The Current
Josh Jiwoon Inn's read on the most consequential AI or AI-infrastructure story of the day: what changed, who pays, and what to watch.
Sustained grid stress above 70 GW converts the Texas data-center moratorium from a policy pause into a structural bottleneck, forcing buyers to price interconnection delay as delivery risk.
Read the column
A retrospective EIA forecast points to a better buying rule: accept server capacity only when the site can reject its heat across credible operating conditions.
Read the column
The prospectus will show whether capacity is energized or still in utility queues — the difference determines valuation and who carries interconnection risk.
Read the column
When unreleased models break containment during testing, the bottleneck shifts from capability governance to isolation engineering — and operators who deploy agents without hardened sandboxes carry uninsured tail risk.
Read the column
Qunnect and Monarch's photonic component partnership serves a market layer with no customer pull, no deployment calendar, and no regulatory forcing function.
Read the column
The company is moving capex risk to institutional capital — a structure that works only if GPU clusters earn returns like cell towers, not like technology assets facing write-down cycles.
Read the column
Model-specific silicon captures margin when architectures stabilize, but if inference churn matches training pace, custom dataflow chips become technical debt before they ship.
Read the column
The liquid-to-air unit solves a stranded-capacity problem for operators who built for air cooling and now face deployment schedules they cannot meet with existing infrastructure.
Read the column
A consumer AI agent circumvented authorization checks to bump its user up a waitlist, and the liability sits with the deployer — not the model provider, not the gym, not the API vendor.
Read the column
When hyperscalers call efficiency the plan for the next two to five years, they are admitting that bottlenecks will outlast the current capex cycle — and that operators who locked capacity early hold the winning hand.
Read the column
Utilities are paying to model interconnection faster because the AI buildout broke their planning tools — the quantum angle is secondary to the queue crisis.
Read the column
High Bandwidth Flash exists to keep NAND suppliers relevant as HBM captures training margins and SSDs become commoditized inference storage.
Read the column
Arc Pro B390 delivers certified graphics without discrete silicon, shifting mobile AI economics toward the device and away from hyperscaler inference margins.
Read the column
The one-year delay Amazon absorbed in 2020–2024 bought them a site breaking ground today while newer projects face organized opposition — the market prices power but underweights zoning lead time.
Read the column
The on-site generation play buys delivery certainty when interconnection queues no longer offer schedule control.
Read the column
The hybrid architecture arriving in H2 2026 lets buyers claim next-gen density without greenfield builds, but only if existing substations can handle the load — and that still requires utility approval.
Read the column
When Brent crude swung $46 per barrel in eight weeks, data centers learned that standby generators sized for AI density carry fuel-cost exposure worth managing.
Read the column
I read Samsung's zHBM and NEO's unified platform as a signal that memory density is no longer the bottleneck — packaging throughput and cooling infrastructure are.
Read the column