The AI boom is still pulling in huge money, but you should pay less attention to the spending race and more attention to where pricing power is actually holding.
The easiest story in AI is the giant number. Microsoft, Amazon, Alphabet, and Meta are now tracking toward the kind of capital spending that used to belong to national infrastructure plans, not software companies. Tom's Hardware, citing Financial Times data from first-quarter earnings, reported that the four companies are on pace for roughly $725 billion in 2026 capital expenditure, up 77 percent from $410 billion last year. Amazon's own plan is about $200 billion.
Those figures are useful, but they can also make you stupid. A bigger data center bill doesn't automatically mean a better business. It means somebody is buying Nvidia chips, land, power, cooling systems, networking gear and memory at a pace the industry hasn't seen before. The harder question is who gets to keep the margin once all that machinery starts serving customers.
Right now, the cleanest answer is the infrastructure layer. The Wall Street Journal reported last month that the same four tech giants spent $133 billion in the first quarter alone, with depreciation charges already up to $41.6 billion for the period. That is the bill behind the magic. If AI revenue doesn't arrive quickly enough, those servers don't stay a futuristic asset for long. They become an accounting weight.
The model layer has a different problem: customers are learning to shop around. GPT-4 entered the market in March 2023 with API pricing of $30 per million input tokens for the standard 8K context version, based on OpenAI's published launch pricing. Prices across the sector have since moved down hard as OpenAI, Anthropic, Google and others compete for developers and enterprise accounts. You don't need a grand theory to understand this. When several models are good enough for the same support ticket, code review, meeting summary or internal search task, procurement starts asking why it should pay more.
Business Insider reported this week that RBC Capital Markets' survey of more than 100 CIOs and technology leaders found enterprise AI spending still rising. Nearly 90 percent of respondents said token budgets were manageable, and 100 percent said they had AI in their budgets. That is a strong demand signal. It is not, by itself, proof that foundation model providers have a beautiful business.
Frankly, this is where a lot of AI commentary gets lazy. It treats usage as profit. They are not the same thing.
Axios reported in March, citing Ramp customer data, that Anthropic was capturing more than 73 percent of AI tool spending among companies buying such tools for the first time. The same Axios piece said OpenAI was still on pace for more revenue, at $25 billion this year versus Anthropic's $19 billion. That tells you the race is real and valuable. It also tells you enterprise buyers are not behaving like loyal fans. They are moving money toward whichever product works best for the job this quarter.
Google has one advantage that deserves more attention than most model benchmark chatter. It owns more of its stack. Its Tensor Processing Units reduce its dependence on Nvidia, and Google used I/O 2026 to put scale on stage: TechRadar's live coverage said Google claimed Gemini 3.5 Flash was processing more than three trillion tokens a day internally. Antigravity 2.0, Google's agent-first coding tool, was part of the same push. You can argue about the demo. You can't ignore the economics of owning silicon, models, cloud distribution and workplace software at the same time.
Where pricing power is actually surviving
The SaaS layer is not one story. Thin AI wrappers are in trouble, and they should be. If your company is a prompt box sitting on a third-party model, a single product release from OpenAI, Anthropic, Google or Microsoft can erase your reason to exist before your next board meeting. Customers don't owe you a category. They owe you money only if you solve a problem they can't solve inside the tools they already use.
Vertical AI companies have a better claim. Sapphire Ventures' software and AI work has pointed to stronger valuations for companies with proprietary data, workflow depth and real annual recurring revenue, especially those that have crossed meaningful scale. That distinction is the whole game. A legal AI product tied into document review, billing codes and firm-specific precedent is not the same thing as a generic chatbot with a nicer interface. One has operational gravity. The other has screenshots.
The same pressure shows up in margins. Traditional SaaS businesses often trained investors to expect gross margins around 70 to 90 percent. AI-native software companies can land lower because inference is a real cost every time a customer uses the product. If usage rises faster than pricing discipline, growth can make the income statement worse before it makes it better.
So when you look at the $725 billion capex race, don't stop at the headline. Ask who owns the chips, who owns the customer, who owns the data, and who can raise prices without getting replaced. The winners in AI won't just be the companies spending the most. They will be the ones that turn all this spending into a product customers can't casually swap out next quarter.
Also read: Ford spent billions learning that AI cannot replace engineers who know where the bodies are buried • Hong Kong's AI-fueled IPO boom is rewriting where Chinese tech capital goes to grow • Cloudflare cut 20% of its workforce while growing its engineering team, and Matthew Prince says every company will do the same
