May 08, 2026

The most telling AI-infrastructure story today didn’t come from a hyperscaler—it came from Washington. In a new partnership, the U.S. Department of Energy and NVIDIA plan two AI supercomputers at Argonne, scaling from 10,000 Grace Blackwell GPUs to a planned 100,000 GPUs (~5,000 exaflops)—and DOE used the moment to talk bluntly about building more power, including gas, nuclear, and even coal. That message lands just as corporate clean-energy promises start to wobble under AI load.

The Big Stories

DOE and NVIDIA Team to Power AI with Energy Leadership lays out the most concrete “state-backed AI factory” blueprint we’ve seen in a while. Argonne’s Equinox targets 10,000 NVIDIA Grace Blackwell GPUs, while Solstice is planned for 100,000 GPUs (DOE frames this at roughly 5,000 exaflops). The subtext is the bigger signal: DOE is pairing compute ambition with a power-supply narrative—explicitly calling for scaling generation (natural gas, nuclear, coal) and pointing to three SMRs going critical by July 4. For anyone underwriting AI data centres, this is the clearest acknowledgement that “more chips” is now inseparable from “more firm power” and faster interconnection work.

Core Scientific to scale Muskogee campus to 1.5 GW is a reminder that the biggest AI plays in the U.S. are increasingly power-and-land plays wearing a data-centre badge. Core Scientific says Muskogee, Oklahoma will reach roughly 1.5 GW gross power capacity, alongside an acquisition of Polaris DS LLC (which controls 440 MW contracted power with Oklahoma Gas & Electric), targeted to close in Q3 2026. The company is also building an unleased 82.5 MW building for Q4 2027 delivery, and pairs Muskogee with a Pecos, Texas 1.5 GW plan—about 3 GW gross aimed at AI workloads. The read-through: the “AI campus” race is turning into a contest for contracted power blocks and multi-year construction sequencing, not just leased capacity.

Thailand approves 958 billion baht digital and data centre investments shows how quickly Southeast Asia is moving from “secondary market” to “capital magnet” when the permitting and policy mood is friendly. Thailand’s BOI approved six projects worth 958 billion baht ($29bn), led by TikTok System (Thailand)’s 842 billion baht ($25bn) data infrastructure expansion; three data-centre projects total 913 billion baht ($27bn). The BOI also discussed electricity readiness, Direct PPAs, and UGT2 to improve clean-energy access. For investors, this is policy-driven demand creation: the country is explicitly trying to align power-market plumbing with digital capex.

Microsoft may delay 2030 renewable matching goal amid AI is the clean-energy reality check the market has been expecting—just not so soon, and not from Microsoft. Bloomberg reports Microsoft is considering delaying or abandoning its 2030 goal to match hourly electricity use with renewables as AI data-centre demand spikes, even as the company points to a 1.2 GW We Energies solar and battery deal (online December 2028) and a prior Constellation/Three Mile Island-linked power deal. If Microsoft—arguably the poster child for ambitious power procurement—needs more time, it raises the obvious question: which operators will be forced into “good enough” procurement, and which grids will tighten their tariff stance on large loads.

Andhra Pradesh SIPB approves ₹2.01 lakh crore investment package bundles India’s data-centre buildout with generation and storage in a way that’s becoming the template. The state cleared ₹2.01 lakh crore of proposals including a Reliance data centre in Visakhapatnam (>₹1 lakh crore), a ₹51,300-crore Reliance solar+BESS project, and a ₹12,297-crore Adani pumped storage hydro project. The signal is less about any single asset and more about packaging: states want the data centre, but they also want the power build alongside it—and they’re willing to fast-track a whole stack of projects to get there.

Behind the Headlines

The NC Clean Energy Technology Center’s Q1 2026 “50 States of Power Decarbonization” report is worth reading as a proxy for how messy the U.S. power build will be under large-load pressure. It logged 509 state and territory actions (49 states + Puerto Rico) and noted 600+ introduced bills, with planned additions including 58,276 MW solar, 54,952 MW natural gas, 30,297 MW storage, and 22,358 MW wind, alongside 30,967 MW of planned coal retirements. Buried in that is the part data-centre investors actually feel: jurisdictions are actively tinkering with large-load/data-centre tariff actions. Translation: even where generation is coming, the “who pays for upgrades, and on what timeline” fight is getting louder—and will increasingly dictate where mega-load campuses can pencil.

CMBlu closes €50m Series C, valuation tops US$1 billion is a small funding round with outsized implications for long-duration storage narratives around data centres. The company raised €50m with participation from Samsung Ventures, pushing valuation above $1bn, and it has a conditional 5GW supply agreement with Uniper (valid until 2037) with deliveries from 2027 in batches of at least 100MWh. CMBlu is marketing its FEOC-safe SolidFlow organic flow battery for data centres and multi-hour LDES. If even a slice of those deliveries materialise, it strengthens the case for storage as a “permitted alternative” when clean firm supply is constrained—especially in markets where gas peakers are politically hard.

Debate over US–Gulf partnerships in AI infrastructure race captures the geopolitical friction sitting underneath the AI buildout. Policy experts argue Gulf states can be key partners because they have capital, energy, and the ability to host large-scale data centres—but the story highlights how slowly chip flows can move when export controls bite: a 2025 proposed agreement to send advanced Nvidia chips to UAE-based G42 has progressed slowly, with only a fraction approved and none delivered. For the industry, this is a reminder that “build the campus where the power is” only works if the compute supply chain can legally follow. Expect more projects to be structured around compliance optics—who operates, who owns, where model training happens—rather than pure economics.