When the Power Grid Starts Pricing Tokens for AI Inference

01 Trigger Event

On July 17, 2026, at the WAIC forum in Shanghai on the “two-way relationship between artificial intelligence and energy,” China’s National Energy Administration sent two concrete signals: first, during the 15th Five-Year Plan period, inference load will surpass training load and become the main source of compute-power consumption; second, of total electricity consumption across society today, nearly four out of every ten kilowatt-hours already come from green power.

During the 15th Five-Year Plan period, inference load will exceed the training load of the past and become the primary consumption of compute power.

On the surface, this looks like the energy system endorsing AI.

But in my view, this is actually the first clear acknowledgment at the policy level that AI’s main battlefield is shifting from one-off training capex to persistent inference operating load.

This is not a rhetorical shift. The resource object itself has changed.

Training is more like a project-based activity; inference is more like a utility load. The former competes on peak GPU capacity; the latter competes on stable power supply, latency, dispatch, and unit Token cost. I have not seen any follow-up supporting documents from the National Energy Administration, so I may be misreading this point. But based on this language alone, the regulatory perspective has already started to treat AI as a new category of power demand, rather than merely a technology-sector slogan.

02 What This Really Means

What truly matters is not “green power supports AI,” but that the energy system is beginning to understand AI through the physical characteristics of inference.

The original statement was blunt about it: latency-sensitive, highly volatile load, concentrated demand. Put those three together and the target is not large-model capability, but serving economics.

Training can be shifted to cheaper, more peripheral power nodes, and can even tolerate longer job cycles.

Inference cannot.

Inference has to stay close to users, close to networks, close to demand peaks, while also absorbing the instant load spikes that come with product breakout moments. The issue is not “whether there are more models,” but “which model’s Tokens can be delivered more reliably by the grid and the datacenter.”

That is what the National Energy Administration is actually talking about.

If inference becomes the primary load, then what gets repriced in the future may not be only GPU rental, but the entire chain: rack location, PUE, green-power matching capability, regional electricity pricing, peak-shaving capacity, and even KV cache hit rate. The last of these sounds like a model-engineering detail, but it ultimately lands on the electricity bill.

I have not run nationwide datacenter dispatch myself, so I may be overstating this. But for API buyers, the correlation between Token pricing and electricity pricing will only strengthen from here, not weaken.

03 Historical Analogy / Structural Comparison

This is closer to AWS in 2014 moving from “selling servers” to “selling infrastructure abstraction” than it is to the product explosion of ChatGPT in 2022.

When the industry talks about training, it focuses on frontier model, scaling law, parameter count, MoE architecture, and capex.

When the industry talks about inference, the center of gravity shifts to an entirely different set of concerns: availability, latency, batching, prompt caching, routing, and region placement.

This is a shift from an “arms race” to a “highway tollbooth.”

The iPhone turning point was putting computing into your pocket. The AWS turning point was turning compute into an on-demand utility call. The next turning point for AI may be turning Token into a continuously consumed good constrained by the power system. What will truly be priced is not model parameters, but deliverable inference.

If this judgment holds, then the moat will move with it: from those best at training models to, at least in part, those best at delivering models to end users at the lowest total cost. I cannot prove this will happen comprehensively within 12 months, and I may be leaning too structuralist here. But at minimum, China’s policy signals are already pushing in that direction.

04 What This Means for AI Builders

For builders, what needs adjusting this week and this month is not PR language, but system design.

First, stop treating model cost as nothing more than the API unit price. Start breaking it down into input Token, output Token, cache hit rate, peak-versus-off-peak traffic spread, geographic distribution, and latency SLA. What you are buying is not a model, but a delivery system constrained by electricity.

Second, routing will matter more than a “single strongest model.” Once inference becomes the primary load, multi-model routing, fallback strategy, asynchronous task pools, and batch API discounts will all move from optimization items to survival items. I have not seen each of your traffic profiles, so I may be generalizing. But if 80% of requests in your product do not require the strongest model and yet all of them are still hitting the most expensive endpoint, then you are feeding your gross margin to the grid.

Third, do not treat the green-power narrative merely as a branding bonus. Once policy pushes “green compute” from concept to implementation, it may enter park admission rules, listing approvals, resource allocation, and even local investment attraction. At that point, platforms that can prove efficient inference utilization will have a smoother path to securing resources.

Fourth, AI infra founders need to rethink distribution. The future advantage is not just access to more GPU, but whether you can orchestrate GPU, datacenters, electricity pricing, networks, and workloads together. MCP is hot. Agent is hot. But if the underlying inference economics do not work, then however elegant the upper-layer protocol may be, it is still only a thin shell.

05 Counterview / Risks

I may be wrong in two places.

First, this may simply be a standard policy statement and not a sign of substantive resource reallocation in the near term. When the National Energy Administration says “inference load will exceed training load,” that does not automatically mean local grids, cloud vendors, and datacenters will immediately rebuild supply around inference priority. Without follow-up documents, subsidies, dispatch mechanisms, and park-level implementation, the significance of this news could be overstated.

Second, China AI’s real bottleneck may not yet be electricity, but rather high-end chips, network interconnects, the software stack, and the density of commercial demand. In other words, even if the power system is ready, model serving may not be able to scale smoothly. I do not have access to the real utilization rates of compute clusters across regions, so I may be projecting a long-term constraint too early onto short-term reality.

So the safest reading is not “energy will become AI’s core bottleneck,” but rather that the policy layer has already started to treat inference as a load type worthy of independent governance.

That alone is important enough.

Because once the government sees it that way, cloud vendors will invest that way, datacenters will be built that way, platforms will sell that way, and application companies will eventually be forced to run their economics that way as well.

Today they are discussing green power.

Tomorrow, what the market may truly reprice is the power path behind every single Token.