The Two-Second Rule: How the Grid Became an Interface Spec

October 21, 2026·11 min read

The most useful sentence NVIDIA published this month is not about GPUs. It is a summary of a specification, from the OCP summit post: large AI infrastructure should stay connected during a grid disturbance, respond automatically, and restore to near-full load in less than two seconds.

That paraphrase is generous, and the document is stricter. On 17 September 2026, OCP published Interface Specification Between Hyperscalers and Utility Grid for Disturbance Performance, version 1.0.0, as a Base Specification. The authors are Google, Meta, Microsoft and NVIDIA. The number in it that I keep coming back to is two seconds, and it is a requirement rather than a target: once voltage at the point of interconnection is back inside the normal band, the computing load has to be back at 90% of its pre-fault active power within two seconds.

I have written about the layers above this one, the runtime that went open source and the rack that became a category, two weeks ago. I am not going to re-report that summit. This post is about the part of the AI factory that is now specified by a document addressed to a utility, and about what changes when a standard starts measuring behavior over time instead of dimensions on a drawing.

What the document asks for

The table is short enough to state plainly. Between 0.90 and 1.10 per unit of nominal voltage the facility runs normally at constant power, and between 1.10 and 1.20 it holds constant power for one second. Between 0.80 and 0.90 it has to ride through for two seconds and may reduce power in proportion to voltage. Between 0.50 and 0.80 the ride-through is half a second. Below 0.50 it may reduce power to any level, including zero, and after 0.15 to 0.25 seconds it may transfer to backup outright. The catch sits in the fourth column, which applies to every case where power was allowed to drop: from the moment voltage is back inside 0.90 to 1.10, the load has two seconds to get back to 90% of where it was.

Two details in that table matter. It applies to computational load only, which the authors say is typically 70 to 90% of a data center’s total load, and they adopt MISO’s convention for measuring it. Cooling and auxiliaries are out of scope. The recovery commitments also closely mirror ERCOT’s proposal in NOGRR 282, with clarifications the authors flag. The consortium published the number a large grid operator had already proposed, so that everyone building for ERCOT and everyone building for MISO builds for one curve.

The document also sets an Active Power Ramp Rate Limit, and it lists the rest of what it does not yet cover: frequency ride-through, rate of change of frequency, phase jump, counter-based protection, finer ramp-rate limits, intra-minute variation and forced oscillations. That list is the next version, and it is harder than this one.

Why a stopwatch at the point of interconnection

Every spec in this series so far told you how to build something. Rack height of 100 inches, a maximum rack weight of 8,000 lb, row density in kW per linear foot, so that a 2027 facility from AMD and a 2027 facility from Meta can be drawn on the same grid. Those are dimensions in ODCS 0.7.0, and dimensions are easy to check at delivery.

This document describes how a building has to behave during someone else’s fault, and the behavior is measured with a clock. That is a different class of requirement. It cannot be verified from a bill of materials. It has to be tested at the point of interconnection, with the utility watching, and it can be lost by a firmware setting in a UPS.

The motivation is in the introduction. In August 2024, an event in North America dropped 1,500 MW of load off the bulk electrical system simultaneously. On 22 July 2026 there was a larger one in the Eastern Interconnection. When a block of load that size trips, frequency and voltage rise, and the system has to absorb the imbalance. A load that comes back slowly is a second event, which is why the spec’s own framing is that post-fault recovery prevents a “secondary” disturbance from sustained loss of load.

Section 4.4.4.1 is the part I find genuinely good, because it is honest about a limit. It says true fault ride-through, staying connected and consuming power at zero voltage, is not achievable with current OEM UPS and PSU architectures without adding capacitance to the compute equipment. Then it says why they did not ask for it: the extra capacitance risks breaching the current injection limits that grid operators such as ERCOT and EirGrid require, so solving for ride-through would create a compliance problem on the way back up. They specified the recovery instead, and wrote down what they had left out. I trust the rest of the document more for that.

The grandfathering clause tells you who wrote it

Section 4.5 is the part to read if you care about adoption. It proposes a separate effective date for the performance requirement, exempts facilities that are already operating or have executed interconnection agreements as of that date for the life of the facility, applies the requirement to growth above an agreement’s contracted volume only for the incremental volume, and offers a best-effort voluntary path for legacy sites. New projects entering load study after the effective date are expected to comply at commissioning.

That is a sensible transition. It is also four companies writing the standard their own buildings will be measured against while carving out every building they have already finished. I do not think that is sinister, and I do think it is worth noticing, because it tells you what the document is for. It is addressed inward, to procurement and to the supply chain, and outward to utilities as an offer. Utilities are not signatories. No regulator has adopted it. The actual constraint on any given project remains whatever that project’s interconnection agreement says, which is why FERC matters more here than OCP does.

FERC wrote the same rule with teeth

On 18 June 2026, FERC issued six tailored show-cause orders under section 206 of the Federal Power Act, one each to PJM, MISO, SPP, CAISO, ISO-NE and NYISO, docket RM26-4, items E-7 through E-12. Each grid operator has 60 days to justify its large-load tariff or file changes, and 30 days to file a report on how it will have enough generation for existing and new large loads. Commission staff read more than 3,500 pages of comment to get there. The five categories include accommodating co-location and behind-the-meter generation, and providing new transmission services for flexible large loads. FERC’s own framing is “speed to power.”

Put the two documents next to each other and they make the same move. A large load is a problem in proportion to how its behavior at the interconnect is defined, so the fix is to define the behavior rather than to refuse the load. One version of that is a base specification with a two-second clause and a grandfathering section. The other is a docket that gives six grid operators 60 days.

For anyone who is not one of the four authors, the fourth category is the one worth following: FERC asks the six operators to consider new transmission services for flexible large loads. That is a question on a 60-day clock, not a rule. Nothing in the docket promises a dialable load a weaker interconnection or a shorter queue. What it does is put flexibility into a tariff proceeding, which is further than four companies writing a Base Specification can take it on their own.

The alternative is to make the load dialable

NVIDIA’s answer is software that treats the facility’s power budget as a pool rather than a reservation per node. DSX MaxLPS allocates that budget dynamically across a managed group. In an evaluation with Nscale in Iceland, on GB300 NVL72 systems running Kimi K2.5 in FP4, the baseline ran 140 GPUs and the MaxLPS configuration ran 192 under the same 264.4 kW provisioned budget. Normalized aggregate throughput rose 49.2%, and efficiency went from 4.10 to 6.12 tokens per second per watt. Median and P75 latency stayed within 5%. P99 time to first token rose 17%.

Today’s GTC Berlin release states the same work differently: the same tokens with up to 30% less power, a Lambda proof of concept at 24% more cluster throughput and 19% more GPU capacity in a fixed budget, and a DSX Flex trial with Emerald AI, EPRI, Nebius, the UK National Grid and NVIDIA in which a Blackwell Ultra cluster cut power draw by up to 40% in response to simulated grid events while critical workloads kept running.

Those claims have different denominators, and NVIDIA’s release runs them together. Capacity inside a fixed budget and the same output at lower power are not the same promise, and the 40% figure came from simulated events, not a real grid emergency. The 17% P99 regression is the number I would put in the first paragraph if I worked at NVIDIA, because for an agent the tail is what the user feels. One stalled decode stalls a tool call, and a stalled tool call stalls the chain behind it.

Still, MaxLPS and the OCP spec are the same idea approached from opposite ends. The spec says behave this way at the boundary. MaxLPS says make the load shapeable enough that behavior becomes a software parameter. I expect the second to be what operators end up buying, because it also saves money in the absence of any disturbance at all.

The agentic half of the argument is the weakest link

The reason this is being written about as an agentic AI story is a set of numbers from the OCP keynotes. AMD’s Lisa Su called 2026 the year of inference, put inference at about 60% of the AI compute market, claimed a roughly 14-fold rise in agentic traffic during the year, and raised her data-center CPU forecast above $200 billion by 2030. Arm’s Mohamed Awad argued agents are virtual users who need three classes of CPU. NVIDIA’s Vera CPU rack, 256 liquid-cooled CPUs, is aimed at tool calling and sandboxed execution.

Those are company projections with no published methodology, and they do not measure what the grid spec cares about. A 1,500 MW block of load is 1,500 MW whether it runs a chatbot or a fleet of agents. Nothing I read measures agentic workloads’ share of any facility’s load, and the spec’s motivation is an electrical event, not a computational one.

The narrower thing I would guess is that agents put more of the work on CPUs and turn one request into a chain of small ones. That is a hypothesis about workload shape, and I have no operator measurement behind it. It also does not tell you what happens to a facility’s load profile over a day. Load could come out flatter, spikier or simply larger, and the two-second rule is written for the size of the block either way.

I would trade all four keynote projections for one operator publishing a measured load profile from a hall running agents at scale. Nobody I could find has published one.

What would make this a contract instead of a press cycle

The first thing I will be watching is whether a utility, a colocation provider or an equipment vendor outside the four authors adopts either spec in a document I can read. The second is whether the separate effective date that Section 4.5 defers ever gets set, since a spec with a deferred date is a spec with an asterisk. The third is whether the frequency ride-through and ramp limits on the future-work list turn up in a tariff, because that is the step where a consortium document stops being a suggestion.

My guess is that the FERC docket gets there first. The spec has the better engineering and the clearer number, but the docket has the 60-day clock, and six grid operators are already rewriting the rules that decide which buildings get connected. If their tariffs land on the same curve the spec picked, the two-second rule becomes an acceptance criterion and every facility built after it inherits the number. That is a large if. The spec cites MISO’s convention and ERCOT’s NOGRR 282 to show that grid operators are already writing load requirements, which is not the same as the six of them settling on one curve.

I have watched that happen in compliance for a year. A voluntary standard becomes a gate the moment a contract references it, and interconnection agreements are next in line.