Read Outs
In-depth explainers on the topics and companies shaping the AI stack — what's happening, why it matters now and over the next few years, how it actually works, and where the real tensions and open questions lie.
- topicCurrent
Storage Gets a Promotion: AI as a Data-Movement Problem
The case that AI infrastructure is shifting from a compute constraint to a data-movement constraint is directionally sound and increasingly reflected in 2026 silicon and cloud roadmaps, but its headline economics rest on undisclosed math and vendor-aligned framing rather than neutral benchmarks. Promising as a reframing, not yet established as a universal law.
Memory & Storage Silicon· NAND / FlashModel Distribution & Serving· Inference PlatformsObject & File Storage· High-Performance Object StoreAug 16, 2026 - topicCurrent
200GW Hiding in the Grid: A Podcast's Optimistic Map of AI's Power Problem
Episode #280 of Moonshots with Peter Diamandis argues that AI's binding constraint is power infrastructure, not compute, and pitches latent grid capacity, cheap sodium-ion storage, and wave-powered datacenters as fixes. The framing is provocative and directionally aligned with real grid-queue pain, but its headline figures rest on secondary summaries rather than verified data, and its splashier solutions remain conceptual.
Power Generation· Renewable EnergyPower Distribution & Electrical· Grid-Scale Storage & MicrogridsData Center Operators· Clean Energy DCAug 16, 2026 - topicAging · 23d
The Capital Allocator's Four Questions: A Capex Sink Betting on a Loop It Cannot Yet Prove
The AI buildout is transferring hyperscaler balance sheets almost directly to hardware monopolies, with the sector's economics resting on a research-automation loop that is empirically real at the coding layer but unproven at the frontier. The cash-flow math is transmission-rich for chipmakers and transmission-negative for model labs and hyperscalers, and markets have begun repricing that asymmetry rather than resolving it.
AI Capital Allocation· Macro ScenariosAI Capability· Frontier ResearchAug 3, 2026 - topicAging · 23d
Project Sopaipilla: How Meta Financed a Gigawatt Off Its Own Books
Meta transferred majority ownership of its ~1 GW El Paso AI campus to BlackRock-managed funds and financed completion with roughly $12.3-12.5 billion in project bonds serviced by its own long-term rent — a structure that shifts capital and equity risk to BlackRock while Meta keeps operational control, though its true off-balance-sheet character remains unproven.
Data Center Operators· Wholesale / Hyperscale LeasingInfrastructure Investment· DC-Focused DeveloperAug 3, 2026 - topicAging · 26d
Three Planes, One Gateway: The 2026 Multi-Model Routing Stack
By mid-2026 enterprise AI infrastructure has coalesced around a centralized gateway that routes requests across multiple models and providers, separating intent-based allocation, governance policy, and serving-capacity dispatch. The layered separation is analytically useful and widely reflected in vendor architectures, but it is not a standardized taxonomy, and the cost and cache-efficiency figures behind it remain vendor-reported rather than independently audited.
Inference Infrastructure· Model RoutingEnterprise Automation· AI Control PlanesAI Models· Multi-Model SystemsJul 31, 2026 - topicAging · 26d
Weights, Not Walls: A Field Guide to Open, Open-Weight, and Closed AI
Openness in AI now runs on a spectrum from fully open-source to open-weight to closed, with the practical dividing line being whether a model's trained parameters can be downloaded. The definitions are settling but not settled, and the policy and safety implications remain genuinely two-sided.
AI Models· Open Weight ModelsInference Infrastructure· Model StrategyJul 31, 2026 - topicStale · 45d
The Gap Is Real But Not a Number: Convergence, Compute, and the Grid Wall
The evidence supports a narrowing but variable U.S.–China frontier-model gap that resists any single "six-month" figure, with American labs holding the hardest reasoning tasks while Chinese models commoditize standard coding at a fraction of the cost. For infrastructure capital, the binding constraint is shifting from model advantage to power delivery and buildout execution — record spending is real, but so are the stalled projects colliding with it.
Frontier Foundation Model Labs· Closed-Source FrontierFrontier Foundation Model Labs· Open-Weight FrontierServerless Inference· Inference API PlatformJul 12, 2026 - topicStale · 77d
Power Down: Two Bearish Notes, One Hair-Trigger Market
A pair of institutional research notes in early June 2026 argued that co-packaged optics and NVIDIA's 800V DC power architecture are ramping slower than consensus, triggering a violent one-day selloff across optical names — but the underlying delay math is contested and rests on secondary summaries of a paywalled report, making this a de-risking of a crowded trade rather than a proven fundamental break.
Optical Interconnect & Photonics· Co-Packaged Optics / Optical I/OGPU & AI Accelerators· Training & Inference GPUPower Distribution & Electrical· Power Intelligence / ControlsJun 10, 2026