challenges are uneven operating quality, large capital requirements, slowing growth signals at selected companies, and a withdrawn outlook at one infrastructure supplier.

The strongest contradictory evidence is company-specific. IREN's record shows revenue growth decelerating across consecutive comparable periods alongside weak profitability-quality and capital-discipline assessments, even as it pursues large infrastructure expansion. Coherent's financial profile identifies weak capital discipline and balance-sheet resilience, while its recent capital spending has been substantial. Amkor's large investment program coincides with negative free cash flow and more net debt, and the record flags declining research-and-development intensity as a factor that can flatter margin expansion. The semiconductor and electrical layers also are not uniformly accelerating. Axcelis maintained an outlook for annual revenue to remain relatively flat versus the prior year, and its reported apparent growth acceleration is qualified by a negative earliest comparison period. Astera Labs' latest product growth rate slowed. EnerSys' Motive Power growth slowed in the latest period. Atkore withdrew its fiscal-year outlook because of its pending transaction, removing a current management forecast from the record. These facts do not disprove the buildout. They do require a narrower conclusion: the infrastructure cycle is supported at the system level, but its company-level beneficiaries face capital-intensity, execution, financing, baseline, and demand-mix risks.

• Capital intensity can weaken cash conversion and increase financing sensitivity.
• Growth evidence at selected anchors has slowed or is difficult to extrapolate.
• A withdrawn outlook reduces forward visibility for Atkore.

What would make this wrong

The thesis would weaken if capacity builders reduce their stated deployment ambitions, if enabling suppliers lose backlog or guidance support, or if grid and facility execution ceases to be a deployment gate.

The clearest observable invalidation would be a reversal in operator commitments. The thesis would weaken if Oracle's infrastructure investment fell materially without replacement evidence elsewhere, or if IREN and Nebius reduced their capacity, power, annualized-run-rate, or revenue objectives. It would also weaken if contracted cloud demand failed to convert into operating revenue and utilization for capacity operators. At the enabling layers, invalidation would appear through reduced backlog, lower revenue or margin outlooks, or evidence that capacity additions have relieved deployment constraints faster than demand requires. Relevant watchpoints include Onto Innovation's backlog and outlook, FormFactor's advanced-test outlook, Vertiv's power-and-thermal outlook, and the revenue and profitability expectations reported by optical suppliers and electrical contractors. A further risk is that a system-wide buildout does not translate into proportionate outcomes for every anchor. Capital-intensive expansion, higher debt, dilution, weak profitability, customer concentration, and non-AI end-market exposure can each break the transmission from broad infrastructure spending to a particular company's results.

• Lower operator capacity and power objectives would challenge the demand side of the thesis.
• Backlog erosion or guidance reductions at enabling layers would challenge supply-chain transmission.
• Company-level financial and execution risks can invalidate a specific transmission mechanism without invalidating the broader narrative.

Catalysts that would change this

The next research state should be driven by updated operator capital plans, contracted-power and capacity milestones, supplier backlog and guidance revisions, and evidence that projects are becoming energized and operational.

The most consequential updates will come from cloud-capacity operators and major infrastructure buyers: changes in Oracle's capital expenditure trajectory; IREN's progress against its AI Cloud objective and contracted demand; and Nebius's revenue, annualized-run-rate, connected-power, and contracted-power milestones. These directly test whether intended capacity is moving from plans to deployed infrastructure. At the supply-chain level, the important watch items are updated backlog and outlook disclosures from Onto Innovation, FormFactor, Vertiv, and Powell; revenue and margin guidance from optical and networking suppliers; and raised, maintained, reduced, or withdrawn outlooks among electrical and construction companies. Progress in Amkor's capital program and evidence of sustained demand for advanced packaging and test are also material to the semiconductor-side constraint. There is no current dated catalyst calendar in the record. Accordingly, the appropriate watch list is disclosure-based rather than event-date-based.

• Operator capacity, power, and contracted-demand milestones are primary demand tests.
• Backlog and guidance revisions test transmission into enabling layers.
• No dated catalyst schedule is available in the current record.

Theme research

AI Fabric Networking and Optical Interconnect

Theme Heat 91 of 100 — very high, measured 2026-10-09. Cluster scaling continues to require more optical bandwidth and lower-latency interconnect, with broad demand confirmation across accelerated-compute architectures. Its transmission through switching, optics, and component supply chains keeps it near the center of AI capex.

High-speed optical, signal-integrity and network components that connect accelerated compute clusters.

Why it matters: Scaling AI clusters raises the value of bandwidth, latency and power efficiency across the network fabric; interconnect quality determines usable cluster performance.

Why now: AI platform roadmaps are increasing network intensity, while suppliers are expanding advanced optical and packaging capacity for next-generation AI infrastructure.

Shared demand driver: Accelerator-cluster spending on high-bandwidth optical links, signal integrity and AI fabric connectivity.

Shared buyer or budget: Hyperscaler and AI-system networking budgets

Companies in this theme (6)

• Applied Optoelectronics, Inc. (AAOI) — Picks-and-shovels supplier · direct exposure. Optical transceivers are direct components of higher-speed data-center network upgrades. Applied Optoelectronics participates through optical transceivers used to increase bandwidth between data-center switches and servers.
• Astera Labs, Inc. (ALAB) — Picks-and-shovels supplier · direct exposure. Connectivity silicon exposure fits AI cluster interconnect infrastructure.
• Arista Networks, Inc. (ANET) — Picks-and-shovels supplier · direct exposure. Supplies high-speed switching and Ethernet fabric gear used to scale AI clusters.
• Coherent (COHR) — Provides optical transceivers, lasers and photonics technologies for data communications · an enabling supplier / picks-and-shovels provider. Coherent supplies photonic technologies that support the optical interconnect layer of large AI compute clusters. Why it fits: AI networking requires scalable optical connectivity and photonic components at increasing speeds.
• Lumentum Holdings (LITE) — Supplies optical components used in high-speed communications and data-center interconnect · an enabling supplier / picks-and-shovels provider. Lumentum participates through optical components that enable high-speed data movement within AI data-center networks. Why it fits: Optical component demand rises as AI clusters require faster links across racks and facilities.
• Viavi Solutions (VIAV) — Provides optical-network test, measurement and network-assurance equipment · an enabling supplier / picks-and-shovels provider. Viavi benefits when higher-speed AI networking raises demand for optical test and network-assurance tools. Why it fits: Faster AI networks require more stringent qualification, testing and performance assurance.

Advanced Packaging and Semiconductor Test

Theme Heat 77 of 100 — high, measured 2026-10-09. Advanced packaging and test remain necessary for heterogeneous AI systems, but current market attention is more concentrated in compute capacity, fabrics, and power. The theme retains strong structural transmission through the semiconductor supply chain.

Assembly, packaging, inspection and test capacity required to turn advanced logic, memory and heterogeneous compute into deployable AI systems.

Why it matters: Advanced packaging and test are critical chokepoints for high-performance compute, memory integration and yield management.

Why now: Fab-equipment spending is projected to grow strongly in 2026 and advanced-node investment is being driven by AI performance and power-efficiency requirements.

Shared demand driver: AI chip and high-bandwidth-memory production requiring advanced assembly, inspection and test

Shared buyer or budget: Foundry, integrated-device-manufacturer and outsourced-semiconductor-assembly budgets

Companies in this theme (4)

• Axcelis Technologies, Inc. (ACLS) — Picks-and-shovels supplier · adjacent exposure. Its installed base participates in the broader domestic semiconductor-capacity and power-chip investment cycle. Axcelis supplies ion-implantation equipment used in semiconductor manufacturing capacity supporting advanced and power-device production.
• Amkor Technology, Inc. (AMKR) — Pure play · direct exposure. Advanced packaging is a direct service line and a critical part of AI-chip supply chains. Amkor provides outsourced advanced packaging and test services needed to assemble complex AI-oriented semiconductor devices.
• FormFactor, Inc. (FORM) — Supplier · direct exposure. FormFactor supplies semiconductor test and measurement hardware used in advanced packaging and chip test.
• Onto Innovation Inc. (ONTO) — Picks-and-shovels supplier · direct exposure. Complex packaging and advanced-node production require tighter yield control and defect inspection. Onto Innovation sells inspection and metrology tools used to control yield in advanced semiconductor manufacturing and packaging.

AI Cloud Capacity and Powered Land

Theme Heat 90 of 100 — very high, measured 2026-10-09. Powered sites and cloud capacity remain scarce relative to AI-load demand, keeping capacity announcements and contracted power highly market-relevant. This is a direct monetization layer of the core AI buildout.

Owners and developers monetizing data-center capacity, GPU cloud infrastructure and sites with access to power and interconnection.

Why it matters: The economic value of AI hardware depends on having available data-center space, electrical capacity and a deployable operating platform.

Why now: Large-load interconnection issues and the scarcity of ready-to-energize capacity are elevating powered land and operational AI cloud capacity as strategic assets.

Shared demand driver: Demand for immediately deployable AI compute capacity backed by secured power and data-center infrastructure.

Shared buyer or budget: Hyperscaler overflow, enterprise AI and GPU-cloud capacity budgets.

Companies in this theme (4)

• APLD — Platform · direct exposure. Its business is directly tied to bringing power, land and data-center capacity together for compute tenants. Applied Digital participates by developing and operating data-center capacity intended for high-performance and AI computing tenants.
• IREN Limited (IREN) — Pure play · direct exposure. Its secured power pipeline and contracted AI cloud capacity are direct expressions of the energized-land scarcity trade. IREN is expanding powered data-center capacity and AI cloud infrastructure to serve large-scale GPU deployments.
• Nebius Group N.V. (NBIS) — Platform · direct exposure. The company sells AI-oriented compute capacity rather than merely supplying adjacent hardware. Nebius participates through AI cloud services that provide customers access to accelerated compute infrastructure.
• Oracle Corporation (ORCL) — Platform · direct exposure. Cloud provider expanding AI capacity and data-center footprint fits delivered AI cloud capacity.

AI Data Center Power and Thermal Systems

Electrical distribution, backup power and liquid-cooling systems required to operate high-density AI clusters.

Why it matters: Rack power density and heat rejection are direct deployment constraints; a GPU order cannot become productive capacity without this physical infrastructure.

Why now: Vertiv reported sharply higher orders and backlog tied to hyperscale and colocation demand, and is expanding power and cooling capacity for AI deployments.

Shared demand driver: Hyperscaler and colocation capital spending on high-density AI data-center electrical and liquid-cooling infrastructure.

Shared buyer or budget: Hyperscaler, cloud and colocation data-center construction budgets

Companies in this theme (3)

• Powell Industries (POWL) — Builds custom electrical distribution and control equipment for mission-critical power systems · an enabling supplier / picks-and-shovels provider. Powell Industries supplies engineered electrical equipment that supports the power-distribution layer of large AI data-center projects. Why it fits: AI campuses require medium-voltage distribution, switchgear and control assemblies before servers can be energized.
• Modine Manufacturing (MOD) — Supplies thermal-management equipment for data-center and industrial cooling applications · an enabling supplier / picks-and-shovels provider. Modine participates through cooling systems used to remove heat from increasingly dense AI compute environments. Why it fits: Higher compute density raises demand for purpose-built heat-rejection and cooling equipment.
• Vertiv Holdings (VRT) — Provides integrated critical power, cooling and infrastructure systems for data centers · a diversified participant (theme is one of several real businesses). Vertiv sells the power and thermal infrastructure needed to commission and operate high-density AI data centers. Why it fits: Its order and backlog data provide a direct read-through on AI physical-infrastructure spending.

Other companies mapped to this narrative

Approved exposures that no canonical theme represents yet, each with the theme its own research named.