Cloud Computing

Hyperscalers are becoming multi-model marketplaces, ending cloud-model exclusivity

Google Cloud began selling OpenAI models and Codex through Model Garden and Marketplace on Oct 21, letting customers burn existing cloud spend commitments on a rival's models. AWS CEO Matt Garman framed it as customers wanting OpenAI models where their data and apps already live, and Google is explicitly selling a competitor's models next to Gemini to capture spend that might go to rival clouds. The exclusive Microsoft-OpenAI era is giving way to distribution deals where the cloud, not the lab, owns the customer relationship. Google is simultaneously pushing new frontier models (Gemini 4 Carbon) to keep its own models competitive on that shelf.

CoreWeave's interest costs ate roughly a quarter of Q2 revenue and it warns customers may need fewer services as inference gets cheaper via smaller models and caching. Physical constraints compound the financial ones: Microsoft pulled out of an Nscale UK data center deal over grid delays. GPU capacity financed against expected demand is exposed if efficiency gains outrun buildout.

Why it matters · Investors should favor neoclouds with contracted demand and power access over those relying on leveraged capacity bets.

CBRE agreed to acquire Onnec Group for $1.1B, and a $3.2B IPO backed by BlackRock and Fidelity and a $151M IPO landed within the same week. Korea's Elice, an AI infrastructure company, fell 32% on debut, a reminder that public markets are discriminating. Digital Edge's reported ~$10B sale exploration points to the same exit window for infrastructure operators.

Why it matters · The exit window for digital infrastructure is open but pricing is uneven, rewarding scale and punishing weaker debuts.

New products target agents rather than humans: Resend Inboxes connects agents to shared inboxes and workflows, and OpenShip Cloud offers managed no-lock-in hosting. Cloudflare launched Clairic for cheap structured extraction, and Lazy AWS targets terminal-driven cloud operations. The sandbox and runtime layer (E2B, Daytona, Fly.io, Modal at $300M ARR) is becoming a distinct category of agent-native cloud.

Why it matters · Developer-infra startups that become default agent primitives can scale quickly, while legacy dashboards lose relevance.

Google Cloud and Cloudflare each offer up to $350K in credits, Microsoft for Startups up to $150K, Datadog up to $100K, and Vercel and Runware up to $30K. Credits are acquisition spend aimed at locking in AI-native workloads early, alongside Nvidia and Accel participating in smaller expansion rounds.

Why it matters · Startup compute costs are subsidized, so early cloud choice reflects incentives and creates switching-cost risk later.