Capex is growing 70%. Operating cash flow is growing 23%. The gap is called: the AI bet. And for SMBs, it means: adoption without capex.
The five largest cloud providers will invest between $600 and $805 billion in AI infrastructure in 2026. For SMBs, this asymmetry means: the most expensive technology bet in history is being financed by somebody else.
How much are hyperscalers investing in AI in 2026?
Core answer: Between $600 and $805 billion — Goldman Sachs at $725B, Morgan Stanley at $805B. Capex is growing three times as fast as the operating cash flow financing it.
| Company | Capex Q1 2026 | Growth YoY |
|---|---|---|
| Amazon (AWS) | $44.2B | +77% |
| Alphabet (Google Cloud) | $35.7B | +90% |
| Meta | $125–145B (full-year guidance) | — |
| Microsoft | $190B (2026 annual plan) | — |
| Oracle | $55.7B | +162% |
Morgan Stanley analyst Brian Nowak already projects $1.2 trillion in global AI capex for 2027. NVIDIA's order backlog is estimated at $500 billion.
Important context: the +70% figure refers to capex growth year over year — not the absolute amount. Capex is growing roughly three times as fast as the funds financing it. The Stargate project, announced with a $500 billion plan as its anchor figure, has developed no active data centers since February 2026, according to Bloomberg and The Information.
Why is this a lever for SMBs?
Core answer: The infrastructure bet is being made by hyperscalers — SMBs pay only for what they actually use. Compute is depreciated externally.
The investment asymmetry sounds like a problem for the tech giants. For SMBs, it's a lever — because the cost curve has inverted.
In 2010, every new software solution meant a server contract, hosting costs, IT staff. In 2026, AI adoption means: variable token costs, no upfront investment.
Anthropic Claude Sonnet 4.6 costs $3 for input and $15 for output per million tokens. An AI phone agent like Ben consumes around 300,000 output tokens per day — under €150 a month.
Microsoft is building the infrastructure for this at $190 billion a year. A mid-sized company pays: nothing upfront.
Compute is depreciated externally — SMB adoption stays cash-flow neutral. That's the structural quirk of the pay-per-token model: upfront investment disappears entirely. What used to be a capital decision (build or lease a data center) is today an operating-cost line item, like phone or electricity.
What are we building at Amazing Agents from this?
At Amazing Agents, we take this structural shift at its word: you don't have to build the data center — you just have to use it.
Ben, our AI phone agent, handles incoming calls from day one: qualification, scheduling, follow-up. For around €150 a month in compute costs. No server, no hardware, no IT department — just a ready-made agent that handles calls.
AI adoption in SMBs isn't a capex issue. It's an opex issue. And opex decisions are same-day decisions, not investment-committee decisions.
What this means for SMB decision-makers
Hyperscalers are building what you use. The only way not to use it is the active decision against it. Three considerations for SMBs:
- Adoption is no longer an infrastructure decision. The question isn't: do we have the server capacity? It's: which process do we want to automate first?
- Token costs are operating costs — and therefore measurable. Anyone who wants to know whether AI pays off can calculate ROI: hours saved × hourly rate versus monthly token costs. That calculation is transparently doable today.
- The capex gap is temporary. As hyperscalers pour $190 billion a year into AI infrastructure, model quality rises — while token prices fall. Whoever adopts today pays less tomorrow for more performance.
FAQ — hyperscaler capex and SMBs
How much are the major cloud providers investing in AI infrastructure in 2026?
Between $600 and $805 billion. Goldman Sachs estimates $725B, Morgan Stanley $805B, MUFG over $600B. Morgan Stanley analyst Brian Nowak already projects $1.2 trillion for 2027.
What does AI adoption cost a mid-sized company?
Thanks to pay-per-token models, the pure compute layer of an AI agent runs under €150 per month. Upfront infrastructure investment is eliminated entirely — SMBs use capacity that hyperscalers have already financed.
What's the difference between capex and token costs?
Capex is upfront investment in infrastructure. Token costs are variable operating costs — comparable to an electricity bill instead of a power-plant bill. SMBs pay only for what they actually use.