MLWhiz Weekly Recsys/ML/GenAI Newsletter # 10 - The week AI infrastructure crossed from a technology story to a financial one

2026-06-10 · Source: MLWhiz: Recs|ML|GenAI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Emerging Technologies & Innovation · Depth: Intermediate, short

Summary

The AI infrastructure landscape is undergoing a significant financial transformation, marked by a \$35 billion private financing deal led by Apollo and Blackstone for Broadcom's AI XPV Platform. This initiative aims to deliver over 20 gigawatts of AI compute capacity by 2028, with Anthropic and OpenAI as key clients, shifting AI compute funding from major tech company balance sheets to broader private capital. Concurrently, new models were released: Google's Gemma 4 12B, an encoder-free multimodal model for consumer GPUs; MiniMax M3, claiming 59.0% on SWE-Bench Pro with a 1M-token context and competitive pricing; and Microsoft's MAI-Code-1-Flash, a 137B-parameter coding model developed without OpenAI data, showing strong performance and integration into GitHub Copilot. Netflix also introduced Mult-DPO, an extension of Direct Preference Optimization for set-wise recommender system data.

Key takeaway

For AI/ML Directors evaluating infrastructure investments, the \$35 billion Broadcom deal signals a new era where private capital will drive compute capacity. You should anticipate potentially lower compute costs due to increased supply, but prepare for rising power and compliance expenses. Re-evaluate your long-term cloud strategy and consider new financing models for large-scale AI deployments.

Key insights

AI infrastructure financing is shifting from tech giants to private capital, impacting compute costs and accessibility.

Principles

• AI compute capacity is now financeable by diverse private capital.
• LLM alignment techniques can significantly improve RecSys performance.
• Investment in dev platforms compounds with AI agent adoption.

Method

Mult-DPO extends DPO to a multinomial formulation, learning that "all items in set S+ should outrank all items in S−" for recommender systems, directly handling set-wise preference data.

In practice

• Test Gemma 4 12B on consumer GPUs or laptops for multimodal tasks.
• Benchmark recommendation models against Netflix's Mult-DPO.
• Evaluate MiniMax M3 or MAI-Code-1-Flash for coding tasks.

Topics

• AI Infrastructure Investment
• Large Language Models
• Recommender Systems
• Multimodal AI
• Code Generation Models
• Direct Preference Optimization

Code references

• RyanCodrai/turbovec

Best for: Machine Learning Engineer, CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Investor

Related on AIssential

• Alphabet Leads as Big Tech Turbocharges AI Infrastructure — Big Tech is rapidly escalating AI infrastructure investment, moving beyond incremental buildouts to hyperscale deployment.
• The State of AI Compute — AI compute capacity is undergoing an exponential expansion, signaling a major shift in economic control.
• MLWhiz Weekly AI/ML Newsletter # 4 — AI buyers are becoming AI owners through equity-based procurement deals, reshaping supplier relationships.
• 😺 Google ran out of cloud — Hyperscalers are investing heavily in AI infrastructure, with demand outstripping current supply and shifting competitive dynamics.
• OpenAI exec says company hopes to burn $50B of somebody else's money on compute this year — OpenAI expects to spend $50 billion on compute in 2026, largely through conditional "investments" from partners.
• The Sequence Radar #816: Last Week in AI: $110B Bets, Nano Banana 2, and the New Economic Reality — AI is rapidly reshaping economic structures and geopolitical dynamics through massive investment and advanced model capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MLWhiz: Recs|ML|GenAI.