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.
ASEAN Sovereign AI in 2026: Models, Compute, and the Regulatory Patchwork
Singapore is buying TPU access while building SEA-LION, Indonesia is shipping Sahabat-AI in five languages, Thailand is scaling Typhoon, Malaysia is funding chips, Vietnam is exporting PhoGPT, and Manila has finally passed an AI Act, but the United States Diffusion Rule and a fragmented data sovereignty regime put a ceiling on the regional ambition.
Across 2025 and into 2026, every large ASEAN economy moved from announcement to delivery on sovereign artificial intelligence. Singapore is executing the December 2023 National AI Strategy 2.0 (NAIS 2.0) under AI Singapore (AISG), with the SEA-LION family now in its third generation and a S$1 billion Strategic AI Compute fund underwriting...
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...
Frontier AI training cost trajectory 2026: the run rate, the deal stack, and the power-bound horizon
Frontier pretraining budgets crossed the half billion mark in 2025 and are heading toward one to three billion dollars per model by 2027, with cluster power, not GPUs, now the binding constraint on the next order of magnitude.
Frontier model training compute scaled at roughly 4x to 5x per year between 2018 and 2024, sat near 10x per year for the leading lab releases, and now confronts a deceleration driven by power, capital, and data, not by silicon. Epoch AI puts GPT-4 near 2e25 FLOPs at a roughly 80 million dollar training cost, Claude 3.5 Sonnet near 3e25, a...
Generative Video AI in 2026: Compute Economics, Hollywood Disruption, and the Copyright Reckoning
Frontier video models now generate 1080p clips in single-digit seconds, training runs cleared USD 100 million per system, and inference still costs cents per output second. The constraint is no longer fidelity but rights, latency, and where the marginal dollar of creative budget lands.
Generative video moved from research demo to production tool over fifteen months. OpenAI shipped Sora 2 in December 2024 with 8 second 1080p clips, Google Veo 2 and Veo 3 added native vertical and longer durations, Runway Gen-4 solved character consistency, and Chinese entrants Kling 2.0 from Kuaishou and Hailuo from MiniMax forced global...
Korea's Memory Cycle 2026: HBM Lead, Logic Gap, Cluster Risk
SK hynix runs the HBM3e 12-Hi book at Nvidia, Samsung Memory chases qualification while its foundry yield gap widens, and the Yongin cluster meets a hard power and water ceiling.
Korea's memory complex enters 2026 with the cleanest revenue setup in a decade and the messiest strategic position in twenty years. SK hynix is the lead supplier of HBM3e 12-Hi to Nvidia for Blackwell Ultra and the Rubin ramp, holds roughly half of HBM revenue, and has a credible HBM4 roadmap pegged to 2026 risk production. Samsung Memory...
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...
Quantum Computing Economics in 2026: Roadmaps, Modalities, and the Post-Quantum Migration
IBM, Google, IonQ, Quantinuum, Atom Computing, PsiQuantum, and Microsoft are converging on fault tolerance from four physical platforms. Quantum equities have repriced sharply, NIST has finalized its post-quantum cryptography standards, and national programs have moved past 20 billion dollars of cumulative public commitment.
Quantum computing in 2026 is four parallel scaling experiments on transmon, trapped-ion, neutral-atom, and photonic platforms, judged against a shared bar set by the surface code. IBM ships Heron R2 at 156 qubits and targets Kookaburra in 2026 and Blue Jay at 4158 qubits. Google's Willow showed below-threshold operation in late 2024. IonQ...
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...