AI Datacenters Executive Overview

Artificial Intelligence (AI) is increasingly viewed as a strategic resource in global power competition. National security experts argue that leadership in AI will confer decisive advantages in defense, intelligence, and economic strengthcnas.org 1afcea.org 2. Indeed, China has explicitly set a goal of becoming the world leader in AI by 2030, backed by state-directed plans and massive investmentsmorganstanley.com 3. Beijing’s New Generation AI Development Plan lays out this ambition, and officials see AI dominance as an “existential” race. Chinese President Xi Jinping has described AI as a key area in which China must achieve self-reliance and superiority for national security. In concrete terms, China is rapidly building AI infrastructure at a national scale: as of mid-2024, China has built or announced over 250 AI-focused data centers across the country under central guidance. (These facilities range from government-supervised supercomputing hubs to cloud clusters powering commercial AI services.) Such a whole-of-nation approach reflects a belief that cutting-edge compute capacity will translate to military and geopolitical clout.

China’s state-led strategy includes constructing dozens of AI supercomputing clusters across all regions of the country (yellow markers on map), guided by initiatives like “Eastern Data, Western Computing.” This map shows the broad geographic distribution of China’s AI data centers in the PRC. The centralized campaign is driven by national policy to secure AI leadership.

The United States, by contrast, has no comparably unified strategy for AI infrastructure, raising concerns that America’s lead in AI could erodecnas.org 4. While U.S. companies (and the Defense Department) are world leaders in AI R&D today, that dominance isn’t guaranteed long-term. China’s focused investments in computing power, talent recruitment, and industrial policy (such as subsidies and targets for AI chips) indicate a willingness to do whatever it takes to win. “Whoever ends up being the first will set the rules of the road for the rest of the world,” one expert noted of the Chinese mindsetafcea.org 5. For U.S. national security, then, maintaining an edge in AI capabilities and infrastructure is seen as vital. This means not only developing advanced AI algorithms, but also ensuring access to the raw computing power and energy needed to train and run those algorithms at scale. The nation that marshals greater compute capacity – effectively harnessing vast arrays of specialized chips in large data centers – will be able to push the frontier of AI (for example, training the most sophisticated defense AI models or crunching intelligence data faster). In short, AI prowess is a new pillar of national power alongside traditional elements like nuclear technology or aerospace, and American policymakers increasingly recognize that “the country with the most robust compute ecosystem will retain key advantages” in setting global norms and prevailing in conflicts.

China’s AI and Power Strategy: A Contrast

China offers a revealing point of comparison in how to support AI growth at a national level. Energy and compute are tightly linked – training advanced AI models demands enormous electricity, as does deploying (inferencing) those models for millions of users. Beijing has proactively planned to meet power and cooling needs for its AI ambitions, treating digital infrastructure as strategic infrastructure. For example, China launched the “Eastern Data, Western Compute” initiative in 2020 to distribute data center growth in a way that optimizes energy usagechinatalk.media 6. Under this plan, eight giant computing hubs in China’s energy-rich western regions (such as Inner Mongolia, Ningxia, and Gansu) will handle data and AI processing tasks, tapping abundant renewable energy and land in those areas. Meanwhile, the more populated eastern provinces focus on latency-sensitive services but can offload power-hungry batch processing to the west. This is essentially a national load-balancing strategy to avoid straining the electric grid in big cities and to utilize remote renewable power sources. The Chinese government touts it as a solution to the energy bottleneck for AI.

Crucially, China is expanding its electric grid capacity at record scale to fuel AI and tech growth. In 2025, China expects to add 500 gigawatts of new renewable generation capacity – an astonishing figure (roughly equivalent to the entire power capacity of Japan added in one year) – explicitly because AI-driven demand is surging for electricity to run computing centersscmp.com 7. Over a quarter of that new capacity (140 GW) is wind power, with similarly aggressive growth in solar. This reflects how seriously China takes the power requirements of AI: it is building out generation (especially wind/solar, but also some coal and potentially new nuclear) so that energy availability does not constrain AI developmentchinatalk.media 8. By government estimates, Chinese data centers consumed about 130 billion kWh in 2022, and will triple to 380 billion kWh by 2030 as AI workloads expand. To manage the environmental and reliability implications, Beijing set a policy goal that 80% of power for new data centers should be from “green” (non-fossil) sources by 2025. This is an extremely ambitious target given that currently around 70% of Chinese data center electricity comes from coal. It has spurred efforts to directly connect renewables to data center campuses and experiment with large-scale battery storage to buffer intermittency. (China is even exploring advanced cooling and nuclear reactor designs to support data centers in energy-scarce inland regions, though officially nuclear is not highlighted in its data center plans.)

The bottom line is that China’s leadership recognizes AI as a national priority and is backing it with coherent policy: strategic plans, infrastructure investments, and energy expansion. It treats power and water for AI facilities as issues of national capacity – for instance, holding high-profile conferences on how to overcome electricity bottlenecks for the “national computing network”. This stands in stark contrast to the U.S., where no equivalent national plan exists to ensure adequate power and sites for AI compute growth. Instead, as we discuss next, the expansion of AI data centers in the U.S. is largely driven by private companies and subject to fragmented local decision-making. This raises the question: Is America inadvertently ceding an advantage by not strategically coordinating its digital infrastructure for AI?

(From a scientific-method standpoint, it’s worth asking if there is evidence against the idea that U.S. AI development is unfocused. One might point out that U.S. private sector investments in AI are world-leading, and that our open market system has generated breakthroughs like advanced AI chips and foundational models without central planning. However, the risks remain that critical national-security AI needs could be underserved – for example, if commercial incentives don’t align with defense requirements. The following sections examine how U.S. AI infrastructure is being built, revealing misalignments.)

Balkanized Local Decisions for Strategic Infrastructure

Despite AI’s national importance, the United States currently leaves the development of AI compute infrastructure to ad-hoc local processes. There is no overarching national strategy guiding where to build data centers or how to prioritize their use – instead, decisions are effectively “balkanized” across hundreds of counties and towns. In practice, when a tech company wants to build a massive data center (which could draw tens or hundreds of megawatts of power), it typically negotiates with a city or county government in secret, then goes through that locality’s standard planning and zoning approval. These processes were never designed to weigh national strategic needs; they focus on local land use, noise, traffic, tax revenue, etc. As a result, strategic digital infrastructure (large AI compute facilities) is being green-lit (or blocked) via siloed local decisions, often with minimal broader oversight.

A quiet stretch of farmland in Kuna, Idaho, with a small public hearing notice, is slated to become a $1 billion data center campus. This Gemstone Technology Park project was advanced under a rezoning vote by the city council in early 2025nucamp.co 9. Such local proceedings (sometimes only signposted by a roadside notice) illustrate how major AI infrastructure can emerge via routine municipal approvals, often with little public scrutiny until late in the process.

Several trends underscore this balkanization. First, major tech firms frequently cloak their data center projects behind non-disclosure agreements (NDAs) and code names, limiting what local officials and residents know. Public officials often sign NDAs at the behest of the company early in negotiations, meaning they cannot reveal the company’s identity or details like power and water usage until a deal is essentially donegoodjobsfirst.org 10virginiamercury.com 11. A recent investigation in Virginia – a state that has become America’s data center epicenter – found NDAs in 25 out of 31 localities with existing or proposed data centers. These NDAs were sweeping, barring disclosure of almost any “non-public” information, well beyond protecting genuine trade secrets. They explicitly allowed the companies to preemptively fight (or narrow) any Freedom of Information Act requests for project details. The effect is that key information about community impact is withheld from the public: for example, one Virginia county’s agreement with Amazon even tried to classify the company’s water and sewage usage for a new data center campus as confidential. Water use can be millions of gallons a day for large cooling systems, yet locals might not learn of that impact due to such secrecy.

Secondly, companies routinely hide behind shell LLCs and code names during permitting. A community might see a proposal from “XYZ Development LLC” for a “Project Emerald,” not knowing whether the end user is Amazon, Meta, Microsoft, or another entitygoodjobsfirst.org 12. This practice, combined with NDAs, means citizens and even council members often do not know who they are really dealing with. In one case in Fredericksburg, VA, the city council created a new zoning district for a data center campus without disclosing the company name until the night of the vote, and even the total number of data center buildings (8–12 planned) was revealed only after the rezoning was approvedvirginiamercury.com 13. Meaningful public debate was stifled – an example of democracy “dying in the dark,” as critics put it.

Finally, local zoning and utility boards are ill-equipped to consider national security implications. Their mandate is typically to assess local effects (noise, traffic, tax base, environmental compliance). If a given data center will consume, say, 100 MW of power, a county may worry about grid strain for local residents – but they cannot judge whether that 100 MW might be better allocated to a different region or a more critical AI project. Yet, cumulatively, these local decisions determine where the nation’s largest computing clusters are built and who gets first access to finite electrical capacity. For instance, multiple rural towns have approved huge data center farms that will instantly become the region’s biggest power users, often negotiating special electric rates for those facilities. In Virginia, regulators created a special tariff class for data centers because their energy draw is so massive and growing. But those interventions happened only after data centers had already proliferated, and mainly to protect other ratepayers – there was no upfront strategic plan.

In short, the U.S. approach to siting AI infrastructure is fragmented and driven by private incentives. Tech companies choose locations based on a mix of cheap land, tax breaks, and available power, then deal with each locality separately. Some jurisdictions eagerly compete to offer incentives (viewing data centers as economic wins), while others have started pushing back with moratoriums or stricter rules due to concerns about noise, water, or landscape impactsgovtech.com 14. The outcome is a patchwork: where an AI supercluster arises might depend on which town council is more amenable or which state waives more taxes, rather than any national prioritization.

It’s worth noting that local officials and companies defend this status quo as normal economic development. They argue NDAs and confidential negotiations are routine for attracting investment and that ultimately these projects bring in revenue with relatively little strain on services (since data centers create few jobs)virginiamercury.com 15. And indeed, secrecy in business deals is not illegal. However, the concern is that what’s expedient for a single county (landing a tech investment quietly) may not be optimal for the nation’s strategic posture. The current model potentially “short-circuits” broader debate about how much of our electric grid should be devoted to various AI uses. We are effectively letting market forces and local politics decide something that has national implications – akin to if each city decided on its own whether to host a piece of the interstate highway system or a military base, without federal input. This balkanization makes it hard to, for example, channel capacity toward truly critical AI projects (Category A-1, defined later) or to ensure resilience (since many data centers cluster in a few regions like Northern Virginia, creating single-points-of-failure risk). Strategic voices are largely absent at these zoning meetings. As one observer noted, decisions about America’s digital infrastructure are “hinging on local land-use processes that were never designed to weigh national strategic needs.”

Idaho Case: An Illustrative “Category B” Build-Out

One stark example of these dynamics is Idaho’s recent data center boom, which has been driven primarily by Category B (commercial, consumer-driven) workloads. In 2022, Meta (Facebook’s parent company) announced it would build an $800 million data center in Kuna, Idaho – a small city outside Boise – marking Idaho’s first “hyperscale” data center for advanced AI and cloud servicesnucamp.co 16. Meta’s facility, now under construction on 485 acres, will span nearly 1 million square feet and is explicitly designed for AI workloads (Meta redesigned its data centers globally to support AI training and recommendation models)datacenterdynamics.com 17. It will consume an estimated 70,000 gallons of water per day for cooling (after implementing an 80% water-use reduction with advanced cooling) and run on 100% renewable energy via new solar projects. Yet, it provides only ~100 long-term jobs. Right next door, a developer (Diode Ventures) proposed the Gemstone Technology Park, a $1 billion, 620-acre data center campus likely to host additional large tech companies. Gemstone was shrouded in mystery (the end user not publicly named, often the case under NDAs) and navigated local approvals in 2025. The project’s anticipated electrical load is 600–800 MW – an enormous draw equivalent to roughly half of Idaho Power’s entire peak demand. This single private campus could require new transmission lines and substations, raising questions about grid impact and water usage for cooling (which sparked local debate).

The Idaho state and local authorities, for their part, have offered generous incentives to lure these data centers (such as sales tax exemptions on servers and property tax breaks)gemstatepatriot.org 18. In hearings, local leaders often framed the projects as economic wins bringing investment and some ancillary infrastructure upgrades (Meta, for example, is funding a $70 million water and sewer system for Kuna as part of its project)nucamp.co 19. Indeed, millions of dollars were pledged to local schools, police, and fire services by Gemstone’s developer to sweeten the deal. The strategic context, however, is that both the Meta and Gemstone facilities are primarily serving commercial “Category B” uses – e.g. Meta’s center will power social media AI systems (like content recommendation algorithms and advertising optimization) rather than defense or scientific research. Idaho’s power grid, which relies heavily on hydropower, will now be partially allocated to these private data farms. By 2026–2030, if both projects fully build out, hundreds of megawatts of Idaho’s electricity will be locked into supporting advertising-driven AI and cloud computing. This illustrates a microcosm of the national trend: growth in AI compute is largely coming from consumer internet giants, and they are securing the necessary infrastructure through local channels that focus on local benefits. (For details on Idaho’s data center incentives, water use, and power planning, see Deep Dive: Idaho Data Centers.)

From a national perspective, Idaho’s case is telling because it’s not an isolated incident: similar stories are playing out in numerous states – from Northern Virginia (where entire counties are being blanketed with data centers for Big Tech cloud and social media companies)virginiamercury.com 20, to Ohio and Iowa (large Facebook/Meta and Google server farms), Texas (where companies like Microsoft and Google have projects, often under codenames like “Project Redbird”), and Alabama (recent proposals for huge data centers in mostly rural areas)govtech.com 21. In each case, the pattern is B-heavy: the computing power being installed serves mainly private-sector web services, advertising, e-commerce, entertainment streaming, and the like (Category B), not national defense or critical public missions. And the decision-makers are diffuse – a town council here, a county board there – each making a narrow call that cumulatively shapes our national computing landscape without strategic coordination.

Classifying AI Workloads: A‑1 vs A‑2 vs B

Not all AI uses are equal in importance or urgency. To inform policy, it’s helpful to classify AI workloads into categories based on their national strategic value. Here we propose three broad categories, explained with examples:

• Category A‑1: National Security & Critical Public Missions. Definition: AI applications that directly serve defense, intelligence, homeland security, or other vital government functions where failure or unavailability would significantly harm national security or public safety. These are the highest-priority uses – they arguably merit priority access to computing power and energy resources during shortages or crises. Examples: AI systems for military use (e.g. Project Maven in the Pentagon, which uses machine learning to analyze drone surveillance footage for threatsapnews.com 22), intelligence agencies’ AI for satellite image analysis or signal intercepts, cybersecurity AI that protects critical infrastructure from cyberattacks, nuclear launch decision-support algorithms, or AI models for pandemic response and biosecurity. In civilian sphere, one could include weather and climate prediction AI that informs disaster response (since it directly protects lives), or AI for managing the power grid to prevent blackouts. (Tags: “Defense AI,” “Intelligence AI,” “Critical Infrastructure AI” – these would all fall under A‑1.) A current example is the Department of Defense’s JAIC (Joint AI Center) initiatives to accelerate AI in warfighting – those systems, once deployed, would be A‑1. Another is NASA and NOAA’s use of AI for satellite data processing to improve hurricane forecasts (arguably critical public safety use). A‑1 workloads tend to require significant compute, but their scale is bounded by the specific mission (e.g., a few strategic AI models or supercomputers for nuclear simulation).
• Category A‑2: Important Civilian, Scientific, and Economic AI (Public Good/High Value). Definition: AI uses that are highly beneficial to society or the economy, but are not directly military or intelligence related. They may not be wartime necessities, but they contribute to national competitiveness, health, and infrastructure. These should be strongly supported, though not necessarily given the same override priority as A‑1 in a crunch. Examples: AI in healthcare (e.g. medical diagnostics AI that can detect diseases from X-rays or MRIs with expert accuracypmc.ncbi.nlm.nih.gov 23), AI for pharmaceutical research (drug discovery, vaccine development), agricultural AI for food security (e.g. optimizing crop yields with computer vision), manufacturing AI that boosts productivity, transportation AI (like systems managing traffic flow or enabling autonomous public transit), and scientific research AI (such as algorithms analyzing particle physics data or astronomy surveys). Also included are enterprise AI applications that make industries more efficient (like supply chain optimization, energy management in smart gridssciencedirect.com 24, or financial fraud detection). A real-world example is Idaho Power’s AI-driven grid modernization project to enhance efficiency and reliability in electricity distributionnucamp.co 25 – this is not military, but it’s critical for energy reliability and thus A‑2. Another example: climate modeling with AI to better project long-term climate risks (vital for policy and infrastructure planning). Or consider education AI tools that personalize learning – beneficial, though a notch below national security in priority. In summary, A‑2 covers public-good AI and mission-critical economic functions. These workloads can also be compute-intensive (training a medical image model on millions of scans, for instance, or running a large language model to aid education), but they are ones where public support or at least awareness is warranted because of their broad benefits.
• Category B: Commercial & Consumer Internet AI (Ads, Social Media, Entertainment, etc.). Definition: AI applications primarily geared toward profit-making services, such as targeted advertising, content recommendation, user personalization, and other consumer-facing products. These are often the largest drivers of AI compute demand today, because big tech companies apply AI at massive scale to billions of users. However, by our hypothesis, they have the lowest direct strategic importance; they are nice-to-have, not need-to-have for the nation. Examples: The algorithms that curate your social media news feed or TikTok video stream (e.g. TikTok’s For You recommendation AI is emblematic – incredibly compute-intensive, but its purpose is to maximize user engagement/ad revenue). Likewise, Facebook/Meta’s advertising ranking algorithms and content moderation AI, Google’s search ads AI and YouTube recommendation engine, Amazon’s product recommendation and e-commerce optimization AI, Netflix’s show recommendation AI, and Snapchat’s filters – all of these fall in Category B. Also included are newer consumer AI services like chatbots for entertainment or coding assistants if they are largely commercial (for instance, OpenAI’s ChatGPT being used to enhance Bing search – mainly a commercial competitor to Google). In enterprise contexts, AI for marketing analytics or personalized shopping would be B. These Category B systems undoubtedly provide economic value and convenience, but if they vanished, the direct national security impact would be negligible (aside from economic ripple effects). Tags: “Ad-Tech AI,” “Social Media AI,” “Consumer Service AI.” A concrete example: Meta’s AI Research SuperCluster with 16,000 GPUs, used to train large models for content understanding – its output improves how Facebook targets ads and shows postsreddit.com 26. The vast majority of AI compute growth in recent years has come from such Category B efforts, as tech giants race to deploy ever-larger models to gain a competitive edge in advertising efficiency or user retentionnucamp.co 27.

Why these categories matter: If the premise is that compute (and the power to run it) is a constrained resource, then we need a way to decide which AI activities get priority when push comes to shove. Categories A‑1 and A‑2 could be candidates for public support or preferential treatment (e.g., expedited grid hookups, subsidies, or dedicated infrastructure), because they align with national interests (security, health, economic resiliency). Category B, in contrast, might be left to the market – or even deliberately de-prioritized if it is hogging resources needed elsewhere.

It’s important to stress that these categories aren’t moral judgments on worthiness; they’re about strategic prioritization. Category B tasks (like serving personalized ads) are not “bad” – they often fund innovation and provide free services to consumers. However, from a national policy view, if there’s a finite amount of high-end chips or power available, a Category A‑1 military AI to deter a hostile missile attack should outrank a Category B app that delivers prettier social media filters. Our hypothesis posits that currently, the vast majority of AI compute is going to Category B, which might warrant rebalancing. We will examine that quantitatively next.

How Much of AI’s Power Goes to A‑1 vs A‑2 vs B? (2028–2030 Outlook)

To test the hypothesis, we look at current trends and projections of AI-related computing and energy use. We then estimate what share is attributable to each category in the near future (late 2020s). This exercise is admittedly challenging – companies do not publicly break out “what fraction of our GPU clusters are training ad models vs. medical models.” However, we can use proxies and industry data to sketch scenarios. All projections here refer to the United States (where data is available, or else using global numbers as a rough guide), and specifically to electricity consumption (MWh) by AI workloads in each category by ~2028–2030. We present a likely scenario, as well as low- and high-bound cases, with our reasoning and sources noted.

Baseline: Data centers as a whole are a major and growing electricity user. In 2024, data centers (worldwide) consumed about 415 TWh of electricity, ~1.5% of global consumptionscientificamerican.com 28. By 2030, that is expected to more than double to 945 TWh. The International Energy Agency (IEA) attributes this largely to the rise of AI computing. In 2024, the IEA estimated that servers handling AI workloads accounted for ~24% of server power draw and ~15% of total data center energy use. In other words, AI was already ~0.22% of global electricity in 2024 (15% of 1.5%). Many experts believe this is an underestimate of AI’s share, given the difficulty of companies separating “AI” servers from others. Regardless, all signs point to AI’s share rising significantly by 2030. If data center energy doubles and much of that growth is AI, then AI could reach on the order of 30–50% of data center energy by 2030. To focus on U.S. share: the U.S., China, and Europe together are 85% of data center energy use, with the U.S. likely the single largest country-share. U.S. data centers (all purposes) might consume on the order of 150–200 TWh by 2030 (very roughly), of which AI might be tens of TWh.

Now, how is that AI power divided among A‑1, A‑2, B?

• Category B (Commercial/Consumer) is, by all indications, the dominant driver of AI compute growth. The world’s largest AI training runs and deployments to date have been done by companies like Google, Meta, OpenAI/Microsoft, Amazon – primarily to enhance products like search engines, ad platforms, social media feeds, and cloud AI services for profit. Meta, for instance, is reportedly spending on the order of $10–$15 billion a year on AI and data center development for its metaverse and ad platforms, building mega-compute clustersthekeyword.co 29emarketer.com 30. Google’s AI effort similarly feeds its advertising and cloud businesses, and Amazon’s AI investments drive both its retail and AWS services (including its own ad network). By proxy, lobbying data reflects these firms’ presence: in 2024, Meta (Facebook) spent $24.4 million lobbying and Alphabet (Google) $12.7 million, mostly defending their internet business interestsopensecrets.org 31. These companies operate tens of massive data centers each, and while not all their workload is AI at any given moment, the trend is that every part of their consumer platforms is increasingly augmented by AI models (from content ranking to language translation). Furthermore, AI inference at scale – serving billions of recommendations or queries daily – is extremely power-hungry in aggregate. For example, every time you use a voice assistant or get a personalized feed, hundreds of machine learning model inferences happen in the cloud. In sum, Category B likely constitutes the bulk of the 15% (and rising) AI share of data center energy reported by IEAscientificamerican.com 32.
• Category A‑1 (Defense/Security), in contrast, is much smaller in scale. The U.S. Department of Defense and Intelligence Community certainly have high-performance computing (HPC) centers and some AI programs, but their compute footprint is modest next to the commercial sector. For perspective, the entire DOD AI budget for 2024 was about $1.8 billionthinkinsider.org 33 – which includes research, not just infrastructure – whereas a single tech giant like Meta or Google spends several times that on AI capex annually. The number of data centers dedicated to classified or military AI is limited (e.g., some at national labs or within agencies). Even major government supercomputers (like DOE’s Frontier or NSA’s clusters) consume on the order of 10–30 MW each, which is a drop in the bucket compared to the thousands of MW being built by the likes of Amazon and Microsoft for cloud data centers. One telling data point: Microsoft’s Azure cloud (which serves many purposes including some government clients) spent ~$9.5 million on lobbying in 2024, while Oracle (another enterprise/cloud provider with gov business) spent ~$11.8 millioninvestopedia.com 34. But these figures are still half of what Meta spent, indicating the center of gravity lies with consumer techopensecrets.org 35. Additionally, a lot of government AI work leverages commercial cloud providers (the Pentagon’s recent $9 billion JWCC cloud contract is split among AWS, Microsoft, Google, Oracle), meaning it rides on infrastructure that is mostly built and operated by those companies. So even if, say, AWS hosts an AI for the CIA, that’s on the same servers that might next minute host an ad-tech training job. It’s hard to precisely extract, but by 2030 we anticipate Category A‑1 will still be only a single-digit percentage of total AI compute demand – perhaps on the order of 5% (likely scenario), with a low-end scenario maybe as little as ~3% (if commercial growth far outpaces government adoption), and a high-end scenario perhaps 10% if government massively invests in AI infrastructure. These numbers align with the notion that government HPC and AI projects, while significant in absolute terms, are vastly outnumbered by commercial ones.
• Category A‑2 (Other high-value AI) sits in between. This category includes many enterprise and scientific uses which, aggregated, could be substantial – but they are diffuse across industries. For instance, every hospital might deploy some AI, every utility might use AI for grid management, etc., contributing to energy use. However, their models are often smaller scale or used less intensively than the always-on consumer internet AIs. One notable chunk in A‑2 is academic and scientific computing. The U.S. has large university and national lab computing facilities (e.g., for climate modeling or genomics). These collectively consume significant power (national labs HPC consume a few hundred MW combined). Another chunk is enterprise data centers running AI for businesses – though many businesses now use cloud providers rather than on-premise data centers, effectively outsourcing that to Category B companies (who then count it in their footprint). An illustrative example: banks using AI for fraud detection – they might run those on AWS or on their own servers. Either way, the energy is counted somewhere. Given that cloud providers report enterprise AI uptake as growing, we include that in A‑2. By 2030, Category A‑2 might account for perhaps 15–30% of AI compute demand. Our likely estimate is around 20%. Low scenario could be ~10% if consumer tech really dominates, and high could be ~30% if industries like healthcare and manufacturing heavily adopt AI (and consumer growth is a bit slower).

Combining these rough estimates for 2030 (U.S. context):

• Likely Scenario: A‑1 ~5%, A‑2 ~20%, B ~75% of AI-related MWh consumption. In this scenario, roughly three-quarters of AI’s strain on the grid comes from Category B – things like serving ads, entertainment, social media personalization – while only one-quarter is split between critical public uses and broader beneficial usesscientificamerican.com 36nucamp.co 37. This aligns with the hypothesis that most of the growth in AI compute is not for national security or essential services.
• High B Scenario (More Extreme): A‑1 ~3%, A‑2 ~12%, B ~85%. This might happen if the commercial race (especially generative AI deployment in consumer apps and advertising optimization) accelerates even more than expected, and if public sector initiatives stagnate. Signs of this scenario would be, say, every social media and e-commerce platform deploying GPT-5 level models that run constantly – ballooning energy usage – while government AI remains limited. It’s a concern for national security advocates, because it implies over eight-tenths of our advanced computing capacity is essentially tied up in chasing clicks and ad dollars.
• Low B Scenario (Optimistic Rebalance): A‑1 ~10%, A‑2 ~30%, B ~60%. This could occur if conscious policy and market trends shift resources. For instance, if federal programs fund a network of AI supercomputers for research and defense (boosting A‑1 and A‑2), and if consumer AI growth is tempered by efficiency gains or regulations (reducing B’s share). It’s still likely B remains a majority (given the profit motive, tech companies will keep expanding AI use), but perhaps closer to ~60%. One could imagine this if, say, by 2030 the U.S. has several “national AI research centers” running models for science and medicine (A‑2), and DOD has fielded AI into many systems (A‑1), while consumer apps hit saturation.

To be clear, these category boundaries blur at times. Some AI uses benefit both commercial and public sectors (e.g., an AI cloud service might serve a mix of clients). But the key takeaway from even the rough numbers is: Category B presently dominates and is expected to continue dominating absent intervention. Multiple independent indicators corroborate this: corporate capital expenditures, energy studies, server shipment trends (hyperscalers account for the majority of server purchases, and hyperscalers largely monetize via consumer services), and lobbying patterns. As one data point, in 2024 four tech giants (Meta, Amazon, Alphabet, ByteDance) – all Category B-centric companies – together spent over $66 million lobbying federal officialsopensecrets.org 38, far outspending any public research or defense AI lobbying presence. This suggests where the influence and growth are concentrated.

It’s worth acknowledging a counter-argument: Some might say even consumer AI has indirect national value (it keeps U.S. companies at the cutting edge, funds AI talent development, etc.). There is truth to that synergy – the U.S. leads in AI partly due to its vibrant tech sector. However, the crux is about marginal allocation: if power or chips become limiting factors, should the next 100 MW go to an extra ad-personalization cluster or to an AI that improves electric grid resilience or analyzes military drone feeds?* Our analysis indicates the status quo will send that 100 MW to the former (Category B) by default**, because that’s where market demand and capital are strongest, unless policy steps incnas.org 39.

Mapping the Major Players: Which Companies Do A vs B (and Their Influence)

To further illustrate the landscape, we map prominent companies to our categories and examine their focus and influence. The AI compute ecosystem is dominated by a handful of U.S. tech companies, each with different mixes of A‑1, A‑2, and B in their portfolios. We also take note of lobbying expenditures as a measure of their political influence, which can shape AI and energy policy in their favor.

• Microsoft (Azure, etc.): Category mix: Mixed A‑2 and B. Microsoft is unique in spanning consumer and enterprise. On one side, Microsoft has Bing search and ads, LinkedIn, and Windows consumer services – these use AI (e.g., Bing’s AI-powered search answers, LinkedIn’s feed algo) which are Category B. On the other side, Microsoft’s Azure cloud serves a broad range of customers, including many enterprise and government (A‑2 and some A‑1). Microsoft also partners with OpenAI to offer AI services; this straddles categories (ChatGPT can be used for business or fun). Microsoft has significant government contracts – e.g. it won part of the Pentagon’s JWCC cloud contract and is working on defense-oriented AI with projects like Azure Government regions (that’s A‑1 usage of its infrastructure). However, the bulk of Microsoft’s revenue still comes from enterprise software and commercial cloud, which are more A‑2, plus its slice of advertising (through Bing) which is B. Estimate: Microsoft’s AI compute might be split roughly 20–30% A‑1 (government/DoD via Azure), 30–40% A‑2 (enterprise, productivity AI, etc.), 30–50% B (consumer-facing like Bing, Office 365 AI features used by consumers, etc.). Lobbying: Microsoft spent about $9.5–$10.4 million in 2024 on lobbyinginvestopedia.com 40issueone.org 41. Its lobbying often focuses on cloud computing regulations, antitrust, and government procurement rules – relevant because Microsoft stands to benefit if government treats Big Tech as partners in national AI (as indeed recent policy suggests, with MS among those awarded contracts for military AI developmentbuiltin.com 42).
• Amazon (AWS and Amazon.com): Category mix: Mixed A‑1, A‑2, B. Amazon Web Services (AWS) is the largest cloud provider, serving countless businesses (A‑2) and some government agencies (A‑1, e.g. AWS hosts parts of the U.S. intelligence community cloud). Meanwhile, Amazon’s retail business uses AI heavily for recommendations, logistics, and its own advertising network – those are B (improving e-commerce conversion and ad sales). Amazon also has consumer devices (Alexa voice assistant is AI-based – Category B). Overall, AWS’s growth means Amazon enables a lot of A‑2 AI (enterprise deployments on AWS SageMaker, etc.), but Amazon’s profit engine includes a big dose of B (ads in search results on Amazon, etc., now a multi-billion-dollar business). Estimate: Perhaps 10% A‑1 (since AWS handles some classified cloud and defense contracts like hosting DoD’s JWCC workloads), 30% A‑2 (AWS for banks, pharma, etc., plus things like Whole Foods inventory AI – not crucial for natsec but economically important), 60% B (retail optimizations, Alexa, and AWS capacity serving ad-tech companies, etc.). Lobbying: Amazon (which includes AWS and retail under one umbrella) spent about $19.1 million lobbying in 2024opensecrets.org 43. Amazon’s lobbying has touched on data center energy (they’ve sought renewable energy deals and favorable rates), privacy, and defending its cloud contracting. Notably, Amazon’s energy deals include buying land next to power plants – e.g., it bought a site adjacent to a nuclear plant in Virginia to ensure a dedicated green power supply for AWSchinatalk.media 44 – indicating how far a Category B/C cloud company will go to secure power.
• Google (Alphabet): We separate Google Cloud vs Google’s own Ads/Services for clarity.

·         Google Cloud: Categories A‑2 (mostly) and some A‑1. Google Cloud Platform (GCP) is similar to Azure/AWS in serving enterprise customers; it has fewer government wins but is courting them (it too is in JWCC for DoD cloud). GCP’s clients use its AI tools for everything from fintech to medical research (A‑2 uses). So, Google Cloud’s compute supports a mix of A‑2 workloads across industries, and a bit of A‑1 if they land intel or defense clients.

• Google Advertising & Consumer Services: Category B dominant. This is core Alphabet: Google Search (world’s biggest AI-powered search engine), YouTube (recommendation algorithms), Google Ads (AI-driven ad targeting), Google Maps suggestions, Android AI features, etc. Alphabet’s revenue is ~80% from advertising, so virtually all those AI models optimizing ad placement and user engagement fall under B. Even Google’s much-publicized AI like DeepMind’s AlphaFold (protein folding), while scientifically valuable (A‑2 output), was funded by the profits of its ad business. So Google as a whole skews heavily B in terms of where its compute is applied (every search query now invokes AI models, every YouTube view triggers recommendation AI). Estimate: Google overall might be 0–5% A‑1, ~15% A‑2, ~80% B. The A‑1 portion is negligible (aside from contracts like some AI work with NOAA or DOD which are small). A‑2 includes things like Google’s contributions to open science (AlphaFold’s compute was significant) and cloud usage for enterprise. B is the rest (ads, etc.). Lobbying: Alphabet (Google’s parent) spent about $12.7 million on lobbying in 2024opensecrets.org 45. Issues include antitrust, AI regulation (ensuring they can self-regulate their AI), and telecom. Given Google’s dependence on power-hungry data centers, it also lobbies on energy and environmental rules (it has pushed for more renewable energy procurement).
• Meta (Facebook): Category mix: Nearly all Category B. Meta’s entire business is consumer social media and the advertising around it. Its use of AI is pervasive – from the algorithms that rank your Facebook and Instagram feeds, to AI that auto-moderates content, to the recommendation engine for Instagram Reels or Facebook Watch videos, and now generative AI chatbots for user engagement. Meta has poured capital into AI infrastructure primarily to improve these products (for example, redesigning its data centers for AI and building the AI Research SuperCluster to train large language models – these models mainly help with content understanding and future AR/VR features)alabamasolutions.com 46reddit.com 47. Meta does not have a cloud service for external customers, nor significant enterprise software – it’s focused on its own platforms. It also has virtually no direct role in national security or critical sectors (aside from being a communications platform). Thus, Category A‑1 for Meta is ~0%, A‑2 maybe ~0–5% (one could argue any altruistic research they open-source, like AI frameworks, are a public good – but that’s minor), and Category B ~95–100%. Lobbying: Meta topped tech lobbying with $24.43 million in 2024opensecrets.org 48. This high spending reflects how policies on data privacy, content regulation, or AI accountability could impact its core business. Meta’s lobbying in the context of our discussion signals that the leading exponent of Category B AI (social media ads) wields significant influence in Washington. It raises a concern: will policy solutions that might restrain power use for non-critical AI face pushback from these well-funded lobbying efforts?
• TikTok (ByteDance): Category mix: Category B entirely. TikTok is a consumer video platform known for its extremely addictive AI-driven recommendation algorithm. In the U.S., TikTok’s infrastructure is a bit unique since data is hosted by Oracle (due to security agreements), but ByteDance (the parent) develops and runs the AI. TikTok’s AI workload is huge – serving content to 100+ million American users – and it’s all for entertainment/ads. There is no enterprise or gov component. A‑1 = 0%, A‑2 = 0%, B = ~100%. Lobbying: ByteDance has ramped up U.S. lobbying given political scrutiny; it spent $10.36 million in 2024. This is notable for essentially a single-app company. Their lobbying likely doesn’t touch energy directly, but rather privacy and security (to avoid bans). Still, from an energy perspective, TikTok contributes to the B category load while being a foreign-owned entity, which is an interesting twist – effectively U.S. grid power can be said to support a Chinese company’s AI operations delivering dance videos. This further underscores the strategic disconnect.
• Oracle: Category mix: Primarily A‑2 (enterprise) and some A‑1. Oracle is a major enterprise software firm that has expanded into cloud services. Oracle’s cloud (OCI) hosts business workloads and has a niche in certain government workloads (it’s the hosting provider for TikTok US data, for instance, and markets itself for secure cloud needs). Oracle doesn’t have a consumer-facing social platform or ad business. So its AI use is in things like database optimization, cloud AI services for enterprise, and possibly some government contracts (it was one of the winners of the Pentagon’s JWCC contract as well). A‑1: Oracle does handle some sensitive U.S. government data (recently partnered with CIA for a classified cloud region, reportedly), so let’s say maybe 10% of its AI footprint could be A‑1 (if that). A‑2: The majority (~70–90%) of Oracle’s AI is serving businesses (financial, retail, etc.) – definitely important but not direct natsec. B: ~0–10%, basically negligible aside from the fact that Oracle’s cloud might host some ad-tech companies as clients. Lobbying: Oracle is very active in DC for its size; it spent about $11.83 million in 2024 on lobbyinginvestopedia.com 49, often on procurement and tech regulation (Larry Ellison, Oracle’s co-founder, has been vocal about competing with Amazon for government cloud business, for example). Oracle’s high lobbying spend (leading its industry category in 2024) shows it’s pushing to shape policy possibly toward more government use of its services – which would boost A‑1 share if successful.
• IBM and Others (Palantir, etc.): While not named in the prompt, it’s worth mentioning a few others:

·         IBM: Historically a big player in AI (think Watson). IBM now provides AI services to enterprise and government (like IBM built the AI system for the U.S. Dept. of Veterans Affairs to help with health records). IBM’s focus is A‑2 (enterprise AI for healthcare, finance, etc.) and some A‑1 (it contracts with DoD and intelligence on AI projects too). IBM no longer runs massive public cloud data centers at the scale of AWS, but it has significant HPC centers for clients. IBM’s lobbying is smaller (a few million/year typically) and it positions itself as a responsible AI provider for societal good (A‑2 emphasis).

·         Palantir: A company squarely focused on A‑1 and A‑2 – it provides AI and analytics platforms to defense, intelligence (A‑1) and also to industries like finance and healthcare (A‑2). Palantir’s compute usage is relatively small though (they often run on cloud infrastructure of others). But policy-wise, Palantir advocates for using AI to strengthen national security – their voice is one calling for prioritizing A‑1.

·         NVIDIA: Not an AI service operator but the supplier of GPUs – they lobby too (NVIDIA spent ~$4.1M in 2024). Their interest is in broad AI growth (sells to all categories). They are relevant because hardware availability can bottleneck A‑1 vs B allocation; currently, companies like OpenAI (B) are soaking up so many NVIDIA chips that government projects have sometimes struggled to procure high-end GPUs. This is a dynamic to watch (the U.S. government may intervene to secure chip supply for A‑1 projects if needed, akin to how it manages access to defense materiel).

In summary, the corporate map shows a heavy skew to B. The world’s largest AI compute footprints belong to firms whose business is primarily consumer data monetization (Meta, Google, etc.). Cloud providers like Amazon and Microsoft have a more mixed client base but still feed a lot of B (since many of their largest cloud customers are themselves in advertising, entertainment, etc.). The total lobbying muscle of the primarily-B companies is enormous – e.g., Meta + Amazon + Alphabet + ByteDance = ~$66 million in 2024opensecrets.org 50. In comparison, companies mostly in A‑1/A‑2 (like defense contractors or healthcare companies) individually spend less on lobbying regarding AI issues, and their voices on AI infrastructure policy are just starting to be heard. The danger is that without conscious effort, policy will be influenced by those with the biggest budgets – which currently align with keeping the status quo of industry-led, B-heavy AI growth. Indeed, industry associations representing these tech firms often resist regulations that might, say, limit data center locations or impose transparency on NDAsazluminaria.org 51govtech.com 52.

As a point of data, in 2024 the top tech companies combined (including those above and Apple, Microsoft, etc.) spent over $85 million on lobbying – a recordaxios.com 53. Issue-wise, a lot of this went to shaping emerging AI governance (e.g., how AI might be regulated for bias or safety – they want light-touch self-regulation) and ensuring infrastructure support (like pushing for government funding of AI research which ultimately flows through their cloud platforms). The risk of regulatory capture is nontrivial: if those defining “critical AI” are influenced by companies whose profit is in B, we might see advertising algorithms being labeled as critical infrastructure to dodge potential energy use curbs. This is why an objective classification (as attempted above) is important.

Aligning Power and Policy: Mechanisms to Prioritize A‑1 (Sketching Solutions)

Given the evidence that market forces alone will channel the lion’s share of compute and electricity into Category B AI, a core premise of this analysis is that policy intervention is warranted to ensure Category A‑1 (and A‑2) get the necessary resources. What mechanisms could achieve this, without unduly harming innovation? Here we outline a few ideas at an executive policy level. (A detailed implementation plan is beyond our scope, but we sketch analogies to existing tools.)

• Invoking Priority Allocation (Defense Production Act & DPAS Analogy): The U.S. has the Defense Production Act (DPA) which includes the Defense Priorities and Allocations System (DPAS). This system allows the government to prioritize certain contracts for national defense above others – for instance, in wartime, contracts for fighter jet components can be prioritized such that suppliers serve those first. We could establish an analogous mechanism for AI compute and data center resources. For example, the government might designate certain AI projects as “Priority AI Programs” (likely Category A‑1 projects: e.g., an Air Force autonomous drone AI or a DHS border security AI). Those programs could then get priority access to scarce resources: if high-end NVIDIA chips are in short supply, suppliers would fill the order for the priority project before, say, selling to a social media company. Similarly, if the electrical grid connection capacity in a region is constrained, utilities (possibly via DOE/FERC guidance) would connect a priority AI data center (say a national lab expansion) before a new crypto-mining or ad-tech data center. In essence, establish a formal queue-jumping right for critical AI. The DPA authority already exists to some degree – in fact, the Biden administration used DPA powers to prioritize semiconductor production for defense. We could extend that to AI infrastructure as critical technology. Policymakers would need to set criteria for what qualifies (to prevent abuse). This ensures that if there’s competition for, say, a 500 MW chunk of power capacity or a batch of accelerators, the satellite imagery analysis AI for intelligence (A‑1) wins over the e-commerce recommendation engine (B) by default of policycnas.org 54.
• “Metered Carve-Outs” of Power Capacity: This idea means reserving a slice of electric grid capacity specifically for A‑1 and perhaps A‑2 projects. For instance, regional grid operators and utilities could be directed (or incentivized) to set aside X megawatts for critical computing facilities in their long-term plans. If a region is planning 1000 MW of new data center load by 2030 (as some areas are), regulators could say “200 MW of that should be allocated to government or public-interest compute centers.” This could take the form of dedicated data center parks for government and research institutions, possibly with federal funding. An analogy is how water rights or radio spectrum are reserved for public use – we might carve out a portion of the “compute grid.” FERC (Federal Energy Regulatory Commission) and state Public Utility Commissions could implement policies such that when a new data center requests interconnection, priority in the interconnection queue (the process of getting grid access) is given to those meeting certain public criteria. In some regions, data center growth has been so rapid that utilities are actually delaying connections – for example, in parts of Virginia, data centers faced wait times because the grid needed upgradesscientificamerican.com 55. If a high-priority AI facility comes along, we’d want it to skip ahead of lower-priority ones in the queue. FERC could issue a ruling that national security-related data centers get expedited interconnection review, similar to how we expedite grid connections for, say, military bases or hospitals.
• Special Energy Rates or Credits for A‑1 and A‑2: Another tool is using pricing to favor the desired uses. Utilities often offer special rate classes for large customers (e.g., data centers often negotiate lower rates or renewable energy deals). Regulators could sanction a preferential electricity tariff for facilities that serve critical AI applications. For example, a data center running a DOE national lab AI could get a slight discount or guaranteed renewable energy supply, whereas purely commercial ones might pay normal or higher rates (reflecting their non-critical status and perhaps compensating for their externalities). This is somewhat controversial (picking winners via rates), but it could be justified under critical infrastructure support. It mirrors how some essential services get lower utility rates (like street lighting, certain government facilities). The state of Virginia has considered “data center tariffs” to ensure residents don’t subsidize data center power costsvirginiamercury.com 56; we could invert that concept to make non-critical AI pay more, subtly encouraging investment to tilt toward critical AI that enjoys better rates.
• Capacity Planning and FERC Queue Reforms: We touched on FERC above – more broadly, integrate AI needs into national grid planning. Today, generation and transmission projects go through FERC-regulated queues and studies. We could incorporate AI data center clusters as a formal category in transmission planning, ensuring that regions earmark the needed substations and lines for not just any data center, but specifically for those tied to A‑1/A‑2 uses (perhaps co-locating them with national labs or bases where appropriate). The government might invest in some publicly owned data center infrastructure (or public-private) to host A‑1/A‑2 work, analogous to how it built national labs for physics research in the 20th century. If such infrastructure exists, FERC and DOE can ensure it’s supplied with adequate power first.
• Transparency and Conditional Approvals: At a process level, we could require that large data center projects disclose the general nature of their workload mix as a condition for certain approvals or incentives. For instance, if a company wants a federal tax break for building a data center (there are federal tax incentives in some cases, and certainly state ones), they might need to report if the facility will predominantly support, say, advertising vs healthcare AI. This transparency could inform local and federal decision-making. Potentially, incentives could be tied to category – e.g., an AI training center for medical research might qualify for an enhanced incentive, whereas one for social media might not. This would steer investment. Of course, companies could try to game this (claiming broader benefits), so criteria must be strict. But it parallels how certain industries (like renewable energy or semiconductor fabs) get special support due to national interest.

All these mechanisms would require careful calibration to avoid heavy-handed control that stifles innovation. The goal is not to “kill” Category B – consumer tech is a huge part of the U.S. economy and a source of AI innovation. Rather, the goal is to ensure Category A‑1 (and key A‑2) is not starved of resources just because it doesn’t generate immediate profit. There is precedent: during WWII, the U.S. curtailed some consumer production to prioritize war production – a drastic example, not analogous to peacetime AI allocation yet, but it shows the principle of prioritization in national interest.

Moreover, coordinating this nationally could actually benefit everyone in the long run. For example, if the power grid is upgraded to support a priority AI project, that infrastructure can also later support commercial uses; the key is sequencing and balancing. Likewise, investments in energy (like new nuclear or renewable plants) made to ensure A‑1 AI gets clean reliable power will increase overall supply, which can help Category B too without conflict. The difference is ensuring those investments happen in time – something the market might not do on its own, as pointed out by a CNAS report: “At home, permitting and regulatory constraints threaten to limit America’s ability to meet the energy demand for large-scale AI data centers… nondemocratic countries… seek to rapidly build frontier-scale AI data centers”cnas.org 57. In other words, if we don’t streamline and prioritize critical AI infrastructure, others (like China) could outpace us simply by virtue of decisive action.

In implementing any such policy, scientific-method rigor suggests we should also monitor for unintended consequences (disconfirming evidence). For instance, if we reserve too much capacity for A‑1 that goes unused, that would indicate over-allocation and need adjustment. Or if Category B companies find ways to claim their usage is “critical” (gaming the system), policies would need tightening. Ongoing data collection on how power is actually used (perhaps requiring large data centers to report anonymized utilization metrics) could help verify if interventions are working – e.g., by 2028 we’d want to see Category A‑1’s share of total AI MWh rising from the current low single digits to maybe double digits, indicating a healthier balance.

Conclusion

The evidence compiled supports the hypothesis: the growth of U.S. AI compute and power usage is overwhelmingly driven by Category B (commercial consumer) workloads, while truly critical (Category A‑1) uses remain a small minority of the total, albeit highly important. This state of affairs is not the result of strategic planning, but rather the byproduct of market incentives and fragmented “balkanized” decision-making. At present, local processes greenlight massive data center expansions with little regard to what the AI inside will be used for – be it serving TikTok videos or accelerating cancer research. The heavy tilt toward B means that, without intervention, we risk scenarios where in a crunch (e.g., an energy shortfall or a geopolitical crisis demanding surge computing), our national security AI could be stuck waiting behind ad engines.

By learning from rivals like China, who treat AI infrastructure as strategic and coordinate accordingly, the U.S. can take steps to rebalance. This does not mean adopting authoritarian controls, but smart policy levers – from prioritization frameworks (like a DPAS for AI) to incentivizing critical uses, to ensuring transparency and preventing one narrow interest from capturing all capacity. The recommendations outlined are initial sketches; they would need refinement and buy-in from stakeholders (including industry – ideally, companies would cooperate if they see that a healthier ecosystem ultimately benefits them too through stable grids and national security).

In closing, the goal is to ensure the “AI National Security Imperative” is met: that the U.S. retains leadership in AI for the things that matter most. Achieving that requires acknowledging the current misalignment (lots of power going to disposable consumer uses) and then deliberately correcting course. The alternative – continuing laissez-faire – isn’t likely to self-correct in time. As one analysis warned, “Without bold investment and streamlined policies, we risk losing our lead before 2030. The clock’s ticking, and China’s not slowing down.” That means we must ask tough questions: Is the next data center someone proposes serving public defense or just private profit? And if it’s the latter, are we comfortable with that when critical needs might go unmet? Our research indicates a need for urgency in addressing these questions. The evidence does not so much suggest scrapping commercial AI (which has its benefits) but rather elevating the national conversation: treating AI infrastructure with the same seriousness as we do critical defense infrastructure. Only then can we escape the trap of balkanized, ad-hoc decisions and instead steer AI development toward bolstering national security and the public interest.

(This report used a range of government, academic, and industry sources to substantiate each claim. For further detailed data and analysis, see the associated deep-dive documents on Idaho Data Centers, China’s Energy Strategy for AI, and Tech Lobbying & Regulatory Capture. The findings here should be revisited as new evidence emerges – for instance, if Category B growth slows or if the government launches major AI infrastructure programs, those would be important factors to update our scenarios.)

Global Compute and National Security | CNAS

afcea.org 58 China’s AI Infrastructure Surge: U.S. Must Catch Up, Experts Say | AFCEA International

morganstanley.com 59 China Quickly Becoming an AI Global Leader | Morgan Stanley

chinatalk.media 60 Powering China’s Data Centers: Batteries or Nukes?

scmp.com 61 China to power grid with record renewable energy as AI spurs demand | South China Morning Post

nucamp.co 62 This Month's Latest Tech News in Boise, ID - Wednesday April 30th 2025 Edition

goodjobsfirst.org 63 What to ask when a data center wants to come to town - Good Jobs First

virginiamercury.com 64 Data centers, non-disclosure agreements and democracy • Virginia Mercury

govtech.com 65 St. Charles, Mo., Moves to Ban Data Centers for 1 Year - GovTech

datacenterdynamics.com 66 Meta's Kuna, Idaho, data center latest to be "paused" in AI redesign ...

gemstatepatriot.org 67 Idaho Datacenter Initiative - Gem State Patriot

apnews.com 68 Pentagon's AI initiatives accelerate decisions on lethal autonomous ...

pmc.ncbi.nlm.nih.gov 69 Artificial Intelligence for Medical Diagnostics—Existing and Future AI ...

sciencedirect.com 70 Artificial intelligence integrated grid systems: Technologies, potential ...

reddit.com 71 Today Mark announced Meta's major AI compute investment. - Reddit

scientificamerican.com 72 AI Will Drive Doubling of Data Center Energy Demand by 2030 | Scientific American

thekeyword.co 73 AI boosts Q2 2025 ad revenue across Google, Meta, and Amazon

emarketer.com 74 $10 billion Meta-Google cloud pact shows scale beats rivalry

opensecrets.org 75 Internet Summary • OpenSecrets

thinkinsider.org 76 Pentagon To Wall Street: How The U.S. Bets On AI - Think Insider

investopedia.com 77 Which Industry Spends the Most on Lobbying?

issueone.org 78 Big Tech Cozies Up to New Administration After Spending Record ...

builtin.com 79 Pentagon Awards Anthropic, Google, OpenAI and xAI Contracts to ...

alabamasolutions.com 80 Meta vs Google: The Billion-Dollar Battle for AI Compute Supremacy –

azluminaria.org 81 Public officials reconsider NDA process amid Project Blue outrage

axios.com 82 2024 tech lobbying in review - Axios