Examine the rapidly developing AI infrastructure stack and determine which companies and technologies are positioned to control the next critical layer of computing.
Executive Summary
The AI boom is no longer simply a race to build better models. It has become an infrastructure race. Behind every AI model sits a massive physical and digital stack: advanced semiconductors, high-bandwidth networking, data centers, electricity, cooling, storage, cloud platforms, and orchestration software.
The four largest U.S. hyperscalers — Amazon, Alphabet, Microsoft, and Meta — are expected to spend approximately $760 billion on capital expenditures in 2026, nearly four times their combined spending in 2022. Combined with the $500 billion Stargate build-out led by OpenAI, SoftBank, and Oracle, total AI infrastructure spending has cleared $1 trillion for the first time. Goldman Sachs estimates cumulative hyperscaler capex could reach as much as $7.6 trillion between 2025 and 2030.
NVIDIA remains the most visible winner. Its fiscal 2026 revenue reached $215.9 billion, with Data Center revenue of $193.7 billion. But the infrastructure economy is spreading well beyond GPUs. Broadcom, TSMC, Arista Networks, hyperscalers, data-center operators, and specialized AI clouds are all becoming critical parts of the system.
At the same time, electricity is becoming a strategic constraint. Gartner forecasts global data center electricity consumption will reach 565 TWh in 2026, up 26% from 2025, with power availability now directly impacting growth.
Who owns the infrastructure underneath AI — and who captures the economics?
The important question for investors, founders, and technology companies is no longer simply: who builds the best AI model?
1. The AI Infrastructure Boom
The first phase of the AI revolution was dominated by models. OpenAI, Anthropic, Google, Meta, and other companies competed to build increasingly capable systems.
The second phase is different. AI workloads are becoming industrial-scale workloads requiring enormous amounts of compute, memory, networking, and electricity.
That changes the economics of the industry. A large AI cluster is not simply a collection of GPUs. It requires servers, networking equipment, storage, power distribution, cooling systems, data-center space, software, and long-term energy contracts.
The infrastructure therefore becomes a system. And when demand for one component increases, pressure spreads throughout the entire stack.
GPU shortages create pressure on networking.
Networking increases power requirements.
Power constraints delay data-center construction.
Data-center shortages push customers toward specialized cloud providers.
Approximately 50% of hyperscaler data center spending is on silicon, and that is also the primary area where wealth creation has been most deeply concentrated. But the spending is cascading through every layer of the stack. The AI economy is therefore becoming an infrastructure economy.
2. Mapping the New AI Stack
The AI infrastructure stack can broadly be divided into several layers.
At the bottom are semiconductors and advanced manufacturing. This includes NVIDIA, AMD, Broadcom, and TSMC.
Above that sits compute infrastructure: servers, accelerators, memory, and specialized rack-scale systems.
Then comes networking, where companies such as Broadcom and Arista Networks connect thousands of processors into large AI clusters.
Above that is the physical infrastructure: data centers, cooling, power systems, and electrical equipment.
Then comes cloud infrastructure, dominated by Microsoft Azure, Amazon Web Services, and Google Cloud, alongside specialized AI clouds such as CoreWeave and emerging providers.
Finally, there is the software and orchestration layer that manages models, workloads, data, and compute.
This matters because value can migrate between layers. If GPUs become more abundant, networking may become more valuable. If compute becomes cheaper, inference demand could increase dramatically. If power becomes scarce, access to electricity could become more valuable than access to land.
The bottleneck determines the economics.
Whoever controls the scarcest layer captures the profit.
3. The Chip Layer: NVIDIA's Dominance and the ASIC Insurgency
NVIDIA remains the central company in the AI infrastructure economy. It holds an estimated 80–85% of the data center AI accelerator market by revenue in 2026, down from approximately 92% in 2023, but still overwhelmingly dominant. Its strategy has evolved from selling chips to selling complete AI factory systems.
At GTC 2026, NVIDIA raised its order outlook for Blackwell and Vera Rubin platforms through 2027 to at least $1 trillion, up from the $500 billion demand estimate provided at GTC 2025. The Vera Rubin platform represents a fundamental architectural shift: rather than a GPU upgrade, Rubin integrates GPU, CPU, networking, and liquid cooling design into a rack-level system.
NVIDIA has also moved aggressively to finance the buildout. In August 2026, the company announced strategic partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent compute financing platforms aimed at mobilizing over $500 billion of third-party capital for AI infrastructure.
But the most significant structural shift in the chip layer is the rise of custom ASICs. CSPs' self-developed ASICs have grown rapidly, reaching 8–11% market share in 2026, with AMD at approximately 5–8% and Intel trailing below 3%. JPMorgan projects global AI accelerator shipments will reach approximately 16.3 million units in 2026, up 62% year-over-year, with NVIDIA and AMD GPUs accounting for about 58%, down from 68% in 2025.
Broadcom has emerged as the primary architect of this bespoke compute era. The company now has six major custom silicon customers, with OpenAI joining the ranks of Google and Meta. Broadcom's Tomahawk 6, the world's first 102.4 Tbps switching silicon, facilitates the interconnection of million-node XPU clusters. Its Co-Packaged Optics (CPO) technology integrates optical engines directly onto the ASIC package, reducing power consumption by 3.5× and lowering the cost per bit by 40%.
4. Advanced Manufacturing and Packaging: TSMC's Chokepoint
TSMC is expanding production capacity five times faster than usual and still cannot meet AI-driven demand. The world's largest contract chipmaker is constructing 25 wafer fabrication and advanced packaging facilities globally in 2026, including 13 in Taiwan — a pace described as unprecedented for the company.
Advanced packaging has become a critical bottleneck. TSMC's chip-on-wafer-on-substrate (CoWoS) technology is used to assemble processors for AI accelerators. The company has committed $56 billion in capex to double CoWoS capacity to a projected 130,000–150,000 wafers per month by late 2026, supporting the Rubin R100 slated for full production in late 2026.
The economics of leading-edge manufacturing are also shifting. A single 2nm fab now costs over $25 billion, approaching three times the cost of a 7nm-era facility. This is driving an "advanced process + advanced packaging" dual-wheel model that pushes the industry toward system-level integration.
5. The Memory Supercycle: HBM as Strategic Resource
High-bandwidth memory (HBM) has become a strategic resource in the AI infrastructure stack. Goldman Sachs forecasts the global HBM market will reach $54.6 billion in 2026, up 58% year-over-year, accounting for nearly 40% of the DRAM market.
The supply-demand imbalance is severe: despite Samsung, SK Hynix, and Micron allocating 70% of new or reallocable capacity to HBM, the HBM capacity gap remains at 50–60%.
SK Hynix, benefiting from deep collaboration with NVIDIA, is expected to maintain over 50% global market share in HBM3E and next-generation HBM4. The company's 2026 HBM capacity is essentially fully booked by customers, signaling a complete seller's market.
Watch this: SK Group Chairman Chey Tae-won has warned that wafer supply shortages could persist until 2030 due to the massive HBM consumption of expanding AI models.
6. Networking: Ethernet vs. InfiniBand
AI clusters require enormous amounts of data to move between processors. That makes networking one of the most important — and least visible — parts of the AI infrastructure economy.
Ethernet switch sales in AI back-end networks more than doubled in Q1 2026, accounting for about two-thirds of data center switch sales in AI clusters. InfiniBand sales more than tripled during the quarter, supported by the ramp of 800 Gbps switches shipping with NVIDIA's Blackwell Ultra platform, but Ethernet maintained a clear lead.
The vendor landscape remains highly dynamic. Celestica regained the leading position in AI back-end Ethernet switches during Q1 2026, followed closely by NVIDIA, with Arista ranked third despite significant deferred revenue.
Broadcom is positioning itself at the center of this convergence — with Tomahawk 6 switching silicon, Co-Packaged Optics, and Jericho 4's "Cognitive Routing" capability. NVIDIA, meanwhile, is pursuing a "copper in, optical out" strategy, using copper within racks for energy efficiency and cost while introducing Co-Packaged Optics for cross-rack scaling.
7. Cloud Infrastructure: Hyperscalers vs. Neoclouds
The cloud infrastructure market has crossed a symbolic threshold: more than $500 billion in annual run-rate revenue. AWS holds 28% market share, with Microsoft Azure at 20% and Google Cloud at 15%.
But the AI-specific cloud market is evolving faster than the broader cloud market. Gartner's 2026 Magic Quadrant for Cloud AI Infrastructure places AWS, Google, Microsoft, and Oracle as leaders, with CoreWeave, Nebius, and Crusoe as visionaries.
CoreWeave has emerged as the leading pure-play AI cloud. The company raised its 2026 capital spending forecast to $35–39 billion, up from $31–35 billion, driven by surging AI demand. Its revenue backlog reached $104.2 billion in Q2 2026, with more than $25 billion in additional net new customer commitments secured in the current quarter.
Meanwhile, Amazon's custom chip business, Trainium, has reached an annualized revenue run rate exceeding $20 billion, growing at triple-digit rates year-over-year. Arm architecture CPUs now hold approximately 50% market share among top hyperscalers, with AWS Graviton and Trainium leading the shift.
8. Power and Cooling: The New Bottleneck
Ask anyone building at scale what actually limits them in 2026, and the answer has shifted away from GPU supply and toward electrons. Server rack loads have risen from around 3 kW for general-purpose compute to as much as 150 kW for AI inferencing — a jump the cooling and power distribution infrastructure built for the last decade was never designed to absorb.
Gartner forecasts global data center electricity consumption will reach 565 TWh in 2026, up 26% from 447 TWh in 2025. AI-optimized servers will account for 31% of total data center power consumption, with their electricity consumption rising to 175 TWh from 95 TWh in 2025. Cooling systems alone are forecast to jump 22.6% this year to 195 TWh.
The power constraint is reshaping where and how data centers are built. A modern data center can be permitted and built in 2–3 years, but the power to run it could take 5–7 years for natural gas, 10+ years for nuclear, and 2–4 years even for solar — while grid interconnection queues stretch beyond five years in many U.S. regions.
Liquid cooling has shifted from optional to essential. Next-generation AI architectures are fully adopting liquid cooling, elevating it from a configuration option to a specification requirement.
9. The Startup Layer: Where Innovation Meets Capital
The AI infrastructure boom has spawned a wave of specialized startups attracting significant capital.
| Company | Raise | Valuation | Focus |
|---|---|---|---|
| Positron AI | $875M | $5B | Inference chips |
| Upscale AI | $190M | $2B | AI infrastructure |
| Velaura AI | $110M | $1B+ | Ultra-low-power silicon |
| Netris | $15M | — | Network automation for AI neoclouds |
Positron AI raised $875 million in September 2026, more than quadrupling its valuation in seven months to $5 billion, with $500 million led by NEA and Jim Clark. Its Asimov chip targets production in the second half of 2027.
Netris, a network automation startup that helps AI neoclouds accelerate deployment, raised $15 million in Series A funding led by Andreessen Horowitz (a16z). The company's platform is now deployed across more than 35 GPU clusters globally, covering approximately 1 million GPUs, with customers including Lightning AI and Foxconn. As the market matures, value is migrating toward companies that solve specific bottlenecks — whether in power efficiency, network automation, or ultra-low-power compute.
10. Big Tech Investment and Strategic Positioning
The scale of Big Tech's AI infrastructure investment has reached unprecedented levels. Based on upper-end estimates as of July 30, 2026:
| Company | 2026 CapEx (Est.) | Strategic Focus |
|---|---|---|
| Amazon | $220B | AWS GPU clusters, Trainium ASICs |
| Alphabet | $195–205B | TPU deployment, Google Cloud AI |
| Microsoft | ~$190B | Azure AI, OpenAI partnership |
| Meta | Up to $145B | Hyperion and Prometheus AI campuses |
Amazon is at the top with a planned $220 billion in 2026 spending, raised from an earlier $200 billion estimate after strong AWS growth. Alphabet increased its forecast to $195–205 billion, Microsoft projects about $190 billion, and Meta expects up to $145 billion.
TrendForce estimates the combined 2026 CapEx of the top nine CSPs will exceed $886.7 billion, with the five North American hyperscalers accounting for nearly 90%. Looking to 2027, the combined CapEx of the top nine CSPs is expected to hit approximately $1.3 trillion.
Risk flag: This spending is increasingly debt-funded. Meta's move to fund its El Paso data center through off-balance-sheet special purpose vehicles pushed bond yields north of 7%, while Oracle raised debt directly and was downgraded to one notch above junk. Morgan Stanley puts the sector's off-balance-sheet obligations at close to $1.65 trillion.
11. M&A and Partnership Activity
The infrastructure race is driving significant consolidation and strategic partnerships.
The $40 billion Aligned Data Centers acquisition — completed in July 2026 by MGX, the Artificial Intelligence Infrastructure Partnership (AIP), and BlackRock's Global Infrastructure Partners (GIP) — is one of the largest private investments in digital infrastructure to date. Aligned's portfolio encompasses 51 campuses with over 6.4 GW of operational and planned capacity. The consortium committed an additional $5 billion in growth capital.
Nscale's acquisition of Anyscale — the AI compute platform built by the creators of the open-source Ray framework — reflects the neocloud land grab. The deal is expected to close in the second half of 2026.
NVIDIA's reported $14 billion acquisition of Hugging Face would be a landmark deal, giving the chipmaker control of the leading open-source AI hub and further integrating the hardware and software layers.
NVIDIA's $5 billion investment in Intel represents a strategic realignment of historic rivals. The two companies will co-develop x86 system-on-chips integrating NVIDIA RTX GPU chiplets, with NVIDIA gaining access to Intel's advanced manufacturing and packaging capacity.
Broadcom's extended partnership with Meta to deploy technology supporting multi-gigawatts of Meta's custom silicon (MTIA) underscores the deepening ties between custom silicon designers and hyperscalers.
12. The Emerging Competitive Landscape
The competitive dynamics of AI infrastructure are evolving along several axes:
- Vertical integration vs. horizontal specialization. NVIDIA is integrating downward into full AI factory systems while Broadcom is building platform-level solutions spanning custom silicon, networking, and optics. Meanwhile, hyperscalers are integrating upward into custom silicon.
- The ARM vs. x86 battlefront. ARM architecture CPUs now hold approximately 50% market share in accelerated servers, overtaking x86 for the first time — driven by AWS Graviton and Trainium, Google's Axion, and the broader ARM ecosystem's efficiency advantages.
- The power-constrained geography. Data center developers are increasingly treating grid capacity as a site selection criterion rather than an afterthought. Securing long-term power purchase agreements has become as important as securing land or capital.
- Sovereign AI and regional diversification. Governments are increasingly investing in domestic AI infrastructure. NVIDIA's partnerships with Australian firms to build AI factories supporting up to 2 GW by 2027, and its collaboration with Japan on national-level physical AI infrastructure, reflect the geopolitical dimension.
13. Value Capture: Who Earns the Returns?
Morgan Stanley estimates that GenAI infrastructure could achieve approximately a 25–50% capital return in a base-case scenario. AI lab-paid inference business margins have surged from low double digits (10–20%) in 2025 to 50–65% or higher in 2026, with gross margins improving by 30–50 percentage points.
But the returns are not evenly distributed. The market currently exhibits a pattern of "upstream earns first, downstream customers spend first."
- NVIDIA, TSMC, SK Hynix, and Broadcom are capturing value at the component level.
- Hyperscalers are converting capex into cloud revenue, with Alphabet's cloud backlog surpassing $240 billion and Microsoft sitting on an $80 billion queue of Azure orders it cannot yet fulfill.
The critical question is whether the massive capital deployment will generate sufficient returns before the debt burden becomes unsustainable. UBS calculates an annual financing gap of approximately $400–500 billion for 2026–2028 AI capex, with debt accounting for $150–200 billion, representing about 10–15% of total revenue and investment.
14. The Bottleneck is the Business
In the AI infrastructure economy, the bottleneck is the business. Wherever the constraint lies — whether in HBM capacity, CoWoS packaging, power availability, or network automation — that is where pricing power and strategic value concentrate.
In 2026, the bottleneck has shifted from chips to power. This shift is repricing the entire infrastructure stack. Companies that control access to electricity — whether through long-term power purchase agreements, on-site generation, or grid interconnection rights — are gaining leverage over those that control land or capital.
The arms race is not about who has the best model. It is about who controls the physical and digital infrastructure that makes models possible.
The next computing layer will not be controlled by a single company or technology. It will be controlled by the companies that solve the most binding constraints across the stack:
- NVIDIA controls the compute layer.
- Broadcom controls the custom silicon and networking layer.
- TSMC controls the manufacturing chokepoint.
- SK Hynix controls the memory layer.
- And increasingly, utilities, energy infrastructure providers, and companies that can deliver gigawatt-scale power are becoming the most critical — and most undervalued — players in the AI infrastructure economy.
In that race, the winners will be determined not by who spends the most, but by who solves the most critical bottlenecks at scale.
Key Takeaways
- AI infrastructure spending has crossed $1 trillion and is accelerating toward $7.6 trillion cumulative by 2030.
- The bottleneck has shifted from chips to power — electricity, not GPUs, is now the binding constraint.
- NVIDIA remains dominant but is losing share to custom ASICs from Broadcom, Marvell, and hyperscaler in-house designs.
- HBM memory is a strategic chokepoint — SK Hynix controls over 50% of the market and is sold out through 2026.
- Networking is the next battleground — Ethernet is winning over InfiniBand, but the vendor landscape is highly dynamic.
- Power and cooling have become specification requirements, not afterthoughts, reshaping data center siting and design.
- Debt-funded capex is a growing risk — off-balance-sheet obligations approach $1.65 trillion.
- Value is migrating toward bottleneck solvers — companies that solve the most binding constraints capture the most durable strategic value.
The CODEW Take
The AI infrastructure economy is entering its most important phase. The first phase was about proving that AI could work. The second phase is about building enough infrastructure to make AI commercially useful at scale.
That requires enormous amounts of capital. It also creates enormous opportunities.
NVIDIA has demonstrated the economics of owning a critical layer of the stack. Broadcom and Arista are showing how networking becomes increasingly valuable as clusters scale. TSMC remains critical because advanced AI chips ultimately depend on advanced manufacturing and packaging.
Data-center operators are becoming strategic infrastructure owners. Cloud providers are becoming capital-intensive AI utilities. Specialized AI clouds are testing whether speed and specialization can overcome the balance-sheet advantage of hyperscalers. And energy is becoming part of the technology strategy.
The biggest mistake would be to view this simply as a GPU cycle. It is much larger. The AI buildout is creating a new infrastructure economy in which compute, power, networking, data centers, storage, and software are increasingly interconnected.
The winners will be the companies controlling the bottlenecks. The losers will be the companies that mistake spending for competitive advantage.
Who owns the infrastructure that makes intelligent machines economically possible?
That is where the next great technology fortunes may be created.
Sources: NVIDIA, AMD, Broadcom, TSMC, Microsoft, Amazon, Alphabet, Meta, Goldman Sachs, Morgan Stanley, JPMorgan, UBS, Gartner, TrendForce, Dell'Oro Group, IEA, McKinsey, CoreWeave, SK Hynix, Reuters, and company filings.
The CODEW Special Report is an in-depth editorial research and analysis series examining the technology companies, markets, transactions, strategies, and emerging trends shaping the next phase of the digital economy.