February 20, 2026

G42 is about to reset the scale conversation in India. The UAE AI group has signed a term sheet with C‑DAC to deploy an 8 exaflop national AI supercomputer cluster hosted in-country — nearly 19x the combined peak of AIRAWAT and PARAM Siddhi-AI — with Cerebras and MBZUAI involved, and an explicit promise that all data stays in India. That’s not just a bigger machine; it’s a statement about where AI compute (and control) is meant to live.

The Big Stories

G42 finalises term sheet for an 8 exaflop AI supercomputer cluster hosted in India with C‑DAC, partnering with Cerebras and MBZUAI, and positioning the system as open to startups plus public/private entities — while keeping data sovereign. The headline number matters because it collapses the gap between “national AI infrastructure” as a policy slogan and as an actual procurement-scale system. If this lands, it raises the bar for domestic hosting expectations for anyone selling AI services into India.

Enterprise cloud infrastructure spending hit $119.1B in Q4 as GenAI drives growth, with Synergy putting the full-year market at $419B and ~30% constant-currency YoY growth; CoreWeave is now posting >$1.5B in quarterly cloud revenue. In the same cycle, Cisco rolled out its Silicon One G300 ASIC (102.4 Tbps) aimed squarely at AI networking bottlenecks. The connective tissue here is simple: demand is no longer the question — the constraint is increasingly the fabric (and the ability to keep expensive GPUs fed).

The grid is becoming the limiting reagent, and the IEA is saying it out loud. The agency forecasts renewables plus nuclear reaching 50% of global generation by 2030, but warns that >2.5 TW of projects are stuck in connection queues and that grid investment needs to rise 50% by 2030 — with up to 1.6 TW potentially unlockable via reforms and grid-enhancing tech (IEA Electricity 2026: queues and grid capex are the bottleneck). That macro view shows up locally too: in Virginia, Dominion is proposing two 230 kV transmission rebuilds near Charlottesville and Gordonsville tied to regional load growth from data centres, with the Piedmont Environmental Council urging SCC comments by March 25, 2026 (Dominion’s 230 kV rebuild proposals draw scrutiny). More wires is the obvious answer — but it’s also where permitting politics meets load forecasts.

Nuclear-for-data-centres is moving from conference talk to consortium slides and early money. Deep Atomic and partners are pitching SMR-powered AI data centre “campuses,” including a DoE-submitted concept for Idaho National Laboratory; the MK60 targets 60MW electric plus 60MW cooling, with series manufacturing ambitions by end‑2029 and an NRC pre-application already referenced (Deep Atomic consortium proposes SMR-powered AI campuses). In parallel, NuCube Energy has raised $13M to push testing, licensing, and demonstration of a high-temperature modular microreactor in the US (NuCube raises $13M for microreactor work). Read these together as a signal that “behind-the-meter baseload” is now an investable narrative — but the timetable is still the story.

In Brief

• UK social licence is getting more concrete: E.ON says 66.8% of respondents view nearby data centres more favourably if waste heat feeds local heat networks, and flags its Silvertown ectogrid project in East London as saving ~4,000 tonnes of CO2 annually (UK survey backs waste-heat reuse).

• Florida lawmakers are taking aim at large AI sites: a Senate committee passed SB 484 covering rate protection, water permitting, local authority, and banning NDAs that obscure impacts (Florida committee advances AI data centre restrictions).

• Microsoft is still experimenting with what “storage” could mean: Project Silica demonstrated borosilicate glass media storing up to 4.8 TB on a thin 120mm square sheet, with claimed durability up to 10,000 years (Project Silica glass storage proof points).

• India’s grid buildout is being procured around new industrial load: RECPDCL issued an RFP for interstate transmission serving proposed green hydrogen/ammonia projects and datacenter loads in Vizag (Phase I), with a 30-month COD target and a 35-year BOOT term (Vizag transmission RFP).

• US environmental regulation is swinging back toward “keep coal online”: the Trump administration said EPA will roll back Biden-era mercury and air toxics rules for coal plants, explicitly framing it as cost relief and baseload support amid rising demand from data centres (EPA to roll back mercury/air toxics rule).

• Grid-adjacent costs are surfacing in retail pricing: Columbia Gas Ohio lifted its default supply rate by 35% effective Feb 1, citing cold weather, high spot prices, and demand from LNG exports and data centers (Ohio gas supply rate jumps).

• The politics of “where you build” keeps getting messier: the Town of Marana rejected referendum petitions seeking to overturn a rezoning for a proposed data center due to missing legal property descriptions (Marana rejects data center referendum petitions).

• Storage economics keep bending in the right direction: BNEF says the global benchmark cost for a 4-hour BESS fell 27% YoY to $78/MWh in 2025 (BNEF benchmark BESS cost falls).

What to Watch

• India compute + sovereignty: the G42/C‑DAC term sheet is explicit on in-country hosting and data residency — expect that posture to echo through how AI cloud and colocation are sold into India (G42’s 8 exaflop India plan).

• The grid backlog becomes the investment thesis: the IEA’s queue numbers and capex call sit alongside on-the-ground transmission rebuild proposals — more projects will live or die on interconnection and permitting, not racks and real estate (IEA queue warning; Dominion rebuild proposals).

• AI networking is the new choke point: cloud spend is accelerating, but the next wave of competitive advantage looks like fabric efficiency and utilization gains as much as GPU supply (Synergy/Cisco AI networking push).