At a glance
New South Wales moved its large-load electricity bill through the lower house on 15 September 2026, with flexibility among the tests a data centre would be scored on.
Jensen Huang says the grid sits at about 60% of peak most of the time, and his trade is availability for speed and price.
NVIDIA was selling software to curtail an AI factory on a grid signal by June 2026, and says Silicon Valley Power has sent one more than 200 signals that all worked.
Firmus is the only Australian operator to have published a curtailment commitment: up to 220 hours a year under a 600MW South Australian supply agreement.
Australian networks sell a single firm connection to large loads, priced the same whether the buyer would accept interruption or not.
NVIDIA chief executive Jensen Huang spent three minutes of a March podcast interview on the electricity grid, and the passage has circulated ever since. He made three asks, addressed to three different parties, that together would rewrite who pays for a data centre being available all the time. He has pressed the case since, and NVIDIA now has a working result behind it.
It matters in Australia this month. New South Wales moved its large-load electricity bill through the Legislative Assembly on 15 September 2026 and its consultation on grid access closed the day before, both proposing to score data centres partly on how flexible they will be.
Huang asked data centres to give up guaranteed power in return for a cheaper and faster connection. Australia is close to requiring the first half of that trade, Firmus has taken it voluntarily, and no network here sells the second half.
An ask each for tenants, operators and utilities
Speaking to the podcaster Lex Fridman on 23 March 2026, Huang put the case for using power the grid already has. The grid is "designed for the worst case condition with some margin", he said, and "most of the time we're nowhere near the worst case condition and we're probably running around, call it 60% of peak". His interest in it was direct: "I just wanna use their excess. It's just sitting there."
The tenant's contract. "The end customer puts requirements on the data centers that they can never not be available," Huang said, and he put the fix at board level: "I bet the CEO doesn't know this. I'm gonna talk to all the CEOs."
The operator's design. "We have to build data centers that gracefully degrade," so that when a utility needs to "back you down to about 80%", the operator moves the workload elsewhere, slows the computing rate and keeps the data intact.
The utility's price list. Rather than telling a customer "it's gonna take me five years to increase my grid capability", Huang wants utilities to offer "more segments of power delivery promises", so a buyer accepting a weaker guarantee connects sooner and for less.
Ten weeks later, at his GTC Taipei keynote on 1 June 2026, Huang put the second ask on stage as a product: NVIDIA's DSX Flex "reads real-time grid signals and dynamically adjusts factory power when the grid needs relief", he said, and AI factories running it "make the grid stronger".
Silicon Valley Power sent a signal to an NVIDIA AI factory on a hot evening in Santa Clara in August 2026, and its draw fell from 4MW to 3MW. NVIDIA published that account on 15 September 2026, saying the utility has sent more than 200 demand signals since and that each one worked. Five days before that account, NVIDIA had named eight Australian partners for up to 2GW by 2027, in a statement carrying no flexibility, grid signal or demand response language.
A hyperscale tenant's uptime clause ends in a network build
A hyperscale tenant writes constant availability into its cloud contract, the cloud provider writes it into the colocation lease, the operator buys a firm connection, and the network answers the only way it can, by building. Networks then recover that build across every connected customer. Huang's proposal interrupts the chain at the top, so the price of a guarantee reaches the party that wants it.
Duke University's Nicholas Institute found in February 2025 that the United States could absorb about 76GW of new load if that load accepted curtailment 0.25% of the time, about 22 hours a year, rising to about 126GW at 1%. The researchers did not model transmission constraints, so the figure sizes system headroom and says nothing about any single line.
It works, but only at small scale so far. The Santa Clara result was a 4MW site, not a campus, and NVIDIA is the only source for it. Whether an Australian operator can do the same depends on one thing: whether it owns the computing jobs it would have to stop.
The first two asks are already moving offshore. Texas legislated a curtailment obligation in 2025 for loads of at least 75MW, Google had 1GW of demand response contracted with five American utilities by March 2026, and in June 2026 the Federal Energy Regulatory Commission ordered six grid operators to justify or reform how they connect large loads. The discounted connection remains the open item, in America as here.
Firmus has contracted up to 220 flexible hours a year in South Australia
Firmus, the Australian AI compute provider and one of NVIDIA's eight partners, signed a 600MW, 12-year South Australian energy supply agreement with commodity trader Gunvor in June 2026. Inside it is a demand response commitment: Firmus will reduce consumption for up to 220 hours each year when wholesale prices exceed agreed thresholds. The agreement is tied to 1.2GW of new renewable generation and 1.5GWh of storage by 2032, and Firmus has contracted the output of the 200MW Koolunga battery to firm it.
Gunvor's contract fires on the wholesale price, and no network operator is in the loop. Firmus steps back when the market is expensive, which is usually when the system is tight, so the grid gets the same relief out of a commercial supply agreement.
Firmus sells compute from sites it runs itself, so it owns the workload it would turn down. Data Centres Australia chief executive Belinda Dennett said on 15 July 2026 that different types of data centre have different limitations, "particularly with regard to location and ability to curtail their workloads at peak times". A colocation operator leases halls under availability terms agreed with its tenants, and an AI data centre serving live inference has less room to slow down than one running a training cluster. A curtailment clause can still go into a connection agreement.
Australian networks sell one connection product to large loads
No Australian network publishes a tariff or a connection product that prices interruptibility. Flexibility is paid for after connection instead, through mechanisms built for someone else: registered capacity in the wholesale demand response mechanism reached 343MW in June 2026, held between two registered service providers, and AEMO's reserve contracts buy blocks of at least 10MW for reliability rather than for access.
AEMO wrote on 1 June 2026 that data centres are "likely to operate as relatively inflexible loads, prioritising uptime and reliability". Its own measurements show the room inside connections already granted: a connected fleet of 2.2GW drew an average 27% of that capacity in the first quarter of 2026, against a queue that reached 17 projects and 9GW.
Powerlink has run dynamic line ratings since 2017
Transmission lines, conductors, sensors and line ratings appear nowhere in Huang's passage, and Australian networks have chased the same headroom through the wires for years. Powerlink has reported dynamic line ratings on its Gladstone 275kV circuits since its 2017 planning report, with the ratings updated inside AEMO's dispatch engine, and TasNetworks told the regulator in 2018 it had implemented dynamic rating on most of its transmission corridor. A Transgrid trial with Infravision, backed by A$732,000 from ARENA, measured 34% more capacity on one 330kV span than the probabilistic static rating allowed.
A static rating already assumes hot, still air, so the extra capacity appears on cool, windy days and shrinks on the calm afternoon that sets the system peak. Britain's National Grid put the average at 8% per circuit across 39 circuits in April 2026. A data centre draws a flat load around the clock and is assessed against the hour the network is tightest, which is the hour the ratings add the least capacity.
NSW proposes to rank grid access partly on flexibility
The three parts of the trade | Australia, 16 September 2026 |
|---|---|
The tenant's contract | NSW asks the proponent, not the tenant, to show it can cut 25% of forecast average load for up to two hours; Commonwealth legislation is to be introduced in early 2027 |
The operator's design | Firmus has contracted up to 220 hours a year; no other Australian operator has published a commitment |
The utility's price list | No Australian network sells an interruptible connection; the AEMC proposed a voluntary framework on 5 August 2026 |
Source: Certified Strategic Editorial, from the Lex Fridman transcript, the GTC Taipei keynote, Firmus, the NSW Data Centre Guidelines, the NSW connection consultation, National Cabinet's 26 August 2026 statement and AEMC advice to energy ministers, 16 September 2026.
The AEMC gave energy ministers its advice on data centres on 5 August 2026. One of its four recommendations would let data centres enter flexible demand agreements with networks voluntarily, with AEMO's direction powers available in market or system stress, and a second would streamline connection for a data centre sitting behind a battery or other firming asset. It is the closest any Australian body has come to pricing the connection itself, on an indicative 24 to 36 months once ministers trigger it.
The state instruments move sooner. The NSW data centre guidelines published in August 2026 ask a proponent to show it can cut grid-supplied demand by 25% of forecast average load for up to two hours, using load shifting, storage or nearby generation, with diesel back-up ruled out. Storage is the route that does not depend on which jobs a tenant is running: 25% of a 300MW campus for two hours is about 150MWh of battery. The consultation that closed on 14 September 2026 proposed an entry bond of A$30,000/MW of requested capacity, which is A$9 million on a 300MW campus, with access rights contested on criteria including demand flexibility and willingness to pay. The large-load bill that passed the Legislative Assembly on 15 September 2026 would let the minister declare access schemes over connections of 5MW and above. Transgrid, reporting more than 20GW of enquiries against 1.5GW of signed agreements in Greater Sydney on 10 September 2026, says connection agreements can incorporate provisions to curtail loads in system stress.
An operator choosing where to build among the data centres in Australia meets flexibility as a condition of consent and a tiebreaker for queue position well before it arrives as a discount. Building the capability early buys rank in a 9GW queue, which is the currency the current instruments actually pay in.
What to watch
Whether any network publishes a tariff for an interruptible connection, and what discount it carries.
Whether DSX Flex appears in any of the eight Australian partner deployments, and whether a network signs up to send the signals.
The NSW large-load bill in the Legislative Council, and the regulations setting the megawatt threshold, expected in the fourth quarter of 2026.
The AEMC's final determination on access standards for large inverter-based loads, due 29 October 2026.
Commonwealth legislation on data centre energy, water and land-use standards, which National Cabinet agreed on 26 August 2026 to introduce in early 2027.