AI Infrastructure — Competitive Dynamics & Investment Implications

Neutral landscape analysis of the AI infrastructure value chain, competitive positioning, and structural investment themes

May 12, 2026 Sector: AI / Data Infrastructure Scope: Public Equities

Research Summary: Hyperscaler AI capex is running at ~$434B TTM (Q1 2026), approaching $500B for CY2026. NVIDIA retains 70-80% of AI accelerator revenue, but the competitive landscape is evolving faster than at any prior point. AMD is gaining credibility, custom silicon is a structural headwind (15-25% of hyperscaler compute today, growing to 30-40% by 2028), and the value chain is bifurcating between winners (HBM memory, networking, power) and segments under margin pressure (commodity servers, general cloud). The bull/bear debate hinges on whether AI workload demand grows fast enough to absorb both custom silicon displacement and capex scrutiny.

1. Competitive Positioning

NVIDIA — Moat Assessment

NVIDIA holds an estimated 70-80% share of AI accelerator revenue and 85-90%+ of the training GPU market specifically. Blackwell (B100/B200) transition began in late 2025; GB300 (Blackwell Ultra) is now in production ramp.

CUDA Software Ecosystem

Very High

15+ years, millions of developers, 3,000+ libraries. PyTorch becoming more backend-agnostic, but ROCm still 2-3 years behind.

Full-Stack Integration

High

NVSwitch, NVLink, DGX/MGX reference designs create deep switching costs. Customers building custom interconnects as response.

Talent & Research Velocity

High

Fastest product cadence in semiconductors (annual architecture updates). Competition for ML engineers intensifying.

Customer Lock-in

Medium

Enterprise SW layers (NeMo, cuDNN, CUDA graphs) create stickiness, but large hyperscalers deliberately multi-sourcing to reduce leverage.

Erosion signals to watch: Microsoft co-design with NVIDIA (but heavy Maia deployment), Meta using NVIDIA AND custom silicon concurrently, DeepSeek efficiency revolution showing model efficiency can compress hardware demand per unit output.

AMD — MI300/MI400 Progress

AMD has captured an estimated 5-10% of AI accelerator revenue by end of 2025, up from sub-3% in 2023. Growth concentrated in inference workloads where MI300X's 192GB HBM3 capacity advantage is decisive for large-model serving.

Product	Timeline	Key Advantage	Adoption
MI300X	Late 2023 - present	192GB HBM3; large KV cache for inference	Microsoft Azure, Meta, Oracle Cloud
MI350 (CDNA 4)	2025 launch	H200-class performance; improved ROCm	Ramping at hyperscalers
MI400	Late 2026 (rumored)	TSMC N3; next-gen architecture	Pre-production

Key constraint: AMD's challenge is software ecosystem, not hardware specs. ROCm 6.x has materially improved PyTorch/JAX compatibility, but CUDA parity remains partial for sparse computation, custom kernels, and debugging tools.

Custom Silicon — The Structural Threat

This is the most structurally important competitive threat to the merchant GPU market over a 3-5 year horizon.

Player	Product	Internal Workload Share	External Availability	Impact on Merchant GPU
Google	TPU v5p / v6 "Trillium"	~30-40% of training	Google Cloud only	~$3-5B/yr diverted from NVIDIA
Amazon	Trainium 2 / Inferentia 3	Growing (Nova models)	AWS Neuron SDK	Rising; cost reduction incentive
Microsoft	Maia 100 / Maia 2	~10-15% of inference	Internal only	Moderate; training still NVIDIA

Aggregate impact: Custom silicon handles ~15-25% of total hyperscaler AI compute (early 2026), concentrated in inference. Growing to 30-40% by 2028 — but the total pie is growing fast enough that absolute merchant GPU demand still increases.

Broadcom — The Custom ASIC Enabler

Broadcom has emerged as the most investable pure-play on custom silicon as the neutral party designing XPUs for hyperscalers who want in-house AI chips.

Metric	Value
AI Revenue (FY2024)	$12.2B (up from ~$2.3B in FY2022 — 5x in 2 years)
FY26E Revenue Growth	+68% (accelerating)
XPU SAM (2027E)	$60-90B from just 2-3 customers
Known XPU Customers	Google, Meta, + third (likely Apple or ByteDance)
Gross Margin (AI)	High-60s to low-70s %

Networking — InfiniBand vs. Ethernet

Vendor	Position	AI Cluster Use Case	Trend
NVIDIA (InfiniBand)	Dominant in training	GPU-to-GPU interconnect; closed system	Stable share in training
Arista Networks	Leading Ethernet AI fabric	400G/800G switching; Meta, Microsoft	Gaining share in inference
Cisco	Losing AI share	Nexus 9000; weaker RDMA support	Declining relevance
Broadcom	Silicon supplier	Tomahawk 5 (51.2Tbps) merchant ASIC	Powers Arista + white-box

The Ultra Ethernet Consortium (AMD, Intel, Microsoft, Meta, Arista, Broadcom) is pushing open alternatives to InfiniBand. Inference scale-out has shifted toward Ethernet for cost/scale reasons. Arista is the prime beneficiary.

Memory — HBM Bottleneck

High Bandwidth Memory (HBM) is the single most constrained component in the AI accelerator stack. HBM pricing runs 3-5x the dollar-per-bit vs. conventional DRAM.

Company	HBM Share	Status	Advantage
SK Hynix	~50%+	Market leader; first to HBM3E	Yield leadership; sole initial H100 supplier
Samsung	~30-35%	Qualified after yield delays	Scale; closing technology gap
Micron	~15-20% (growing)	HBM3E qualified; gaining allocation	US manufacturing; CHIPS Act; geopolitical hedge

2. Value Chain Analysis — Where Value Accrues

Segment	Gross Margin	Moat Durability	Margin Pressure
AI Accelerators (NVIDIA)	70-75%	Very High	Low (pricing power intact)
Custom ASIC Design (Broadcom)	65-70%	High	Low (long design-win cycles)
HBM Memory (SK Hynix, Micron)	50-55%	High	Medium (cycle risk by 2027)
Advanced Packaging (TSMC CoWoS)	~53% (blended)	Extreme	Low (capacity constrained)
AI Networking ASICs (Broadcom)	60-65%	High	Low
AI Networking Switches (Arista)	63-65%	High	Low (software moat)
Power Infrastructure (Vertiv, Eaton)	35-42%	Medium-High	Medium
Server OEMs (SMCI, Dell, HPE)	10-15%	Low	High (commodity assembly)
Generic Cloud Compute (IaaS)	Varies	Medium	High (pricing pressure)

Key insight: Value concentrates at the design layer (NVIDIA, Broadcom, Marvell) and at physical bottlenecks (TSMC fabrication, nuclear power, HBM memory, liquid cooling). Assembly/integration (SMCI, Dell) captures the least margin despite high revenue throughput.

Most Durable Competitive Advantages

CUDA ecosystem lock-in (NVIDIA): Software switching costs are enormous — 15+ years of developer tools and libraries
Advanced node manufacturing (TSMC): Cannot be replicated quickly; 2nm and below remains TSMC's domain
HBM IP and yield know-how (SK Hynix): Proprietary packaging and stacking expertise with years of lead
Hyperscaler XPU relationships (Broadcom): Long design-win cycles (2-3 years) create multi-year revenue visibility
Nuclear operating licenses (CEG, TLN): Irreplaceable 24/7 baseload power; permitting takes 5-10 years

3. Bull vs. Bear Debates

Bull Case: Still Early Innings

Hyperscaler capex commitments unprecedented and growing — $500B CY2026E
Inference demand accelerating, not replacing training — each AI query = GPU cycles
Sovereign AI programs ($100B+ globally) barely started
DeepSeek efficiency = Jevons Paradox: cheaper AI expands applications, drives MORE compute
Physical AI (robotics, AV) opens entirely new TAM wave
TSMC advanced nodes remain sold out through 2026

Bear Case: Cycle Peak Risk

Capex-to-revenue correlation weakening; ROI scrutiny accelerating
Custom silicon + efficiency gains could shrink GPU TAM by $20-30B/yr from peak
Model efficiency (distillation, quantization, MoE) compresses hardware per workload
Inventory build risk if demand slows even temporarily
China restrictions block $10-15B/yr potential market for NVIDIA
~5 companies represent majority of GPU demand — concentration risk

Cycle-Turning Metrics to Watch

Metric	Bull Signal	Bear Signal
Hyperscaler capex guidance	Sequential increases; margin of beat	Flattening or cuts; pull-forward language
NVIDIA GPU lead times	6+ months for new product	Normalizing to <12 weeks
HBM pricing	Spot > contract pricing	Spot < contract; declining QoQ
TSMC CoWoS utilization	>95%	Below 85%
Hyperscaler AI revenue growth	100%+ YoY (Azure AI, Bedrock)	Deceleration to <50%
NVIDIA gross margin	Sustained >70%	Compression toward 65%
Arista/Broadcom book-to-bill	>1.1x	Sub-1.0x
AMD AI GPU share pace	Gaining 1-2%/quarter	Stalling below 10%

4. Key Risks

Geopolitical (US-China Chip Restrictions)

Export controls progressively tightened (Oct 2022, Oct 2023, 2024-2025); H20 chip under additional restrictions
NVIDIA China revenue fell from ~20-25% to ~5-10% of data center revenue
Huawei Ascend 910C achieves ~60-80% of H100 performance; filling domestic gap
Taiwan risk: TSMC manufactures >90% of advanced AI chips. Any disruption catastrophic with no near-term alternative. Arizona fab (N4/N3 by 2026) provides partial mitigation only.

Demand Sustainability

Hyperscaler "digestion periods" after large GPU buildouts — utilization rates matter
Enterprise AI spend slower than hyperscaler; ROI proof points still needed
AI startup capex meaningful but not immune to funding cycles
Open source model commoditization may reduce training intensity

Technology Shifts

Neuromorphic / photonic / analog: All in R&D or early commercial — not competitive at scale yet, but 5-10 year uncertainty
Quantum computing: Not a meaningful AI infrastructure threat in 2026-2030 window
Sparse/MoE architectures: Already deployed (GPT-4, Mixtral, Gemini) — structural model efficiency trend compressing per-token compute

Supply Chain Concentration

Chokepoint	Concentration	Geographic Risk
Advanced foundry (TSMC)	~90%+ of leading-edge AI chips	Taiwan (geopolitical)
EUV lithography (ASML)	100% sole supplier	Netherlands (export controls)
HBM memory	3 suppliers only	South Korea (SK Hynix, Samsung)
CoWoS packaging	TSMC dominant	Taiwan (expanding to Japan)

5. Investment Themes

Training vs. Inference — The Structural Shift

2023-2024 was training-dominant. 2025-2026+ is shifting toward inference as models deploy to mass markets.

Dimension	Training	Inference
Compute density	Very high (weeks-long runs)	Lower per query; massive parallelism
GPU preference	NVIDIA B200/GB300 (NVLink critical)	Flexible: AMD MI300X, custom ASICs, edge chips
NVIDIA moat strength	Strongest	Narrower (cost/perf competition opens)
Networking	InfiniBand dominant	Ethernet viable; Arista benefits
Memory	HBM3E/HBM4 critical	HBM for large models; GDDR for smaller

Beneficiaries of inference shift: AMD (MI300X capacity advantage), Arista (Ethernet), Broadcom (custom inference ASICs), Qualcomm (edge), Micron (HBM for inference servers).

Thematic Expressions

Theme	Best Expression	Thesis
AI spending continues	NVDA, AVGO	Direct beneficiaries of every AI dollar spent
Power is the bottleneck	TLN, CEG, BE	Physical constraint creates irreplaceable pricing power
Inference > Training	ANET, DELL, AMD, APH	Broader beneficiary set; non-NVIDIA options viable
Under-the-radar AI infra	WCC, APH, FIX	Industrial names with AI growth at value multiples
Picks and shovels	TSM, LRCX, ONTO	Win regardless of which chip/architecture wins
HBM / memory cycle	SK Hynix, MU	Constrained supply, premium pricing, CHIPS Act
Custom silicon enabler	AVGO, MRVL	Benefit from the exact trend threatening NVIDIA

Underappreciated Areas

Area	TAM	Key Players	Why Overlooked
Optical interconnect / CPO	$10B+ by 2027	Coherent, Lumentum, Marvell, Ayar Labs	Copper limits not yet widely understood
Power semiconductors (VRMs)	High NVIDIA content/socket	Monolithic Power (MPWR), TI	Classified as analog semis, not AI
AI storage / NVMe	Growing with dataset sizes	Pure Storage, Western Digital	Storage seen as legacy
Cabling & connectors	400G/800G fiber in tight supply	Amphenol, TE Connectivity	"Boring" industrial classification
Liquid cooling infrastructure	$30B+ by 2028	Vertiv, Asetek, Green Revolution	Transition from niche to mainstream underway
Semiconductor equipment (yield)	Complex multi-die packaging	KLA (KLAC), Lam Research	Seen as cyclical semi-cap

6. Highest Conviction Positions

NVIDIA (NVDA) — 26x Fwd P/E

Best risk/reward in AI hardware. CUDA moat + Blackwell cycle are real. 65% growth at $216B scale with 60% margins. Paradoxically the cheapest major AI stock.

Broadcom (AVGO) — 37x Fwd P/E

Custom silicon enabler benefiting from the exact trend that threatens NVIDIA. Revenue accelerating to 68%. Networking + VMware optionality. More predictable XPU ramp.

Arista Networks (ANET) — 38x Fwd P/E

Best pure-play on Ethernet AI cluster networking. 35% growth, 43% operating margins. Winning the InfiniBand-to-Ethernet transition at inference scale.

Vertiv (VRT) — 54x Fwd P/E

Most direct power/thermal play on AI data center density increases. $15B backlog. Under-owned by tech investors. Liquid cooling transition doubles content per rack.

Micron (MU) — US-listed HBM exposure

Cheapest US-listed HBM play. CHIPS Act beneficiary. Geopolitical hedge vs. Korea-concentrated SK Hynix. Growing NVIDIA allocation.

Tactical / Higher Risk-Reward

Ticker	Thesis	Key Risk
Marvell (MRVL)	Custom ASIC + optical DSP for hyperscalers; re-accelerating	Execution risk; customer concentration
Coherent (COHR)	Optical components for AI cluster interconnect; early innings	Volatile; transition timing
Talen Energy (TLN)	Nuclear power at 17x; Amazon 1.92 GW PPA	Regulatory; nuclear operational
WESCO (WCC)	DC segment +70% at 22x blended; undiscovered	Industrial cyclicality; margin

Avoid / Underweight

Ticker	Reason
Cisco (CSCO)	Losing AI networking share; Splunk distraction; no custom silicon bet
Intel (INTC)	Gaudi not gaining traction; foundry losing ground; turnaround uncertain
Super Micro (SMCI)	Revenue real but margins thin (10-15%); governance/audit risk is binary
Palantir (PLTR)	87-102x P/E; no margin of safety despite great growth

Confidence & Data Notes

Category	Confidence	Basis
NVIDIA market share, HBM dynamics, Broadcom AI revenue	High	Multiple public earnings disclosures
Custom silicon share estimates, AMD precise share	Medium	Hyperscalers don't disclose internal allocation
China restriction specifics, CoWoS utilization rates	Lower	Rules change frequently; TSMC doesn't disclose granular util