Chips, Watts, and Diplomacy: AI’s new power struggle

⚡ TL;DR
AI’s voracious appetite for power is turning kilowatt-hours into geopolitical currency. As companies like OpenAI quietly diversify cloud suppliers and Washington eases chip restrictions on China, Wall Street is buying up electric utilities in anticipation of surging demand. The race for computing power—once about silicon alone—now hinges just as much on securing reliable energy sources.

The coming "cold chain" for AI infrastructure will reshape economies, diplomacy, and sustainability for decades to come.

1. Compute Diplomacy Is Back

1.1 OpenAI Goes Multi-Cloud

OpenAI quietly added Google Cloud as an infrastructure supplier in May, ending its de facto exclusivity with Microsoft Azure and signaling that “who owns the flops” is now a board-level risk factor. The updated partner list on OpenAI’s site shows it now relies on services from Google in addition to Microsoft, Oracle and CoreWeave. The Google deal – finalized after months of talks – underscores how massive computing demands to train and deploy AI models are reshaping competitive dynamics, and marks OpenAI’s latest move to diversify its compute sources beyond primary backer Microsoft. Earlier this year OpenAI partnered with SoftBank and Oracle on the $500 billion Stargate infrastructure program and signed multi-billion-dollar agreements with CoreWeave to hedge against chronic GPU shortages. By moving to a multi-cloud strategy (including a new deal with Google that had been blocked until its prior Microsoft lock-in expired in January), OpenAI is ensuring that “who owns the FLOPs” doesn’t become a single point of failure for its business.

1.2 Nvidia’s H20 Chip Re-Enters China

The race for advanced AI chips has made GPU export controls a major geopolitical bargaining chip.
On July 16, U.S. officials signaled that licenses will again allow Nvidia’s downgraded H20 GPUs into China – folding the decision into broader rare-earth mineral negotiations. The planned resumption is a reversal of export controls imposed in April to keep top-end AI chips out of Chinese hands. Commerce Secretary Howard Lutnick confirmed the H20 move was “put in the trade deal with the magnets,” referring to an agreement by President Trump to restart rare-earth exports from China. The U-turn illustrates how export controls have become a bargaining chip rather than a hard barrier. It could restore billions in Nvidia sales (the company had estimated a $15 billion revenue hit from the curbs) and has already set off a scramble among Chinese tech firms to order the chips. U.S. lawmakers voiced concern, noting the H20 still provides substantial AI capability (albeit legally pared-down for export) and works with Nvidia’s dominant software ecosystem. Nvidia’s CEO Jensen Huang, visiting Beijing, argued that completely cutting off China would only accelerate domestic alternatives (Huawei, etc.) and erode U.S. leadership. The significance of this policy shift ultimately depends on volume: “If China is able to get a million H20 chips, it could significantly narrow – if not overtake – the U.S. lead in AI,” one Washington analyst noted.

1.3 Federal Megabucks for Electro-Compute

At a Pittsburgh summit on July 15, President Trump touted a package of $90 billion in AI-and-energy projects across Pennsylvania, bundling data-center campuses with new power plants and grid upgrades. The inaugural Pennsylvania Energy & Innovation Summit saw executives from energy and tech giants (Blackstone, SoftBank, Google, ExxonMobil, etc.) announce tens of billions in planned investments. Taken together, Trump said, these moves show the state “reclaiming its industrial heritage” at the forefront of an AI-powered industrial revival. Among the headline commitments: Blackstone pledged $25 billion for new data centers in northeast Pennsylvania alongside natural gas generation to power them; Westinghouse promised to build 10 new nuclear reactors (a plan estimated at $75 billion economic impact) beginning in 2030; FirstEnergy announced $15 billion to upgrade the electric grid; and Google unveiled a $3 billion program to modernize hydroelectric plants in partnership with Brookfield, adding 670 MW of renewed capacity. In addition, state officials highlighted projects converting retired coal facilities into gas-fired “AI campuses” and efforts to fast-track permits for transmission lines. Political fanfare aside, some of these investments will take years to materialize – stretching into the 2030s – but they underscore a strategic alignment of compute and kilowatts. Notably, private capital is dovetailing with policy: Blackstone’s utility JV with PPL, for example, aims to supply power under long-term contracts to the very hyperscalers building AI data centers. Take-away: Access to both advanced chips and predictable electrons is emerging as a key lever of statecraft and corporate strategy.

2. Kilowatts—The Next Scarce Resource

2.1 Demand Curve

Soaring demand for AI compute is driving a wave of new data centers, expected to consume more power than some countries within a few years.
The International Energy Agency now projects global data-center electricity use to double to ~945 TWh by 2030, more than Japan’s current total power consumption. Generative AI and other machine-learning workloads are the dominant driver of this growth. In fact,AI-specific data centers could quadruple their energy draw by 2030 as companies race to train ever-larger models. Already, an estimated 1,240 U.S. data centers are built or under construction – nearly four times the number a decade ago – with hundreds aimed squarely at AI applications. Business Insider finds these facilities could soon consume more electricity than Poland (population 37 million) uses in a year. If current building trends continue, their power generation needs may triple over the next three years. The public-health externalities are likewise non-trivial: emissions from the plants powering these data centers could cause between $5.7 and $9.2 billion in annual health costs from pollution, by one estimate. In short, AI’s appetite for watts is growing at a breakneck pace – and scaling up electricity supply is now as critical as scaling model size. Some tech firms are responding by co-locating data farms directly next to power stations (even buying decommissioned plants), while others pursue renewables and efficiency gains to offset the surge. But as detailed below, even optimistic efficiency improvements might only bend the curve, not break it; total demand is still set to soar.

2.2 Utilities Become Tech Assets

Wall Street is racing to lock in generation and grid capacity for the AI era. Blackstone’s recent $25 billion bet on Pennsylvania (for data centers plus gas plants) is intended to catalyze another $60 billion in follow-on investment from partners and clients. In essence, data centers, power stations and pipelines are being treated as a single integrated asset class. Broader investor appetite has even turned stodgy utility-sector analysts into unlikely stars, as kilowatt-hour forecasts start moving stock prices. The trend hasn’t gone unnoticed by energy regulators: for example, BlackRock’s attempt to acquire a major Midwestern utility (Minnesota Power) to secure a data-center energy pipeline drew pushback from state officials concerned about rate hikes. Nonetheless, the buying spree continues. Private equity firm KKR recently took a stake in Nextera’s grid business, and Blackstone itself agreed in May to purchase TXU New Mexico’s electric utility serving 800,000 customer. The goal, as one report put it, is simple: profit from the rising demand for electricity from AI data center. This convergence is blurring the line between Big Tech and Big Power. It’s now common to see hyperscale cloud operators signing 20-year power purchase agreements or even funding new power plant construction. Investor presentations by utilities increasingly highlight data-center connections and AI load growth alongside traditional metrics. In the words of Google’s CFO: “We’ve got to get better on power – it’s a limiting factor for the AI economy until then”. Expect more unconventional tie-ups between tech giants and energy firms in the coming months, as each seeks to hedge the other’s critical inputs.

2.3 Cooling the Beast

Density keeps spiking: Nvidia’s latest reference design for AI compute racks runs at 132 kW per rack, far beyond the limits of traditional air cooling. At these power densities (on the order of 30 kW per server chassis), chilled liquid is no longer optional – it’s essential. NVIDIA and Schneider Electric, for instance, have co-developed liquid-cooled “AI-ready” data center blueprints that can handle these loads, promising ~20% cuts in cooling energy usage and 30% faster build-outs thanks to prefab designs. Immersion cooling – submerging servers in specialized fluids – pushes the envelope even further. Studies suggest this approach can yield up to 50% energy savings on cooling and allow 2–3× higher rack density compared to air-cooled setups, though maintenance remains a challenge (e.g. handling fluid evaporation and component servicing). Major operators like Digital Realty are already retrofitting existing facilities with liquid loop systems and exploring immersion for their highest-density clusters. These companies are also hunting for renewables and waste-heat reuse opportunities to meet corporate carbon targets. In Scandinavia and Canada, a few data centers now pipe excess heat to nearby towns or greenhouses – a symbiotic model that groups like the IEA say could meet 5–10% of heating needs in colder regions by 2030. While exotic cooling tech alone can’t solve AI’s energy hunger, it certainly helps: every watt not spent on chilling is one that can go into computation (or reduce emissions). And as high-performance AI silicon (like Nvidia’s upcoming GB200 “Blackwell” GPUs) pushes towards 140 kW+ per rack, the industry’s willingness to embrace novel cooling solutions is no longer in doubt. The real question is how quickly these innovations can be deployed at scale – and whether gains in efficiency can keep pace with the ballooning demand.

3. The “New Cold Chain”: From Silicon to Electrons

Layer	Key choke-points	2025 signals (examples)
Fabs & advanced packaging	TSMC capacity; U.S. CHIPS subsidies	UAE allowed to import 500k of Nvidia’s top AI GPUs annually under new U.S. licence deal.
GPU supply	Export controls; multi-vendor hedging	Nvidia H20 re-admitted to China; Oracle to buy $40 bn of Nvidia GB200 chips for OpenAI’s Texas center.
Cloud & colocation	Multi-cloud neutrality vs sovereignty	OpenAI–Google deal; CoreWeave $6 bn expansion; SoftBank’s 5 GW Stargate campus in Texas 2027.
Thermal management	Liquid and immersion cooling	132 kW/rack designs; immersion cooling cuts ~50% energy, boosts density; waste-heat reuse pilots in EU.

Why “cold chain”? In logistics, a cold chain is the end-to-end system that keeps perishables (like vaccines or food) refrigerated from factory to fridge. AI now demands an analogous continuous chain of “cold” infrastructure: spanning semiconductor fabs, export licenses, cloud data-center buildouts, dedicated power plants and cutting-edge cooling tech. A break in any link – be it a GPU quota or a grid outage – can spoil the whole batch, stalling product roadmaps or model deployments. The strategic advantage will go to those who can vertically integrate this AI cold chain and ensure each link is robust.

4. Strategic Playbook

For Enterprises

Map your exposure – Assess your dependence on both compute and electricity markets. Are your AI workloads tied to a single cloud provider or region? Hedge these risks via multi-cloud arrangements and direct power purchase agreements (PPAs) for critical facilities.
Co-design for cooling – Align your AI hardware procurement with data-center cooling upgrades. High-density GPU clusters planned for 2024–25 might require retrofits (liquid cooling, etc.) to avoid stranded capacity. Audit your roadmap: will your facilities handle 30 kW+ per rack if needed?
Scenario-plan export risks – If your ML training pipeline depends on specific high-end chips, model the impact of geopolitical disruptions. Develop contingency plans (alternative suppliers, longer lead times, model optimization for lesser hardware) in case export controls tighten unexpectedly.

For Investors

Blend infra and energy – Treat data centers, power plants, and even transmission lines as a unified investment thesis. The best alpha may lie in the intersections (e.g. financing behind-the-meter generation for cloud operators). Consider utility acquisitions or partnerships to secure long-term energy supply for tech tenants.
Watch policy catalysts – Keep a close eye on emerging incentives or requirements at the state and federal level. New AI-focused legislation or energy infrastructure bills can create arbitrage opportunities in site selection. For example, generous tax breaks for “AI hubs” or fast-track permitting zones could significantly boost ROI for projects in those regions.

For Policymakers

Synchronize energy & tech rules – Streamline the permitting process for both data centers and power infrastructure, and do it in tandem. Bottlenecks in one area threaten progress in the other. A modernized grid that can’t get permits for new lines, or an AI park waiting on a substation, are equally problematic. Consider “one-stop” regulatory frameworks for AI infrastructure.
Tie incentives to efficiency – When offering subsidies or fast-track status, require best-in-class efficiency measures. For instance, mandates for liquid cooling or waste-heat reuse on large AI data centers can help mitigate environmental impact. Public support should come with strings that drive innovation in sustainability – ensuring new facilities are not just bigger and faster, but also greener.

Conclusion

The scramble for AI supremacy is no longer just about smarter algorithms; it is equally a race for chilled megawatts delivered at the right geopolitical coordinates. Chips and power are becoming two sides of the same coin. Companies that treat compute, kilowatts and cooling as an integrated cold chain will out-innovate those still operating in silos. We are witnessing the rise of a new strategic discipline – call it compute diplomacy or energy-aware AI development. The next 12 months will reveal whether public- and private-sector leaders can build this chain fast enough to meet exploding demand, or whether energy will become the new silicon shortage. One thing is clear: in the age of AI, the competitive edge will belong to those who can ensure the flops keep flowing and the servers stay cool.

FAQ

Q1: Why are AI data centers consuming so much power?
A1: Training and running advanced AI models (especially large language models and generative AI) requires tens of thousands of high-performance GPUs or TPUs running 24/7. This makes AI data centers far more power-intensive than traditional corporate data centers. For perspective, global data-center energy use is expected to double by 2030 largely due to AI. Unlike typical cloud services, AI workloads can’t easily “throttle down” – a big model either runs at full tilt or not at all. Additionally, cooling all that equipment takes significant power. In sum, AI’s appetite for computation directly translates into huge electricity draw, which is why companies like Google, Microsoft, and OpenAI are investing in new power plants and grid upgrades alongside their AI projects.

Q2: Can’t we just use renewable energy for all this new AI demand?
A2: Renewables are certainly part of the solution – tech firms are among the world’s largest buyers of wind and solar. Google, for example, is investing $3 billion to upgrade hydroelectric plants to power its AI data centers. However, the scale and 24/7 reliability needed for AI workloads pose challenges. Peak AI demand may not align with when the sun shines or wind blows. That’s why there’s a push for “firm” power (e.g. natural gas, nuclear) that can provide uninterrupted supply. We’re also seeing hybrid approaches: some AI campuses pair on-site solar farms and big battery banks with grid power as backup. In the long run, a mix of renewable generation, energy storage, and maybe new nuclear will likely be required to sustainably meet AI’s power needs. Policy incentives (and corporate climate pledges) are nudging things in this direction, but it’s a transition that will take time.

Q3: What is the Stargate project mentioned in this report?
A3: Stargate is an ambitious initiative led by OpenAI, SoftBank, Oracle, and others to build massive AI supercomputing centers. The idea is to invest in dedicated infrastructure for AI research and services on U.S. soil (and abroad), on a scale much larger than a typical cloud region. The first Stargate data center is under construction in Texas (near Abilene) and is expected to host on the order of 400,000 Nvidia GPUs once fully equipped. The project’s backers have pledged up to $500 billion (presumably over many years) to roll out multiple such sites. Think of Stargate as creating “AI hubs” – hyperscale campuses with their own power plants, ultra-fast networks, and cutting-edge cooling, all optimized for AI. It’s both a response to exploding AI demand and a strategy to keep the U.S. at the forefront of AI compute capacity.

Q4: How do export controls on AI chips factor into this cold chain?
A4: Export controls are like valves in the AI cold chain – governments can open or restrict the flow of advanced chips between countries. For instance, the U.S. had banned sales of Nvidia’s top GPUs to China, but recently agreed to allow the slightly scaled-down H20 model under a new deal. These policies directly affect who has access to cutting-edge hardware. In our cold chain analogy, if countries can’t import the “brain” of the operation (advanced AI chips), it doesn’t matter how much power or cooling they have – their AI efforts will be hamstrung. That’s why we saw China investing in domestic GPU alternatives and the UAE negotiating to import 500k Nvidia chips per year. In short, export controls are a geopolitical tool that can strengthen or choke off parts of the AI supply chain. Companies need to monitor these rules closely; in some cases, they’re even designing new chips to comply with export limits (Nvidia’s A800, H800, etc. for China market) so that the chain keeps moving albeit with slightly lower performance.

Q5: What are companies doing with the waste heat from AI systems?
A5: All those GPUs generate immense heat, and traditionally data centers have just released that heat into the atmosphere via cooling towers or chillers. Now, with sustainability in focus and energy costs high, companies are finding creative ways to reuse that heat. Some are using heat exchangers to capture it and pipe hot water to nearby buildings – for heating offices, homes, or even greenhouses. This is already happening in parts of Europe: for example, in Scandinavia, some cloud data centers feed district heating systems. The IEA estimates that by 2030, reusing data-center heat could supply about 10% of Europe’s space heating demand.Another approach is integrating absorption chillers or heat-driven refrigeration, essentially using waste heat to run cooling systems (a virtuous cycle of sorts). While heat reuse won’t apply to every data center (it needs the right location and partners), it’s an increasingly popular idea to improve the overall efficiency and community impact of these AI facilities.