NVIDIA just posted the largest quarter any company has ever reported. Broadcom and AMD are closing the gap faster than expected. And Gartner's own numbers show why some of this spending won't survive to 2028.
- Gartner now projects $2.59 trillion in worldwide AI spending for 2026, up 47% year over year — revised upward from its own $2.52 trillion estimate published just four months earlier.
- NVIDIA closed fiscal 2026 with $215.9 billion in revenue (up 65%), then opened fiscal 2027 with an $81.6 billion quarter, up 85% year over year, pushing its market cap to roughly $5 trillion.
- Broadcom's AI chip revenue grew 143% year over year in its most recent quarter and the company has now guided to $56 billion in AI revenue for the full year, with a $73 billion order backlog behind it.
- Gartner also says AI is sitting in the "Trough of Disillusionment" through 2026, and that roughly 40% of current AI projects won't make it to 2028 — a gap between infrastructure spend and proven return that every buyer below needs to plan around.
Compute Has Replaced the Model as the Story
Generative AI's first two years were a story about models — who had the best chatbot, the sharpest image generator, the fastest-growing app. In 2026, the story has moved down a layer, to the physical infrastructure that makes any of it possible. GPUs, custom AI accelerators, high-bandwidth memory, networking gear, and the power systems that keep it all running have become the part of the AI economy that Wall Street, hyperscalers, and governments are now fighting hardest to secure.
The scale involved keeps surprising even the analysts tracking it. Goldman Sachs Research's baseline model puts annual AI capital expenditure across compute, data centers, and power at approximately $765 billion in 2026, climbing to roughly $1.6 trillion a year by 2031 — around $7.6 trillion in cumulative spend over that stretch. Separately, Goldman's equity analysts raised their combined 2025–2030 capex forecast for the four largest hyperscalers to $5.3 trillion, up from an already-large $4.5 trillion estimate just months earlier. That pattern — big numbers getting revised even bigger — has become the defining feature of this cycle.
AI Computing Power Sector Snapshot — 2026
Sources: Gartner, Goldman Sachs Research, NVIDIA/Broadcom/AMD earnings filings
$2.59T
Worldwide AI spending, 2026 (+47% YoY)
$765B
Goldman Sachs AI CapEx estimate, 2026
$215.9B
NVIDIA fiscal 2026 revenue (+65%)
~$5T
NVIDIA market capitalization
Global AI Spending by Segment, 2026
| Segment | Approx. Share of $2.59T |
|---|---|
| AI Infrastructure (servers, IaaS, network fabric, chips) | 45%+ |
| Software (including GenAI models) | ~25% |
| Cloud & IT Services | ~18% |
| Devices & Professional Services | ~12% |
Source: Gartner Worldwide AI Spending Forecast, published May 19, 2026 — up from a $2.52T estimate the same firm published in January.
AI Infrastructure Market Growth, 2026–2031
$101B
$116B
$134B
$154B
$177B
$202B
Compound annual growth rate ≈ 14.9%. Source: Mordor Intelligence AI Infrastructure Market Report, January 2026 update.
NVIDIA: The Largest Quarter Any Company Has Ever Reported
NVIDIA closed its fiscal 2026 (ended January 2026) with $215.9 billion in revenue, up 65% year over year, and full-year Data Center revenue of roughly $197 billion — the segment that now makes up the overwhelming majority of the company's business. The fourth quarter alone brought in $68.1 billion, a record, with Data Center revenue of $62.3 billion, up 75% from a year earlier.
Then the company topped it. Its first quarter of fiscal 2027, reported in late May 2026, brought in $81.6 billion in revenue, up 85% year over year, with Data Center revenue of $75.2 billion, up 92%. That single quarter pushed NVIDIA's market capitalization to roughly $5 trillion, making it the most valuable company in the world. Jensen Huang has told investors AI infrastructure buildouts will absorb "at least" a trillion dollars of NVIDIA's chips through 2027, and for now demand has kept pace with that claim — the company says Blackwell-generation GPUs are essentially sold out.
NVIDIA's AI accelerator market share is still estimated at roughly 70–80%, but that figure has started drifting down for the first time in this cycle, for reasons covered in the Broadcom section below.
Broadcom: The Custom-Chip Challenger Doubling Every Few Quarters
Broadcom isn't trying to out-GPU NVIDIA. Its strategy is custom AI accelerators (ASICs) co-designed with a small number of hyperscalers — Google, Meta, and others — that want chips built for their specific workloads instead of general-purpose silicon. That bet is paying off faster than even Broadcom projected: AI semiconductor revenue hit $8.4 billion in its first fiscal quarter of 2026, up 106% year over year, then jumped to $10.8 billion the following quarter, up 143%. Management has now guided to roughly $56 billion in AI revenue for the full 2026 fiscal year, backed by a $73 billion order backlog, and reiterated a target of approaching $100 billion in annual AI chip revenue by 2027.
A multiyear partnership with Meta to co-develop its MTIA chips — an initial commitment exceeding 1 gigawatt of capacity running through 2029 — illustrates why this matters beyond Broadcom's own income statement: custom ASIC shipments are projected to grow roughly 44.6% in 2026, nearly triple the growth rate of general-purpose GPUs, as more of the largest AI buyers try to reduce their dependence on any single supplier.
AMD: Building a Second Full Stack, Not Just a Second GPU
AMD remains well behind NVIDIA in raw AI accelerator share, but its first quarter of 2026 showed the gap narrowing on more than one front. Total revenue reached $10.3 billion, up 38% year over year, with Data Center segment revenue of $5.8 billion, up 57%, driven jointly by EPYC server CPUs and the ramp of Instinct MI350 Series GPUs. CEO Lisa Su described the results as a structural shift in the business, and the strategic logic behind that claim came into focus with a new agreement: Meta has committed to deploying up to 6 gigawatts of AMD Instinct GPUs across multiple product generations, including a custom chip built on AMD's next-generation MI450 architecture.
What's notable is the framing. AMD isn't pitching a single chip as an NVIDIA replacement — it's positioning EPYC CPUs, Instinct GPUs, and its Helios rack-scale system together as an alternative full-stack platform for hyperscalers wary of depending on one vendor for their entire AI infrastructure.
AI Compute Leaders — Head to Head
| Metric | NVIDIA | Broadcom | AMD |
|---|---|---|---|
| Position | Market leader (GPUs) | Custom ASIC specialist | Full-stack challenger |
| Latest annual/run-rate revenue signal | $215.9B FY2026 | $56B AI revenue guided, FY2026 | $10.3B total, Q1 2026 alone |
| Latest YoY growth | +85% (Q1 FY2027) | +143% (AI revenue) | +57% (Data Center) |
| AI accelerator share | ~70–80% | Leader in custom ASIC co-design | ~5–7%, rising |
| Signature product | Blackwell / Vera Rubin GPUs | Custom XPUs (Google, Meta) | Instinct MI350/MI450 + Helios |
Why Semiconductor Demand Looks Historically Unusual
Consumer electronics used to set the pace for the chip industry. That's no longer true. IDC's semiconductor trackers have pointed to global chip market growth exceeding 50% for 2026, with data center semiconductors alone approaching $500 billion — a demand shift toward AI servers, accelerators, networking silicon, and memory that has no real precedent in the industry's history. TSMC, which fabricates the advanced chips behind nearly every major AI accelerator on this list, has become as strategically important as any of the companies designing the chips themselves; its capacity constraints, not chip design, are increasingly the limiting factor on how fast this entire market can grow.
Inference Is Quietly Becoming the Bigger Business
Training headlines get the attention, but running AI models day to day — inference — is where a growing share of compute spending actually goes, by some industry estimates now approaching two-thirds of AI-related infrastructure use. Training clusters run hard for finite stretches; inference runs continuously, every time someone opens a copilot, an AI search result, or an automated support chat. That's a meaningful distinction for anyone modeling long-term demand: even if the pace of new model training slows, inference volume has its own independent growth curve tied to how many people and systems are actually using AI day to day.
The reality check inside Gartner's own numbers
The infrastructure spending in this piece is real and accelerating. What's less discussed is that the analyst firm behind the headline $2.59 trillion figure is also the one warning that AI sits in the "Trough of Disillusionment" through 2026, and that roughly 40% of current AI projects will not survive to 2028. Gartner's own analysts note that AI is now more often sold to enterprises by their existing software vendor than bought as a standalone initiative — a sign that buyers want proven ROI before committing new budget, not another platform pitch.
None of this contradicts the infrastructure numbers above — hyperscalers and chipmakers are, if anything, under-building relative to demand. But it does mean the enterprises paying for all this compute are under more pressure than ever to show it's producing something measurable, not just accumulating capacity.
Three Things That Could Actually Slow This Down
Power, not chips, is now the binding constraint. Analysts including Goldman Sachs and Morgan Stanley now describe a US data center capacity shortfall exceeding 11 gigawatts today, a gap some project could widen toward 49 gigawatts by 2028. In many regions, securing grid interconnection and transformers now takes longer than building the data center itself.
Manufacturing remains concentrated at a single chokepoint. TSMC produces the advanced nodes behind nearly every serious AI accelerator, from NVIDIA's GPUs to Broadcom's and AMD's custom silicon. Its capex plans keep rising, but so does demand — advanced-node supply is expected to stay tight through at least mid-2027 on most industry estimates.
The financing model is shifting under the buildout. Hyperscalers are increasingly funding data centers through debt and off-balance-sheet leases rather than pure cash flow; one widely cited estimate puts over $600 billion in future data center lease commitments not yet reflected on the major hyperscalers' balance sheets. That doesn't threaten near-term spending, but it does mean the AI buildout is now exposed to credit markets in a way it wasn't two years ago.
How AI Capital Flows Through the Ecosystem
↓
GPU & Custom Chip Vendors (NVIDIA, Broadcom, AMD)
↓
Chip Manufacturing (TSMC)
↓
Data Center Construction & Power Infrastructure
↓
AI Services, Copilots & Applications
Advisory: What This Means for You
Infrastructure spend and proven return are moving at different speeds right now. Here's how that gap should shape decisions by role.
Revenue growth alone no longer differentiates these names — all three chipmakers above are compounding fast. Watch backlog quality (Broadcom's $73B is contracted, not speculative), customer concentration risk, and how much of forward capex is debt-funded versus cash-funded, since that changes how a spending slowdown would actually hit the stocks.
Gartner's own 40%-project-failure estimate is a planning input, not a scare statistic. Before expanding an AI contract, require a baseline metric and a measured after-state — cost per resolved ticket, hours saved per workflow, error rate — the same way you'd underwrite any other infrastructure investment.
Custom ASIC demand growing nearly triple the pace of merchant GPUs signals that hyperscalers want multiple suppliers, not one dominant vendor. Companies positioned to co-design for a second or third hyperscaler customer — the way Broadcom and AMD both are — carry less concentration risk than single-customer plays.
Power, not chip supply, is the near-term constraint on this entire market. Regions that can offer fast grid interconnection and firm power commitments are effectively competing for the same data center investment dollars as any other economic development incentive — and increasingly winning or losing projects on that basis alone.
Inference economics, not training-cluster access, is where a growing share of durable value is likely to accumulate as usage scales. Building around lower-cost, higher-volume inference workloads may be a more defensible position than competing for scarce training capacity against hyperscaler-scale budgets.
Frequently Asked Questions
Estimates vary by scope. Gartner puts total worldwide AI spending at $2.59 trillion for 2026. Goldman Sachs Research's narrower AI capex model — covering compute, data centers, and power — puts the figure at approximately $765 billion for the year.
Yes, with an estimated 70–80% share of AI accelerator compute, but that share is narrowing for the first time as Broadcom's and AMD's custom-chip and GPU businesses grow faster, in percentage terms, than NVIDIA's own already-large base.
The honest answer is mixed. Physical demand for compute currently exceeds supply — GPUs are described as sold out and data center power is the binding constraint, not lack of demand. At the same time, Gartner's own research shows AI projects at the enterprise-adoption level failing at a high rate, and a growing share of data center capacity is being financed through debt and leases rather than cash. Those two things — real physical scarcity and financially stretched adoption — can both be true simultaneously.
Core Insights Review contributors publish research-based analysis and editorial insights on commercial real estate, PropTech, smart infrastructure, sustainable construction, industrial real estate, and emerging technologies shaping the future of the built environment.
➡️ Also Read: Forecasting AI Data Centers Electricity Demand: How Much Power Will Artificial Intelligence Consume by 2030?
➡️ Also Read: AI Data Centers and the Global Electricity Surge: Why Power Is Becoming the New Bottleneck of Digital Infrastructure
Check for more information: Core Insights Review
Follow us at: LinkedIn, Facebook, and X
