Big Tech Is Spending $650 Billion on AI This Year. Bubble or Revolution?

I spent last weekend trying to explain to my dad why Meta is planning to spend somewhere between $115 billion and $135 billion this year. Not on employees. Not on acquisitions. On AI infrastructure. He asked me if that was more than the entire country of Croatia produces in a year. It is. By a lot. Croatia's GDP is about $75 billion.

The thing is, Meta is not even the biggest spender. Microsoft is reportedly earmarking around $145 billion for AI-related capital expenditure in 2026. Google and Amazon are filling in the rest, with estimates putting the combined Big Tech AI spend at roughly $635 billion this year. That is a 67% spike from 2025, according to analyst reports from Goldman Sachs and Sequoia Capital.

I've been tracking this space closely, and I still can't fully wrap my head around numbers this large. For context, $635 billion is more than the GDP of Sweden, more than the entire global music industry generates in a decade, and roughly what the U.S. federal government spends on defense. All of it flowing into data centers, chips, and the belief that artificial intelligence will reshape every industry.

Here's what I think is actually going on, who's winning, who's losing, and whether this is the dot-com bubble all over again.

Where Exactly Is the Money Going

The spending breaks down into four major buckets: data centers, chips, talent, and foundation models. The biggest chunk goes to physical infrastructure. Microsoft alone is building or expanding data centers in over 30 countries. Google broke ground on six new facilities in Q4 2025.

Chips are the bottleneck everyone is throwing money at. NVIDIA's H100 and B200 GPUs remain the gold standard for AI training, and demand outstrips supply by a wide margin. Micron and SK Hynix, the two dominant DRAM manufacturers, are effectively sold out for all of 2026. DRAM prices have surged 80-90% year over year.

Then there's the talent war. Senior AI researchers at top labs command $5-10 million annual compensation packages. OpenAI, Google DeepMind, and Anthropic are bidding against each other for a pool of maybe a few thousand people worldwide who can actually push the frontier forward.

The Number That Should Make You Nervous

McKinsey's 2025 State of AI report found that only about 1 in 50 enterprise AI investments deliver what the consultants call "transformational value." Gartner's numbers are similarly bleak: roughly 30% of generative AI projects get abandoned after the proof-of-concept stage.

So we have $635 billion flowing into a technology where the vast majority of corporate deployments fail to deliver meaningful returns. That disconnect is either a sign of irrational exuberance or a sign that the big players know something the average enterprise doesn't.

The optimistic read is that Big Tech is not the average enterprise. Meta and Google are not buying AI chatbots for their HR departments. They are building the underlying infrastructure that everyone else will rent. Different game entirely.

The Bull Case: This Is Infrastructure

The strongest argument for this spending spree is historical analogy. In the late 1990s, telecom companies laid enormous amounts of fiber optic cable. Much of it went unused for years. Many of those companies went bankrupt. But the infrastructure they built became the backbone of the modern internet.

If AI is truly a general-purpose technology on the scale of electricity or the internet, then overspending on infrastructure now could look prescient in a decade. The companies that survive the shakeout will own the rails that everything runs on.

Jeff Bezos made this exact argument about AWS in the early 2010s, when Amazon was spending billions on cloud infrastructure that critics said would never generate returns. AWS now brings in over $100 billion in annual revenue. Sometimes the crazy bet pays off.

The Bear Case: Revenue Doesn't Match the Hype

The counterargument is simpler and more grounded. AI revenue across the entire industry was estimated at around $90-100 billion in 2025. The spending planned for 2026 is over six times that. Good: the technology is improving rapidly. Bad: the revenue gap is widening, not closing.

Sequoia Capital's David Cahn published an analysis arguing that AI companies would need to generate $600 billion in annual revenue just to cover the infrastructure costs being incurred. That is roughly the entire revenue of Apple and Microsoft combined. We are nowhere close.

The dot-com parallels are uncomfortable. In 2000, telecom companies spent over $500 billion (adjusted for inflation) building out fiber networks. The crash wiped out $5 trillion in market value. The technology was real. The timeline was wrong.

What Happened Last Time (and the Time Before That)

The fiber optic boom of 1999-2001 is the most obvious comparison. Companies like WorldCom, Global Crossing, and JDS Uniphase became Wall Street darlings. Then they became cautionary tales. WorldCom's fraud was discovered in 2002. Global Crossing filed for bankruptcy the same year.

But the cloud computing spending wave of 2013-2017 offers a different lesson. Critics said Amazon, Microsoft, and Google were overbuilding. The spending looked irrational. Within five years, cloud infrastructure became a multi-hundred-billion-dollar market, and the early spenders dominated it.

The pattern is consistent: transformative technologies attract excessive investment, a shakeout follows, and the survivors reap disproportionate rewards. The question is never whether the spending is rational. It's whether you're one of the survivors.

The Winners So Far

NVIDIA is the clearest winner. The company's data center revenue hit $115 billion in fiscal 2025, up from $47 billion the year before. Every dollar of Big Tech AI spending flows through NVIDIA's supply chain at some point. Jensen Huang is having the business quarter of a lifetime, every quarter.

Cloud providers are the second-tier winners. AWS, Azure, and Google Cloud are all reporting AI-driven revenue acceleration. Microsoft's cloud revenue grew 33% year over year in Q4 2025, with management attributing the majority of the upside to AI workloads.

Infrastructure picks-and-shovels plays are also thriving. Arista Networks, Vertiv (data center cooling), and Eaton (power management) have all seen their stock prices double or more since early 2025. Building the physical layer of AI is lucrative even if the application layer disappoints.

The Losers So Far

The losers are harder to spot because nobody wants to admit they wasted money. But the pattern is clear if you look at enterprise surveys. Bain & Company's 2025 tech survey found that 65% of companies that purchased AI tools in the prior 18 months described adoption as "below expectations."

Good: these companies are experimenting and learning. Bad: most of them bought expensive solutions before understanding the problem they were solving. The classic technology adoption mistake, buying the tool first and looking for the use case second, is playing out at enormous scale.

The companies losing the most are mid-market enterprises spending $1-5 million on AI initiatives without the data infrastructure, talent, or organizational readiness to make them work. They are the equivalent of small businesses buying Salesforce licenses and using 10% of the features.

What This Means for Startups

If you're raising money for a startup right now, your pitch better involve AI. CB Insights data shows that AI-focused startups captured 48% of all venture funding in Q4 2025, up from 28% two years earlier. The remaining 52% is being split among every other category of technology.

Non-AI startups are getting squeezed. Several VC partners I've spoken with describe a two-track market: AI companies get meetings in 48 hours, everyone else waits weeks. The valuation premiums are equally stark. AI startups are raising at 40-80x revenue multiples, while SaaS companies in adjacent spaces struggle to get 15x.

This creates a perverse incentive. Companies are shoehorning AI into their pitch decks even when their core value proposition has nothing to do with machine learning. That never ends well, but it's happening at scale.

The DRAM Bottleneck Nobody Is Talking About

Here is something that should get more attention. The AI spending boom has created a genuine semiconductor supply crisis. High-bandwidth memory (HBM), the specialized DRAM that powers AI accelerators, is in severe shortage. SK Hynix and Micron have pre-sold their entire 2026 production capacity.

DRAM spot prices have increased 80-90% since mid-2025. That cost gets passed through the entire chain. Training a frontier model that cost $100 million in 2024 now costs $180-200 million purely due to hardware inflation. The irony is thick: the more money Big Tech throws at AI, the more expensive AI becomes.

This is not a problem that gets solved quickly. New semiconductor fabrication capacity takes 2-3 years to come online. Samsung, TSMC, and SK Hynix are all expanding, but the supply-demand gap won't close before late 2027 at the earliest.

So Is It a Bubble

Honestly, it's both. The underlying technology is real and transformative. The spending levels are almost certainly unsustainable at their current trajectory. Those two things can be true at the same time.

Good: AI is genuinely improving productivity in coding, content creation, data analysis, and customer service. Bad: the gap between what companies are spending and what they're earning from AI is growing, not shrinking.

My best guess is that we're in the "overshoot" phase of a legitimate technology revolution. Too much money is chasing too few proven use cases. A correction will come. Some companies will get hurt. And the technology will still change everything over the next decade.

Tracking the Signal Through the Noise

Whether you're bullish or bearish on AI spending, one thing is clear: the pace of announcements, deals, and pivots is accelerating. A single week in February 2026 saw Meta announce its capex guidance, NVIDIA release new chip benchmarks, and three AI startups raise billion-dollar rounds. twixb's AI & Machine Learning newsfeed tracks the signal through the noise -- funding rounds, product launches, and the moves that actually matter.

However you follow this space, don't rely on headlines alone. The nuance matters more than ever. A $135 billion spending announcement means something very different depending on where the money goes, what returns are expected, and what timeline management is working with.

Quick Reference

Winners: NVIDIA, cloud providers (AWS, Azure, GCP), infrastructure companies (Arista, Vertiv), HBM manufacturers (SK Hynix, Micron).

Losers: Mid-market enterprises buying AI without a plan, non-AI startups competing for VC attention, anyone who needs cheap DRAM.

The uncomfortable truth: $635 billion is being spent on AI infrastructure in 2026 while only about 2% of enterprise AI projects deliver transformational value. The technology is real. The spending is real. The returns are mostly theoretical.

What to watch: NVIDIA's quarterly earnings for demand signals, enterprise adoption surveys from McKinsey and Bain for reality checks, and DRAM pricing as the canary in the coal mine for whether the hardware boom is sustainable.

The honest take: We've seen this movie before. Massive investment in transformative infrastructure, followed by a bust, followed by the survivors building trillion-dollar businesses on the wreckage. The script hasn't changed. The budget just got a lot bigger.