In February 2026, Alphabet (Google’s parent company) announced a record-breaking $84.75 billion equity financing plan—the largest single-round equity raise ever by a global tech company, surpassing any previous fundraising event by any enterprise. The financing structure consists of three layers: a public offering of $34.8 billion (Class A/C common stock and mandatory convertible preferred shares), a $40 billion At-the-Market (ATM) program, and a $10 billion private placement by Berkshire Hathaway. The simultaneous arrival of these three capital sources has transformed the "AI compute arms race" from a technical narrative into a true capital war.
Capital Logic: Why Did Alphabet Launch Unprecedented Financing in 2026?
To understand this financing, it’s essential to break down both its scale and timing.
In terms of scale, the $84.75 billion corresponds to Alphabet’s 2026 capital expenditure (CAPEX) budget, which ranges from $180 billion to $190 billion. The vast majority of this will be allocated to AI data centers, server clusters, and related infrastructure. Alphabet’s announcement explicitly stated that CAPEX in 2027 will "increase significantly"—indicating the company is still in the early stages of its expansion cycle. For a clearer perspective, consider this comparative data: the four major tech giants (Alphabet, Amazon, Microsoft, Meta) collectively invested about $410 billion in AI infrastructure in 2025, with projections jumping to around $650 billion in 2026—a surge of more than 58%. Alphabet alone accounts for roughly 30% of this total.
From a timing perspective, the first half of 2026 coincides with the intersection of three key factors:
- Supply Constraints: AI chips and data center construction remain supply-constrained, with market bottlenecks in production capacity rather than demand. Alphabet must secure critical equipment and capacity ahead of time to avoid being squeezed by other cloud providers.
- Capital Structure Shift: Ultra-large cloud computing firms are transitioning from a "self-balancing cash flow model" to a "moderate leverage model." Historically, these companies financed CAPEX through internally generated cash flow, but the current pace of AI investment has outstripped the natural growth of free cash flow. Debt and equity financing have become necessary supplemental channels.
- Berkshire’s Signal Effect: The $10 billion private placement includes participation from Berkshire Hathaway—a value-investing powerhouse that has rarely added major tech stock positions since 2011. Berkshire’s involvement signals a highly conservative validation of Alphabet’s large-scale CAPEX plans from a long-term asset return perspective.
It’s important to note that Alphabet is not alone. Amazon’s 2026 AI investment target is about $200 billion, Microsoft’s is projected at $120 billion, and Meta has revised its target upward to $125–$145 billion. The five ultra-large cloud providers (Amazon, Alphabet, Microsoft, Meta, Oracle) are expected to spend a combined $660–$690 billion in CAPEX in 2026, nearly doubling the 2025 total. This isn’t just a single company’s financing story—it’s a financial reflection of an industry-wide paradigm shift.
Google Data Center Expansion: From Regional Architecture to Compute Power Map
To grasp the underlying logic of this capital expenditure cycle, we need to look at the asset side—specifically, the global deployment of physical infrastructure. Alphabet’s CAPEX ultimately translates into data centers, fiber networks, and power infrastructure around the world. Over the past 12 months, Google has announced several large-scale international data center projects, forming a pattern of "deepening in Europe + expansion in Asia-Pacific + penetration in Latin America":
| Region | Investment Scale | Time Span | Core Details |
|---|---|---|---|
| Germany | €5.5 billion (about $6.3 billion) | 2026–2029 | New Dietzenbach data center, expansion of Hanau campus |
| Belgium | €5 billion (about $5.8 billion) | 2026–2027 | Expansion of St. Ghislain campus, addition of 110MW+ carbon-free power |
| Sweden | Undisclosed | Starting 2026 | First Horndal data center, air cooling + waste heat recovery |
| UK | £5 billion (about $6.8 billion) | 2025–2028 | Waltham Cross data center operational since September 2025 |
| India | About $15 billion | 2026–2030 | Partnership with AdaniConnex, GW-scale AI data center cluster |
| US | Not disclosed as a whole | Ongoing | Multiple state-level data center expansion projects |
A common feature of these deployments is site selection based on the abundance of renewable energy and grid stability. In Belgium, Google’s expansion locks in more than 110MW of carbon-free power from suppliers like Eneco and Luminus. In Sweden’s Horndal project, Google uses air cooling and waste heat recovery systems. All these designs point to a central theme: energy constraints for AI infrastructure are shifting from secondary to primary bottlenecks.
AI Infrastructure Value Chain: The Logical Progression from Chips → Power → Cooling
Breaking down capital expenditure into physical assets reveals a clear value chain:
Chip Design and Manufacturing: From GPU Monopoly to ASIC Multipolarity
Chip investment—especially in GPU/ASIC and supporting network chips—accounts for the largest share of AI infrastructure spending, roughly 40%–50% of total CAPEX. NVIDIA still dominates the market, with its Blackwell platform holding over 70% of high-end GPU shipments. The GB200/B200 series, fueled by 2025 orders, is expected to supply through the second half of 2026. To address capacity shortages and reduce reliance on a single supplier, major cloud providers are accelerating their in-house ASIC strategies—reshaping the AI chip supply landscape.
Google’s TPU strategy is a hallmark of this shift. Its latest AI accelerator, TPU V7P, is set for mass production in 2026, designed by Broadcom and manufactured using TSMC’s 3nm process. At the same time, Google’s first Arm-based Axion CPU has been deployed, replacing the traditional x86 host processors and forming a fully self-developed compute ecosystem. Broadcom is benefiting from this trend: its AI chip revenue reached $20 billion in fiscal 2025, and its Q1 AI semiconductor revenue for fiscal 2026 is projected at $8.2 billion.
However, risks remain. Macquarie recently downgraded Broadcom from "outperform" to "neutral," citing Google’s accelerated in-house chip development and supplier diversification—MediaTek has entered Google’s AI chip supply chain. Broadcom’s share of Google TPU-related revenue is expected to drop from about 95% in 2026 to 65% by 2028. The ASIC sector is moving from Broadcom’s relative dominance to multi-vendor competition, posing potential pressure on Broadcom’s long-term margins.
Additionally, Google’s internal ASIC strategy directly impacts external suppliers. While Broadcom will benefit from Google’s CAPEX growth in the short term, dilution of its market share is already drawing attention.
Power Layer: Revaluing Energy Infrastructure
The average power demand per 10,000 square feet of AI data centers has tripled in the past three years. This nonlinear increase in energy density means power supply is shifting from an "operational cost item" to a "scalability constraint." Data center PUE (Power Usage Effectiveness) has improved from about 1.8–2.0 in the early days to below 1.2, but absolute power growth driven by compute expansion keeps pushing up the share of power costs.
Google’s signed renewable energy power purchase agreements in Europe now total over 4.5GW of installed capacity, plus additional carbon-free power contracts for data centers in Belgium and Sweden. At least three levels of beneficiaries can be observed: engineering contractors upgrading the grid, traditional and new energy power generators with long-term supply contracts, and equipment suppliers for storage and grid optimization solutions. The US Energy Information Administration (EIA) expects US grid infrastructure CAPEX growth to accelerate further in 2026–2027, with 20%–25% attributable to AI data center power demand.
Cooling Layer: The Inflection Point for Liquid Cooling Penetration
The power density of AI servers—especially NVIDIA Blackwell high-end GPU clusters—has far exceeded the cooling capacity limits of traditional air cooling. The NVIDIA GB200 NVL72 cabinet consumes over 120kW per rack, making liquid cooling not just an option, but a necessity.
Google’s TPU v7 clusters and latest Axion CPU-driven servers have deployed fourth-generation liquid cooling systems, marking a rapid breakthrough in liquid cooling penetration for ultra-large-scale deployments. The liquid cooling supply chain includes coolant providers (such as 3M and Fluorocarbon), cooling distribution unit manufacturers, server cold plate designers, and data center cooling system integrators. Compared to the GPU chip sector’s concentrated landscape, this track is more fragmented, but technical barriers and certification cycles provide a temporary moat for early movers.
Risk Framework: Sustainability of the CAPEX Cycle
Every capital-intensive investment cycle must face the test of return assumptions. AI infrastructure investment carries three levels of potential risk:
Demand Sustainability Risk: The marginal growth rate for AI model training compute demand is expected to slow in 2026–2027, shifting from annual doubling in 2023–2025 to a 30%–50% growth range. If inference-side compute demand doesn’t fill this gap in time, data center utilization rates may experience temporary declines.
Oversupply Risk: The five ultra-large cloud providers are expected to spend $650–$690 billion in CAPEX in 2026, while combined AI cloud service revenue for the same period is projected at $250–$300 billion. This gap means CAPEX payback periods will be measured in years, with realization highly dependent on large-scale commercialization of inference applications.
Technology Substitution Risk: Google’s internal ASIC strategy has already diluted Broadcom’s market share, demonstrating that customers can sustainably capture value from the supply chain through technological substitution. Any supplier currently enjoying excess profits in the AI infrastructure value chain may see those windows gradually close as clients pursue vertical integration.
Conclusion
Alphabet’s $84.75 billion financing in 2026 marks the transition of AI infrastructure construction from the first stage of "racing to build capacity" to the second stage of "full-chain efficiency optimization." The current investment expansion is still accelerating, but the market’s focus is shifting from "who is spending" to "who is spending most effectively"—namely, unit compute cost, unit energy output, and capital return cycles.
For ecosystem participants, this shift means the duration of benefit windows varies significantly by position: upstream players at key technology nodes (such as high-end GPUs and advanced process foundries) enjoy the longest windows; midstream general hardware suppliers face higher substitution risks; downstream energy and cooling infrastructure providers have more stable demand visibility.
From a longer-term perspective, the current annualized CAPEX of about $650 billion reflects AI’s deep penetration as a general-purpose computing platform in the industrial economy. If this assumption holds, we are still in the relatively early stages of the infrastructure investment cycle—making this the central reference point for understanding the capital race in the era of super data centers.
