AI’s Next Infrastructure Wave is Power, Cooling, Networking, and Memory – Issue #32
Spotlight: AI’s Next Infrastructure Wave is Power, Cooling, Networking, and Memory
This week’s Spotlight is: AI’s Next Infrastructure Wave is Power, Cooling, Networking, and Memory
I covered why we’re in the AI infrastructure gold rush in this piece, and recently had the opportunity to sit down with Amin Vahdat from Google at the Silicon Catalyst Summit 2026. One thing became very clear: the AI race is no longer just about faster AI accelerators or bigger models. The real contest now is to solve the underlying physics of how intelligence is built and delivered.
Three themes stood out:
Moving bits is now more expensive than the compute itself. Memory and data movement are becoming bigger bottlenecks than compute.SRAM, HBM, DRAM, SSDs, and the networks connecting them are all becoming constraints. The next frontier is memory architecture: lower latency, 3D stacking, direct GPU-to-storage connectivity, and fundamentally rethinking how compute connects to memory. Linear extrapolation will not get us where we need to go. Radical reinvention will.
Power and cooling are becoming the defining long-term constraints of the AI era. Google alone is projecting roughly $185B in capex this year, with a significant portion going toward power and cooling. The opportunity goes beyond generating more power. The biggest levers are using existing power more intelligently and cooling infrastructure. Hyperscalers historically provisioned for 99.999% availability, leaving massive amounts of power sitting idle, while cooling is increasingly becoming the limiting factor in extracting full performance from modern AI systems.
Networking is being rebuilt from the ground up. Google recently announced million-node non-blocking clusters and TPU architectures connected through a 3D torus, where chips can read and write one another’s memory at terabytes-per-second speeds. Co-packaged optics and photonic interconnects are no longer theoretical — they are becoming foundational. The entire networking stack — within the rack, within the data center, and across data centers — is being reinvented simultaneously.
Bottom line: The picks-and-shovels moment in AI infrastructure is expanding beyond GPUs into memory, networking, power, and cooling. These are the foundational layers where the next generation of category-defining infrastructure companies will be built.
The power law is already reshaping the infrastructure stack. This week, Micron and SK Hynix each surpassed a $1 trillion valuation as demand for AI memory continues to surge, while Samsung began shipping next-generation HBM4E memory chips to support future AI accelerators and data centers. At the same time, SK Hynix unveiled a new cooling architecture for AI memory systems, underscoring how performance is increasingly constrained not just by compute, but by thermals, power, and data movement. The winners in the next phase of AI may not only be the companies building models, but also those solving the underlying infrastructure bottlenecks that make intelligence possible at scale.
Full Weekend Edition below. 👇
Signals Shaping the Future of AI:
Infrastructure
SK Hynix unveils a new cooling architecture for AI memory systems. The company’s iHBM technology integrates cooling elements directly into high-bandwidth memory packages, reducing thermal resistance and supporting next-generation AI accelerators and data centers. Click here
Micron and SK Hynix each surpass a $1 trillion valuation as AI memory demand surges. Rising demand for high-bandwidth memory chips used in AI servers and accelerators continues to drive growth across the memory semiconductor market. Click here
DG Matrix introduces an 800V solid-state transformer platform for NVIDIA MGX AI factory racks. The launch targets the power-conversion layer of AI data centers, where higher-voltage architectures are becoming necessary to support denser GPU systems and reduce electrical losses. Click here
Microchip launches 3.3kV silicon-carbide power modules for AI hyperscale data centers. The release points to power electronics becoming a critical AI infrastructure layer as operators move toward solid-state transformers and higher-voltage data-center architectures. Click here
Samsung ships samples of its next-generation HBM4E AI memory chips to customers. The new high-bandwidth memory product delivers higher performance, capacity, and efficiency, supporting growing demand for AI accelerators and data center infrastructure. Click here
Enterprise
Anthropic releases Claude Opus 4.8 with dynamic workflows for large-scale agentic coding and professional work. The release pushes enterprise AI from single-agent task completion toward coordinated subagent execution, effort controls, and longer-running software workflows. Click here
Snowflake expands its AWS partnership with a $6 billion commitment to accelerate enterprise agentic AI adoption. The agreement deepens collaboration around cloud infrastructure, data platforms, and AI deployment for enterprise customers. Click here
Wipro expands its ServiceNow partnership to embed agentic AI workflows across IT, HR, procurement, and cybersecurity. The partnership integrates Wipro Intelligence with the ServiceNow AI Platform, pushing enterprise AI from isolated pilots toward governed workflow execution across core operating functions. Click here
Visa invests in Replit to support agentic payments for developers. The partnership explores integrating Visa’s payment infrastructure into Replit, enabling AI agents and applications to securely accept and process transactions. Click here
Ultimo launches digital workers for industrial maintenance teams across planning, technician workflows, and safety operations. The release brings agentic AI into enterprise asset management, embedding specialized workers into Microsoft Teams and maintenance systems for governed industrial workflow execution.Click here
Capital Flows
Anthropic raises $65 billion at a $965 billion post-money valuation. The round concentrates capital around frontier-model deployment, enterprise adoption, and compute expansion at a scale that pushes private AI financing into public-market-sized territory. Click here
Cognition raises more than $1 billion at a $25 billion pre-money valuation for its AI software-engineering platform. The financing shows capital continuing to concentrate around agentic coding systems as software development becomes one of the first enterprise labor markets exposed to autonomous AI substitution. Click here
OpenRouter raises $113 million to expand its AI model exchange as usage reaches 25 trillion tokens per week. The round reflects growing demand for inference routing, model governance, failover, and cost optimization as enterprises shift from single-model adoption to multi-model production architectures. Click here
Utilidata extends its Series C to scale AI-driven power optimization for data centers. The financing supports the deployment of NVIDIA-powered grid-edge modules designed to uncover unused electrical capacity and improve data-center power utilization. Click here
Pace raises $46 million from Thrive and Sequoia to scale AI agents for insurance operations. The round shows vertical workflow automation attracting growth capital, where agents can execute regulated back-office tasks across submissions, claims, servicing, and document processing.Click here
Research
Researchers introduce SiDP, a memory-efficient inference architecture for large-scale offline LLM serving. The work reframes inference scaling around shared-weight and bandwidth-aware execution strategies that increase usable GPU memory capacity and throughput without proportional hardware expansion. Click here
Researchers develop a self-aligning photonic material that processes light directly on silicon chips. The breakthrough could enable more efficient optical components for AI data centers by reducing the need for energy-intensive signal conversions between light and electronics. Click here
Zuckerberg’s Biohub unveils an AI “world model” of protein biology. The release pushes AI-enabled science toward programmable biology, combining protein-structure prediction, protein language modeling, and large-scale biological mapping into a research platform. Click here
Nimble publishes research showing retail AI agents waste most of their web-reading context. The findings point to retrieval efficiency and web-interface design as practical bottlenecks for commercial agent deployment. Click here
TELUS Digital releases a GenAI safety benchmark for enterprise AI applications. The benchmark adds a deployment-oriented lens to AI safety evaluation as enterprises move from model access toward application-level risk testing.Click here
Policy
Illinois lawmakers pass a frontier AI safety bill requiring third-party audits of major model developers. The bill would move U.S. AI governance beyond self-reported commitments by requiring independent verification of safety practices for frontier labs. Click here
OpenAI publishes its Frontier Governance Framework for advanced AI systems. The framework outlines how the company approaches risk assessment, model oversight, security, incident response, and compliance with emerging AI regulations in the U.S. and Europe. Click here
Coachella considers a data center moratorium after local backlash over a major AI infrastructure campus. The move shows AI policy pressure shifting into local permitting and land-use fights that can slow compute buildouts before state or federal rules intervene. Click here
Los Angeles and Riverside courts begin testing AI systems to accelerate legal-document review and case processing. The deployment signals growing institutional adoption of AI inside state judicial infrastructure, where workflow automation, document triage, and operational efficiency are becoming public-sector implementation priorities. Click here
Ohio suspends a key data-center tax break amid AI infrastructure costs and local opposition, pressuring state policy. The pause turns AI data center incentives into an operating risk signal for compute developers, showing how state tax policy, grid costs, and community resistance can directly affect infrastructure siting economics. Click here
Global AI Strategy
The EU drafts a tech-sovereignty strategy with proposed cloud and AI infrastructure measures. The plan shifts European AI strategy from regulating foreign platforms to building domestic cloud, AI, semiconductor, and data center capacity. Click here
China plans to embed AI across national energy infrastructure to optimize power-grid operations, forecasting, and electricity-system coordination. The initiative ties AI deployment directly to grid management, renewable-energy integration, and power-system efficiency, reflecting how sovereign AI strategy is increasingly converging with national energy infrastructure planning. Click here
Alibaba Cloud launches an agentic AI ecosystem for global enterprise customers. The release packages Qwen models, agent tools, MCP-compatible cloud skills, execution sandboxes, and enterprise agent suites into a full-stack platform for agent deployment. Click here
Eleveight AI launches Armenia’s first Blackwell-powered AI factory. The deployment positions Armenia and the South Caucasus as an emerging regional AI infrastructure node, with Blackwell capacity becoming a sovereign and regional competitiveness signal. Click here
Canada will release a refreshed national AI strategy next week. The plan is expected to focus on sovereign AI infrastructure, AI adoption, workforce training, safety, and scaling domestic AI companies. Click here
Talent Signals
Each week, we spotlight key roles tied to the themes shaping this week’s AI headlines, connecting talent to the companies driving the news.
OpenLight delivers integrated silicon photonics platforms that bring optical I/O closer to AI accelerators and data center switches, enabling the high-bandwidth, energy-efficient connectivity roadmap the AI era requires. OpenLight is hiring across photonics design, packaging, test, PDK development, and platform engineering. Open roles are listed on its careers page. Click here
Fluidstack delivers high-density GPU clusters for frontier labs, governments, and sovereign AI initiatives. As nation-state compute becomes a strategic asset and AI labs look for large-scale infrastructure beyond traditional hyperscalers, Fluidstack is expanding across data center engineering, supply, platform, and operations. Open roles are listed on its careers page. Click here
Oklo is developing advanced nuclear power plants designed to deliver reliable, clean baseload power for industrial customers, including AI data centers. As power availability becomes one of the biggest constraints on AI infrastructure growth, Oklo is hiring across nuclear engineering, licensing, project development, manufacturing, and operations. Open roles are listed on its careers page. Click here
You can see all the opportunities at Mayfield-backed AI companies here.
Social Signals
The most important conversations in AI are unfolding across social media, where top voices are shaping the next wave of signals and strategy. Here are some of the top social signals and their takes from the past week.
Aravind Srinivas (Click here) — “What matters is that you do it with incredible velocity. Fail faster, succeed faster. Never slow down. Velocity is not our moat. It’s our superpower. The secret to success in life is who has the best questions. Your ability to have the best questions will be the defining skill of your life.” In a commencement address that generated nearly 10,000 reactions and more than 150 reposts on LinkedIn, Srinivas argued that AI makes curiosity, learning, and speed more important, not less. His message to graduates was that in a world where answers are increasingly abundant, the advantage shifts to those who can learn continuously, compound knowledge, and ask better questions than everyone else.
Garry Tan (Click here) — “The bottleneck has never been compute or capital. It’s taste and judgment about what humans actually want. Infinite compute just makes the great founders faster and the confused ones more confused.” Tan agrees with the idea that technology alone does not create great products. As AI and compute become more abundant, the advantage increasingly shifts to founders who understand customer needs, make strong product decisions, and can translate new capabilities into things people genuinely want.
David Lieb (Click here) — “Thought experiment: if every company suddenly had infinite free compute, what new products would emerge? My take: with very few exceptions, not much would change. The bottleneck is figuring out what people want, and it’s not so easy to apply compute to solve that.” In a post viewed 170K+ times, Lieb argues that the limiting factor in innovation is not access to technology, but understanding customer needs. As AI capabilities become more abundant and accessible, product judgment, user insight, and distribution increasingly determine which companies create lasting value.
