AI cloud costs for the next few years are being set by a company that just told the market it plans to spend $220 billion on infrastructure this year, and the market cheered. Amazon's quarterly numbers landed on 30 July.
The reaction is the interesting part, and TechCrunch framed it well in Investors love AI, as long as you're a cloud host. Investors are rewarding the layer that rents compute and punishing the layer that only buys it. If you build software from a small team here, that split is worth more to you than the headline number.
📊 What the market actually rewarded
Here are the figures from the quarter, as reported:
| Metric | Amazon, latest quarter |
|---|---|
| AWS revenue | $42 billion, up 37% year-over-year |
| Total net sales | up 20% |
| Capex, fiscal year ended 30 June | $173 billion (up from $107.65 billion) |
| 2026 capex forecast | raised from $200B to $220B |
| Cash reserves | down $7.6 billion year-over-year |
| Free cash flow | negative for the first time in 2026 |
| Stock, after hours | up nearly 10% |
Now the contrast. Meta reported in the same window, spent heavily on the same kind of buildout, and its stock fell 8% — investors fixed on the cash flow crunch and the continued spending without an equivalent revenue line to point at.
Same spending behaviour. Opposite reaction.
Key takeaway: The market is not pricing AI enthusiasm. It is pricing whether your AI revenue appears on the same page as your AI spending. Amazon can show a 37% growth line next to the bill. Meta, for now, cannot.
💡 The most useful sentence in the whole story
Amazon's CEO said something that I think matters more to a two-person team than it does to a fund manager:
"We see the AI business following very much the same margin trajectory we saw in the core business before...AWS and Amazon Bedrock can have a wildly successful business without its own frontier model."
Read that again with your own product in mind. The largest AI infrastructure business on earth is openly saying it does not need to own a frontier model to win.
If that is true at their scale, it is far more true at yours. I keep meeting local developers who treat "we didn't train our own model" as a confession. It isn't. The durable parts of a small software business have never been the model weights:
- Distribution — do people already come to you for this problem?
- Boring proprietary data — the messy local dataset nobody else bothered to clean.
- Workflow fit — knowing that a Sri Lankan payroll run has EPF, ETF and APIT in it, and that a generic tool will get it wrong.
- Trust and support — a real person who answers when it breaks.
None of those four require a GPU cluster. All of them are hard to copy.
📉 Negative free cash flow is a signal about your future pricing
This next part is my reading, not a claim the article makes — flagging that clearly.
Capex does not stay on the balance sheet quietly. $173 billion of data centre spending becomes depreciation, and depreciation lands on the income statement for years afterwards. Amazon just posted its first negative free cash flow period of 2026 while raising the forecast anyway. That is a company betting the revenue arrives before the depreciation bites.
For you, the practical consequence is simple:
| Scenario | What happens to your inference bill |
|---|---|
| AI demand keeps compounding | Scale economics improve, prices likely keep drifting down |
| Demand plateaus mid-buildout | Providers must recover depreciation — free tiers shrink first |
Warning: Do not build a product whose margin only works at today's promotional API price. Free tiers and aggressive per-token pricing exist partly because everyone is buying market share during a buildout. Buildouts end.
The article makes the same point from the other direction: cloud host revenue depends on AI spending being sustainable across the whole industry. If demand doesn't justify the buildout, AWS and its competitors are exposed too.
🌐 What a $220 billion bill looks like in rupees
Here is where it gets local. Your compute is priced in USD. A meaningful share of small-team revenue here is priced in LKR. That mismatch is a real risk sitting quietly inside your unit economics.
Three things I'd actually do before shipping anything AI-powered:
- Model the cost per active user, not per request. A single call is cheap. A chatty user across a month is not. Our AI agent cost calculator exists for exactly this — multi-turn cost is where budgets die.
- Know your token counts before you know your price. Prompt bloat is the most common cause of a surprise bill. Measure with the token counter instead of guessing.
- Price your LKR invoices with FX headroom. If you quote a fixed rupee retainer against a dollar API bill, you have written an unhedged currency position into your contract. The USD-LKR earnings calculator is a blunt but honest way to see the gap.
The cheapest optimisation remains the unglamorous one: don't call a large model when a small one, a cache, or a regex will do. Route by difficulty. Cache aggressively. Most "AI features" are 80% deterministic logic wearing a model as a hat.
🚀 What this means for you
The headline says investors love AI if you're a cloud host. The more useful version for a small builder is this: the market is rewarding proven revenue attached to spending, and nothing else. That standard is being applied to trillion-dollar companies. It will absolutely be applied to you by your first serious client or investor.
So:
- Stop apologising for not owning a model. Amazon says it doesn't need one either.
- Attach every rupee of AI spend to a revenue line you can name. If you can't name it, it's a hobby, and that's fine — just label it correctly.
- Assume today's inference price is a floor, not a ceiling. Build the margin to survive a 2× price move.
- Own the layer they can't buy — local context, distribution, the data nobody cleaned.
The buildout is being financed by people with $220 billion to deploy. You will never win that game, and you don't need to. You're renting from them. Rent carefully, price honestly, and keep your costs attached to something real.
Commentary on reporting by TechCrunch. Figures cited are from the linked article.
Top comments (0)