How MAG7 CAPEX Will Lift a Constellation of Small Strategic Techs
Event Horizon The MAG7 (the handful of mega-tech firms driving the current wave of AI, cloud, and infrastructure investment) have moved from being mere software platforms to becoming enormous industrial actors. Their capital expenditures (CAPEX) for AI compute, data centers, optical and networking upgrades, and supply-chain resilience have ballooned to levels that reframe the economics of entire technology stacks. This CAPEX wave will not only build bigger datacenters and more GPUs; it will alter demand curves across materials, photonics, interconnect IP, domestic foundries, energy storage, and specialized edge compute.
That is where the opportunity lies for smaller, structurally strategic equities: companies that supply specific, sometimes deep-tech inputs to the MAG7 buildout. Firms such as POET (POET; photonic integration), Lightwave Logic (LWLG; modulators and electro-optic materials), Eos Energy Enterprises (EOSE; grid-scale storage), AXT Inc. (AXTI; compound semiconductor substrates), Arteris (AIP; network-on-chip interconnect IP), GSI Technology (GSIT; specialized memory/AI accelerator architectures), and SkyWater Technology (SKYT; domestic foundry services) are not casual beneficiaries. They are infrastructure multipliers.
This scroll maps the CAPEX flows, explains the industrial linkages, and argues (strategically, not as financial advice!!) that the MAG7 CAPEX surge will materially accelerate demand and strategic relevance for the named small caps — some in obvious ways (more data center interconnects ? POET/LWLG), others in structural, systemic ways (chip sovereignty and substrate demand ? AXTI/SKYT; energy resilience ? EOSE). The endgame: a durable re-ranking of what counts as infrastructure and which companies are considered strategic enablers of the AI era.
1 — The MAG7 CAPEX Moment: scale, cadence, and intention To understand the opportunity set, we must first measure the size and character of the MAG7 CAPEX wave. CAPEX no longer describes a routine plant upgrade or a single factory expansion; it describes industrial campaigns; multi-year, multi-hundred billion dollar programs to build compute farms, specialized fabs, fiber backbones, power substations, and energy storage at a continental scale.
The headline is as follows: major cloud and AI vendors pushed CAPEX to record levels in 2024–2025 and signaled still-higher budgets for 2026. Alphabet raised its 2025 CAPEX estimate materially, Microsoft reported record quarter CAPEX driven by AI compute, Meta increased its capex guidance to reflect large datacenter programs, and Amazon’s capital plan approached the $100 billion mark. The narrative is rather consistent: hyperscalers are allocating huge sums to secure GPU clusters, next-generation networking (photonics), and the power and thermal infrastructure those machines demand.
Why now? Two technical trends explain the magnitude I suppose:
- Model scale growth — LLMs and next-gen models are orders of magnitude more compute-intensive, requiring many more GPUs and supporting hardware. Training and retraining cycles have shortened even as model parameter counts grew; meaning constant incremental CAPEX to keep pace.
- Operational density and replacement cadence — AI training infrastructure ages faster (and is refreshed faster) than traditional servers; useful lives for some AI servers have been shortened as organizations prioritize latest-generation accelerators.
If we project these two technical facts across global hyperscale datacenter builds and the push for sovereign or ally compute capacity (friend-shoring of AI infrastructure), we get a CAPEX number that dwarfs earlier cloud cycles. That spending creates both direct demand and indirect demand (upgrade power distribution; install better optics; buy more advanced chips and substrates).
2 — Channels of transmission: how CAPEX flows create new demand nodes CAPEX is not a single, fungible bucket. It flows through industrial layers and creates specific, measurable demand shocks at distinct points in the technology stack. For investors (us) thinking systemically, the most important point is not more servers but where the money touches supply chains and technology nodes.
Let’s consider five channels:
- GPU/Accelerator Procurement — Large purchases of GPUs, AI accelerators, and custom chips create demand for chip packaging, interposer materials, and customized silicon components.
- Data-Center Buildouts — New sites require advanced optical interconnects, fiber modules, cooling systems, and resilient power (that drives demand for photonics, POET/LWLG, and energy storage EOSE).
- Networking & Interconnect Upgrades — Moving more data faster pushes upgrades in photonic integration, modulators, lasers, and interconnect IP (benefitting POET, LWLG, AIP).
- Power & Energy Resilience — Datacenters demand stable, low-latency power; the rise of gigawatt scale deployments drives battery storage procurement and microgrid architecture (EOSE).
- Sovereignty/Onshoring Policies — CAPEX tied to domestic or allied capacity lifts demand for local fabs, substrate supply, and domestic foundry services (AXTI, SKYT).
These are the levers. Big CAPEX announcements create purchasing commitments and long-term procurement roadmaps that filter down through OEMs, ODMs, component suppliers, and material producers. The smaller names we are watching occupy those specific layers.
3 — Photonics and interconnect: the immediate uplifts (POET, LWLG, AIP)3.1 Photonics as a first-order effect As compute density explodes, the limiting factor for many hyperscale systems becomes bandwidth per watt and thermal efficiency; not raw transistor density. The physics are rather simple: moving bits with electrons across copper becomes energetically inefficient at large scale; photons are the solution for long links and high-bandwidth interconnects inside datacenters and between nodes across campus and region.
This creates two adjacent, high-leverage markets:
- Chip-level photonic integration (silicon photonics, hybrid integration), where companies like POET Technologies (POET) aim to integrate optics onto chips to reduce cost and improve link density.
- Electro-optic modulators and materials (LWLG) — efficient modulators reduce both latency and energy consumption of fiber links, directly reducing datacenter operating expenses (OPEX) over time.
When a hyperscaler decides to deploy new optical spine designs across dozens of datacenters, the demand pattern looks like this: optical engines ? photonic modules ? modulators ? substrate materials ? interconnect IP. That’s a chain of purchases where small suppliers (POET, LWLG) can scale orders quickly if adoption begins. The important dynamic is that once hyperscalers commit to a photonics standard (or experiment with it at scale), procurement volumes move from pilot orders to production orders quickly; a classic hockey stick for component makers.
Arteris (AIP) sits in a complementary place: as chips move from general-purpose processors to heterogeneous SoCs full of accelerators and photonic I/O, efficient network-on-chip (NoC) architectures are required. Arteris’ interconnect IP helps SoC designers manage on-chip dataflows efficiently. As more hyperscalers commission custom ASICs and SoCs (for inference, telemetry, security, or domain-specific workloads), demand for Arteris’ IP licensing and integration services should scale.
3.2 Mechanism: prototyping ? standardization ? volumeThere’s a three-step cadence to how small photonics firms go from niche to mass demand:
- Proof of concept — hyperscalers fund pilots to test latency, thermal profiles, and integration ease.
- Standardization — successful pilots become internal standards; procurement teams write specs and issue multi-year supply contracts.
- Production ramp — suppliers scale manufacturing capacity (often with their own CAPEX or via contract manufacturing partners) to meet multi-exabyte link demand.
Given the scale of MAG7 CAPEX, pilots can be funded at literally enormous expense, and the probability that such pilots move to standardization is crucial. That is the core structural reason POET and LWLG matter: they are at the points where hyperscalers need suppliers.
4 — Power, grid services, and energy storage (EOSE)Datacenters are energy-intensive. The MAG7 CAPEX wave therefore generates a parallel energy CAPEX wave: substations, power feeds, on-site energy storage, backup systems, and on-site generation (solar/wind + batteries or fuel cells) must be built or upgraded.
While lithium-ion batteries have been the dominant paradigm for grid and backup systems, alternative chemistries like zinc-based flow batteries and other safer chemistries are attractive for long-duration, large-scale deployments. Eos Energy Enterprises (EOSE) operates in this space: grid-scale batteries tailored for utility and datacenter deployments. The reasons smaller hyperscalers and hyperscale datacenters care include:
- Safety and environmental footprint — certain sites prefer chemistries with lower fire risk and more domestic supply chains.
- Duration economics — longer discharge durations (multiple hours) are becoming valuable as renewables account for a greater share of power supply to reduce curtailment and manage peak loads.
- Resilience & regulatory preference — local authorities and utilities may prefer storage systems with domestic supply chains supported by local industrial policy.
From a CAPEX linkage perspective, when the MAG7 builds new datacenters, utilities and site operators often coordinate to co-finance microgrids and on-site storage. Those procurement waves can generate multi-year orders for storage suppliers, and the timing is lumpy: datacenter ground-breaking ? power planning ? storage procurement ? installation. EOSE and similar firms sit squarely in that procurement funnel.
5 — Substrates, specialty materials, and foundries (AXTI, SKYT)The MAG7 CAPEX wave does more than buy racks and GPUs; it demands raw materials and manufacturing capacity. Two interconnected supply chains at play here:
- Compound semiconductor substrates — materials like gallium arsenide (GaAs), indium phosphide (InP), and others are used in photonics, RF, and certain accelerators. AXT Inc. (AXTI) produces such substrates; higher photonics and RF demand from datacenters and telecom will raise substrate needs.
- Domestic foundries & specialized fabs — when hyperscalers seek to secure supply chains for mission-critical ASICs, they often prefer allied domestic fabrication runs or trusted contract fabs. SkyWater Technology (SKYT) is a US foundry focused on specialized nodes. As MAG7 firms commission custom chips and edge ASICs for latency-sensitive inference, domestic foundry demand increases.
Recent policy actions (CHIPS Act and allied programs) explicitly incentivize domestic chip capacity. When hyperscalers align with these policy pushes ; e.g., build friend-shored capacity in allied countries, they shift some demand away from globalized vendors to domestic suppliers. Substrate producers and smaller, specialized foundries are the direct beneficiaries.
Mechanistically, capex leads to demand for packaging and materials in two ways:
- ASIC programs: A company (say, Meta or Amazon) commissions custom inference ASICs. That requires substrate material, packaging, and foundry slots; benefiting AXTI and SKYT.
- Photonics & RF modules: As optical links proliferate inside datacenters, module vendors buy upstream substrates for lasers, modulators, and detectors.
Small specialized material producers can thus see large order increases with relatively small changes in hyperscaler architecture choices.
6 — Edge architectures and specialized accelerators (GSIT)As the MAG7 invests in global AI deployments, they also decentralize compute. Not every inference call should travel to a centralized mega-datacenter; latency-sensitive and bandwidth-constrained services will run on edge nodes. Those nodes need efficient, specialized architectures; often smaller, lower-power accelerators or memory systems optimized for pattern matching and associative retrieval.
GSI Technology (GSIT) plays into that need with specialized memory and associative processing approaches tailored for low-latency pattern matching. The MAG7 CAPEX wave builds many edge nodes: telco central offices upgraded for compute, content delivery nodes with near-AI inference capabilities, and on-prem solutions for enterprise customers. Those nodes demand component-level specialization that GSIT and similar firms can supply.
Importantly, the edge isn’t only about tiny scale: it’s a distributed system problem. Upgrading thousands or tens of thousands of nodes can produce material aggregate demand. As hyperscalers design architectures that offload certain inference tasks to edge nodes for latency and cost reasons, component suppliers positioned for that edge will see growth.
7 — Demand elasticity, procurement strategies, and the pull factor A common investor mistake is to assume CAPEX is a simple one-to-one demand multiplier. It isn’t. The real dynamics depend on procurement strategies and technology choices by the MAG7:
- Do they standardize on an existing supplier, or open procurement to competitive bidding? Hyperscalers sometimes prefer a few trusted vendors (e.g., for density, reliability), which means winners can scale dramatically.
- Are they vertically integrating? Some MAG7 firms build their own chips or modules; others rely on suppliers. When they vertically integrate, smaller vendors can still benefit via spinouts, joint ventures, or as secondary suppliers.
- Are they pushing for domestic sourcing due to policy and geopolitical pressure? Where government incentives exist, domestic small caps disproportionately benefit.
The key point is that CAPEX opens doors but adoption depends on product fit, timing, and procurement politics. Small suppliers that align early with standardization efforts may find themselves awarded long multi-year contracts. The suppliers that miss the standardization window often remain niche.
8 — Pathways to scale: three realistic scenarios for small caps Per usual with our scenarios:
Scenario A — Strategic Supply: Direct procurement & contracts Hyperscalers standardize a technology in pilots (e.g., silicon photonics modules). Procurement teams award supply contracts to qualified vendors. Small caps (POET, LWLG) win tiered supply agreements. Result: multi-year revenue ramps, larger manufacturing contracts, and margin expansion due to scale.
Likely? Hyperscalers prefer rapid rollouts when power and bandwidth gains are real. Photonics pilots have historically moved to production given strong cost/OPEX improvement claims.
Scenario B — Policy-backed Onshoring: Industrial policy multiplies demand Governments incentivize domestic suppliers for strategic reasons — e.g, subsidy to build photonics fabs, tax credits for local battery storage, CHIPS Act–style credits for domestic packaging. This reduces cost disadvantages for small suppliers and accelerates domestic adoption. AXTI and SKYT benefit from friend-shoring and domestic procurement, while EOSE benefits from grid-resilience grants.
Likelihood? Industrial policy is in the zeitgeist; many countries want domestic control over critical infrastructure.
Scenario C — Ecosystem Leap: partnerships & IP adoption Smaller companies form deep technical partnerships or get acquired. For example, an accelerator designer (GSIT) may partner with an SoC company, embedding memory or associative compute IP into customized chips which hyperscalers then use for edge inference. Or Arteris’ NoC IP becomes part of a standard SoC template used by multiple hyperscalers.
Potential? Big tech often acquires or partners with specialized firms to speed up adoption, and successful partnerships often scale faster than organic adoption.
9 — Why each named small cap matters POET Technologies (POET) — photonic integration Role: Integrates optics and electronics at the chip level.
Connection to CAPEX: As datacenters expand bandwidth requirements, more optical interconnects will be deployed; driving demand for integrated photonics that lower per-bit cost and energy use. POET’s integration approach makes photonic components denser and cheaper to deploy. If a hyperscaler chooses to adopt POET’s integration tech at scale, orders move from 10s to 1,000s of units quickly.
Lightwave Logic (LWLG) — modulators & electro-optic materials Role: Electro-optic materials and modulators that push energy efficiency and speed for fiber links.
Connection to CAPEX: Datacenter operators seek to reduce OPEX by replacing energy-hungry links. LWLG’s materials reduce power per link, enabling large savings over hyperscale deployments. CAPEX budgets that prioritize long-term OPEX reduction will favor efficient modulators.
Eos Energy Enterprises (EOSE) — grid & datacenter storage Role: Zinc-based grid-scale battery systems.
Connection to CAPEX: New datacenters require power resilience and sometimes integrate on-site storage. EOSE can win part of the storage procurement wave, especially in jurisdictions favoring non-Li chemistries or where domestic supply chain resilience matters.
AXT Inc. (AXTI) — compound semiconductor substrates Role: Produces substrates for lasers, photonics, and RF devices.
Connection to CAPEX: As optical and RF modules proliferate, substrate demand rises. MAG7 procurement of optics/modules cascades into demand for substrates.
Arteris (AIP) — network-on-chip IP Role: On-chip interconnect IP for SoCs.
Connection to CAPEX: Custom SoCs for inference and telemetry require robust NoC IP. As MAG7 increases production of mixed SoC/photonic chips, Arteris’ IP becomes a licensable piece of the supply chain.
GSI Technology (GSIT) — specialized memory and accelerators Role: Memory architectures and associative processors for low-latency tasks.
Connection to CAPEX: Edge nodes and specialized inference tasks favor efficient associative compute. If MAG7 architectures offload certain workloads to such accelerators to save power, GSIT benefits.
SkyWater Technology (SKYT) — domestic specialty foundry Role: U.S. contract manufacturer and foundry for specialized nodes.
Connection to CAPEX: As MAG7 onshores critical SoC production or needs trusted subcontract capacity, SkyWater’s foundry slots and secure process IP become valuable.
10 — Non-linear nature of adoption Important caveat fellow readers: CAPEX waves create conditions but not guaranteed customer wins for small caps. Adoption is non-linear: a single design win with a hyperscaler (or large OEM) can multiply revenue many times over; conversely, failure to meet spec or timing issues can mean being bypassed.
Key risks:
- Execution risk for small suppliers — manufacturing scale, yield, and logistics can be bottlenecks.
- Standardization risk — if hyperscalers choose alternate technical standards, winners and losers flip quickly.
- Vertical integration risk — MAG7 firms may choose to vertically integrate successful technologies, reducing supplier margins or eliminating suppliers.
- Policy risk — subsidy programs can change with politics; a quick policy reversal could reduce domestic procurement.
That said, certain demand drivers mitigate risk:
- Multiple hyperscalers often run parallel pilots; winning one can create FOMO.
- Diverse procurement across regions (US, EU, APAC) increases the number of potential buyers.
- Long lead times for specialized materials and foundry capacity produce multi-quarter visibility once procurement decisions are made.
Timing matters because CAPEX converts to orders in predictable lags: R&D ? pilot ? procurement specification ? production ramp. That process can take 12–36 months for deep-tech components, sometimes faster for modular components.
11 — Geopolitics, policy, and friend-shoring as multipliers CAPEX is not politically neutral. Governments now view data, AI, and chip capacity as strategic national assets. That means procurement decisions are influenced by political objectives and industrial policy. The result is friend-shoring and allied procurement frameworks that create additional demand pools for domestic suppliers.
Two implications for the named small caps:
- AXTI and SKYT benefit from CHIPS-style and allied procurement — domestic substrate and foundry capacity is explicitly preferred in many procurement frameworks.
- EOSE and other grid suppliers could benefit from energy resilience grants tied to datacenter siting — local governments want resilient power near large compute sites.
Thus, geopolitical alignment is now a feature, not a bug. Companies that can credibly present themselves as trusted, allied suppliers can access procurement pipelines with high switching costs. In this sense, policy acts as a demand accelerator.
12 — The macro multiplier: beyond single ordersBuy one rack of accelerators and you buy hundreds of interconnect modules, tens of substrate lots, and gigawatt hours of energy storage capacity. The total economic impact of CAPEX is multiplicative. For example:
- A single new hyperscale campus with 100k racks requires fiber trunks, photonic modules, power substations, backup storage, and sometimes new on-site power generation.
- Each custom SoC program can create orders for substrates, IP licensing (Arteris), packaging, and foundry slots.
This is where the MAG7 CAPEX wave becomes transformative: it’s not merely a direct revenue push for one supplier; it’s a demand multiplier across adjacent, downstream suppliers. Small specialized firms occupying those adjacencies have outsized leverage: even a 1% share of final procurement in a large hyperscaler roll-out can equal multiples of their current revenue.
13 — Cautionary Tail Downside scenario 1 — Standardization on incumbents Hyperscalers prefer single large suppliers (e.g., incumbent optical module vendors) for reliability; small innovators fail to scale on manufacturing or reliability measures, and incumbents capture most of the rollout. Result: limited order flow for new entrants.
Downside scenario 2 — Vertical integration and insourcingMAG7 firms invest CAPEX into internal R&D and build their own photonics or SoC capabilities, choosing to insource the tech over time. Suppliers get either acquisition offers (some upside) or are sidelined.
Both are plausible. That’s why execution matters: small caps need to demonstrate manufacturability, yield, and reliability at hyperscaler scale to convert CAPEX waves into durable revenue.
14 — Re-industrialization of technology What the MAG7 CAPEX wave catalyzes is not just a one-off tech upgrade; it is a re-industrialization of technology. The last two decades saw software eat the world; the next decade sees software demand industrial hardware like never before. The economic implications:
- Capital flows into fabrication, materials, and energy sectors in ways not seen since the microelectronics mobilization of the 1970s–1980s.
- Labor demand shifts back toward manufacturing engineers, photonics specialists, and system integrators.
- National industrial policy becomes central again; the private sector is an execution arm for geopolitical strategy.
The small caps listed are therefore not quaint sidelined plays; they are nodes in a re-industrialization that will define the competitive balance of the 2030s.
15— End states and structural re-rankingsIf the MAG7 CAPEX wave continues, the long-term structural outcomes include:
- New supply-chain hierarchies: materials and module suppliers will gain bargaining power as demand is concentrated on a smaller set of specialized suppliers.
- Higher technical thresholds: SoC and photonics integration becomes the default for efficiency, raising the bar for entry and elevating incumbents that scale early.
- Fragmentation and friend-shoring: multi-bloc supply networks will persist, and suppliers aligned with allied procurement will be preferred.
- Hybrid winners: companies that can combine IP with manufacturability (design + production) will be most valuable.
These outcomes reshape what counts as strategic in corporate portfolios. Small caps that can demonstrably embed into the MAG7 supply chain can achieve outsized growth relative to their current scale.
16 — Singularity: CAPEX creates the scaffolding; small caps supply the bricks Large CAPEX programs are scaffolding: they create the architectural demands; massive compute farms, optical spines, resilient power plants, on-site storage, and bespoke SoCs. Small, nimble, deep-tech firms supply the bricks and mortar: photonic modules, modulators, substrates, NoC IP, edge accelerators, storage chemistries, and domestic foundry capacity.
POET, LWLG, EOSE, AXTI, AIP, GSIT, and SKYT are not random micro-bets; they are targeted exposures to the material components that hyperscalers must buy to realize their CAPEX plans. The next decade of technology will not be decided solely by software architecture or model design; it will be decided by who can reliably supply the hardware that undergirds AI infrastructure.
The MAG7’s CAPEX wave is therefore a structural economic event; one that will push demand for specialized inputs, re-orient industrial policy, and create winners among well-positioned, small strategic companies. The uplift is both market and mission driven: operational economics plus geopolitics. In that confluence, the small technical names we’ve discussed have a plausible path to becoming indispensable elements of 21st-century infrastructure. |
