Your tech portfolio could be on the wrong side of the AI boom
AI is not a single trade. It is a cascading capital cycle reshaping silicon, networks, power, software, and the economics of the cloud. The headlines and index returns make it look simple: own “AI” and prosper. But the boom is uneven, bottleneck-driven, and already crowded in some places. If your tech portfolio is overweight yesterday’s beneficiaries or tomorrow’s commoditized layers, you could be underperforming precisely while “AI” is winning.
What the market has already priced
– Concentration risk: A small set of megacaps has captured the lion’s share of AI gains. If you own broad tech exposure, you likely own them. That doesn’t mean you own the rest of the stack that will capture incremental dollars.
– Capex rotation: Hyperscalers are redirecting budgets from general-purpose compute and classic SaaS to AI infrastructure. That’s good for certain chip and power vendors, but it pressures vendors tied to legacy workloads.
– Narrative premium: Many companies have appended “AI” to their story without durable unit economics, distribution, or defensible IP. The market has rewarded some of these with multiples that assume persistence of early-cycle windfalls.
Where the value is accruing right now
The near-term value capture skews to constrained inputs and must-have workflows.
– Compute bottlenecks: Accelerators, advanced packaging, and high-bandwidth memory (HBM) remain the tightest points in the system. The further you are from these constraints, the less certain your upside.
– Networking and optics: AI clusters are network-bound. Ethernet/InfiniBand switches, optical transceivers, PAM4 DSPs, and cabling are foundational and in upgrade cycles (400G to 800G to 1.6T).
– Power and cooling: Data center power delivery, switchgear, transformers, UPS, liquid cooling, and thermal management are gating deployment. Long lead times create durable backlogs.
– Workflow-embedded software: Application vendors that sit in daily workflows with proprietary data and pricing power can monetize AI features. Horizontal model providers face price pressure; “wrappers” with no moat face extinction.
Who’s at risk of being on the wrong side
– Legacy software without an AI-native roadmap: Products that were once point solutions risk becoming features inside platforms that can bundle AI at low marginal cost.
– Model-agnostic middleware without moats: Abstraction layers that route between LLMs are easy to replicate and face vendor-native competition. Switching costs are low and pricing tends toward zero.
– Hardware assemblers with windfall margins: System integrators and server assemblers benefit in the scramble, but customer concentration, supply dependency, and normalization of margins are real risks when supply catches up.
– General-purpose cloud and classic compute: AI capex is crowding out some spend on non-AI servers and storage. Vendors tied to shrinking slices of the data center may see slower growth than the “AI boom” headlines suggest.
– Consumer internet with thin moats: AI-generated content and AI search can compress discovery economics and ad yields for sites dependent on organic traffic or commodity content.
– Content libraries with unresolved IP leverage: Copyright and training-data disputes create uncertainty for both content owners and model deployers; smaller players bear disproportionate legal and compliance risk.
Anatomy of the AI stack and where margins may migrate
– Accelerators: GPU leaders have captured outsized rents via software ecosystems and networking moats. Competition is rising (alternative GPUs, custom ASICs), but the key is not “chips” in general—it’s CUDA-scale software lock-in, developer mindshare, and networking integration.
– Foundry and packaging: Advanced nodes and 2.5D/3D packaging (e.g., CoWoS-style) are capacity constrained. These are long-cycle, capital-intensive bottlenecks with durable pricing.
– Memory: HBM is a chokepoint. Leaders with yield advantages enjoy pricing power today; memory is cyclical, but AI mix may structurally lift margins and ASPs even through cycles.
– Networking and optics: Moving from 400G to 800G and beyond drives content per rack. Vendors with strong positions in switches, optical modules, and DSPs have multi-year tailwinds as AI topologies densify.
– Thermal and power: Liquid cooling, busways, transformers, breakers, and backup generation are in shortage. Grid interconnection queues and long equipment lead times create multi-year revenue visibility for suppliers and data center REITs that can secure power.
– Application layer: Value accrues where AI is deeply integrated into workflows with measurable ROI and defensible data: productivity suites, design tools, cybersecurity telemetry, fintech compliance, service management. Pure “chat with your docs” tools with no distribution will struggle.
– Model providers: Foundation model margins are already under pressure from open-source and from customers training smaller, domain-specific models. Without proprietary data, distribution, or platform lock-in, sustained economic rents are uncertain.
Don’t forget the power system
The AI boom is as much about electrons as it is about tokens.
– Grid constraints: Power availability is a hard cap on model deployment. Transformers, transmission, interconnection, and permitting are gating. Utilities with data center exposure, grid equipment suppliers, and developers who can secure power are critical.
– Thermal limits: High-density racks require liquid cooling, retrofits, and specialized facility designs. Vendors with proven deployments and service footprints have the edge.
– Generation mix: Near-term, diesel and gas peakers backstop reliability; longer term, nuclear uprates, SMRs, wind/solar with storage, and demand response will matter. Policy and rate cases will shape returns for utilities and data center landlords.
What could go wrong with the consensus AI trade
– Oversupply whiplash: A step-change in accelerator supply or a pause in hyperscaler spending can compress margins and expose inventory. Semiconductor cycles still exist.
– Price wars at the model layer: As costs fall and open-source improves, API prices compress. Application vendors pocket savings; model vendors see shrinking gross margins.
– Regulatory and export controls: Restrictions on selling advanced chips to certain geographies and compliance regimes like the EU AI Act raise cost and limit TAM for some vendors.
– Reliability ceilings: If agentic systems and autonomy plateau short of promised reliability, some revenue ramps push out. Early pilots may not translate to enterprise-wide rollouts on the expected timeline.
– Power delays: Interconnect queues and local opposition can delay new capacity, shifting spend timing and causing batchy revenue.
How to reposition: a practical playbook
Barbell your exposure
– Infrastructure bottlenecks and “picks and shovels”
– Advanced packaging and HBM supply chain
– High-speed networking and optical interconnects
– Power equipment, transformers, UPS, switchgear
– Liquid cooling and thermal management
– Data center landlords with secured power and high pre-leasing
– Grid technology and transmission-enabling equipment
– Workflow-embedded applications with defensible moats
– Deep integrations and proprietary data advantages
– Measurable ROI and pricing power (not just feature parity)
– Distribution, renewals, and high net dollar retention
Underweight or avoid
– Model-agnostic middleware without lock-in or data gravity
– Standalone foundation model vendors without distribution
– Legacy point solutions susceptible to being bundled by platforms
– Hardware assemblers with temporary mix uplift and single-customer dependence
– Ad-dependent consumer sites vulnerable to AI search and content commoditization
Due diligence checklist for any “AI beneficiary”
– Revenue attribution: What percentage of revenue is directly tied to AI deployments? Is it recurring (subscriptions, maintenance) or project-based?
– Unit economics: If usage-based, what are gross margins per token/inference? How sensitive are margins to rapid cost declines?
– Supply chain and lead times: Where are the bottlenecks? What’s the visibility on capacity adds (HBM, packaging, optics)?
– Customer concentration: How much revenue comes from two or three hyperscalers? What is the duration and cancellation risk?
– Valuation versus durability: Are you paying peak-cycle multiples for windfall margins? What normalizes when supply meets demand?
– Regulatory exposure: Export controls, data residency, compliance costs. Who bears them and how do they impact gross margin?
– Capital intensity and FCF: What capex is needed to meet demand? How quickly does incremental revenue convert to cash?
Key scenarios to monitor in 2025–2027
– Compute scarcity persists: Bottlenecks shift from GPUs to HBM to power to networking; infrastructure names continue to outperform while application adoption is gated by capacity.
– Inference cost collapse: Costs fall faster than expected; value shifts from model providers to applications and end-users. Model API pricing compresses; application gross margins expand.
– Open-source parity: Open models match closed models on key benchmarks. Enterprises adopt small, domain-tuned models; training spend concentrates in a few hyperscalers.
– Power-limited buildouts: Grid constraints slow new region launches. Data center REITs with power win; utilities with constructive regulation re-rate; some software ramps delay.
– On-device AI: Edge inference improves, shifting some workloads from cloud to devices. Beneficiaries include edge silicon, RF, and sensor stacks; cloud inference TAM still grows but with different mix.
Red flags that your portfolio is on the wrong side
– Heavy exposure to vendors whose core value proposition is now a single LLM feature
– Rising revenue but falling gross margin as usage scales
– Marketing-led “AI” narratives with minimal R&D intensity or IP
– Customer wins concentrated in pilots and POCs with slow conversion to enterprise-wide deployments
– Dependence on geographies under export restrictions for advanced compute
– Balance sheets stretched to chase an inventory or capacity cycle that could turn
A few illustrative angles
– The accelerator leader may keep share longer than bears expect because software and networking moats matter, but rent extraction invites competition; watch total cost of ownership for customers and the speed of software portability.
– Foundry and packaging leaders are structural winners but carry geopolitical and cyclicality risks; these are long-duration assets, not momentum trades.
– HBM suppliers enjoy rare pricing power; the risk is not whether AI needs more memory, but when supply catches up and how much of the uplift is structural versus cyclical.
– Networking winners ride topology shifts to ever-faster speeds; their risk is architectural change and hyperscaler insourcing.
– Liquid cooling and power distribution may be earlier in the rerating cycle than semis; their backlogs are tied to physical constraints with long lead times.
Practical portfolio hygiene
– Replace “own AI” with explicit theses: name the bottleneck or moat, define the customer, and model how dollars flow.
– Expect rotation: bottlenecks move. Be ready to recycle gains from over-earning segments into the next constrained layer.
– Build dispersion into sizing: early-cycle high-beta names warrant smaller weights; structural compounders can be core.
– Track real-time signals: GPU lead times, hyperscaler capex guides, HBM capacity announcements, transformer backlogs, AI API price changes, and enterprise adoption surveys.
The bottom line
Owning “tech” is not the same as owning the AI value chain. The boom is real, but it is selective, bottleneck-driven, and already well owned in some places. Portfolios overweight legacy software, commoditizing middleware, and windfall hardware margins could stagnate while the AI economy grows. Tilt toward constrained infrastructure and workflow-embedded applications with data and distribution moats, keep an eye on power as a core part of the story, and be prepared to rotate as bottlenecks—and profits—shift.
This article is for informational purposes only and is not investment advice.
