Good morning, AI enthusiasts!
Marvell’s proposed $115 billion Lumentum deal puts AI’s optical connections in the spotlight. Today’s briefing also looks at model-training feedback, Anthropic’s computing costs and the infrastructure needed to bring AI into everyday business.
Your AI Edge briefing:
THE QUICK SCOOP
MARVELL
Summary: Marvell has agreed to acquire optical-components maker Lumentum in an all-stock transaction valued at about $115 billion, including debt. The proposed combination would bring more of the hardware that moves data between AI chips under one roof.
The Quick Scoop:
Optical connections are a critical part of large AI computing systems.
The companies expect the deal to close in the first half of 2028, subject to approvals and other conditions.
The agreement expands Marvell’s challenge to Broadcom across AI infrastructure.
Your AI Edge: AI capacity depends on the network connecting processors as well as the processors themselves. Treat this as a proposed transaction, not a completed acquisition.
ANTHROPIC
Summary: Anthropic’s updated IPO filing shows quarterly revenue of roughly $17 billion, alongside a $567 million operating loss. Demand is growing quickly, but the expense of securing computing capacity remains central to the business.
The Quick Scoop:
Quarterly revenue rose from about $1.4 billion a year earlier.
The reported $23.3 billion net loss largely reflects non-cash accounting charges.
Spending commitments total about $895 billion, mostly for computing power.
Your AI Edge: Revenue, operating loss and accounting-driven net loss measure different things. Keep them separate when assessing claims about the economics of AI.
SPONSORED BY MAMMOTHCLUB
Turn your automation ideas into practical workflows with Mammoth Club’s Build AI Agents with LLMs in n8n course.
Build LLM chains and summarization workflows.
Connect your knowledge with a RAG question-and-answer workflow.
Add web search, memory and external tools to agents.
TRAJECTORY
Summary: Trajectory introduced Continual Grading, a research preview that learns how specialists judge AI outputs and creates grading criteria for new tasks. The aim is to reduce the manual work needed to train models for domains where there is no simple right answer.
The Quick Scoop:
Experts provide a small collection of graded examples.
A learned metagrader produces task-specific criteria for unseen work.
The resulting scores feed reinforcement-learning training.
Your AI Edge: For specialized AI, start by defining what a good result looks like. Evaluate learned graders on held-out expert judgments before trusting their scores.
OPENAI
Summary: OpenAI launched its Grid Protectors Program to help future compute campuses reduce electricity demand when local grids are strained. The company plans to work with utilities and regulators on location-specific arrangements.
The Quick Scoop:
Flexible demand would become part of future campus planning.
Its Georgia Power agreement includes up to 1,000 megawatts of demand response.
The planned Effingham County campus is intended to serve as a pilot.
Your AI Edge: Look for signed commitments and measured performance as projects come online. This is a program for future facilities, not evidence that those capacity reductions are operating today.
MICROSOFT & LUMEN
Summary: Microsoft and Lumen expanded their co-selling collaboration to connect enterprise data, applications and infrastructure for AI workloads. The initiative brings network modernization into Azure planning earlier.
The Quick Scoop:
The collaboration targets connectivity across offices, data centers and clouds.
Lumen intends to bring more networking capabilities to Azure.
Additional capabilities and joint go-to-market details will follow over time.
Your AI Edge: Before adding another model or agent, map where its data lives and how it can access that data securely. Network and governance decisions can determine whether a pilot scales.
SPONSORED BY MAMMOTHCLUB
Bring Higgsfield footage into CapCut and shape it into shareable content with Mammoth Club’s practical video-editing course.
Import clips and build your timeline.
Improve pacing with audio, text and transitions.
Work with brand kits and export your finished video.
AI POLICY
Summary: A senior Chinese financial-regulatory official warned that global AI investment is becoming overly concentrated, Reuters reports. The official argued that the resulting pressure on borrowing costs could crowd out traditional industries.
The Quick Scoop:
The official made the remarks in Hong Kong.
The concern extends beyond technology companies to the wider economy.
The official called for policy frameworks addressing AI-related economic and financial risks.
Your AI Edge: This is a policymaker’s warning, not a finding that every AI investment is unproductive. Watch for concrete policy measures and evidence behind the concentration concerns.
OPENAI
Summary: Former U.S. Army secretary Dan Driscoll is joining OpenAI as senior counselor for national security modernization. He will advise leadership and partners on practical uses of advanced AI in national security and cyber resilience.
The Quick Scoop:
The position is newly created.
His work includes engagement with governments, industry and allied institutions.
He will help OpenAI understand national-security organizations’ operational needs.
Your AI Edge: The appointment signals continued attention to public-sector deployment. It does not by itself announce a new product, contract or change in model capabilities.
THE MAMMOTH LOWDOWN
⚙️ TOOLS FOR AI EDGE
Five tools to explore for giving AI applications better access to knowledge.
Weaviate - Build vector search, retrieval-augmented generation and AI memory.
Milvus - Store and search vector embeddings with an open-source database.
Pinecone - Connect AI applications to relevant knowledge through managed retrieval.
pgvector - Add vector similarity search to PostgreSQL.
Vespa - Build search and recommendation applications with retrieval and ranking.
📢 What Else Is Buzzing In AI
Garry Tan’s token bill - TechCrunch reports he spends up to $200,000 annually on AI tokens.