Where the math is defensible.
Long-form research on live enterprise decisions. Publication is selective. Every number traces to a named source. No takes without evidence.
The African data center buildout 2026: Lagos, Nairobi, Cape Town, and Cairo
Africa hosts roughly 1.5 percent of global colocation capacity but is on track to triple installed megawatts by 2028, with Lagos, Nairobi, Cape Town, and Cairo absorbing the bulk of new builds and forcing fresh thinking on power, regulation, and inference latency.
African data center capacity is finally inflecting in 2026. Installed colocation power across the continent now sits near 700 megawatts, still a small share of the global 55 gigawatt market, but contracted pipelines suggest a path to roughly 1.8 gigawatts by 2028. South Africa, Nigeria, Kenya, Egypt, and Morocco anchor that growth. The st...
The Custom Silicon Insurgency Against Nvidia in 2026
AWS Trainium 2, Google TPU v5p and Trillium, Microsoft Maia, Meta MTIA, and a possible OpenAI ASIC are reshaping where AI compute margin lives, but the binding constraint sits one layer down at HBM and CoWoS-L.
The hyperscalers spent 2024 and 2025 telling investors that custom silicon would relieve their dependence on Nvidia, and 2026 is the year those claims start meeting empirical scrutiny. AWS has stood up Project Rainier for Anthropic at a publicly disclosed scale of more than one million Trainium 2 chips. Google has placed its TPU v5p gener...
Johor's Data Center Boom: Singapore's Spillover and Malaysia's Grid Bet
Singapore's moratorium pushed roughly 1.6 GW of latent demand across the causeway, and Johor is now Southeast Asia's most concentrated hyperscaler buildout. The binding constraints are grid, water, and sovereign data law.
Between Singapore's 2019 moratorium and the partial 2022 reopening under the Pilot Call for Application, roughly 1 to 2 GW of regional hyperscaler demand had no clean home. Johor absorbed most of it. Sedenak Tech Park, Iskandar Puteri, Kulai, and Pengerang now host announced commitments from Microsoft, AWS, Google, ByteDance, Equinix, GDS...
Small Modular Reactors in 2026: Order Books, AI Off-take, and the Capex Curve
The first wave of SMR designs has moved from licensing slides to concrete pours and signed power purchase agreements, but the NuScale UAMPS cancellation still anchors investor memory and the load is now hyperscaler, not municipal.
Advanced nuclear is no longer a slide deck category. TerraPower has broken non-nuclear ground at Kemmerer, X-energy has signed a definitive engineering contract with Dow Seadrift, Holtec is preparing the SMR-300 around the Palisades restart, Kairos has the first Hermes test reactor under construction with a Google off-take for 500 megawat...
PJM capacity market in 2026: what the next auction is telling us
After the 2025-26 delivery year auction shocked the market with a more than ninefold price jump, the upcoming PJM capacity auctions will determine whether the largest U.S. power market can reconcile data center demand with a thinning generation stack.
PJM Interconnection runs the Reliability Pricing Model, a forward capacity construct that procures resource adequacy three years ahead for 65 million customers across 13 states and the District of Columbia. The July 2024 base residual auction for delivery year 2025-26 cleared at $269.92 per megawatt-day, roughly nine times the prior year,...
Texas, ERCOT, and the AI Siting Reset
Senate Bill 6, the ERCOT large-load study, and a 30 to 40 GW interconnection queue are forcing hyperscalers to rethink West Texas, behind-the-meter gas, and the price of speed.
Texas became the default home for marginal United States AI compute through 2024 and 2025 because ERCOT offered the only grid in North America that could absorb gigawatt-scale loads on a multi-year horizon rather than a multi-decade one. That window is closing on its own terms. The ERCOT December 2025 long-term load forecast now carries r...
AI capex met the grid: when the megawatt curve breaks
Hyperscaler capital spending crossed 500 billion dollars across 2025 and 2026 while the average US interconnection wait sits above 4 years. The constraint is no longer chips. It is megawatts on a calendar.
AI infrastructure capex has cleared 500 billion dollars across 2025 and 2026 between the four hyperscalers, NVIDIA, Oracle, and the new wave of neoclouds. The chips are arriving. The grid is not. Lawrence Berkeley National Lab puts the active US interconnection queue above 2,600 gigawatts, with median wait times above 4 years and rising. ...
Quebec hydropower and the new gating of AI compute
Quebec spent two decades selling itself as the cheapest, greenest place on the continent to plug in a data center. In 2026 Hydro-Quebec is throttling new connections, redesigning industrial tariffs, and forcing hyperscalers to rethink where the next gigawatt of AI training capacity actually lands.
Quebec sits on roughly 37 gigawatts of installed hydro capacity, historically the cheapest large-scale clean power in North America. From 2018 the province first banned new bitcoin mining connections, then welcomed AI campuses, and by 2024 began throttling all new large industrial loads above 5 megawatts. Hydro-Quebec's 2035 strategic pla...