The $800 Billion Question: Can the AI Infrastructure Supercycle Prove Its Returns?

2026's biggest bet is AI infrastructure. Hyperscalers' total spend this year is estimated at ~$775–800B. The numbers are dizzying; but even AI leaders calling for a 'slowdown' last week is a reminder that the real question is returns.

The spending side

• Alphabet announced $175–185B total capex for 2026, with $70–74B for data-center construction.
• Microsoft committed over $80B for fiscal 2026; Meta expects $80–100B.
• US data-center electricity demand rose from 23 GW (2023) to 42 GW (2026); AI racks draw 50–100 kW (a classic rack draws 5–10 kW).

Nuclear is now a requirement, not a preference

Meta's nuclear PPA, Amazon's expanded nuclear offtake and Microsoft's Three Mile Island restart show nuclear is an operational requirement for AI at scale. Constellation plans to bring an 835 MW unit online in 2027 under a 20-year deal with Microsoft.

The real question: returns

Spending at this scale is only sustainable if it produces concrete value. That's exactly where the 'AI bubble' debate sits: buying raw compute is easy; extracting business value from it is hard. The difference comes not from the chip but from the software layer that uses it efficiently.

The remot angle

remot is the layer that turns capex into returns: it routes each request to the most suitable model, eliminates needlessly expensive calls, and measures cost and quality. It lets enterprises extract the most value from existing models without investing directly in billion-dollar infrastructure — capturing the supercycle's benefit, not its bill.