In 2026, Google plans to spend $40 billion on data centers and AI infrastructure—an amount larger than the GDP of Iceland—all before anyone knows whether the demand for generative AI will justify it.
Key Takeaways
- Google committed $40 billion to infrastructure in 2026, dwarfing most national economies.
- Amazon, Microsoft, and Meta are matching spending despite uncertain usage curves.
- Power consumption is becoming a hard limit—some projects are stalling over electricity access.
- The race isn’t about better models anymore; it’s about who can deploy the most silicon, fastest.
- If demand doesn’t materialize, we could see a cloud infrastructure glut by 2027.
Google’s $40 Billion Gamble
Google’s 2026 infrastructure budget isn’t just aggressive—it’s asymmetric. The company isn’t merely expanding. It’s overbuilding. The $40 billion figure, confirmed by CFO Anat Ashkenazi in a February earnings call, includes new data centers in Ohio, Tennessee, and Malaysia, all optimized for AI training and inference workloads. These facilities will run on dedicated power grids, some tied directly to nuclear microreactors under development with X-energy. Each data center is designed to house over 100,000 GPUs, primarily NVIDIA’s H100s and the upcoming B100s, arranged in ultra-dense configurations that require custom power delivery and cooling architectures.
What’s striking isn’t the scale, though that’s hard to ignore. It’s the timing. Generative AI adoption has plateaued outside enterprise workflows. Most developers still use open-source models or repurpose existing cloud instances. There’s no hockey stick in utilization. Yet Google is acting like one’s inevitable. The company is not waiting for customer demand to catch up. It’s banking on a future where every search, every ad impression, every video recommendation runs through a generative model in real time. That future would require 10x to 100x more compute than today’s peak loads. Google is building for that scenario now, even if it means years of underutilized capacity.
That’s the new logic of the AI race. It used to be about model size, then inference speed, then cost per token. Now it’s about pre-emptive saturation. Build it before they come. Hope they come.
The Shift From Innovation to Construction
Three years ago, the AI headlines were about parameters: who trained the biggest model, who cracked multimodal reasoning, who released a foundation model that didn’t hallucinate. Today, the milestones are cubic feet and megawatts.
Microsoft is leasing 500 acres in Wisconsin for a new AI campus, with plans for 12 buildings and a 1.2-gigawatt power draw—enough to run a mid-sized city. The project, located near the Mississippi River for access to water cooling and barge transport of equipment, is being built in partnership with Bechtel and expected to go live by Q4 2026. Amazon Web Services has fast-tracked 17 new availability zones globally, each outfitted with tens of thousands of H100s and their successors. In Dublin, AWS is constructing a 160-megawatt facility that will draw power from Ireland’s national grid, a move that’s sparked regulatory scrutiny over renewable supply. Meta has shifted its entire 2026 capex to “AI-adjacent infrastructure,” according to CFO Susan Li, including fiber backbones and edge compute pods for AI-driven ad targeting. In Singapore, Meta is deploying AI inference nodes within 10 milliseconds of 90% of Southeast Asia’s internet users, a latency play aimed at real-time content personalization.
This isn’t R&D. It’s civil engineering with GPUs.
Power Is the New Bottleneck
Compute was the limit in 2022. Then it was memory bandwidth. Now it’s electricity.
In Virginia’s Loudoun County—dubbed “Data Center Alley”—utilities have stopped approving new connections. The grid can’t handle more load. Google halted construction on a $1.8 billion facility in Jackson, Tennessee, after the local power co-op said it couldn’t guarantee supply beyond 2028. The Tennessee Valley Authority, which supplies power to much of the region, has capped new data center draws at 50 megawatts per site, a fraction of what Google and Microsoft are requesting. In Oregon, Amazon had to scrap plans for a 300-megawatt facility after Portland General Electric said it couldn’t source enough carbon-free energy to meet state mandates.
“We’re hitting physical constraints,” said Brian Lesser, CEO of Global Infrastructure Partners, in a panel at the original report. “You can’t just wish for more megawatts. You have to build the grid, the substation, the fuel source. That takes years.”
The Cooling Problem No One Talks About
More power means more heat. And liquid cooling systems can’t scale fast enough.
Most new data centers are switching to direct-to-chip cooling, but the pumps, manifolds, and fluid distribution units aren’t being manufactured at the pace required. One supplier, Allied Control, says lead times have stretched from 8 weeks to 6 months. Some operators are reusing decommissioned oil drilling pumps to move coolant—a jerry-rigged fix at trillion-dollar scale. In Ohio, Google is testing immersion cooling with 3M’s Novec fluid, which submerges entire server racks. But the fluid is expensive—about $250 per liter—and supply is limited. 3M has paused expansion of its Novec production due to EPA scrutiny over PFAS chemicals.
And water? Don’t count on it. In Chile’s Atacama Desert, a planned AI cluster was scrapped after local farmers protested the use of scarce groundwater for cooling towers. Sustainability claims only go so far when you’re pulling 800 megawatts in a drought zone. Even in water-rich regions like Scandinavia, local governments are pushing back. Norway’s energy minister recently stated that “AI data centers will no longer get priority access to hydropower” unless they contribute directly to national innovation goals.
The Cloud Glut That Could Come
We’ve been here before. In the early 2000s, telecom companies laid hundreds of thousands of miles of fiber, betting on an internet usage boom. Most went bankrupt. The capacity was real. The demand wasn’t there yet.
Now, history could repeat—not with bandwidth, but with AI compute.
Consider the numbers:
- Global AI infrastructure spending in 2026: $180 billion (AI Business estimate)
- Year-over-year growth: 47%
- Percentage of new data centers dedicated to AI: 68%
- Utilization rate of AI-optimized instances (Q1 2026): 39%
That last one matters. If less than four in ten GPUs are in active use, what happens when all this new capacity comes online by Q3?
Prices will drop. Margins will shrink. Smaller cloud providers may not survive. And the big three—AWS, Azure, Google Cloud—will be stuck with underused assets they can’t decommission easily. These data centers aren’t like office buildings. They’re specialized facilities with 15-year depreciation cycles. Once built, they need to run at high utilization to break even. If they don’t, cloud providers will have to slash prices to fill capacity—potentially triggering a price war that could last years.
It’s not just capital at risk. It’s credibility. These companies told investors that AI would drive exponential revenue growth. If usage doesn’t spike, the narrative collapses. Alphabet’s stock, for instance, has traded at a premium since 2023 based on AI monetization potential. A sustained miss on utilization could knock 10–15% off its valuation, according to analysts at Bernstein.
Who’s Actually Using All This Compute?
Enterprises are buying AI services, yes—but mostly for narrow tasks: document summarization, customer support routing, basic code generation. The moonshot applications—real-time video generation, autonomous supply chains, generative design at scale—aren’t live at meaningful volume.
Startups are even more conservative. Most can’t afford sustained access to AI cloud tiers. They’re using cached models, distillation, or hybrid local-cloud setups to stretch budgets. A typical AI startup spends $20,000–$50,000 per month on cloud inference, which limits how often they can retrain or scale. Some, like Hugging Face and Anyscale, are offering model optimization tools to help startups reduce token usage, but that only delays the cost crunch.
And consumers? They’re not paying for AI. They’re using it embedded in apps, for free. Which means no direct revenue stream to justify infrastructure costs. Google’s AI-powered Search Generative Experience (SGE) is used by 120 million people monthly, but doesn’t generate incremental ad revenue—yet. Meta’s AI chatbots in WhatsApp and Messenger are free. OpenAI’s GPT-4o is bundled into Microsoft 365, with no standalone consumer pricing.
So who’s the customer here? At this point, the answer seems to be: the future.
The entire industry is betting that by 2028, there will be demand for 100x more inference than today. That multimodal agents will run constantly in the background of every digital interaction. That every enterprise workflow will be AI-mediated. No one can prove it. But everyone’s building as if it’s certain.
The Bigger Picture: Geopolitics and Energy Security
The AI infrastructure race isn’t just a tech story. It’s a geopolitical one. The U.S., China, and the EU are treating AI compute as strategic infrastructure, like railroads or 5G. In China, the government is funding a $60 billion AI industrial park in Hangzhou, set to house 5 million servers by 2028. It’s being built by Alibaba, Tencent, and state-owned China Mobile, with power supplied by new nuclear and hydro plants. The goal isn’t just economic—it’s about reducing reliance on U.S. chip exports and maintaining control over domestic data flows.
In the U.S., the Department of Energy has quietly fast-tracked permitting for small modular reactors (SMRs) in Idaho and Wyoming, specifically to power AI data centers. The first, a 30-megawatt X-energy reactor, will power a Google facility in Wyoming by 2027. The Pentagon is also investing in AI-ready infrastructure, with the Defense Innovation Unit funding mobile microgrids that can support field-deployed AI systems in conflict zones.
The EU is taking a different path. It’s capping data center power use at 100 megawatts per site unless operators can prove 90% renewable sourcing. France has banned new AI data centers in Île-de-France (including Paris) until 2027 due to grid strain. These policies could slow European AI development but may force more efficient computing practices. Startups like Mistral AI in France are focusing on smaller, specialized models that use 10% of the energy of GPT-4—potentially giving Europe a long-term edge in sustainable AI.
What Competitors Are Doing Differently
Not every company is betting on brute-force infrastructure. Some are trying to outmaneuver the giants with efficiency. Tesla, for example, isn’t building traditional data centers. It’s using its Dojo supercomputer—designed for autonomous driving—to train AI models at lower cost. Dojo’s custom D1 chips and tiled architecture reduce energy per inference by 30% compared to GPU clusters, according to internal benchmarks. Tesla plans to open Dojo to external users by 2027, offering AI training at half the price of AWS.
Oracle is taking a niche approach. It’s partnering with Anthropic to offer AI services optimized for enterprise workloads, with strict data privacy controls. The infrastructure is smaller—just 40,000 GPUs by 2026—but tightly integrated with Oracle’s database and cloud stack. This appeals to regulated industries like banking and healthcare, where data sovereignty matters more than raw speed.
Meanwhile, open-source efforts are gaining ground. The AI collective EleutherAI is developing modular models that can be mixed and matched, reducing the need for massive centralized training. Projects like Llama 3 and Mistral’s Mixtral allow companies to run capable AI locally, bypassing cloud costs entirely. Red Hat has started offering “AI-in-a-box” servers pre-loaded with open models, targeting mid-sized firms that can’t afford cloud-scale AI.
These alternatives won’t stop the infrastructure boom, but they might limit its fallout. If efficient AI becomes viable at scale, the glut could be less severe. The winners may not be the ones with the most data centers—but the ones who use them best.
What This Means For You
If you’re a developer, this spending spree means cheaper compute is coming—eventually. Once supply outstrips demand, cloud providers will slash prices to fill capacity. That could make high-end inference affordable for indie builders and small teams. But it may take 12 to 18 months, and only after bigger players absorb the losses.
If you’re building AI products, don’t assume the infrastructure will stay this abundant. Power constraints could lead to regional shortages. Some countries may cap AI energy use. And if a cloud glut triggers a price war, smaller vendors might disappear—taking your APIs with them.
What if the demand never comes? What if AI’s real limit isn’t compute—but usefulness?
Sources: AI Business, The Information
