AI's Insatiable Power Appetite Rewrites the Energy Playbook

A Climate Week NYC panel revealed just how fast the ground is shifting beneath energy planners: Bloomberg NEF's forecast for U.S. data center capacity has jumped 50% in six months, and the race to power artificial intelligence is now colliding head-on with clean energy goals, consumer affordability, and a domestic natural gas supply that can't stretch far enough to satisfy everyone.

• Bloomberg NEF projects 118 gigawatts of U.S. data center capacity by 2030 — 12% of national electricity demand, rising to 20% by 2035
• IRA repeal through the One Big Beautiful Bill Act triggers a clean energy deployment race before tax credits phase out, followed by a likely installation lull
• 70% of new data center generation would be natural gas if lowest-cost sources prevail, clashing with decarbonization commitments
• PG&E argues every gigawatt of data center load could cut rates by 1% by filling a grid running at just 45% utilization
• Princeton Digital Group targets 75–100% hourly clean energy matching by 2036 across Asia, pioneering granular tracking in India and Indonesia
• Tech giants have struck roughly 40 GW of nuclear partnerships — but nearly all remain non-binding letters of intent

---

The Numbers That Changed in Six Months

The basics: Bloomberg NEF, the energy research arm of Bloomberg, tracks global electricity demand and generation investment. At a Climate Week NYC fireside chat moderated by Adrian Varga, Senior Energy Strategist, ACT Group, Stephanie Diaz, Technology and Innovation Senior Associate, Bloomberg NEF, walked through the firm's latest modeling — and the numbers have moved dramatically.

Why it matters: AI-driven data center demand has overtaken transportation electrification as the single largest driver of rising U.S. electricity consumption through 2050 in BNEF's models. That shift is forcing utilities, regulators, and clean energy developers to rewrite plans in real time.

Where things stand: BNEF now forecasts 118 gigawatts of U.S. data center capacity online by 2030 — a figure 50% higher than the firm's projection from just six months earlier. That translates to roughly 12% of average U.S. electricity demand by 2030, climbing to approximately 20% by 2035.

"In our latest forecast, we foresee 118 gigawatts of data centers coming online by 2030," said Stephanie Diaz, Senior Associate, Bloomberg NEF. "That roughly translates to 12% of average US electricity demand by 2030."

The concentration is even more dramatic in specific regions. In the PJM interconnection — home to Virginia's "Data Center Alley" — data centers could account for a third of electricity demand. Globally, the U.S. accounts for about half of projected data center capacity.

The energy trilemma no one can solve: Diaz identified a fundamental policy collision at the heart of the current administration's agenda. The Trump administration is simultaneously pursuing AI dominance (requiring massive electricity growth), energy dominance (including expanded LNG exports), and energy affordability for consumers. BNEF's modeling shows these three goals cannot coexist: pursuing all three creates a projected 12 billion cubic feet per day deficit in domestic natural gas supply.

"You can't have all three of those. They actually don't work," Diaz said.

---

IRA Repeal Creates a Race, Then a Lull

The basics: The One Big Beautiful Bill Act (OBBBA) repeals the Inflation Reduction Act's tax credits for solar, wind, and battery storage on varying timelines, removing the policy framework that drove record U.S. clean energy deployment.

Why it matters: The phase-out structure creates a near-term sprint — developers are racing to qualify projects before credits expire — followed by a projected installation lull as the economics shift. But the story is more nuanced than a simple clean energy setback.

Despite cutting its clean energy installation forecasts due to the IRA repeal, BNEF actually increased its overall clean energy forecast by 26%. The reason: the sheer scale of new electricity demand from data centers requires so much new generation that clean energy still grows — just not as fast as it would have under the IRA.

The catch is stark. "In our modeling, if we look at just the lowest-cost generation source, it'd be something like 70% of the marginal additions just for data centers would be natural gas," Diaz said. That creates a direct tension between powering AI and meeting decarbonization targets.

What's next: The near-term deployment race means 2025–2027 could see accelerated renewable installations as developers lock in expiring credits. After that, the trajectory depends on whether demand growth alone — absent federal incentives — can sustain clean energy investment at scale.

---

Who Pays for the Grid AI Needs?

The Contrarian Case: Data Centers as Rate Relief

Why it matters: The question of who pays for grid expansion to serve data centers will directly affect utility bills for residential and commercial customers in every major U.S. market. The answer is far from settled — and one major utility is making a surprising argument.

Aaron Johnson, Chief Sustainability Officer, PG&E, presented a counterintuitive case: data centers don't just consume electricity — they could make it cheaper for everyone else.

"Every gigawatt of data center we add has the opportunity to reduce our rates, which have been growing very quickly, especially as we adapt to the wildfire risk, by 1%," Johnson said.

His logic: PG&E's electric grid is utilized only about 45% of the time. Data centers, which run around the clock, fill that underutilized capacity and spread fixed costs across more kilowatt-hours. PG&E currently has 5 GW of committed data center projects and another 10 GW of expressed interest — on a 20 GW system.

The Unresolved Question

Diaz cautioned that the industry hasn't yet grappled with the harder question. "How do we make sure data centers pay their share? I don't think we have collectively really wrestled enough with what is their share exactly, where do we draw the lines on that," she said.

The conversation has shifted from "what's the cheapest technology?" to "who bears the cost of new capacity?" Data centers are willing to pay premiums for power, but there is growing societal concern about whether residential ratepayers will end up subsidizing infrastructure built primarily for tech giants.

Speed to Power

On the supply side, panelists described creative approaches to accelerating generation: repurposing retired power plant sites with existing grid connections, deploying batteries for load shifting, and virtual power plant partnerships. The Google-Voltus-PJM collaboration for 100 MW of demand response was cited as a concrete model.

Johnson also advocated for a fundamental shift in how utilities plan their grids. Traditional "deterministic" planning assumes worst-case scenarios — "every outlet has a hair dryer plugged into it, and they're all operating at full bore," he said. He argued that sensor technology and software now enable probabilistic approaches that could unlock significant capacity without compromising safety. PG&E is piloting virtual power plant concepts not just for grid operations but for planning assumptions.

Supply chain constraints add another layer. Even PG&E, a major buyer, struggles to procure critical equipment. "We send folks to Korea and we can't get moved up in the queue," Johnson said.

---

Asia's Clean Energy Lab: Hourly Matching Across Eight Markets

Why it matters: Annual renewable energy matching — the standard corporate practice of buying enough clean energy certificates to cover a year's consumption — is giving way to hourly granular tracking, a far more rigorous standard. Data center operators are driving this shift, and the most ambitious work is happening not in the U.S. or Europe but across Asia.

Preeti, Group Director for Sustainability and Technology, Princeton Digital Group, described PDG's effort to match 75–100% of its electricity consumption with clean energy on an hourly basis by 2036 across its 2 GW, eight-market Asia-Pacific portfolio.

"We've set a target for ourselves to be about 75 to 100% matched with 24/7 clean energy by 2036. This is very challenging in Asia, where every market has its own dynamics," Preeti said.

PDG has pioneered hourly granular energy tracking in India through a partnership with Flexidao and is extending similar work to Indonesia. In India, the economics already favor clean energy: "You can procure green energy at a significant discount to brown. So it's commercially viable," Preeti said.

But markets like Malaysia have only recently opened direct access to corporate consumers, and contractual structures for clean energy procurement remain undeveloped in much of the region. PDG's strategy centers on being an anchor customer — the large buyer whose commitment drives scale and makes renewable projects financeable.

"We can't be passive consumers of energy. We have to be an active member of the energy ecosystem, ensuring that we are anchor customers leading to more clean energy joining the grid," Preeti said.

---

40 Gigawatts of Nuclear Hype — and Some Real Deals

Next-Gen Bets

Why it matters: Large energy consumers are evolving from buyers to technology development partners, absorbing first-of-a-kind risk in exchange for the promise of firm, clean power. Their willingness to invest early could accelerate cost curves for nuclear, geothermal, and long-duration storage — or produce expensive dead ends.

Diaz reported that BNEF has tracked approximately 40 GW of partnerships between data centers and nuclear companies. But she was careful to calibrate expectations: "I'm going to emphasize the word partnerships there because this is a lot of letters of intent and memorandums of understanding" — non-binding expressions of interest, not firm commitments.

More tangible is Google's Free Rates Clean Transition Tariff, which is helping fund Fervo Energy's enhanced geothermal project in Nevada. Google is absorbing some grid connection costs and first-of-a-kind technology risk — a model Diaz argued large tech companies could replicate across emerging clean technologies, much as they pioneered early power purchase agreements for conventional wind and solar.

Johnson offered a reality check from the procurement side. In California, geothermal has significant potential but has not been competitive in solicitations — solar plus batteries continues to win. At some point, he acknowledged, battery capacity will need to match solar capacity, and a complementary firm clean technology will be needed.

PG&E is also exploring a pragmatic bridge: using data center backup generation — currently Tier 4 diesel transitioning to natural gas — as a peaking resource, potentially avoiding the need to build new peaker plants.

---

EVs: The Next Demand Surge — and a Massive Grid Battery

Heather McGorry, Vice President, Energy and Sustainability, CoreWeave, flagged a demand wave that could dwarf data centers. Bloomberg research suggests EV electricity demand could overtake data center demand around 2033 — and unlike data centers, EV load is distributed and mobile, a fundamentally different challenge for grid planners.

Johnson said PG&E has actually scaled back distribution-level grid investment because EV adoption hasn't matched earlier forecasts. But he expressed strong optimism about vehicle-to-grid technology: "The potential for EVs to be batteries sitting in everyone's driveway or in the garage in your apartment building is a really spectacular technology that will tap into an amazing amount of resource to smooth out the grid."

The complication is ultra-fast chargers capable of delivering 400 miles of range in 15 minutes. These create punishing load profiles for the distribution grid, and the shift from home charging to public fast charging reduces utility control over charging timing. If sufficient battery capacity — including EVs — is eventually deployed, Johnson suggested time-of-use rate structures could become unnecessary, fundamentally changing utility business models.

---

Engineering the AI Factory

Jeff Schmidt, Senior Vice President, Industrial Segment, UL Solutions, described the safety and engineering challenges of powering AI at scale. Direct current power — increasingly favored by hyperscalers — poses unique risks: "Unlike alternating current that crosses zero, direct current does not. And so if you get an arc with direct current, it can lead to a very catastrophic fire."

UL Solutions is a founding member of CurrentOS, a direct current consortium originally focused on buildings that has pivoted toward data centers. Schmidt is also involved in the Open Compute Project as hyperscalers push power density and thermal management to new limits.

Despite the risks, Schmidt was bullish on the broader payoff: "The innovations that are going into making these AI factories — and trust me, we understand there are a lot of societal risks and issues that have to manage it. But on an engineering level, the amount of innovation that this is creating is going to really proliferate and make the electrification of everything go faster and more effectively."

---

Minor Items

• Richard Tarberton of the World Resources Institute asked panelists whether the natural gas build-out for data centers is inevitable and whether large AI product users could pressure hyperscalers toward more renewable energy. Panelists acknowledged the tension but pointed to commercially viable clean energy pathways in some markets.
• CoreWeave, represented by McGorry, provides cloud compute for AI training and inference to Google, Meta, Microsoft, OpenAI, and Anthropic.
• Adrian Varga framed the session around the "old energy trilemma" — reliable, affordable, and sustainable power — arguing the challenge is whether AI can be powered in a way that satisfies all three.