A new round of reporting is putting energy use back at the center of the AI buildout. According to The Korea Times, a study found that advanced AI systems can use 136.5 times more electricity than standard chatbots, a gap that underscores how quickly power demand can rise as vendors move from simple text generation to more capable reasoning and multimodal features.

A separate report from Crypto Briefing points to the same broader trend from a different angle. It said Meta’s newer internal model effort, referred to as Watermelon, uses roughly ten times more compute than an earlier model called Avocado. Taken together, the two reports suggest that the AI industry’s next wave is being shaped not only by model quality and product adoption, but by how much infrastructure vendors and customers are willing to finance and operate.

The available evidence in this story is limited. Neither source provides the full underlying study or technical methodology in the material available here, and the Meta-related report appears to rely on secondary coverage rather than a primary technical disclosure. Even so, the claims matter because they align with a pattern already visible across enterprise AI, where each step up in capability often brings a disproportionate increase in hardware demand, energy use, and operating cost.

What the reports appear to show

The Korea Times describes a study warning that “advanced AI” consumes 136.5 times more electricity than standard chatbots. Without the full paper, key definitions remain unclear: the report does not specify which models were tested, what counted as “advanced,” whether the comparison measured training or inference, or how workloads were standardized.

Those details matter. A lightweight chatbot that answers short prompts from a compact language model has a very different hardware profile from a system that uses large-context reasoning, tool calling, search, image analysis, or multi-step agent behavior. In practical terms, many products marketed as AI assistants now bundle several expensive operations behind a single response. That makes energy comparisons highly sensitive to product design.

The Crypto Briefing item adds another signal on the compute side. It says Meta’s Watermelon model uses ten times more compute than Avocado, framing that increase as evidence of a stronger internal push toward proprietary AI. Here too, the sourcing visible in the evidence is thin. There is no accompanying research paper, benchmark pack, or infrastructure note in the materials provided. But the direction of travel is familiar: larger or more advanced model programs tend to require more chips, more power, and more capital.

For product teams using OpenAI, Anthropic, Google, or Meta models through APIs or cloud services, these back-end differences are often hidden. What shows up instead is pricing, latency, capacity limits, and availability tiers. For teams training or fine-tuning models directly, the tradeoffs are much more immediate.

Why the electricity gap matters now

The timing is important. The AI market is moving beyond basic chatbot deployments toward AI agents, long-context workflows, coding assistant tools, multimodal search, and enterprise copilots that sit inside daily operations. Those products can improve usefulness, but they also stack more computation into each user interaction.

That raises a basic business question: when does a better answer stop being worth the added cost? If a high-end model or agent architecture needs dramatically more electricity than a standard chatbot, the economics change for both vendors and buyers. Cloud bills go up. Hardware utilization becomes more important. Regional deployment decisions may shift toward locations with cheaper or more reliable power.

The issue is especially relevant for enterprise AI buyers that want predictable unit economics. A pilot may look attractive when usage is low, but a broad rollout across customer support, search, internal knowledge, software development, and automation can expose just how expensive heavy inference workloads become. In that context, “advanced AI” is not just a capability label. It is also a budget category.

The reports also land amid wider concern about AI data center expansion. Utilities, regulators, and hyperscalers have all been pulled into debates over whether power generation and transmission can keep pace with demand from GPU-heavy infrastructure. Even without the full methodology behind the 136.5-times figure, the core warning fits that larger market discussion.

Evidence, caveats, and who is making which claims

The strongest factual constraint in this story is that the evidence is indirect. The Korea Times report references a study but, in the material available here, does not include the paper itself, its authors, or the exact experiment design. As a result, the 136.5-times figure should be treated as a reported finding rather than a settled industry baseline.

The same caution applies to the Meta-related claim. Crypto Briefing says Watermelon uses ten times more compute than Avocado, but the source evidence available here does not include a Meta technical note, model card, or public engineering documentation. That means the compute comparison should be read as media-reported, not as a directly verified disclosure from Meta.

There is also a measurement problem that frequently complicates AI energy stories. Compute is not the same thing as electricity, and electricity is not the same thing as carbon emissions. A model can use more compute but run on more efficient hardware, in a better-optimized data center, or in a region with different power sources. Likewise, a chatbot can appear cheap in one benchmark and costly in production if it relies on repeated calls, retrieval layers, or guardrail systems.

Still, uncertainty cuts both ways. Thin sourcing means readers should avoid over-reading the exact multipliers, but it does not erase the underlying pattern. Across generative AI, capability gains often come with steep resource costs, especially when vendors move to larger models or heavier reasoning paths.

What it means for builders and enterprise teams

For builders, the immediate takeaway is architectural discipline. Teams do not need the most expensive model for every request. Routing simple tasks to smaller models, reserving premium inference for hard cases, trimming context windows, and limiting tool use can materially reduce cost and power demand. In many cases, better orchestration beats defaulting everything to a frontier model.

For enterprise AI teams, the story is a reminder to evaluate products on workload shape, not just demo quality. A standard chatbot that handles high-volume FAQ traffic may be cheaper and more reliable than a richer assistant that invokes multiple models and external tools on every turn. That does not make advanced systems a bad investment. It means buyers need visibility into inference patterns, latency, and pricing under real usage.

The implications extend to coding assistant and workplace automation deployments as well. These systems can generate measurable productivity gains, but they often involve sustained, repeated inference rather than occasional queries. If usage grows from a few hundred employees to tens of thousands, infrastructure intensity becomes a procurement issue, not a back-office detail.

This is also relevant to open-source and self-hosted strategies. Some companies adopt open models to reduce dependence on API pricing, but large deployments still require substantial GPU capacity and energy planning. The real question is not cloud versus self-hosted in the abstract. It is which combination of model size, latency, privacy, and utilization makes sense for a given workflow.

Competition is moving from model quality to infrastructure efficiency

The reports point to a competitive shift that is easy to miss in feature-focused coverage. AI vendors are no longer competing only on benchmark scores or product breadth. They are competing on how efficiently they can deliver advanced capabilities at scale.

That matters for Meta, whose AI strategy has spanned both open-weight releases and major internal model investment. If the Watermelon-versus-Avocado compute comparison is directionally correct, it suggests Meta is willing to absorb significantly higher infrastructure costs in pursuit of stronger proprietary performance. The same pressure applies to rivals across enterprise AI: model progress is constrained not only by research, but by available power, chips, and capital.

For customers, this may lead to a more stratified market. Basic chatbot services could keep getting cheaper as models and serving stacks improve. At the same time, premium AI agents and advanced multimodal systems may remain costly because they consume much more compute per useful action. That gap could shape pricing tiers, product packaging, and buyer expectations over the next year.

What to watch next

First, watch for publication of the underlying study cited by The Korea Times. The credibility of the 136.5-times claim depends heavily on how the researchers defined advanced AI, what baseline they used, and whether they measured training, inference, or both.

Second, watch for direct disclosures from Meta on Watermelon and Avocado. A technical paper, model card, or executive comment would help separate market interpretation from documented engineering facts.

Third, track whether major vendors begin emphasizing efficiency metrics more explicitly. Enterprise AI buyers are increasingly likely to ask not only which model performs best, but which model delivers acceptable quality per watt, per token, or per task completed.

Finally, pay attention to product design. The winners in AI agents and workplace automation may not be the systems with the most raw capability. They may be the ones that use advanced reasoning selectively, keeping costs and electricity demand within a range that enterprises can actually sustain.

Creati.ai perspective

The most important point in this story is not the exact multiplier in either report. It is that AI economics are becoming inseparable from infrastructure economics. As products evolve from a standard chatbot into richer assistants and AI agents, compute intensity rises fast, and electricity becomes a practical constraint on product strategy.

For founders and product teams, that creates a clear mandate: design for capability efficiency, not just capability maximums. In enterprise AI, the durable advantage may come from smart routing, smaller specialized models, and workflow-aware orchestration as much as from the biggest frontier systems. The market is still rewarding model ambition, but over time it is likely to reward disciplined deployment even more.