Follow the Megawatts: What the 2026 AI Spending Numbers Are Really Saying
Updated: 8 minutes ago
PARKLAND, FL | Joaquin "Jack" Fagundo on what the 2026 AI and data center spending numbers really mean once you look past the headlines.
Put a few CIOs in a room lately and the conversation finds its way to the same question, dressed up a dozen different ways. Is all this AI spending real? And what does it mean for everyone who isn't a hyperscaler?
I hear it from board members, from clients, and from neighbors here in Parkland. My answer has settled into two parts. The money is real, and the forecasts keep climbing. But the companies that come out ahead won't be the ones that spend the most. They'll be the ones with access to power, the patience to build in stages, and the discipline to make every phase earn the next one.
A forecast that won't sit still
Gartner now expects worldwide AI spending to reach $2.67 trillion in 2026, up 49.5% from 2025. Infrastructure is the largest piece, about $1.48 trillion, or roughly 56% of the total. In January, Gartner's number for 2026 was $2.52 trillion. It has raised the forecast twice since then, adding about $143 billion, and most of that increase landed in infrastructure. Its outlook for 2027 is $3.64 trillion.
I pay more attention to revisions than to headlines. When a forecast moves up every time it's published, demand is running ahead of the people paid to model it.
Four companies, one enormous line item
The clearest signal comes from the hyperscalers' own guidance. After third quarter earnings at the end of October, the full year 2026 picture looks like this: Amazon at about $220 billion in cash capex, Alphabet at $195 billion to $205 billion, Microsoft at about $170 billion for calendar 2026, and Meta at $135 billion to $145 billion. At the midpoints, that's roughly $730 billion from four companies in a single year. Each counts capex a little differently, so I treat the total as directional.
Two details interest me more than the total. Meta spent $34.9 billion on capex in the third quarter, more than its operations generated in cash, which left free cash flow at about negative $2.1 billion for the quarter. And Microsoft trimmed its calendar year outlook because third party data center capacity arrived late.
Sit with that for a moment. The best capitalized companies on earth are waiting on buildings and power. The constraint has moved off the balance sheet and onto the construction site. Alphabet and Amazon have both said they expect to spend significantly more in 2027.
The constraint has moved off the balance sheet and onto the construction site.
Reading the long range forecasts
The five year estimates point the same way. McKinsey projects that data centers will need about $6.7 trillion of capital worldwide by 2030, $5.2 trillion of it tied to AI workloads, and expects global demand for capacity to nearly triple, to about 219 gigawatts, with roughly 70% of it coming from AI. Goldman Sachs models $5.3 trillion of capex from the four largest hyperscalers between fiscal 2025 and 2030.
I read those as scenarios, not promises. McKinsey's own range for AI data center investment runs from roughly $3 trillion to $8 trillion, which is the honest way to frame something this uncertain. So my advice to CIOs is simple. Plan for the middle of the range. Design every program so you can pause at the end of any phase. Make each phase prove its return before the next one gets funded. McKinsey, for what it's worth, lands in the same place.
The real budget is power
Every data center conversation I have now ends at the utility. The International Energy Agency estimates global data center electricity use at about 485 terawatt hours in 2025 and projects roughly 945 TWh by 2030 in its base case, close to what Japan uses in a year today. In the United States, data centers are on track to drive nearly half of electricity demand growth through 2030, and the IEA warns that about one in five planned projects could be delayed by grid constraints.
Some perspective from the operator's chair. Supporting The Walt Disney Company, I ran a footprint of five data centers with more than 50 MW of capacity. That was a serious enterprise estate. McKinsey now puts a typical AI optimized facility at 20 to 30 MW on its own, against 5 to 10 MW for a traditional one. The unit of planning has changed underneath all of us. Racks and square footage used to drive the budget. Now it starts with megawatts, interconnection dates and cooling. This fall that same power story reached household bills, which is one of the AI warning signs I wrote about recently.
What I'd tell a CIO this quarter
Start by planning capacity in megawatts, and ask your colocation and cloud partners for power delivery dates in writing, because that is where AI programs slip. Treat accelerators as short lived assets. Microsoft has said roughly two thirds of its current capex goes to short lived assets, mostly GPUs and CPUs, and your depreciation schedule and refresh plan should assume the same.
Put cost discipline on inference early. Training gets the headlines, but the recurring bill comes from running models every day, and that's where FinOps pays for itself. Fund one phase at a time. And keep the hybrid option open. Owning, leasing and renting compute each make sense at different points on the curve, and the right mix will shift as supply and pricing do.
The view from South Florida
I live and work in Parkland, in northern Broward County, so I watch the regional story closely. South Florida has always mattered for connectivity. Many of the subsea cables linking Latin America to the rest of the world come ashore here, which makes Miami the network gateway to the region. That's a very different role from the giant training campuses rising wherever land is cheap and the grid is big.
New capacity is coming anyway. Iron Mountain broke ground on MIA-1, its first Miami data center, a 16 MW facility planned for 2026, and the Tampa Bay Times has reported more AI data centers in the works in Miami-Dade and Palm Beach counties. Local governments are paying attention. Pembroke Park, here in Broward, moved this year to require a special exception, with a public hearing, before any data center can go into its industrial zone.
My bet is that South Florida's opportunity sits in inference, interconnection and low latency service to Latin America, more than in massive training sites. Power, water and hurricane resilience will shape every project here. The operators who plan around those constraints from day one are the ones whose projects will actually get built.
So, is it real?
It is. Roughly $730 billion of hyperscaler capex this year, a $2.67 trillion AI market and a data center power curve that nearly doubles by 2030 are not a mirage. But for everyone outside those four companies, the job looks the way infrastructure work always has. Know your constraints. Secure capacity before you need it. Make every dollar prove itself.
If you're working through an AI infrastructure plan, or just want to compare notes on what you're seeing, I'm always glad to talk. You'll find me on LinkedIn.
Joaquin "Jack" Fagundo writes about data centers, cloud and the business of AI from Parkland, FL. Over 20+ years he has run the data center portfolio supporting The Walt Disney Company, led data center work for Carnival Corporation across its shoreside facilities and the onboard data centers of its fleet, managed engagements at Google Cloud, and directed the migration or build of more than 20 on premises sites, the largest above 100 MW. Earlier he served as CTO of Birch Communications and Director of IT at Georgia-Pacific. He holds an MBA in IT Management and the BICSI RCDD, PMP and Google Cloud Digital Leader credentials, is a Google AI Ambassador, and works in English, Spanish and Portuguese. Read more about Jack or find him on LinkedIn and about.me.
Sources
