IBM Says Two Nighthawk r3 Chips Will Reach 1,040 Qubits in 2027

IBM said Wednesday that two Nighthawk r3 processors, each with 520 programmable qubits, will be linked into a single 1,040-qubit system in 2027. The chip is planned, not built: what shipped is Nighthawk r2, a 120-qubit part that IBM says now runs 100,000 circuits a second.

The figures come from IBM's Nov. 11 announcement from Yorktown Heights, New York. Jay Gambetta, director of IBM Research and an IBM Fellow, framed the release around two measures: "Useful quantum computing hinges on whether results can be delivered with accuracy and speed, both are fundamental."

What the Nighthawk r3 plan specifies

IBM describes r3 as designed to run 10,000 quantum operations across 520 programmable qubits, with 10 chip-to-chip links it calls l-couplers. Square-lattice connectivity carries over from r2. The headline of the release says "more than 1,000 qubits"; the body gives 1,040.

The number that matters here is not the qubit count but the link between chips. In the release, four l-couplers joined Nighthawk r1 processors and reached 99.3 percent quantum state transfer fidelity against a 99.9 percent target. That is 0.6 percentage points short of the goal, and r3 needs 10 of those links, not four.

	Nighthawk r2	Nighthawk r3 (planned)
Programmable qubits	120	520
Operations per circuit	7,500+	10,000
Chip-to-chip links	none	10 l-couplers
Status	available now	2027

Speedups are measured against Heron, and on different things

IBM's release reports speedups of up to 19x over earlier IBM processors, all against Heron, its previous processor line. Clients and partners including E.ON, Algorithmiq and Q-CTRL have run 34 billion circuits on r2 since its September introduction, according to IBM.

1. E.ON, quantum machine learning for grid optimization15 x
2. IBM and Oak Ridge, magnetic materials simulation12 x
3. Algorithmiq, circuit runtime8 x

Source: IBM newsroom release, Nov. 11, 2026 (vendor-supplied)

All three are IBM or partner figures, and the release does not say who audited them. They also use a different denominator from the 25x claim in an Aug. 31 IBM blog post, which compares raw circuits per second: more than 100,000 on r2 against roughly 4,000 on Heron. Circuit throughput and time to a finished result are separate quantities; the blog's own early advantage-candidate tests showed up to 10x faster runtimes with no loss in accuracy.

The mechanism is reset. IBM's blog says the effective qubit lifetime drops from a median of about 200 microseconds to roughly 25 nanoseconds when the reset coupler is active, which lets the idle gap between circuit runs shrink to as little as 1 microsecond.

The r2 die has 120 programmable qubits, 218 tunable couplers and 120 separate reset elements, one per qubit, which IBM's blog totals at 458 physical quantum elements. The 7,500-operation result used a method called Probabilistic Error Amplification for observable estimation. NAND says the figure counts two-qubit gates and rises from about 5,000 on the first Nighthawk revision, and that IBM had set 7,500 as its end-of-2026 target, so it arrived about a quarter early.

Two published numbers that disagree

IBM's blog says r2 cuts initialization error about 25-fold across the device. NAND Research, whose principal analyst Steve McDowell wrote a Sept. 2 review, reports an 11-fold reduction and attributes it to IBM. Neither document explains the gap, and the two may use different baselines.

Access is also described two ways. IBM's blog says r2 is on the IBM Quantum Platform as the system ibm_phoenix and that new users can create a free account. NAND says r2 is initially limited to Premium and Flex customers. IBM's November release says only that r2 is available now.

What IBM moved and what an analyst expected

NAND's September note said a long-range coupler was expected to push past 1,000 qubits in 2028. IBM's November headline puts a 1,040-qubit system in 2027. The two documents may be describing different things, since NAND was writing before the r3 details existed, but the date moved a year in IBM's favor on paper.

NAND also lists where IBM is weakest. It says Quantinuum leads on two-qubit fidelity and all-to-all connectivity, and that Google has more peer-reviewed error-correction results. NAND's reading is that IBM's lead rests on fleet scale, throughput and managed services, not per-gate fidelity.

IBM is pairing the hardware with classical machines. A co-design platform with Oak Ridge National Laboratory and AMD places Nighthawk beside ORNL's Umbra system, which uses AMD Instinct MI355X GPUs and 5th Gen EPYC processors. The classical half of that pairing is covered in our reporting on AMD's EPYC 9996 cache latency and Nvidia's Vera CPU memory bandwidth.

What would change this read

The l-coupler transfer fidelity is the variable to watch. If the 99.3 percent figure reaches the 99.9 percent target before r3 silicon is fabricated, the 2027 date is credible. If it stalls, a 1,040-qubit system would be two 520-qubit chips that behave as two machines. IBM's release gives no date for r3 samples, and no price for either processor.