MIT QMIT Quantum Initiative: IBM and Google Plans

MIT's QMIT quantum initiative pairs IBM and Google roadmaps, with fault-tolerant targets and error-correction evidence explained. See what is real today.

What QMIT is and how it fits MIT's quantum history

QMIT is an MIT-wide quantum science and engineering initiative that spans the institute's schools, the MIT Schwarzman College of Computing, and MIT Lincoln Laboratory, according to the launch event transcript supplied for this article. Its founding framing ties the initiative to decades of MIT quantum work rather than a single discovery.

The clearest historical reference point in that framing is the summer of 1981, when MIT hosted the first Physics of Computation conference. The transcript describes it as roughly 50 physics and computing researchers exploring whether quantum mechanics could be used for computation, and calls it the unofficial kickoff for quantum applications.

The initiative is built around problem-driven missions, with shared intellectual and eventually shared physical space. MIT faculty named in the transcript as champions include Ian Waites, Anantha Chandrakasan, and Nergis Mavalvala, alongside the QMIT vision presented by Donna. MIT has not published a full list of QMIT programs in the material available here, so specific mission timelines remain unverified.

IBM's fault-tolerant quantum computing roadmap

IBM's stated goal, presented at the QMIT launch, is a fault-tolerant quantum computer by the end of this decade, with a 2029 target of 200 logical qubits running 100 million operations on a system called IBM Starlink. That target is a plan, not a delivered machine.

The architecture described relies on LDPC codes rather than the surface code IBM investigated until about 2019. IBM reported a memory experiment in 2024, a full set of logical operations and an architecture in 2025, and a real-time belief-propagation decoder for the more complex LDPC decoding problem. Each of these is a step in an error-correction program, not proof that a complete fault-tolerant computer exists.

The hardware argument rests on packaging rather than physics alone. The transcript describes qubit connections over a few millimeters up to 14 millimeters with no measurable degradation of quantum gates, multi-level wiring through a low-loss dielectric, qubit reset at roughly one part in 3 x 10^-5 in about 10 nanoseconds, and a design that separates qubits, readout, and I/O onto different wafers.

IBM's stated system-level direction is modular cryogenic design, with about 3,000 qubits per cryogenic environment as a goal, superconducting connections replacing coax cables, cold CMOS control that IBM says uses 50 times less power, and decoding on AMD FPGAs in a few nanoseconds. Those figures are vendor-reported engineering targets and demonstrations from the talk, not independent measurements.

Google Quantum AI's roadmap and Willow error-correction results

Google Quantum AI, the quantum computing group founded and led by Hartmut Neven, is following a published six-step roadmap that ends at roughly one million physical qubits and, in the group's estimate, about a thousand or a few thousand well-protected logical qubits. Neven said at the QMIT event that the team is around the halfway point of that roadmap.

The 2024 Willow chip result is the centerpiece of the hardware story. Neven described the earlier 2023 error-correction demonstration as disappointing because the logical error rate dropped only about 4 percent, then said the 2024 experiment reduced the logical error rate by a factor of about 2 for each increase in code distance while also improving per physical qubit.

A separate single-dimension experiment reported a logical error rate down to 10^-10, which Neven framed as encouragement that the two-dimensional case should follow. Google reported roughly 20 microseconds of T1 coherence in 2022 and said it now routinely exceeds 300 microseconds, with logical error rates at distance-5 code approaching 10^-4, and an error suppression factor lambda moving from just above 1 in 2023 toward 4. All of these are first-party Google results, and Neven explicitly noted that Google emphasizes system-level experiments because other groups hold records for qubit count or coherence separately.

Quantum Echoes and out-of-time-order correlators

Google's October 2025 Nature cover result, described at QMIT as quantum echoes, is the group's claim to a computation that is both classically expensive and useful. Neven said the team invested 10 person-years of red-teaming to test whether the best classical algorithms could reproduce it, and that the protocol includes an experimental hardness check by inserting dephasing gates mid-circuit.

The mechanism is a forward unitary, an inverse, and a butterfly operator applied mid-circuit, with the effect measured on a distant probe qubit. The single-pass version is an out-of-time-order correlator; the team's repeated version, called OTOC squared, showed greater sensitivity to dephasing and, Neven said, more opportunity for long-range interference. Nobel laureate Michel Devoret argued in review that the weakness Nature editors flagged was a feature, comparing it to multi-pass interferometry.

The companion application is quantum machine learning of molecular structure from nuclear magnetic resonance data. Neven described it as a new NMR protocol based on OTOCs. The claim that this is classically hard rests on Google's own red-teaming, so it is vendor-reported evidence, not an independent hardness proof.

Where IBM and Google differ on quantum strategy

IBM and Google are pursuing the same end goal with different architecture bets and different public milestones. IBM's plan centers on efficient LDPC codes, modular cryogenic hardware, and a 2029 fault-tolerant system target; Google's roadmap counts six milestones and tracks code-distance scaling on a path the group says is roughly half complete.

Dimension	IBM	Google Quantum AI
Stated end goal	Fault-tolerant system by 2029, 200 logical qubits, 100 million operations	Six-step roadmap to about 1 million physical qubits and 1,000+ logical qubits
Error-correction code emphasis	LDPC codes with real-time belief-propagation decoding	Surface-code distance scaling with lambda error suppression
Recent reported results	2024 logical memory; 2025 full operation set and decoder	2024 Willow distance-scaling result; 1D logical error rate to 10^-10
Hardware direction	Modular cryostats, cold CMOS control, multi-wafer packaging	Larger modular cryostats, cryogenic control via Atlantic Quantum
Application focus	Quantum as a subroutine in classical chemistry workflows	Quantum echoes, NMR-based molecular structure, quantum AI

Both groups acknowledge that wiring and cryogenic engineering are the practical bottlenecks, not qubit physics alone. IBM describes separating qubit, readout, and I/O wafers; Google describes the coupler-per-qubit wiring as an "engineering abomination" and points to its acquisition of MIT spinout Atlantic Quantum for cryogenic control electronics.

Why error correction is still an open research problem

Neither roadmap treats error correction as solved. IBM's Jay Gambetta told the QMIT audience that error correction remains one of the most important areas for students, and argued that concatenated inner and outer codes OpenAI large design space, including algebraic codes layered on LDPC codes and post-selected error detection.

Post-selection is the concrete near-term example. IBM described post-selected error detection codes producing 140 logical qubits and GHZ states with twice the improvement of the decoded version, plus spacetime post-selected codes showing larger gains on Clifford-driven graph states. These are research demonstrations of a technique, not a replacement for full fault tolerance.

The implication for anyone tracking quantum computing is that logical qubit counts and logical error rates are not directly comparable across groups unless the code, distance, decoder, and measurement method match. Comparing IBM's LDPC architecture figures with Google's surface-code scaling numbers would mix unlike configurations.

Security, quantum networking, and the post-quantum timeline

Google's cryptanalysis estimate is the most decision-relevant number in the talk for security planning. Neven said a 2025 paper by team member Greg Gibney showed that 2048-bit RSA integers using Peter Shor's algorithm need only about one million physical qubits, which maps onto Google's milestone six, and that elliptic-curve cryptography, including Bitcoin's 256-bit ECC, is weaker.

Neven's stated conclusion was that organizations should move to post-quantum cryptography because the timeline is moving inward, not because a cryptographically relevant quantum computer exists today. A one-million-qubit requirement is a resource estimate within Google's own architecture and error-correction assumptions, so it does not by itself prove how soon such a machine will be built.

IBM's networking work runs in a different direction. The company has described a quantum networking unit that connects its best processors to a tunable device, an L coupler for joining multiple modules, a project with Fermi on microwave-to-waveguide coupling at 4 Kelvin, and projects with Caltech and Q-NEXT at Argonne on microwave-to-optical transduction. These are active research programs, not deployed quantum networks.

Quantum AI, quantum neuroscience, and the limits of the claims

Neven's closing argument was that quantum resources could improve AI by generating training data cheaply. He cited AlphaFold as an example that depended on the Protein Data Bank, which he said started in 1971, and claimed that producing one second of fusion-relevant plasma data with current technology would cost about one trillion dollars. That figure is a motivating estimate, not a measured cost from a published study.

The quantum neuroscience proposal goes further. Neven said he and external colleagues formulated a conjecture that quantum processes are needed to explain conscious experience, and described an experimental program linking human brain to quantum computer. The concrete early step is exposing biological systems, including cortical brain organoids and Drosophila fruit flies, to isotopically varied xenon, an anesthetic gas, and observing whether the flies recover differently by isotope.

Fruit flies falling at about three atmospheres of pressure and recovering within seconds when the plunger is released were described as an isotope-sensitivity demonstration. That is an early experimental result in an exploratory program; it does not establish a theory of consciousness or a practical quantum-AI advantage, and the article treats both as open research directions rather than findings.

FAQ

• What is QMIT at MIT? QMIT is an MIT-wide quantum science and engineering initiative spanning the institute's schools, the MIT Schwarzman College of Computing, and MIT Lincoln Laboratory. It is organized around problem-driven missions and was introduced with talks from IBM's Jay Gambetta and Google Quantum AI's Hartmut Neven. MIT has not published a detailed public program list in the material reviewed here.

• Has IBM built a fault-tolerant quantum computer yet? No. IBM's target is a fault-tolerant system by 2029 with 200 logical qubits and 100 million operations, called IBM Starlink. The company has reported component results including a 2024 logical memory experiment, a 2025 architecture and full operation set, and a real-time LDPC decoder, but the complete machine remains a stated goal.

• How many qubits does Google need to break RSA-2048? Google researcher Greg Gibney reported in 2025 that 2048-bit RSA integers could be factored with Shor's algorithm using about one million physical qubits under Google's architecture assumptions. That is a resource estimate tied to a specific error-correction design, not a date for when such a machine will exist.

• What did Google's Willow chip achieve in 2024? The Willow experiment reduced logical error rate by roughly a factor of two for each increase in code distance while improving per physical qubit performance. Google frames this as milestone two of its six-step roadmap; the next milestone, milestone three, is one well-protected logical qubit at a 1-in-a-million error rate.

• What are quantum echoes in Google's October 2025 result? Quantum echoes is Google's name for a protocol that runs a unitary forward, inverts it, and applies a butterfly operator mid-circuit, measuring the effect on a probe qubit. The measurement is an out-of-time-order correlator, and a repeated version, OTOC squared, showed stronger sensitivity to dephasing and supported a companion NMR application.

Source video