The $1 Trillion Data Center Market Is a Budget Redistribution, Not Just Growth

The Number Behind the Number

The headline is that global data center capex will exceed $1 trillion in 2026. But the more important story is what sits inside that number and what gets displaced to make room for it.

Dell'Oro Group raised its 2026 data center capex outlook above $1 trillion as hyperscale AI deployments and server pricing accelerated. The research firm increased its forecast following stronger-than-expected spending in the first quarter, noting that the four largest U.S. cloud providers — Amazon, Google, Meta, and Microsoft — increased data center capital expenditures by 78% year-over-year. That growth is not distributed evenly across the infrastructure stack. It concentrates in one layer: AI accelerators.

Accelerated servers for AI training and domain-specific workloads could account for approximately two-thirds of total data center infrastructure spending by 2030. This implies a sharp structural reordering of budget priorities away from general-purpose compute, storage, and networking — not necessarily a contraction in absolute dollars, but a clear demotion in relative importance. Traditional infrastructure vendors should not mistake a growing market for a friendly one.

The Semiconductor Layer

Zoom into the silicon and concentration becomes stark. In 2024, the total semiconductor addressable market for data centers reached $209 billion, spanning compute, memory, networking, and power. By 2030, that figure is projected to grow to nearly $500 billion.

Within that pool, GPUs dominate and NVIDIA dominates GPUs. GPUs remain the cornerstone of AI infrastructure, with NVIDIA capturing 93% of server GPU revenue in 2024. Yole Group forecasts GPU revenue will grow from $100 billion in 2024 to $215 billion by 2030. That is a doubling of an already enormous market — while the broader semiconductor TAM also expands. The GPU segment is not just growing with the market; it is taking share from it.

The operational consequences are material: NVIDIA's gross margins run approximately 75% on data center silicon, reflecting both architectural advantages and CUDA software lock-in. Those margins flow directly from concentration. When one vendor controls 93% of a $100 billion market, pricing becomes unilateral.

The CUDA Moat Is the Real Constraint

Most operators frame NVIDIA dependence as a hardware procurement problem. It is actually a software problem. NVIDIA's strongest defense remains CUDA, the programming ecosystem with over 5 million active developers and two decades of library optimization. Every major ML framework — PyTorch, JAX, TensorFlow — runs on CUDA natively. Switching means rewriting production workloads against immature alternative toolchains, not just signing a new contract.

The exit routes being discussed in the market — AMD's ROCm, Google's JAX/XLA, Amazon's Neuron SDK — all demand meaningful engineering investment to reach production parity. AMD's challenge with the MI300X and MI350 series is not hardware; it is ROCm software ecosystem maturity. Hardware parity is necessary but not sufficient for displacement.

Hyperscalers Are Playing a Different Game

The large cloud providers understand this, which is why custom silicon programs function as leverage plays, not cost optimizations. AI ASICs are gaining momentum, with Google, Amazon, and Microsoft investing in domain-specific silicon to optimize performance and reduce dependence on NVIDIA. AI ASICs will reach almost $85 billion in 2030 as hyperscalers pursue vertical integration and cost control.

In 2026, this shift is accelerating past merchant GPU shipments for the first time. TrendForce projects 44.6% ASIC growth against 16.1% for merchant GPUs in 2026. These programs are now in production at scale. Microsoft's Maia 200 claims 30% better performance per dollar than the best hardware in its existing fleet. Maia 200 currently serves GPT-5.2 models for OpenAI and powers Microsoft 365 Copilot workloads from its Des Moines data center. Google's TPU program is further along: Google's TPU v7, codenamed Ironwood, was announced at Cloud Next in April 2025. Each chip delivers 4,614 FP8 TFLOPS with 192 GB of HBM3E memory at 7.37 TB/s bandwidth.

The catch: these investments are not transferable to the enterprise market. A mid-size company cannot threaten to build its own TPU to negotiate better NVIDIA pricing. The bargaining chips that Google and Amazon have built over five years are unavailable to enterprises without the scale for custom silicon. Those enterprises are price takers.

The Secondary Constraints

Capex forecasts model silicon spend. They underweight the physical constraints that will bind deployments before chips arrive. The trillion-dollar threshold is supported by well over 10 million high-end accelerators as the primary capex driver, along with related infrastructure. That density requires power and cooling most enterprise data centers were not built to provide.

Alongside investments in AI clusters, hyperscalers continued expanding general-purpose cloud infrastructure to support public cloud growth, agentic AI workloads, and rapidly growing storage requirements. Dell'Oro projects spending growth will accelerate further in the second half of 2026 as NVIDIA Rubin-based systems enter production deployments and hyperscalers refresh custom AI accelerator platforms. Rubin is the next forcing function on power density. Operators planning GPU deployments in 2026 and 2027 should audit facility power and cooling capacity before signing purchase commitments.

Outside hyperscale, enterprise data center investment remains constrained by tariffs, monetary policy, and uncertain AI returns. The trillion-dollar story is largely a hyperscaler story. Enterprises are watching utilization metrics carefully before extending commitments — the right posture given the capital intensity.

What to Watch

1. NVIDIA Rubin deployment data in 2H 2026. Capex growth is expected to accelerate further in 2H26, driven by the ramp of NVIDIA Rubin systems and refresh cycles for hyperscaler custom accelerator platforms. Rubin's thermal profile and rack density requirements will determine whether current facility infrastructure can absorb the next generation without a full power retrofit.

2. AI ASIC revenue share crossing 20% of accelerator spend. AI ASICs are projected to reach almost $85 billion in 2030. Watch for the quarterly crossover where hyperscaler custom silicon displaces enough merchant GPU spend to show up in NVIDIA's data center revenue growth rate as deceleration.

3. CUDA alternatives reaching production readiness. OpenAI's Triton has emerged as an off-ramp, allowing developers to write hardware-agnostic kernels in Python. Triton now features backends for Google's TPU, AWS Trainium, and AMD's MI350 series. When Triton or similar abstraction layers achieve production-grade reliability across multiple backends, NVIDIA's software moat narrows — and enterprise switching costs drop with it.

4. TSMC capacity as the binding constraint. Every chip in the hyperscaler ASIC market — Google TPU, AWS Trainium, Microsoft Maia — manufactures at TSMC, which produces approximately 92% of advanced AI chips at 7nm and below. The hyperscaler ASIC market, the independent ASIC market, and the NVIDIA GPU market all rely on the same manufacturing base. Any disruption at TSMC reprices every forecast in this piece simultaneously.