Miner Weekly: Can Flexible AI Load Steal Bitcoin Mining’s Grid Pitch?
This article first appeared in Miner Weekly, a weekly newsletter by BlocksBridge Consulting, curating the latest news in energy, bitcoin, and AI compute from TheEnergyMag. Subscribe to receive in your inbox once a week.
Today’s featured EIF speaker: Blake King, Director and Head of Power at Galaxy (Nasdaq: GLXY), where he leads power strategy for the company’s data center operations and has helped guide its transition from bitcoin mining into large-scale AI and compute infrastructure. He brings nearly a decade of power-sector experience across ERCOT, Oak Ridge National Laboratory, and RES, spanning transmission planning, power systems R&D, and grid modeling.
EIF Dallas (July 23) is a high-signal, curated conference convening leaders across energy, compute infrastructure, and capital to map the next decade of power and data center buildout.
Get your tickets today and use code MNRW30 for 30% off exclusively for Miner Weekly subscribers. You can also explore sponsorship opportunities to elevate your brand and connect with leaders across energy, infrastructure, and capital.
Bitcoin miners have long defended their power use with a simple argument: mining may be energy-intensive, but it is unusually interruptible.
A mining facility can curtail quickly when power prices spike or the grid is under stress. It can monetize surplus electricity when power is abundant, then shut down without disrupting households, factories or customer-facing software. That flexibility has become central to the industry’s pitch to utilities, regulators and investors, especially in markets such as Texas where demand response has become part of the mining business model.
But a new line of research raises a provocative question for the AI-infrastructure boom: what if GPU data centers can learn a version of the same trick?
A paper posted to arXiv in late June, titled “Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute,” argues that AI data centers should not always be modeled as static, inflexible power loads. The authors describe a software architecture that combines grid signals, workload scheduling and power telemetry to reduce, defer or shift AI workloads in response to power-system conditions.
In a real-world deployment on a 130 kW GPU cluster, the paper says the system demonstrated rapid load reduction, sustained curtailment, carbon-aware operation and geographically distributed load shifting while preserving priority service levels.
The headline results were notable for the speed and range of the claimed response. The authors said the system reached 100% compliance across more than 200 National Grid power events in the UK test environment. In a simulated emergency dispatch modeled on a historical grid contingency, the cluster reduced power use by about 30% within 40 seconds. In another fast-response test, it cut load by 40% within roughly one minute. The paper also reported sustained curtailment of 10% to 40% for periods ranging from two to 10 hours, as well as a U.S. geo-shifting test in which 10% of live inference traffic was shifted from Oracle GPUs in Virginia to Illinois.
The paper is not yet peer reviewed. The arXiv version says it has been submitted to IEEE for possible publication, so its findings should be treated as early technical evidence rather than settled academic consensus. Still, the idea is not emerging in isolation. The June paper builds on a related Nature Energy article, “AI data centres as grid-interactive assets,” which reported a field demonstration of software-based demand response at an AI cluster in Phoenix, Arizona. A Nature Reviews Electrical Engineering summary of that work said the researchers showed that data centers can function as flexible, grid-aware loads by adjusting computing workloads in real time rather than relying on new hardware or on-site batteries.
The Phoenix trial is important because it moves the discussion beyond simulation.
According to the preprint version of the Nature Energy work, the test was conducted at an Oracle Phoenix Region Cloud data center on a 256-GPU cluster built with NVIDIA A100 Tensor Core GPUs. The system used Databricks MosaicML for workload orchestration, Weights & Biases for telemetry, and Amperon’s grid-demand forecasting tools. The project involved Emerald AI, NVIDIA, Salt River Project and the Electric Power Research Institute, and was conducted through EPRI’s DCFlex initiative.
The cluster ran representative AI workload mixes, including training, inference and fine-tuning jobs. The researchers classified jobs into flexibility tiers based on how much performance reduction users could tolerate: strict jobs with no performance reduction, and other tiers allowing average throughput reductions of up to 10%, 25% or 50% over a three- to six-hour period. That matters because the system did not simply shut everything off. It selectively slowed, paused or power-capped jobs depending on their tolerance for delay.
The reported result was a 25% reduction in cluster power usage for three hours during peak grid events while maintaining quality-of-service guarantees. The authors said the reduction was achieved without hardware retrofits or battery storage, using workload orchestration and GPU power-limiting techniques such as dynamic voltage and frequency scaling, combined with job pausing.
That does not mean an AI data center is suddenly as flexible as a bitcoin mine.
The Phoenix test involved a 256-GPU cluster, not a 100 MW or 1 GW AI campus. The later arXiv paper’s 130 kW deployment is also small relative to the scale now being discussed by hyperscalers, utilities and public miners. Scaling this approach from a cluster to an entire campus would require coordination across power delivery, cooling, networking, customer contracts and service-level agreements. It would also depend heavily on workload mix.
The researchers acknowledge that limitation. Not all AI workloads are temporally flexible. Batch training, fine-tuning and some inference jobs may be slowed or paused. But strict “Flex 0” jobs — such as real-time inference, streaming and model serving — were not modified in the Phoenix demonstration. Future flexibility may require geographic load shifting, where workloads move between regions to avoid grid stress without creating unacceptable latency or performance penalties.
Even with those caveats, the research cuts into one of bitcoin mining’s most important narratives.
For years, miners have argued that they are uniquely useful to power grids because they can behave less like a conventional industrial load and more like a controllable energy buyer.
AI data centers are more complicated. They are built around expensive GPUs, customer commitments, uptime expectations and high utilization targets. Operators do not want billions of dollars of accelerators sitting idle. But the research suggests the choice may not be binary. An AI facility may not need to run flat out or shut off entirely. It may be able to preserve high-priority workloads while throttling lower-priority jobs, delaying flexible training runs, power-capping GPUs or shifting traffic to another region.
That is the miner-relevant point. Yet there are still reasons for skepticism.
The research has commercial backers like Emerald AI, NVIDIA, Oracle and power-sector partners, with a clear interest in showing that AI infrastructure can be made more grid-friendly. The evidence remains early, and the field tests are far smaller than the campuses now being planned. Utilities will also need market rules, telemetry standards and compensation structures before AI data centers can become meaningful demand-response resources. Customer willingness is another open question: not every AI user will accept slower jobs in exchange for lower power costs or faster interconnection.
But the debate is worth watching.
Bitcoin miners helped popularize the idea that digital infrastructure could act as flexible load. Now AI data centers are beginning to move into that same communication, policy, and engineering conversation.
If the research scales, mining’s grid edge may not disappear, but it may become less unique. The flexibility pitch is no longer just a bitcoin mining story. It is becoming an AI-infrastructure story too.
Regulation News
- DOE Lets PJM Tap AI Data Center Backup Power During Heat Wave
Hardware and Infrastructure News
- Bitdeer Signs Conditional Lease for Norway AI Data Center Site
- Blockfusion Touts Up to $5.4B From AI Lease as SPAC Deal Advances
- Cargo thieves have set their sights on data center supplies
- Bloom, Brookfield Expand AI Power Financing Partnership to $25 Billion
Corporate News
- IREN, Keel Join Russell Indexes as AI Infrastructure Stocks Draw Broader Attention
- Ionic Digital Raises $400M Ahead of Nasdaq Listing as AI Revenue Overtakes Bitcoin Mining
- Sphere 3D Seeks DarkHorse Rebrand as It Pivots Toward AI Infrastructure
- IREN Adds Oracle, Google Veterans to AI Cloud Leadership Team
Financial News
- American Bitcoin Sets Reverse Stock Split as Sub-$1 Nasdaq Risk Builds
- Applied Digital Expands Credit Facility to $430M as AI Data Center Buildout Continues
