AIX Global Innovations Claims Fault-Tolerant Quantum Computing on Rented Hardware. Their Own Paper Says Otherwise.

June 19, 2026 — On June 15, 2026, AIX Global Innovations, a Los Angeles-based company, issued a press release on BusinessWire announcing what it described as “a major fault-tolerant quantum computing breakthrough quietly achieved in April 2026 on rented IBM Quantum hardware.” The announcement was timed to coincide with The Economist’s Commercialising Quantum Global 2026 conference in London, held June 16–17.

The company claimed its proprietary software, Seed IQ, had achieved fault-tolerant quantum computing on standard IBM Heron r2 and r3 processors accessed via IBM’s public cloud subscription service. The press release stated that AIX had cleared “all four strict fault-tolerant quantum computing requirements simultaneously” at “approximately one physical qubit per logical qubit.” Across an eight-week campaign on five IBM Heron processors, they reported 22,500 primitive-composition circuits with “zero detected logical errors” and twenty-two chemistry runs all landing inside chemical accuracy.

“The question is no longer how many more qubits are needed before FTQC becomes possible,” said Denise Holt, Founder and CEO of AIX Global Innovations. “Seed IQ makes it possible today through governed execution rather than massive hardware scale.”

A self-published 103-page report on Zenodo accompanied the announcement. A companion Jupyter notebook was uploaded to a separate Zenodo record on June 14.

The Coverage

Quantum Zeitgeist published the claims with minimal independent analysis, largely restating the press release. Quantum Computing Report covered the announcement using hedging language (“AIX reports,” “the company utilized”) but included no independent expert commentary. The press release was also picked up by Yahoo Finance, Morningstar, StreetInsider, and AP News via BusinessWire’s distribution network.

As of June 20, I found no coverage in Nature, Science, IEEE Spectrum, MIT Technology Review, Physics Today, or Ars Technica, and no public technical response from IBM, whose hardware was used for the experiments. No quantum computing research group has publicly acknowledged or commented on the claims.

My Analysis

Who Is Making This Claim

Denise Holt is described on LinkedIn as “Founder & CEO, AIX Global Innovations” and “Voting Member — IEEE Spatial Web Protocol.” Her documented professional history centers on technology evangelism and content creation. From roughly 2022 through 2025, she positioned herself as the leading educator on Active Inference AI (a neuroscience-derived framework based on Karl Friston’s Free Energy Principle), producing blog posts, podcasts, and paid certification programs about VERSES AI. Her Substack was originally called “Spatial Web AI by Denise Holt.” After VERSES AI experienced significant difficulties, Holt founded AIX Global Innovations and published a post explicitly distancing the new company from VERSES.

Denis Ovseyenko is described as having “decades of experience designing and deploying mission-critical systems for global financial institutions, major technology companies, and government entities” with expertise in “AI, distributed infrastructure, event-driven architectures, and physics-informed system design.”

I found no peer-reviewed quantum computing, quantum error correction, or quantum information publications under either founder’s name. Their prior quantum work consisted of QuTiP-based simulations published on their own blog in January 2026, which Holt herself described as “preliminary and simulated, and real hardware validation remains the goal.” Five months later, the company announced the most significant result in quantum computing history. The AIX advisory board includes a former FCC CIO, an IP lawyer, a cybersecurity CEO, and a cognitive science researcher. The paper itself lists no quantum physicists or QEC researchers among its advisors or contributors.

What Their Own Paper Says

The AIX paper is available on Zenodo. It has not been submitted to a peer-reviewed journal. It is published under “All Rights Reserved” with an explicit prohibition on “redistribution, republication, modification, commercial use, extraction, or reuse of proprietary methods.” I have reviewed it in detail. The claims do not survive contact with the paper’s own methodology.

The d=1 Problem

Consider what the press release claims: “approximately one physical qubit per logical qubit.” It also claims “distance-3 and distance-5 surface-code QEC.” Read together, these suggest fault-tolerant computation at a 1:1 encoding ratio with distance-5 error correction. The paper tells a different story.

The paper reports separate QEC demonstrations at d=3 and d=5 in Stage 1 (April 9). These are standalone validation experiments. They are not the register on which the FTQC composition and chemistry claims rest. All composition work, all chemistry workloads, and all “FTQC primitive” demonstrations run on what the paper calls a “150-qubit governed encoded register held at substrate distance d=1.”

Distance d=1. One.

A quantum error-correcting code of distance d can correct floor((d−1)/2) errors. At d=1, that is floor(0/2) = zero. A distance-1 code can correct exactly zero errors. This is not a property of surface codes specifically. It follows from the quantum Hamming bound and the definition of code distance in quantum information theory. At d=1, there are no redundant qubits carrying the information needed to detect or correct anything. A distance-1 “encoded qubit” is a physical qubit.

AIX acknowledges this and claims it as a feature: “Seed IQ governance acts as an operational substitute for code distance.” This sentence claims that software can substitute for the physical redundancy that makes quantum error correction mathematically possible. It cannot, for the same reason that software cannot recover a file from a stolen hard drive that was never backed up. The backup copies are the mechanism by which errors are detected and corrected. Without them, there is nothing to correct with.

The press release’s juxtaposition of d=3/d=5 QEC with a 1:1 qubit ratio is misleading. The d=3/d=5 results and the 1:1 ratio come from entirely different parts of the campaign. The QEC demonstrations at d=3 used 13 physical qubits; at d=5, the paper reports 41 physical qubits. The 1:1 ratio applies to the d=1 computation register, where no error correction is happening.

Post-Selection Disguised as Fault Tolerance

AIX reports “zero detected logical-error events” across 22,500 circuits (45,000 total executions across replicated calibration windows). The standard metrics for demonstrating fault-tolerant quantum computing are logical error rate per round at each code distance, and a demonstration that logical error rates decrease as code distance increases (the below-threshold criterion). Google reported these metrics for Willow. Quantinuum reported them for their trapped-ion work. AIX reports LER-like figures for its standalone d=3 and d=5 tests, but does not demonstrate below-threshold scaling for the d=1 register on which all FTQC and chemistry claims depend.

Instead, AIX describes a multi-stage filtering pipeline. Circuits are compiled with “calibration-aware synthesis” (selecting the best qubits based on current hardware calibration data, a standard Qiskit technique). Runs pass through “runtime admissibility checks.” A “three-stage admissibility certification” filters results at compilation, execution, and commit time. Measurements undergo “heralded post-selection on the parity stabilizer,” which discards shots that fail a parity check.

AIX claims “Data rejection / post-selection: None” for the H₂ workload, but this refers to per-run rejection only. Per-shot post-selection is happening within each run. The paper reports “typical retention rates on a healthy substrate exceed 97%” for H₂, which is plausible for a 2-qubit circuit on good hardware.

For the more complex molecules, the filtering is more aggressive. The notebook’s own audit data reveals that the H₂O run on IBM Fez had 45 out of 61 commit-stage accepts, with 16 vetoes. That is a 26% rejection rate at the commit stage alone, before accounting for per-shot parity filtering. Other runs show lower veto counts, but the pattern is clear: substantial filtering is happening, and the “zero detected logical errors” metric applies only to the results that survived the filtering pipeline.

When I asked Denis Ovseyenko directly on LinkedIn what percentage of total shots were rejected by the admissibility checks before arriving at the reported results, he did not answer. He blocked me.

That question has a number. They have it. If the number supported their FTQC claim, there would be no reason to withhold it.

The Twelve-Decimal Trick

AIX emphasizes that different IBM Heron chips produce results “bit-identical to twelve decimal places” and presents this as proof that the system works. The phrase appears repeatedly throughout the abstract and paper.

The paper’s own methodology (§13.10, §14) explains why this happens. The H₂O calculation, for example, lives on a two-dimensional trial-state subspace parameterized by two angles (θ, φ). The paper describes a Newton fixed-point projection: take noisy QPU measurements, project them onto this low-dimensional subspace using Newton’s method (a standard classical optimization algorithm), and evaluate the energy at the fitted parameters (θ*, φ*).

Two noisy measurement vectors from different chips, if they land in the same convergence basin of the Newton fit, will produce bit-identical parameters and bit-identical energies. The classical math is deterministic. Different noise in, same fixed point out, same twelve decimals. This demonstrates reliable classical optimization. It demonstrates nothing about quantum error correction.

AIX frames this cross-chip identity as evidence that the “admissibility-and-projection contract” works. It does work — as a classical fitting procedure. The twelve-decimal agreement is a property of Newton’s method converging on a smooth, low-dimensional surface defined by known physics. It would produce the same agreement regardless of whether the underlying quantum computation was fault-tolerant, error-mitigated, or completely unprotected.

The Chemistry Results Are Standard NISQ Work

AIX reports chemical accuracy on five molecular workloads: H₂, LiH, H₂O, BeH₂ equilibrium, and BeH₂ transition state. The H₂ calculation is a 2-qubit, 6-Pauli-term Hamiltonian (five non-identity measured expectation values plus one identity scalar term). A classical laptop solves it exactly in microseconds.

Strip the proprietary terminology and the pipeline is: prepare trial states on the QPU, measure Pauli expectation values via bitstring counting, filter through admissibility checks, run Newton optimization on a parameterized subspace, report the converged energy. The paper explicitly denies this is a variational quantum eigensolver (VQE), arguing that Seed IQ avoids arbitrary-angle rotations on the hardware by using magic-state-injected T gates and a closed Clifford+T gate set. That is a legitimate semantic distinction about gate compilation. It does not change the functional reality: projecting noisy quantum measurements onto a parameterized trial-state subspace via classical Newton iteration and reporting the energy at the converged parameters is a variational optimization method, regardless of how the gates were compiled. This is much closer to NISQ-style variational chemistry with error mitigation, post-selection, and classical projection than to fault-tolerant computation.

Groups have been achieving chemical accuracy on these same small molecules using error-mitigated NISQ methods since Peruzzo et al. (2014) and Kandala et al. (2017). The AIX results (all 22 runs inside 1.6 mHa) show tight engineering on the hardware used, but they do not require or demonstrate fault tolerance by any standard definition.

The Notebook

Denis Ovseyenko stated on LinkedIn that “a Jupyter notebook with telemetry, datasets and IBM Quantum evidence is available as a supplement to the paper.” I examined this notebook.

The notebook contains 12 hardcoded references to paths on Ovseyenko’s local filesystem (/Users/denis/... across local quantum_ibm_hardware directories). It depends on a custom Python module (parse_runs.py) that is not included in the Zenodo upload. It contains zero pd.read_csv() calls and zero json.load() calls for external data. The core data is either loaded by the missing module from local files, or manually typed into the notebook as pandas DataFrame constructors repeating the paper’s own numbers.

Its final cell explicitly describes itself as a “lite distribution” and states that the full audit ledger and circuit archive are “access-gated” behind AIX’s standard NDA. The paper confirms this: “the full 32-character IBM Quantum workload identifiers are available in supplementary material under NDA.”

As distributed, the notebook is not independently executable. It is the paper’s claims reformatted into charts and tables, not standalone reproducible evidence.

The AI Writing Question

The paper has a prose problem worth noting. It repeatedly uses phrases such as “twelve-decimal” and “admissibility” in numerous variations, offers no limitations section or discussion of future work, and often treats the output of its own admissibility pipeline as evidence that the pipeline is valid. The title is unusually long by scientific publishing standards. This does not prove AI authorship. It does, however, read less like a paper that has been pressure-tested by QEC reviewers and more like a promotional technical narrative optimized to defend a claim.

The Escalating Claims

Beyond the FTQC assertion itself, the language surrounding Seed IQ follows a pattern worth examining. Each claim is larger than the last, each is harder to falsify, and each positions the company further beyond the reach of normal evaluation.

The press release states that FTQC “is now an execution problem, a governance problem, and a commercialization opportunity.” This reframes the hardest open problem in experimental physics as a business problem that AIX has already solved.

On LinkedIn, Holt went further: “Our ‘AI’ is not like any AI you see currently. This is not deep learning. This is not predictive optimizers.” She described Seed IQ as having “a multi-agent belief propagation ability that no one else in the classic active inference field has been able to achieve.” The claim is no longer confined to quantum computing — Seed IQ’s AI itself is presented as unprecedented.

Then scope inflation: “We have taken this far beyond just achieving FTQC. We are well into the ability to compute ground state molecules using our own primitive compute pipeline.” The original claim (FTQC on rented hardware) was already extraordinary. Now the company has gone beyond FTQC.

And the press release describes Seed IQ as having “applications across quantum computing, data centers, autonomous systems, financial models, infrastructure systems, and other complex domains.” A technology that solves the hardest problem in quantum physics and also governs data centers, financial models, and autonomous systems is not a specialized engineering achievement. It is a marketing pitch.

Each layer of the claim is designed to make evaluation harder. If the FTQC claim is challenged on physics grounds, the response is that Seed IQ operates by a mechanism no one else has seen. If the AI claim is challenged, the response is that it’s an entirely different kind of AI. If credentials are questioned, the response is that existing credentials are irrelevant because nothing like this exists. The structure is self-sealing: every objection is preemptively framed as a failure of the objector’s understanding.

The Public Response

When technical questions were raised in a LinkedIn discussion, the founders’ responses confirmed the pattern.

Denise Holt responded to criticism by stating: “For the people making comments on here as if you understand what we are doing, I guarantee you don’t because you’ve never seen anything like this. It doesn’t exist outside of our company.” She added: “We are not just error-correction, we are actually steering the system into viability.”

The claim that Seed IQ steers the system “into viability” beyond error correction raises an obvious follow-up: what does the error correction component consist of, given that the paper’s own computation register operates at code distance 1? At d=1, there is no error correction happening. There is nothing to go beyond. When pressed on this in a subsequent comment, Holt clarified: “I didn’t say error correction wasn’t happening. But it is only part of the governing process.” The paper provides no mechanism for how error correction operates at a code distance that, by mathematical definition, can correct zero errors.

Denis Ovseyenko responded to a commenter asking for technical proof with: “You are gonna look very silly soon enough.”

Neither founder engaged with any of the specific technical objections raised: the d=1 problem, the shot rejection rate, the Newton optimizer producing the twelve-decimal agreement, or the quantum Hamming bound. Every response has been a variation of “you don’t understand,” “it’s all there if you look,” or future vindication. These are not technical arguments. They are requests to be exempt from scrutiny.

What This Is, and What It Is Not

Based on a detailed review of the paper, the notebook, and the public claims, AIX appears to have built an error mitigation and post-selection pipeline with classical variational optimization, applied to small molecules on IBM cloud hardware, wrapped in proprietary terminology and marketed as fault-tolerant quantum computing.

The component techniques (calibration-aware qubit mapping, parity-based post-selection, admissibility filtering, dynamical decoupling, zero-noise extrapolation, Pauli twirling, classical Newton projection) are published, documented, and available as standard library functions or established techniques in IBM’s own Qiskit framework. The quantum content in all this follows established NISQ variational and classical hybrid patterns. The paper does not contain a single result that requires, or demonstrates, fault tolerance by the definition used by the quantum error correction community, by Google, by IBM, by Quantinuum, or by any QEC textbook.

Yet the press release calls this an “FTQC Breakthrough.” The paper operates at code distance 1, which corrects zero errors. The gap between those two statements is the entire problem.

For context, consider what actual progress toward fault-tolerant quantum computing looks like. Google’s Willow chip demonstrated below-threshold error correction at distances 3, 5, and 7 on a purpose-built 105-qubit processor, with results published in Nature with a team of over 100 researchers and more than a decade of development. Quantinuum demonstrated record logical qubit fidelity using color codes on their H2 trapped-ion system, also published in Nature. Microsoft announced Majorana 1 and published related topological-qubit work in Nature, though the claim remains contested among some physicists. IBM’s roadmap targets its first fault-tolerant system (Starling) for 2029, using architectures and interconnects that do not yet exist.

Those efforts involve hundreds of PhD physicists, billions of dollars of investment, custom-designed hardware, and publication in the world’s most selective peer-reviewed journals. The claim that a two-person startup with no quantum physics publications achieved FTQC on rented cloud hardware in a few months, publishing the results on Zenodo under NDA, is not consistent with the state of the field.

What This Means for the Industry

I have been warning about quantum winter risks for months. Announcements like this one accelerate that risk. When claims this detached from physical reality circulate through trade media and LinkedIn without challenge, they create two problems simultaneously. They raise expectations that cannot be met, setting the industry up for a credibility collapse when reality catches up. And they train the CISOs, investors, and policymakers who need to make real decisions to dismiss all quantum claims as hype, including the legitimate ones.

This is the Q-FUD cycle I keep writing about. The hype merchants inflate claims. The backlash produces cynics. Both groups are wrong, and both make the useful work harder.

The quantum trade media bears responsibility here. At least two publications ran this announcement without contacting an independent QEC expert. A single phone call to any quantum error correction researcher at any university or national lab would have identified the d=1 problem within minutes. The trade press functioned as a pass-through for marketing, and the people who pay the price are the readers who rely on these outlets for accurate information.

A Note on Process

Here’s how honest scientists act when they think they uncovered something: submit the paper to Nature, Science, or Physical Review Letters, where independent experts in quantum error correction will evaluate the methodology, reproduce the key results, and determine whether the “d=1 inversion” represents a genuine advance or a mischaracterization of error mitigation as fault tolerance. That is how results of this magnitude get validated. BusinessWire press releases are not a substitute for that process, and “you don’t understand” is not a response to specific, answerable technical questions.