The AI Data Center Bill Keeps Rising. These 7 Stocks Could Collect.
Bloomberg sees 2027 capacity rising about 10% while its projected buildout cost jumps roughly 70%
Listen to the audio version of this article (generated by AI).
What cost $9 billion last year and $14.6 billion today – without getting one watt bigger?
A gigawatt of data-center capacity, according to Bloomberg Intelligence’s freshly revised modeling. The gigawatt didn’t change. What goes inside it did.
Bloomberg raised its 2027 global capacity forecast from 108 gigawatts to 119 – about a 10% bump. But its estimate of what that capacity will cost jumped from about $1 trillion to $1.7 trillion – a 70% leap.
Every new gigawatt is being stuffed with more accelerators, more high-bandwidth memory, more storage, faster networking, and far heavier power and cooling equipment than the model assumed a year ago.
The original AI trade revolved around a handful of chip companies. Bloomberg’s new math highlights how that money spreads across the entire building.
Here’s where I think it flows next…
Why the AI Data Center Buildout Just Got 70% More Expensive
Bloomberg sees the global data-center market growing from roughly $630 billion in 2025 to $1.06 trillion in 2026, $1.74 trillion in 2027, and $2.8 trillion by 2030.
Capacity climbs alongside it – 119 gigawatts in 2027 and 229 by 2030, up from a prior 2030 estimate of 189.
And the buyer base keeps widening. Hyperscalers started this boom, neoclouds followed, and now sovereign AI programs, governments, and large enterprises are building capacity of their own.
More buyers are chasing a more hardware-intensive buildout. Put those forces together, and Bloomberg’s 2027 spending estimate is roughly $700 billion higher than before. So, where does all that extra money go?
Across nearly every layer of the data center.
1. Compute: Nvidia Still Gets the Biggest Check
One category collects the largest check by far.
Bloomberg expects compute servers and GPUs to absorb roughly $1.1 trillion of the 2027 market – nearly two-thirds of the total.
Nvidia (NVDA) remains the clearest anchor for that spending. Its latest quarterly data-center revenue hit $89 billion, up 117% year over year, and the Vera Rubin platform is now ramping across the major cloud providers.
What’s driving the growth? Frontier models keep getting bigger. Reasoning systems spend more time on each request. Agents keep working long after the initial prompt, and more companies and countries are building AI at once. Every one of those trends consumes more compute.
Nvidia remains my core holding for this layer of the boom. But Bloomberg’s data shows the opportunity gets much bigger once you look past the processor.
2. Memory: Small Models, Giant Appetites
Included in Bloomberg’s model is the finding that might surprise you most: when it comes to high-bandwidth memory (HBM), small models aren’t small.
Under its modeled workloads, even a modest 7-billion-parameter model could require roughly 1.6 terabytes of high-bandwidth memory – surprisingly close to what far larger models need.
That’s because inference memory goes well beyond storing the model itself. The system needs working memory – the key-value cache – which balloons as context windows stretch and more users pile on simultaneously, plus headroom to run reliably. A compact model serving thousands of long-running agents can chew through enormous amounts of memory.
That makes Micron Technology (MU) one of my favorite plays here. It’s already shipping HBM4 for Vera Rubin, more than 75% of its fiscal 2027 output is committed, and customers are discussing allocations into 2028 and beyond.
And every big usage trend – more users, longer conversations, agents that never log off – pulls on that same memory supply.
3. Storage: AI’s Quietest Spending Explosion
AI has to remember far more than most investors realize – training sets, model checkpoints, video and sensor data, agent logs, customer records, generated files, and everything needed to audit a system when something goes wrong.
Bloomberg sees total storage growing from roughly $125 billion in 2025 to $612 billion in 2030 – and the data-center slice rises much faster. Nearline hard drives and data-center flash were about $39 billion of spending in 2025; Bloomberg’s 2030 mix implies roughly $360 billion. That’s nearly a ninefold increase.
Seagate Technology (STX) shines in this category. Its high-density drives give hyperscalers an economical way to retain massive pools of training data and AI-generated content, and its latest HAMR platform is already qualified with major cloud customers.
Sandisk (SNDK) is the faster, more aggressive flash angle. Hard drives preserve the data lake; flash serves whatever needs answers right now. The AI economy needs plenty of both.
4. Networking: More AI Chips Create More Traffic Between Them
An AI cluster can hold tens of thousands of processors. As clusters grow, networking content grows with them – and persistent agents turn that traffic from bursts into a constant flood. Custom accelerators add even more connections among CPUs, GPUs, memory, and storage.
That’s Marvell Technology’s (MRVL) sweet spot. It participates in custom silicon and supplies the high-speed optical and networking technology that moves information through these systems. A hyperscaler can design its own accelerator; but it still needs that chip wired to everything around it. Marvell is my preferred higher-growth networking play on that traffic.
5. Power and Cooling: The GPU Is Useless Without Electricity
AI racks pull extraordinary amounts of electricity into a tiny footprint and convert most of it into heat. That creates two immediate jobs: get enough power to the rack, and carry the heat away.
Vertiv Holdings (VRT) sells the equipment for both, spanning power distribution, backup, liquid cooling, thermal management, and modular infrastructure. And it’s expanding into microgrids and behind-the-meter power so data centers can make their own electricity instead of waiting in the grid’s years-long connection line.
The old bottleneck was getting the GPU. The next one may be getting that GPU enough electricity to switch on. Vertiv is my favorite power-and-cooling stock for this phase.
6. Copper and the Grid: AI Spending Doesn’t Stop at the Data Center
A data center’s bill of materials doesn’t stop at its own walls.
Bloomberg estimates direct data-center copper use could rise from roughly 390,000 metric tons in 2025 to about 1.3 million tons by decade’s end – and the generators, substations, transformers, and transmission lines feeding those campuses could require another 770,000 tons. Combined, that’s potentially more than 2 million tons of copper by 2030.
Rather than guessing at daily copper prices, I prefer the operating play: Quanta Services (PWR), which designs, builds, and maintains the electrical infrastructure connecting new generation to enormous loads like data centers. Its backlog hit $53.4 billion in Q2, including $43.8 billion of electric-infrastructure work.
The chips sit inside the building; Quanta makes sure the electricity reaches them.
Why I Prefer the AI Infrastructure Suppliers
A $1.7 trillion bill raises an obvious question: who actually pays it?
Bloomberg’s model shows U.S. hyperscalers swinging from a modest 2025 cash surplus to funding gaps of $130 billion in 2026, $297 billion in 2027, and $140 billion in 2028.
That doesn’t stop the boom. But it does mean operators will lean harder on debt, private credit, infrastructure funds, and joint ventures, and maybe trim buybacks along the way.
Returns will be uneven, too: Bloomberg models AWS at roughly 24% ROIC in 2026, while newer infrastructure providers start much lower or below zero as they spend ahead of revenue.
This is exactly why I prefer the suppliers.
The operator has to finance the project, fill the capacity, and eventually earn a return. The companies selling chips, memory, storage, networking, cooling, and electrical gear get paid while the campus is going up.
The pattern running through Bloomberg’s entire model – operators financing while suppliers collect – is the same one at the center of my XPANSE research.
There’s a reason this buildout has Washington’s attention, too. One high-ranking government official has warned that America’s dependence on foreign-made chips could end in what he calls “an economic apocalypse.”
The race you just read about – trillions of dollars, flowing through every layer of the data center – is also the race to prevent that outcome. And Elon Musk’s next great project sits right in the middle of it.
In my latest briefing, I lay out the three steps I’d take today to get positioned – and give away the name and ticker of my favorite stock for the theme.
Every gigawatt is getting more expensive. Own the companies sending the invoices.
