AI ML Infrastructure Platforms

Neocloud debt stress is testing the AI buildout financing model

Neoclouds such as CoreWeave, Nscale and Nebius are scaling on debt-financed projects backed by Nvidia chips and equity and by private credit from Apollo and Blackstone. CoreWeave's bonds yielded ~15% and its shares fell 16%, and its debt briefly traded at distress levels. Nscale has filed to go public at a $35B valuation built on Microsoft and Anthropic contracts, with a large gap between promised megawatts and capacity actually online. Nvidia is the common thread, the top investor in 50 of 78 recent deals, which makes the supply chain increasingly circular. Weekly capital swung from $24.2B in mid-August to under $1B in early October, showing how lumpy mega-rounds are.

OpenRouter data is read as showing open-source share rising from under 30% to over 60% this year. Bending Spoons reportedly uses open models 99% of the time. That favors model-agnostic routing, gateway and serving layers such as OpenRouter, LiteLLM, Fireworks, Baseten, Together AI and SGLang over any single frontier vendor.

Why it matters · Value is shifting to neutral orchestration and inference-optimization layers that arbitrage model price and quality.

Uber burned a year of AI budget in 4 months, and GitHub's compressed coding-agent output still raised the total bill. This creates demand for token compression (Edgee), KV-cache inference optimization (Tensormesh), compilers (Luminal) and prompt/config optimizers (Promptic). Developer-productivity measurement platforms like DX are being adopted at Snowflake, Sony and BNY.

Why it matters · FinOps for inference is becoming a budget line item, opening room for tooling startups that cut cost per token.

Evaluation platforms such as Vals, Braintrust and Kolena, and persistent-memory layers such as Mem0 and GBrain, are being positioned as required production infrastructure. Garry Tan's GBrain Tools launched as a credential layer connecting agents to Gmail, GitHub and Stripe. Inference leaders Fireworks and Baseten are themselves buying customer-discovery tooling, and Vals agents are being used in materials discovery.

Why it matters · Trust, memory and observability layers are sticky and less capex-heavy than compute, offering better software margins.

Series D+ rounds took $39.6B across 23 deals in 90 days, against about $12.3B across 37 seed deals, and a $750M Series D at a $15.75B valuation closed this week. An IPO of $330M priced alongside Nscale's filing and Cerebras trading 31% below its IPO valuation show public markets repricing. Option-like structures are also being priced into top-of-market rounds.

Why it matters · Exit windows are open but selective, so late-stage entry prices carry growing downside risk.