Computing-Power Coordination Standard: AI's Real Bottleneck Isn't in Chips

01 Trigger Event

The "Technical Specification for Monitoring and Energy Management Systems of Data Center Microgrids," submitted by China to the International Electrotechnical Commission (IEC), has recently been approved as a new work item. A Chinese expert serves as the project leader, working alongside experts from France, Germany, Spain, and other countries to jointly develop the standard. This is the world's first international standard focused on the "data center computing-power coordination" domain, according to CCTV News.

02 What This Really Means

On the surface, this looks like "China leads an IEC standard"—a soft power PR win for any standards-based announcement.

But the question worth asking is: why does this standard appear at this specific moment?

Over the past decade, the AI infrastructure bottleneck narrative has centered on three things—chips (NVIDIA), interconnect (NVLink, InfiniBand), and cooling (liquid cooling). This narrative has made the "upper layers" of compute appear scarce, letting GPU vendors capture pricing power.

But starting in 2024, more and more signals point to a fourth dimension: electricity.

H100 draws 700W per card; B200 jumps to 1000W+. A 100MW data center equals the peak load of a mid-sized city. In parts of the US, grid interconnection queues are already pushed past 2030. AWS, Microsoft, and Google are all scrambling for nuclear + geothermal + solar direct-supply PPAs. Oracle has directly tied 1GW of nuclear capacity to STMicroelectronics. Data centers have shifted from "buying electricity" to "making their own."

When AI compute itself becomes a dispatchable load on the grid, who defines this interface?

The traditional positioning of data centers is "pure electricity consumer." Computing-Power-Grid Coordination upgrades them to "electricity participants"—peak shaving, frequency regulation reserves, balancing renewable output fluctuations, even reverse power flow (VPP, virtual power plant). Once the interface is standardized, whoever defines the reference architecture first locks in the upstream position in AI infrastructure procurement for the next decade.

That's what this IEC standard is fighting for. Soft power PR is the surface; interface definition rights are the substance.

03 Historical Analogy

The closest parallel is the 3GPP standards war during the 4G/5G era.

Around 2016, Huawei pushed Polar Code; Qualcomm pushed LDPC. The two sides battled for 18 months over 5G NR channel coding schemes. On the surface it was technical; in reality it determined telecom equipment market share and patent royalty distribution for the next decade. Ren Zhengfei personally took the stage—it was a genuine national-level industrial strategy. China Mobile's proposal on the TDD frame structure followed the same logic.

OpenStack is the counterexample—AWS completely ignored this "open cloud standard" and defined the de facto standard through its own market share. So standards battles are often won not by technical superiority, but by who runs the reference design first + who locks in customers first.

Computing-power coordination sits between these two cases:

• On one hand, the data center ↔ grid interface must have a standard, because this is physical infrastructure, not software (unlike cloud, where one player can dictate terms)
• On the other hand, whoever first demonstrates the end-to-end integration of AI training clusters + energy storage + solar + grid dispatch will define the reference

China's move here aims to seize the "interface definer" ecosystem position. I haven't worked inside IEC, so I'm not sure how much P-member voting weight a new work item can command, but project leadership + first proposer status typically translates into agenda-setting power in international standards battles. This is a playbook validated in the 3GPP era.

04 What This Means for AI Builders

Short-term (3-6 months)—virtually zero direct impact. The standard is still in NP (new proposal) stage; your Cursor sessions calling Claude API today remain completely unaffected. During this window, token gateways like opcx.ai don't need to adjust their product roadmaps either.

Medium-term (12-24 months)—the real watershed. Three things will happen.

First, data center siting will reshuffle. The past decade followed "build where the customers are": Northern Virginia, Phoenix, Northern California. The next wave will be "build where you can plug into grid dispatch": the Middle East (Saudi NEOM, UAE G42), Northern Europe (Iceland geothermal, Norway hydropower), and China's western provinces (Ningxia, Inner Mongolia's mega wind-solar bases) may form new "compute-power hubs." Legacy IDC clusters will be marginalized, or forced to retrofit energy storage + direct-supply agreements.

Second, token economics may bifurcate. Arbitrage window: "sovereign compute" (China/EU/Middle East self-built standards) vs. "market-based compute" (US-led hyperscalers). The same model, different jurisdictions, may see 30-50% price differences—not because the model is expensive, but because the bundled cost of electricity + standards + compliance differs. For model router products, this marks the inflection point where routing dimensions evolve from "model + provider" to "model + provider + jurisdiction."

Third, capital flows. Liquid cooling, energy storage, VPP software, smart grid dispatch—companies in this chain will receive an additional valuation logic: not just "AI concept stocks," but "AI infrastructure essentials." I may take a closer look at the repricing of companies like Stem, Fluence, and Sungrow in the AI era.

05 Counterarguments / Risks

I may be overestimating the substantive impact. Three counterarguments—I rebut each in turn.

First, IEC standards typically take 3-5 years from NP (new proposal) to final IS (international standard) publication, plus another 3-5 years for large-scale commercial deployment. Within that window, what actually determines reference architecture is TSMC's process nodes, NVIDIA's GPU roadmap, and AWS Trainium / Google TPU custom silicon. IEC's position is far weaker than 3GPP's—3GPP is directly tied to mobile communications spectrum (a national resource that must be followed); IEC faces an open market with diverse stakeholders, and its agenda-setting power is much less binding. The "must follow" attribute of standards is far lower in the electricity domain than in telecommunications.

Second, the very term "computing-power coordination" has marketing undertones. China is constrained by export controls in AI chips and GPU clusters, so pivoting to "electricity + standards" soft narratives is a rational national strategy. But that doesn't mean the technical content is solid—it might mean the technical shell is moving ahead of thin content. I can't read this news without carrying that suspicion.

Third, this news has only appeared in domestic Chinese media (CCTV, 36Kr) so far; there's almost no international reaction. Either Chinese media is amplifying it, or international peers haven't caught on yet. I lean toward the former—genuinely strategically valuable standards proposals don't surface in only domestic sources.

Where I might be misjudging: if Google or Microsoft one day proactively announces joining this IEC working group, I've underestimated; if 6 months from now this standard is still being drafted internally by Chinese experts with no substantive international P-member participation, I've overestimated. I'll be watching this window.