Zephyr built a bottom-up model of DRAM/HBM capacity and AI demand, forecasting that accelerator memory, agentic CPU memory, and other AI workloads could create a major DRAM shortfall by 2030. This is a forecast/model, not a confirmed market outcome.

If the model is directionally right, AI’s next bottleneck may sit in the memory layer. The race is not only about building more GPUs; providers also need enough HBM and DRAM to feed accelerators and increasingly memory-hungry agents.

According to TechCrunch, Satya Nadella argues that companies pay for AI through their subscriptions, then pay again through the proprietary knowledge they leave behind in prompts, tool calls, and corrections.

That argument puts workflow traces and evaluation data in the category of strategic assets. If those traces become part of a provider’s learning loop, companies need to understand what they own, what they share, and what advantage they may be giving away when they outsource AI.

Gavin Baker | cheaper models could make AI infrastructure grow faster: Gavin Baker offers a conditional thesis: if demand shifts from frontier labs with high inference margins toward cheaper models, users may get more intelligence per dollar, creating additional token demand and moving margin toward infrastructure providers. Michael Burry has expressed a similar view, but this remains an investor thesis rather than a confirmed market forecast.

Meta expands its Louisiana AI data center to 5GW and more than $50 billion: Meta announced an expansion in Louisiana to 5GW alongside more than $50 billion in total investment. That is announced investment/capacity, not proof that the full operating capacity is already online, but it shows the scale at which capital and power are being committed to AI.

Nous Research is reportedly raising for Hermes: TechCrunch reports that Nous Research is reportedly discussing a new funding round at a $1.5 billion valuation for Hermes, potentially worth at least $75 million. If completed, it would add to the signal that open/local agent products with built-in skills are attracting capital outside the frontier labs.

PixVerse raises $439 million and crosses a $2 billion valuation: TechCrunch reports that video-generation startup PixVerse has raised $439 million and passed a $2 billion valuation. Capital continues to gather around video generation and world models, with companies targeting both consumer and enterprise distribution.

Space data centers are still far from economic reality: TechCrunch reports that experts generally view orbital AI data centers as an economically distant idea: launch costs would need to fall sharply, while high-powered satellites would need to be mass-produced. The signal is a useful reality check against space data-center promises.

Apple is rumored to be planning an M7 Ultra with 1.5TB of RAM: AppleTrack reports a rumor that Apple’s M7 Ultra, expected in 2029, could support 1.5TB of RAM. If accurate, that memory capacity could expand the chip’s on-device AI capabilities; for now, the specification remains reported/rumored.

Claude Artifacts move toward collaborative workflows: TestingCatalog reports that Claude Artifacts now support public sharing, multiplayer editing in Claude Code, and creation through Claude Tag. Bringing the workflow into Slack also reduces the number of times users need to leave the place where work is already happening to use AI.

Anthropic starts localizing Claude pricing in India: TechCrunch reports that Anthropic has started displaying Claude pricing in rupees in India, its second-largest market, while UPI support is not yet available. Frontier-model competition is moving into payments and distribution: reducing payment friction can matter almost as much as improving the model.

Joon Lee shared results from 3,000 FrontierCode 1.1 sessions: Fable 5 with a cheap sidekick scored 60.7 at $1.86 per run, while Opus 4.8 with the same sidekick scored 54.6 at $2.04. The main lesson is how the lead model decides what to do itself and what to delegate.

Practical guidelines:

A quick test is to choose a coding task with a clear output, separate planning/review from implementation/testing, and compare a single-model run with a lead-plus-sidekick run. If lead turns fall, corrective edits do not rise, and quality stays the same or improves, the architecture may be worth expanding.

Simon Willison introduces a shift from zero trust to agent trust: agents need permissions, boundaries, and behavioral controls. It is a useful lens for building agents that can do real work without receiving more authority than the task requires.

Flexential’s vendor-sponsored report looks at power, network, compute, and cost as physical constraints on AI infrastructure. Its value is in pulling the AI conversation back toward the factors that can bottleneck an entire roadmap: electricity, connectivity, and operating cost.

This essay argues that companies need to retain ownership of their data, traces, evaluations, and model orchestration instead of merely buying access to a model. It offers a useful way to read the AI competition through ownership of the learning loop.

The Stanford Digital Economy Lab report focuses on measuring enterprise AI deployment inside real workflows rather than simply counting pilots or enabled accounts. Its framework helps separate AI that changes operations from AI that is still stuck at the experimentation stage.