Wall Street is leaning into a simple thesis for the coming year: the AI infrastructure buildout is still in the early innings, and semiconductors sit squarely at the center of that capex supercycle. A fresh call from Cantor frames 2026 as an outperformance year for chips, underpinned by hyperscaler spending, tight supply in advanced memory, and the continued shift from pilots to production in AI workloads. For investors, the opportunity is not monolithic—different rungs of the stack may lead at different points—but the direction of travel remains clear: as long as compute demand compounds, the AI infrastructure buildout should keep funneling dollars to silicon, packaging, and the networks that bind it all together.
Why 2026? The spending wave crests into deployment
After a frenetic 2024–2025 characterized by GPU shortages and scramble-mode procurement, 2026 is shaping up as the year the AI infrastructure buildout meets deployment at scale. Hyperscalers are guiding to another leg up in data-center capex to support training and, crucially, high-throughput inference. Multiple independent outlooks now converge on a step-function increase in AI-related spend next year, with a larger share earmarked for custom accelerators, networking silicon, and high-bandwidth memory (HBM). The practical implication: order books across several semiconductor sub-sectors have multi-quarter visibility, even as delivery windows lengthen and power constraints complicate regional timelines.
Sub-sector winners in the AI stack
1) Accelerators (GPUs and custom silicon).
The compute heart of the AI infrastructure buildout continues to be accelerators. Merchant GPUs will remain the reference platform for frontier training and increasingly for enterprise inference, thanks to robust software ecosystems and total-system performance. At the same time, custom AI ASICs are gaining ground with mega-buyers seeking cost and power efficiency at scale. Expect both tracks to grow in 2026: GPUs for flexibility and time-to-market, ASICs for unit economics once models and workloads standardize.
2) High-bandwidth memory (HBM).
No HBM, no AI. The tightness in supply is not just about wafers—it’s advanced packaging yields, substrate availability, and the thermal envelope. As model sizes expand and sequence lengths grow, HBM stacks climb and bandwidth-per-socket rises, prolonging the HBM upcycle. That dynamic puts memory among the clearest beneficiaries of the AI infrastructure buildout in 2026, with pricing power and mix both supportive.
3) Advanced packaging & foundry services.
CoWoS-like 2.5D/3D integration, chiplets, and advanced interposers are moving from specialized to essential. Foundry partners and outsourced semiconductor assembly and test (OSAT) providers with capacity in advanced packaging are likely to see persistent tightness through 2026. As compute vendors race to new architectures, the packaging bottleneck remains one of the gating factors for the AI infrastructure buildout.
4) Networking (optics, switches, DPUs).
Training clusters are only as fast as their fabric. As node counts climb, demand for high-radix switches, 800G/1.6T optics, and acceleration at the NIC/DPU layer intensifies. Expect networking silicon to be a second-derivative winner of the AI infrastructure buildout, particularly as cloud providers push toward Ethernet evolution and mixed-fabric topologies to balance cost, performance, and vendor diversification.
5) Equipment (WFE) and metrology.
If the AI infrastructure buildout keeps lifting advanced-node demand, wafer fab equipment should benefit—both at logic (N3/N2 ramps) and at memory, where HBM’s layer count and cell architectures require more deposition, etch, and inspection steps. Even if industrial/consumer end-markets stay uneven, AI-centric tool demand can cushion the cycle and support 2026 order momentum.
What’s new in Cantor’s framing
Cantor’s sector view highlights three reinforcing pillars for 2026 outperformance: (1) the secular runway in AI compute demand, (2) memory tightness—especially HBM—as a structural, not transitory, constraint, and (3) the broadening of beneficiaries beyond accelerators to include networking, packaging, and select equipment names. The emphasis on memory supply constraints is notable; it counters the reflexive fear that a faster supply response will crush pricing. Instead, structural bottlenecks along the packaging and substrate chain should keep the market tight enough to preserve mix and margin for well-positioned suppliers through the next several quarters.
Addressing the key bear arguments
“AI capex is peaking.”
Evidence suggests sequencing rather than peaking. Training clusters built in 2024–2025 are transitioning to revenue-generating inference in 2026, while fresh training demand continues for next-gen models. In parallel, enterprise adoption is beginning to scale beyond pilots. That creates a multi-track capex footprint: ongoing training builds, inference fleet expansion, and new vertical/edge deployments—each additive to the AI infrastructure buildout.
“Multi-sourcing dilutes winners.”
Vendor diversification is real, but the market is expanding faster than any one supplier’s share erosion can offset. Moreover, software and system-level performance still govern workload placement. In 2026, we’re more likely to see mix shifts within growth than outright revenue contractions for leaders in GPUs, custom ASICs, and networking.
“Policy risk will choke demand.”
Export controls, content rules, and localization mandates will continue to shuffle the deck regionally. Yet policy uncertainty often delays rather than destroys demand; it reroutes orders and reshapes product stacks. Companies that preemptively design for compliance (e.g., location attestation, region-specific SKUs) can keep participating in the AI infrastructure buildout while reducing headline risk.
Macro tailwinds—and constraints
Lower rate volatility and a gradual glide path toward policy easing improve the multiple backdrop for long-duration growth assets, including semiconductors. Meanwhile, power availability has become the new gating factor for data centers. Utilities, grid upgrades, and on-site generation are now part of the procurement conversation. That constraint could elongate rather than truncate the AI infrastructure buildout, smoothing demand across more quarters as projects phase to align with power and real-estate timetables.
What to watch in 1H 2026
- Hyperscaler capex guides: Look for confirmation that year-over-year budgets are rising and skewed toward AI accelerators, networking, and storage tuned for AI inference.
- HBM capacity additions: Clarity on new lines and yields will inform how long the pricing/mix tailwind lasts for memory suppliers.
- Packaging lead times: Signals from foundry/OSAT partners on advanced packaging capacity will dictate delivery cadence for AI systems.
- Networking upgrades: Orders for 800G/1.6T optics and next-gen switches will confirm cluster scale-outs tied to the AI infrastructure buildout.
- Enterprise attach: Proof points that large non-tech enterprises are committing to multi-year AI programs—especially with on-prem or co-lo inference—would broaden demand beyond hyperscalers.
Portfolio construction: mapping the stack
A barbell approach can capture the AI infrastructure buildout while managing risk. On one side, own the platform leaders in accelerators and system software that monetize both training and inference growth. On the other, hold select “picks-and-shovels” tied to HBM, advanced packaging, and high-speed networking—areas with structural bottlenecks and potentially longer-tailed pricing support. Rounding out the core, equipment names with exposure to AI logic and HBM process steps can provide cyclical diversification if industrial and consumer recovery remains uneven.
Risks to the call
- Overbuild in inference capacity if model efficiency leaps faster than expected or if AI monetization lags at the application layer.
- Supply overshoot in HBM or substrates should yields surprise to the upside—most likely a 2H 2026 story.
- Policy and geopolitics, including export-control volatility and incentives that redirect rather than expand total demand.
- Power constraints that delay cluster turn-ups, shifting revenue recognition to later quarters.
- Macro shocks that tighten financing or curb end-demand for AI-enabled services.
Bottom line
The AI infrastructure buildout is the defining capital cycle of this era, and 2026 looks set to deliver another year of semiconductor leadership. Cantor’s stance aligns with a growing body of data: hyperscaler budgets remain elevated, HBM and packaging constraints extend pricing power, and the beneficiary set is broadening beyond GPUs to the connective tissue—memory, networking, and equipment—that turns compute into scaled services. Execution and policy will continue to inject volatility, but the direction of travel remains favorable. For long-term investors, leaning into the stack—selectively and with an eye on bottlenecks—remains one of the most compelling ways to compound through the next leg of AI adoption.
FAQ
What exactly is driving semiconductor outperformance in 2026?
Hyperscaler and enterprise demand for training and inference capacity, tight HBM supply, and sustained investment in advanced packaging and networking fabrics that are essential to the AI infrastructure buildout.
Is the opportunity limited to GPUs?
No. While accelerators are core, memory, packaging, optics, switches, and wafer fab equipment are critical beneficiaries as the AI infrastructure buildout scales.
Could supply catch up and crush margins?
Some normalization is possible later in 2026, but packaging and substrate constraints should keep conditions constructive near term. Memory and advanced packaging remain the key swing factors.
What’s the biggest near-term risk?
Timing. Power availability, policy decisions, and delivery logistics can push recognized revenue between quarters, adding volatility even when multi-year demand is intact.
Disclaimer
This article is for information and education only and does not constitute investment advice or a solicitation to buy or sell any security. Investing involves risk, including the possible loss of principal. Do your own research and consider consulting a qualified financial advisor before making investment decisions.
