The Capex Arms Race Orphans Small Workloads

Around $700 billion. That's the combined capital expenditure the big four hyperscalers have signaled for 2026, up from roughly $410 billion the year before, with the overwhelming share going to AI datacenters, accelerators, and the power to run them. Now hold that number next to a mundane one: the monthly bill for a typical early-stage product, a container, a managed database, some preview environments, which lands somewhere between a cinema ticket and a car payment.

The claim of this post is about what happens between those two numbers. When a vendor's capital, attention, and executive incentives stampede toward one class of customer, every other class becomes a maintenance obligation, and the ordinary small workload, the vast silent majority of what actually runs on the internet, is quietly becoming an orphan. That neglect is not a scandal. It's arithmetic, and it's the opening that every developer-focused cloud, ours included, is built on.

Where the attention went

Follow the incentives, because they're not hidden. Analysts note the capex plans consume close to all of the hyperscalers' operating cash flow, against a historical norm near 40%, which means these companies have bet the balance sheet on AI infrastructure, and boards don't bet the balance sheet on something and then staff their best people elsewhere. The keynote minutes, the org charts, the launch cadence all follow the money; count the developer-experience announcements at any recent cloud keynote against the AI ones and you get a ratio that would have been unthinkable in 2019.

None of this means the small-workload services break. They persist, in competent maintenance mode, which from the inside feels responsible and from the outside feels like a product aging in place: consoles that grow another menu layer every year, defaults still tuned for 2019's instance families, quotas and small-instance availability that quietly degrade in busy regions because scarce power gets allocated to whoever signed the nine-figure commitment. When capacity is rationed, the rationing follows revenue. A startup's container fleet does not win that auction.

The bundle economics make orphanhood structural rather than accidental. Small workloads were never the product; they were the cross-subsidy donors, paying sticker prices and idle-hour billing that helped fund the enterprise discounts. The AI era didn't create that asymmetry. It just removed the strategic reason to pretend otherwise, because the next trillion dollars of hyperscaler revenue is not coming from developers with side projects, and everyone's roadmap knows it.

The opening

Markets route around neglect, and this one visibly is. The existence proof is the row of platforms serving exactly the orphaned buyer, Render, Fly, and yes, us, each built on the observation that a developer deploying an ordinary app deserves a product designed for that act rather than a 240-service console designed for a procurement committee. The unbundling logic applies at the bottom of the market just as it did at the GPU top: when the bundle stops competing for a customer class, specialists collect it.

What the specialists offer isn't magic; it's focus applied where focus left. Deploy paths measured in minutes, pricing a human can compute, preview environments that cost cents because someone bothered to make idle mean zero. Every one of those is a small-workload feature that a hyperscaler could build and won't prioritize, for the same reason a container ship doesn't compete for kayak customers.

There's a historical rhyme worth noticing here. The hyperscalers themselves were born from exactly this dynamic: AWS emerged because the enterprise IT vendors of the 2000s were busy servicing their biggest accounts while developers wanted a credit card and an API. The neglected buyer of one era funds the giants of the next, which is why "too small to matter" has such a poor track record as a strategy, and why the orphanage keeps producing founders.

The steelman: the tide lifts kayaks too

The counterargument deserves its due. Hyperscalers still ship developer-facing improvements constantly; capex is not attention, and the same buildout that serves AI contracts eventually cheapens the compute underneath everyone, the way GPU-driven datacenter builds also refresh CPUs, networks, and storage. Small workloads ride infrastructure they could never fund alone, free tiers remain genuinely generous, and declaring neglect from launch-cadence vibes is exactly the kind of unfalsifiable claim this blog complains about elsewhere. Fair, all of it, and the trickle-down is real.

But spillover is not stewardship. Riding on shared infrastructure is what orphans do; being designed for is what customers get, and the difference shows up in a thousand small product decisions that spillover can't fix: which defaults rot, whose quota request waits, which console flow was last rethought when the team that owned it still existed. The honest version of the hyperscaler value proposition for small workloads in 2026 is "world-class infrastructure, secondhand attention", and for plenty of teams that's a fine trade. The teams for whom it isn't are why we exist.

So run the attention audit on your own stack. Take the service you deploy to most days and find the last meaningful improvement that made your specific work better, not a new AI capability, not a keynote demo, the boring path you actually walk. If you're struggling to name one from the last two years, you're not a customer anymore. You're an installed base, and the difference between those two words is what an entire upcoming post is about. Installed bases get repriced.

Related: Power Is the New Region, where the capacity your workloads compete for actually goes. More about what we're building at light-cloud.com.