What the AstraZeneca research found about IonQ that almost nobody has reported
Recently, I directed my team, which includes researchers with advanced backgrounds in chemistry, drug discovery, and medical research with a broad market-looking quantum research-related project. In order to proceed we decided to use the IonQ - AstraZeneca findings to help build a base-line template for our upcoming work. What I share here is not the research project we're focused on, but what we realized during the baseline build. And it is meaningful for the IonQ community. Specifically, IonQ had substantial and meaningful 'beats' or 'advantages' over there competitors in certain key areas, including IBM, and these advantages open up significant pathways for future partnerships and projects. What was notable and consistent was that IonQ's own research publications were extremely conservative in their reporting, and the broader picture tells a much more compelling story (as cited in the data below).
Tomorrow, September 8th, is IonQ’s Investor Day on the NYSE.
It is likely no surprise to anyone who reads my reports that due to my and my team’s thorough research, I am optimistic about the company’s short- and long-term growth — with my views often leaning, as the balanced evidence I present does, toward a view that IonQ is positioned for quantum market dominance. This view I hold is based solely upon the substantial research my team and I do.
Am I excited? Yes. Substantially. And the reasons are very clear: I believe that the research conducted - research that challenges the good and the not so good - points to a very clear picture that IonQ's quantum platform offering is THE REALITY that could enhance the human condition in ways other technologies fall short. And I want to be upfront about this - fiercely, my team and my views are independent. We come at the quantum sector as a whole with two unique but different objectives. The first is straightforward: is our significant investments worthy of the stake we own and what are the considerations going forward? The second objective is always the driving force: does this investment meet the philosophical and spiritual approach that we as a team embrace life: in other words is there a greater good at hand, particularly with a distinct possibility to come to the aid of those with healthcare challenges?
So it probably is no surprise to some of you who read these reports that I have a deep focus on quantum healthcare initiatives. That focus is what led my very competent team us to study the AstraZeneca partnership carefully, and what we found is extremely positive, far-reaching, and — to the best of my view — understood by very few.
With Investor Day tomorrow, I thought I would share some of these insights, because they really matter. And let me be clear about my view: much of what I share here I expect will bear on the company’s future endevours.
A note on interest, stated up front rather than buried: I hold a long equity position in IonQ via assorted platforms. The full disclosure is at Section 20.
CONTENTS
- The inversion at the centre of the story
- The customer wrote the test
- Who co-signed the standard
- The second class, rated higher, currently open
- Slow gates as a fault-tolerance asset
- Nine codes, one device
- Energy: the measurement, not the superlative
- The cryogenic argument that needs no vendor
- Peer review as an evidence class
- What execution at scale actually looked like
- The 9× is a live-system improvement
- Silicon, not a count
- Ownership instead of assurance
- Two doors nobody is opening
- The five worth putting on stage
- Where this leads
- Honest concessions
- Source appendix
- Disclosure of interests
EVIDENCE TIERS USED IN THIS ESSAY
Every load-bearing claim below carries one of four tags. The legend is the series standard; the worked examples are drawn from this essay.
FACT — established by a primary source read directly, and checkable by any reader with the citation. Worked example: IonQ completed the SkyWater acquisition on 31 July 2026. Both parties’ own releases. [R18]
INFER — a conclusion drawn from facts, where the reasoning is stated and the reader can test the step. Worked example: Because AstraZeneca’s published criterion places binding affinity outside the accuracy-limited set, and because IBM’s flagship biomolecular result is a binding problem, IBM’s flagship sits outside the class its customer’s own scientists prioritised. The two facts are checkable; the join is inference. [R5, R13]
ARG — advocacy. A position the author holds, built on evidence but not compelled by it. Worked example: That the pattern found at AstraZeneca is reproducible at other pharmaceutical companies. One account is not nine.
UNDISC — the record does not settle it, and it is named rather than filled. Worked example: Whether IBM’s 210× characterisation of its own protein result is supported. The figure does not appear in the paper’s abstract; only the 40× is verifiable from the published record. [R13]
1. The inversion at the centre of the story
There is a peculiar problem at the heart of IonQ’s public story, and it is the opposite of the one most people assume. The prevailing argument about this company concerns whether it overclaims. The evidence, read closely and in full, says the reverse: IonQ’s scientists systematically underclaim, and the gap sits in transmission rather than in substance.
Three IonQ papers volunteer limits nobody asked for. The Walking Cat fault-tolerance architecture discloses an intrinsic error floor no decoder can beat. [FACT · R14] The real-time decoding result states plainly that it does not generalise to competing architectures. [FACT · R15] The April 2026 benchmarking framework sets a chemical-accuracy criterion of 1 mHa against the field’s roughly 1.6 mHa convention — stricter than the convention — and then publishes that IonQ’s own hardware does not clear it across the range, with the noise regime plotted rather than averaged away. [FACT · R19] IonQ’s published resource tables run two to three times more cautious than the company’s own roadmap slide. [FACT · R14, R20]
For an institutional reader, that is not a weakness to be managed. It is the single most valuable and least used fact about the company, because it reframes every other claim for a sceptical audience. A vendor whose papers are more conservative than its marketing is a vendor whose papers can be trusted — and a customer who checks will find exactly that. [INFER]
The rest of this essay is what that inversion has been hiding.
Why it matters. A vendor whose published limits run stricter than its own marketing is a vendor whose next claim can be believed without independent verification. For a sceptical institutional reader that reverses the usual burden of proof, and it is the frame every other section here depends on.
2. The customer wrote the test
The strongest finding in the AstraZeneca research requires no performance claim, no hardware comparison, and no inference at all.
AstraZeneca’s own quantum group published a single criterion in Drug Discovery Today in March 2025 for deciding where quantum computing is worth applying: is the problem accuracy-limited? Two conditions must hold. Higher accuracy must change the answer, and the link from the computed quantity to the decision must survive downstream uncertainty. [FACT · R5]
They then worked both examples themselves.
Binding affinity fails. Linking interaction strength through signalling to a disease endpoint carries too many unknown biological influences, and compounds with similar binding energies produce very different responses. Their assessment is that value from further accuracy has plateaued and the problem is not accuracy limited in that context. [FACT · R5]
Reaction barriers in transition-metal catalysis pass. Their worked example is enantioselective hydrogenation with an iridium-based Pfaltz-type catalyst — a transition-metal complex with strong electron correlation. The link to the decision, in their words, is straightforward: lower barriers give higher reaction yield, and nothing downstream obscures it. [FACT · R5]
Binding affinity is the class of IBM’s largest published biomolecular result. [FACT · R13] Reaction barriers are the class of IonQ’s joint result with AstraZeneca. [FACT · R1] The customer wrote down the test, and IonQ’s problem class passes it. [INFER]
The confirmation is not even indirect. AstraZeneca’s Anders Broo is quoted in IonQ’s own June 2025 release describing the work as an important step toward accurately modelling activation barriers for catalysed reactions relevant to route optimisation in drug development. [FACT · R17] That is not a proposition being sold to a customer. It is the customer saying it first.
Why it matters. Fit to the customer’s own published criterion is the one competitive argument that cannot be answered with a bigger machine. A competitor can out-scale IonQ on any axis it chooses and still sit outside the problem class AstraZeneca’s scientists said was worth the effort.
3. Who co-signed the standard
What makes that finding unusually durable is its authorship.
This is not IonQ describing itself, and it is not an analyst’s inference. It is the customer’s published selection criterion — six authors, four from AstraZeneca Gothenburg. And two of the six are IBM-affiliated: Ivano Tavernelli at IBM Quantum Zürich, and Jason Crain at IBM Research, Hartree. [FACT · R5]
The competitor’s own scientists co-signed the standard against which the competitor’s flagship pharmaceutical problem class is judged to have plateaued. The class judgement was published in March 2025 and predates the flagship result it bears on by fourteen months. [FACT · R5, R13] Nothing about that arrangement was designed to favour IonQ, which is precisely why it carries weight.
Why it matters. Authorship is what makes this finding survive a hostile read. A standard co-signed by IBM’s own scientists, and published fourteen months before the result it bears on, cannot be waved away as an IonQ talking point or an analyst’s convenient framing.
4. The second class, rated higher, currently open
The same paper names a second accuracy-limited class, and rates it above the first.
Nuclear quantum effects — proton tunnelling through classically inaccessible barriers, isotope effects, and tunnelling differences between enantiomeric transition states. AstraZeneca describes this class as strongly accuracy-limited, above reaction barriers, on the grounds that treating nuclei quantum-mechanically compounds an already prohibitive full-configuration-interaction cost. [FACT · R5]
It is adjacent chemistry rather than distant. Tunnelling differences between enantiomeric transition states is the same enantioselective-catalysis problem one layer deeper — the same chemistry, the same scientists, the same decision. [INFER] AstraZeneca has published its own algorithmic work in the class, applying the nuclear-electronic orbital framework to proton-coupled electron transfer, and that work ran on the IBM side of the account. [FACT · R6, R7, R8, R9]
This is the item worth being precise about, because the temptation is to overstate it. IonQ has published nothing here, and no assessment exists of whether trapped-ion hardware suits these methods. It is an identified opening, not a demonstrated capability. [UNDISC] The correct move is a question rather than a claim. But it is a named, uncontested, higher-rated opening in the customer’s own words, and that is a rare thing to find sitting unattended.
Why it matters. This is the only place in the account where the customer has named a higher-priority need that no vendor is currently serving. An opening the buyer has already written down and nobody has claimed is the cheapest business development available — and it will not stay unattended.
5. Slow gates as a fault-tolerance asset
Then there is the argument that nobody appears to be making at all, and it is the most intellectually interesting thing in the file.
IonQ’s slow gates are a fault-tolerance asset. Trapped-ion cycle times of one to five milliseconds are a throughput liability against superconducting microseconds, and that trade-off is real and should be stated in the same breath. But the same slowness is precisely what buys decoding headroom. IonQ has decoded 408 logical qubits and over a million logical operations in real time, with under 0.3 percent stretch, on a single commodity CPU of roughly MacBook class. [FACT · R15]
The comparison can be sourced symmetrically, which is what makes it usable. IBM states in its own publication that its decoder is suited to FPGA or ASIC implementation for real-time decoding. [FACT · R23] So the contrast is not IonQ characterising a competitor; it is each vendor describing itself, and a reader placing the two sentences side by side. A physics disadvantage converting into a systems advantage is a rare shape of argument, and it survives the scrutiny that vendor-on-vendor claims do not.
The downstream consequence is an infrastructure bill rather than a benchmark. IonQ’s own provisioning rule is that decoding resources scale with code blocks per core — so you add commodity CPUs, not custom silicon. [FACT · R15] For a customer modelling the total cost of a fault-tolerant deployment five years out, that is a materially different line item, and it has never been argued publicly. [INFER]
Why it matters. Fault tolerance is where the sector’s five-year value sits, and IonQ’s route there runs on commodity CPUs rather than custom silicon. That is a materially different total cost of ownership, and it turns the most-cited criticism of trapped ions into a reason the architecture scales.
6. Nine codes, one device
The architectural evidence underneath that claim is experimental rather than asserted, and it too is under-reported.
IonQ has run nine qLDPC error-correcting codes with starkly different connectivity requirements on a single device, without hardware reconfiguration — codes that on a fixed lattice would require long-range couplers. [FACT · R16] That is what all-to-all connectivity actually buys, demonstrated rather than specified on a data sheet.
The encoding efficiency follows from the same architecture: 22 logical qubits in 102 physical, where the same distance in surface codes would require 1,782. [FACT · R20] The comparison is IonQ’s own and rests on physical fidelity at 99.99 percent against 99.9 percent — a qualifier that must travel with it, and one IonQ’s paper supplies. The 99.99 percent figure is not aspirational: Oxford Ionics, now an IonQ company, published two-qubit gate fidelity above that threshold without ground-state cooling in October 2025. [FACT · R21]
And break-even quantum error correction has been demonstrated with qLDPC codes on a Tempo engineering test system — error correction on hardware IonQ operates and sells, not in simulation. [FACT · R16, R24] That result was largely buried behind a qubit-count headline in the same quarter’s communications, which is a recurring pattern rather than an isolated one.
Why it matters. Connectivity claims are normally specifications; this one is an experiment. Nine codes on one unreconfigured device, plus break-even error correction on hardware IonQ actually sells, moves the fault-tolerance argument from roadmap to record.
7. Energy: the measurement, not the superlative
The energy story is the clearest case in the file of a strong result being crowded out by a weaker claim, and it deserves its own treatment.
Knitter et al., working across IonQ, QuantumBasel and the Center for Quantum Computing and Quantum Coherence, measured energy-to-solution on a production IonQ Forte Enterprise — 36 trapped ytterbium-171 ions in a surface linear Paul trap. The system carries electrical monitoring and logging at the level of macroscopic component groupings. The team recorded power draw and elapsed time and computed energy in joules for every submitted job. The classical comparator, GPU statevector simulation on a single NVIDIA L4, was measured with CodeCarbon. [FACT · R22]
That methodology is the strongest thing in the paper, and it is what almost nobody has reported. Most published quantum energy figures are estimates — execution time multiplied by an average system power. These are direct electrical measurements on a production system in a commercial data centre. [FACT · R22]
The result: energy-to-solution on the QPU scaled roughly linearly with qubit count, while classical simulation of the same circuits scaled exponentially, with least-squares fits on both. Projected break-even sits at approximately 34 qubits for 600-shot debiased inference circuits. The debiasing procedure itself — 25 executions per circuit with varied qubit mapping, then non-linear filtering — delivered a 24 percent error reduction against purely classical models, including in noisier qubit zones. [FACT · R22]
Now place that next to the comparator position. IBM’s published figure at scale is 2 MW for Blue Jay, a system stated on the IBM Technology Atlas for 2033 and beyond. [FACT · R25] IonQ holds a measurement. IBM holds a projection. Under any serious evidence-tiering discipline those are different classes of claim, and at an account whose own criterion paper is built on distinguishing what is computed from what is asserted, holding the only measured number is the position worth occupying. [INFER]
Which is why the superlative is a mistake. “Lowest cost and lowest energy footprint per logical qubit on the market” is comparative, unmeasured, and invites the question against whose logical qubit, measured how — a question with no available answer. [FACT · R24] It sits on top of a genuinely measured result that no competitor currently matches. The stronger claim is being crowded out by the weaker one, and the fix costs nothing.
The qualifier that must travel with the measurement, and it is not optional: the crossover is against classical simulation of the quantum circuit, not against a classical solution of the underlying problem. One workload, one hardware configuration, quantum inference only. The authors write that the exact crossover point is specific to the hardware, circuits and hyperparameters chosen. [FACT · R22] Stated as energy-to-solution for this workload against this comparator, the finding holds. Stated as quantum uses less energy than classical computing, it does not. Quoting the 34 qubits with the authors’ own scope sentence attached is what makes it credible to a scientific audience.
Why it matters. Measured data beats a projection in any serious evaluation, and IonQ holds the only measured energy figure in the comparison. Leading with the measurement rather than the superlative costs nothing and removes the single question a competitor could ask that has no answer.
8. The cryogenic argument that needs no vendor
Alongside the measurement sits an architectural argument that is genuinely under-used, and its strength is that it does not depend on any vendor’s characterisation of itself.
Trapped-ion systems operate without a dilution refrigerator. IonQ’s Forte runs a cryostat below 10 K; superconducting processors require millikelvin dilution refrigeration. Both are cryogenic — the claim is not “room temperature,” and describing it that way is wrong — but the requirement differs by roughly three orders of magnitude in temperature. [FACT · R24]
The consequence is quantified in independent literature that has nothing to do with either vendor. Dilution refrigerators draw on the order of 10 to 25 kW. Superconducting systems draw on the order of 30 kW peak during cooldown. And cooling is identified as the dominant term in quantum data-centre power usage effectiveness. [FACT · third-party technical literature, consolidated at §10.11 of the source research]
None of that comes from IonQ. An architectural advantage that survives without the vendor asserting anything about itself is rare, and it is worth more than a superlative precisely because a sceptical reader can verify it without trusting either party. [INFER] The related point is a compilation one: all-to-all connectivity removes SWAP routing overhead, eliminating compilation cost that constrains fixed-lattice architectures. Fewer gates for the same circuit is an energy argument as much as a fidelity one. [FACT · R22]
Why it matters. This is the one argument in the essay that requires trusting neither vendor. An advantage a reader can confirm from independent literature is worth more than any first-party claim, especially to an audience that has learned to discount quantum marketing.
9. Peer review as an evidence class
There is a distinction in the record that has gone almost entirely unremarked, and it runs in IonQ’s favour on exactly the axis the industry claims to care about.
The IonQ–AstraZeneca work went to preprint on 27 June 2025, was submitted to journal on 15 August 2025, was accepted on 17 June 2026, and published as Zhao et al., Physical Review Research 8, 033061 on 15 July 2026 — open access under CC BY 4.0. Roughly ten months in review; twelve and a half months from preprint to publication. It cleared peer review. [FACT · R1]
IBM’s 12,635-atom protein–ligand result is arXiv:2605.01138, posted 1 May 2026. It is a preprint, and it has not been peer reviewed. [FACT · R13] That is not a criticism of the work, which is a real engineering achievement at unprecedented scale. It is a statement about evidence class, and evidence class is the entire basis on which this customer’s scientists decide what to believe.
One further detail belongs in the record. IBM’s own headline characterisation of that result at 210× does not appear in the paper’s abstract; only the greater-than-40× system-size increase is verifiable from the published record. [UNDISC · R13] The larger figure may well be supported in the full text. But the asymmetry is worth naming, because the criticism most often levelled at IonQ — a number travelling ahead of its documentation — turns out on inspection to be an industry condition rather than an IonQ one. [INFER]
Why it matters. This customer’s scientists decide what to believe on evidence class, and IonQ’s chemistry result cleared peer review while the competitor’s flagship has not. The criticism most often aimed at IonQ — a number travelling ahead of its documentation — turns out to describe the sector rather than the company.
10. What execution at scale actually looked like
Nobody reports the operational numbers, and they are the ones a procurement function would actually ask about.
The QC-AFQMC demonstration ran 300,983 circuits — including trial runs and post-selection discards — collecting 179,858 matchgate shadows across three molecular species, over several weeks on a shared production system. [FACT · R1] Each shadow circuit ran for a single shot. Error mitigation used leakage-detection gadgets on alternating qubits, two ZZ gates and one ancilla each, eight ancillas in total, with flagged shots discarded by post-selection. The system was IonQ Forte, accessed through Amazon Braket, driven by CUDA-Q. [FACT · R1]
That is not a performance claim. It is an availability, reliability and throughput fact about a commercial machine sustaining a research campaign of a third of a million circuit executions on shared access. For a customer evaluating whether a quantum vendor can support a multi-week programme rather than a demonstration, it is more informative than any fidelity figure — and it has never been used. [INFER]
The chemistry itself deserves stating plainly, because the scope is what makes the result honest: a nickel-catalysed deformylative Suzuki–Miyaura cross-coupling, oxidative addition step, 77 atoms truncated to 41, an (8-electron, 8-orbital) active space selected at single-orbital entropy above 0.2, in a minimal STO-3G basis, on 24 qubits — 16 for the trial state and 8 leakage-detection ancillas. [FACT · R1] Small, and precisely described. Both qualities are the point.
Why it matters. A third of a million circuit executions on shared production access answers the question procurement actually asks: does the machine stay up for a multi-week programme. That is an availability fact, and no fidelity figure substitutes for it.
11. The 9× is a live-system improvement
The quietest number in the whole record is the strongest one scientifically, and it describes something the industry rarely manages at all.
Median matchgate shadow circuit execution time fell from 9.9 seconds to 1.1 seconds — a 9× improvement — delivered through application-specific control-software and firmware tuning on the production system during the engagement. Projected campaign duration fell from 34.5 days to 3.8 days. [FACT · R1]
IonQ made a commercial machine roughly nine times faster on the customer’s own workload, mid-engagement, through software. The 9× is measured, against a measured comparator, on identified hardware. The better-known 20× is a conservative estimate against a baseline implementation that cannot execute molecules of this size at all, with the majority of the gain coming from GPU-accelerated classical post-processing rather than the QPU — all four qualifiers supplied by IonQ’s own engineers in the technical blog. [FACT · R2] The two figures are not equally grounded and should never be presented as though they were.
The related attribution point is worth correcting once and for all. The 656× post-processing advance — reducing local energy evaluation from O(N^8.5) to O(N^5.5) — is IonQ’s own scientific contribution, and it is precisely the improvement the baseline authors had named as the one their method needed. [FACT · R1, R10] It is usually presented as a deduction from the 20× headline, which gets the attribution backwards. The 656× is the science. The 20× is the estimate that followed from it.
Why it matters. Making a customer’s workload nine times faster mid-engagement, through software, is the clearest available evidence that IonQ improves systems already in the field. It is also the better-grounded of the two headline numbers, and the one that will survive scrutiny.
12. Silicon, not a count
The manufacturing story has been crowded out by a number, and the number is the less impressive half.
Three tape-out rounds completed on the first semiconductor-based quantum chip by February 2026. First chip prototypes received back in May 2026, stated to demonstrate the critical quality metrics required for production-grade 256-qubit chips. And in the second quarter of 2026, fully featured, fully integrated QPUs received back from SkyWater and now in test at College Park — described as consolidating individually validated capabilities into a single unified chip architecture. [FACT · R24]
That is not a qubit-count milestone. It is a change in evidentiary state: design, then silicon, then integrated system. [INFER] The chip exists as a physical, integrated prototype rather than a roadmap row, and the closed design-to-fab-to-chip loop — Oxford Ionics design, SkyWater fabrication, IonQ integration and test, inside one company — is the structural claim competitors cannot match. [FACT · R18, R24]
The qualifiers belong with it, and they cost nothing to state. Commissioning targets the first half of 2027; the system is in test, not in service. 256 physical qubits produce no logical qubits, and the distinction should be stated rather than left to be discovered. Management also confirmed under analyst questioning that the 256-qubit chip uses Oxford Ionics microwave excitation sources and does not integrate Lightsynq photonic or quantum-memory technology, which sits in the Boston operation serving networking and QKD work. [FACT · R24] Anyone assuming the 256 is a networked or memory-enabled part is assuming something management explicitly denied — and saying so first is more credible than any nearer claim.
There is also a delivery fact sitting at the same site as the energy measurement, and it is almost entirely unused: two consecutive generations of quantum computer deployed side by side in a commercial setting. A fifth-generation system in final assembly at QuantumBasel, beside the fourth-generation machine already installed there, with fifth-generation machines also deploying at KISTI in Korea. [FACT · R24] That is delivery cadence shown rather than asserted, which is the cleanest available evidence for a capability the market habitually discounts.
Why it matters. Silicon in a test lab is a different order of evidence from a row on a roadmap, and the closed design-to-fab-to-chip loop is a structural position no competitor can assemble quickly. Stating the commissioning date and the physical-versus-logical distinction first is what makes the rest credible.
13. Ownership instead of assurance
IonQ can now answer a supply-chain question with ownership rather than assurance, and the case is structural rather than promotional.
The SkyWater acquisition is complete, not pending: definitive agreement 26 January 2026 at $35.00 per share, cash-and-stock subject to a collar, roughly $1.8bn total equity value; stockholder approval 8 May 2026; final regulatory approval July 2026; completed 31 July 2026. SkyWater operates as a wholly owned subsidiary retaining its name, led by CEO Thomas Sonderman reporting directly to Niccolò de Masi, and retains its status as a pure-play DMEA-accredited Category 1A Trusted Foundry with facilities in Minnesota, Florida and Texas. [FACT · R18]
A customer contemplating a 2028-horizon capability asks who makes the chips and what happens if that supplier reprioritises. Owning the foundry answers that question rather than deferring it. [INFER] IonQ additionally states the acquisition pulls forward functional testing for 200,000-physical-qubit QPUs — supporting 8,000-plus logical qubits — to 2028 from the 2029 row on the published roadmap, by eliminating merchant-foundry queue delays and shortening wafer iteration, with co-located cryogenic wafer testing in Minnesota removing the need to ship fragile wafers to third-party labs. That is IonQ’s characterisation, and it is checkable: it should appear in the roadmap at Investor Day or it did not happen. [FACT for the claim · R18]
There is an ecosystem position underneath it that is never discussed strategically. SkyWater has historically supplied foundry services to D-Wave, EeroQ, Silicon Quantum Computing and PsiQuantum, and IonQ states it will maintain the merchant model. IonQ now owns the foundry that serves several of its competitors. [FACT · R18] That is genuinely double-edged — a strong position and a plausible reason for those customers to seek alternative capacity — and the first observable consequence would be any of the four announcing a foundry change.
The combination is what no competitor in the research matches: a US parent with federal credentials and DoD programme alignment including the Microelectronics Commons; a DMEA Trusted onshore foundry; a UK acquisition cleared by the Investment Security Unit at $1.075bn; a university installation in Cambridge, the city of AstraZeneca’s global headquarters; and a Swedish subsidiary, IonQ AB, inside the customer’s own BioVentureHub incubator in Gothenburg. [FACT · R18, R24, R26] For a customer straddling US, UK and Swedish operations, that answers sovereignty from every direction at once. It also sits on the working export-control path: the United Kingdom is on the BIS 4A906 IEC destination list. [FACT · R27]
And the balance sheet gives IonQ a five-year conversation to have: $3.0bn in cash and investments at 30 June 2026, roughly $2.0bn pro forma after the SkyWater close, second-quarter 2026 revenue of $80.1m up 287 percent year on year, and full-year 2026 guidance raised to $280–290m on an IonQ standalone basis. [FACT · R24]
Why it matters. Owning the foundry converts a supply-chain assurance into a supply-chain fact, and the US–UK–Sweden footprint answers sovereignty from every direction this customer operates in. The balance sheet then funds a five-year conversation most of the field cannot afford to have.
14. Two doors nobody is opening
Two positions in the portfolio are stated in IonQ’s materials but connected to nothing.
The first is security. ID Quantique — majority stake taken April 2025 — ships the Clavis XG, described as the first certified QKD system in the world at TRL 9, with telecom, financial and government customers already deployed. [FACT · R24] The Cambridge–Bristol network-node work is live. This is the commercially mature line in the portfolio, ahead of the chemistry offer in revenue terms, and it reaches a different buyer with a different clock — a CIO and a security organisation rather than a computational chemistry group. It is almost entirely absent from how IonQ is discussed in a pharmaceutical context.
The second is Cambridge. The March 2026 IonQ–University of Cambridge Quantum Innovation Centre carries a stated intent from both parties to deploy the chip-based 256-qubit system at the Cavendish Laboratory. [FACT · R26] It is forward-looking language on both sides, and commissioning targets the first half of 2027 — the system is not in service. But stated deployment intent in a major customer’s headquarters city is a fact worth connecting to the account argument, and at present it is connected to nothing.
The wider version of the same observation: the full-stack roll-up is one story and it is only ever told in pieces. Oxford Ionics, Lightsynq, ID Quantique, Qubitekk, Vector Atomic, Skyloom, SkyWater — plus federal contracted work including AFRL networking, which is a credential distinct from US parentage and Trusted Foundry status. [FACT · R24] Assembled, it is a platform claim. Disassembled, it reads as acquisition activity. IonQ’s own Analyst Day materials flag integration risk, and that flag is legitimate — but the absence of the assembled version is a communication choice, not a constraint imposed by the evidence. [ARG]
Why it matters. Both positions reach buyers on shorter clocks than computational chemistry — a security organisation, and a research institution in the customer’s headquarters city. They are already acquired, already announced, and currently connected to nothing.
15. The five worth putting on stage
If the material above had to be reduced to five items, these are they. Each carries the qualifier that must travel with it.
One. The customer wrote the test, and IonQ passes it. Reaction barriers in transition-metal catalysis pass AstraZeneca’s own published criterion; protein–ligand binding, the class of the competitor’s flagship, does not, by the customer’s own assessment. Requires no performance claim, no hardware comparison, and no inference. Qualifier: this is a technical-author view, not a stated enterprise position. There is no ranked table and no scoring rubric. [R5]
Two. Lead the energy argument with the measurement, not the superlative. Directly measured electrical data from a production system in a commercial data centre, against a competitor figure that is a projection for a system dated 2033 and beyond. Measured against projected is a durable asymmetry. Qualifier: the crossover is against classical simulation of the quantum circuit, for one workload, on one configuration, and the authors say so. [R22, R25]
Three. Lead the 256 story with the silicon, not the count. Three tape-out rounds, prototypes returned, fully featured integrated QPUs back from IonQ’s own foundry and in test. The closed design-to-fab-to-chip loop is the structural claim no competitor can match. Qualifier: commissioning targets 1H 2027, and 256 physical qubits produce no logical qubits. [R18, R24]
Four. Name the comparator. The “other approaches” baseline on Analyst Day slide 34 traces to IBM’s own peer-reviewed Nature paper — Bravyi, Cross, Gambetta, Maslov, Rall and Yoder, the bivariate-bicycle [[144,12,12]] work. Saying so converts an unattributed comparison into one sourced to the competitor’s own publication. The same move is available on decoding, where IBM states its decoder suits FPGA or ASIC implementation. Qualifier: different code families; the logical-qubit ratio is not a like-for-like unit and should not be presented as one. [R20, R23]
Five. Delivery cadence, shown rather than asserted. A fifth-generation system in final assembly at QuantumBasel beside the fourth-generation machine already on that site. It is the cleanest delivery evidence available, and it sits at the same site as the measured energy work. [R24]
Why it matters. These five carry their qualifiers, which is what lets them be repeated in front of a technical audience without being taken apart. Each survives a hostile question; the versions circulating without qualifiers do not.
16. WHERE THIS LEADS
Everything above is diagnostic. It establishes what is true and under-reported, and stops there. This section is the other half — what these positions make possible — and it is written as advocacy rather than analysis. A reader who accepts Sections 1 through 13 is not obliged to accept any of what follows.
The criterion is portable, and that is the largest implication in this essay. AstraZeneca’s accuracy-limited test was published in an open journal. It is not an AstraZeneca artefact; it is a screening instrument. Any pharmaceutical organisation weighing where quantum computing is worth applying can run the same two questions — does higher accuracy change the answer, and does the link to the decision survive downstream uncertainty — across its own portfolio. The finding at this one account is therefore a method rather than a result. Whether the pattern repeats at other companies is an open question and one account is not nine. [ARG] But the instrument for asking is public, and nobody appears to be using it.
The environmental framing is sitting unconnected on both sides of the table. AstraZeneca has framed computation in environmental terms repeatedly, including a reduction in carbon and compute time reported to shareholders at Annual Report level as an achievement of algorithm redesign. IonQ has been building Energy-to-Solution as a metric since September 2025, and now holds directly measured data. Neither party appears to have joined the two. [FACT · R22, R24] An efficiency argument reaches a sustainability function and a computational chemistry function at the same time, on a shorter clock than a capability argument, and it does not require anyone to claim advantage over classical methods.
The second problem class is an invitation rather than a proposition. AstraZeneca named nuclear quantum effects as strongly accuracy-limited and rated it above reaction barriers. IonQ has published nothing there. The right move is a question about what that class requires, asked scientist to scientist, not an assertion of capability — and the value of asking is that the answer determines whether it is worth pursuing at all. [ARG]
The security portfolio reaches a different buyer on a different clock. ID Quantique ships a certified product to telecom, financial and government customers today. Inside a large pharmaceutical organisation that is a CIO conversation and a data-protection conversation, not a research one, and it does not wait on a 2027 commissioning date. [FACT · R24] The same is true of the Cambridge centre, which places stated deployment intent inside the customer’s headquarters city. [FACT · R26]
Foundry ownership is a commercial argument, not only a technical one. A customer contemplating a capability five years out can be answered with ownership rather than assurance, and the merchant model that continues to serve other quantum companies is a revenue line and an ecosystem position at the same time. [FACT · R18] It is double-edged, as Section 13 says, and the honest version states both halves.
And the decoding result is a procurement argument nobody is making. If fault tolerance on this architecture runs on commodity CPUs rather than custom silicon, that is a line in a total-cost model that a competitor cannot match by spending more. [FACT · R15, R23] Capability arguments reach scientists. Cost-of-ownership arguments reach the people who sign.
Why it matters. The findings in this essay are usually treated as a scorecard, and a scorecard changes nothing. Read as openings instead, most of them point at buyers, budgets and clocks that the current framing never reaches.
17. HONEST CONCESSIONS
None of the above rests on beating classical chemistry, and none of it should be read that way.
No vendor has demonstrated advantage over classical methods in this problem class. Not IonQ, and not IBM. IBM’s flagship consumed 94 qubits across two 156-qubit processors, 9,200 circuits, over 100 hours of quantum time and 1.3 billion measurement outcomes, against Fugaku and Miyabi-G — and the underlying result trails classical DMRG. [FACT · R13] That observation cuts both ways and should be presented both ways. With this concession we should also note that as IonQ's technology advances, particulary with the expected upcoming release of 256 followed by the 10,000 system, quantum advantage of classical methods could occur quickly.
IonQ’s own hardware result was off, and directionally wrong. Against a CCSD(T) reference of 53.3 and 45.4 kcal/mol for the two barriers, QC-AFQMC on an ideal simulator returned 57(4) and 44(4). On Forte, it returned 43(3) and 55(3) — the barrier heights shifted in opposite directions, placing the product lower in energy than the reactant and reversing the ordering given by both the reference and the ideal simulation. Trial-state particle number, which should be exactly 8, came back between 10.514 and 10.663 on hardware. The paper states all of this plainly and calls for better error mitigation. [FACT · R1] This is the honest state of the art, and it must be in the room before anyone else brings it there. In addressing this issue, we note that IonQ has made significant strides from the time of the published paper with respect to error mitigation and believe that under a combination of advancements, they are able to significantly and meaningfully reduce error mitigation issues.
The 20× requires four qualifiers to be defensible, and it was announced eighteen days before the preprint containing its supporting figures was public. The derivation is sound and appears in IonQ’s own technical blog; the version that travelled did not carry it. [FACT · R2, R3]
18. SOURCE APPENDIX
Every claim above traces to an entry below. Reference numbers are stable within this essay. Retrieval date 2 September 2026 unless otherwise stated.
A. Peer-reviewed literature and preprints
Ref | Citation | Identifier |
R1 | Zhao, L., Goings, J. J., Aboumrad, W., et al. (41 authors; IonQ / AstraZeneca Gothenburg / NVIDIA / AWS). “Quantum-classical auxiliary field quantum Monte Carlo with matchgate shadows on trapped ion quantum computers.” Phys. Rev. Research 8, 033061 (15 July 2026). Received 15 Aug 2025, accepted 17 Jun 2026. Open access CC BY 4.0 | DOI 10.1103/n1tf-8kr7 · arXiv:2506.22408v1 |
R5 | Kovyrshin, A., Tornberg, L., Crain, J., Mensa, S., Tavernelli, I., Broo, A. “Prioritizing quantum computing use cases in the drug discovery and development pipeline.” Drug Discovery Today 30(3), 104323 (March 2025) | DOI 10.1016/j.drudis.2025.104323 |
R6 | Kovyrshin, A., Skogh, M., Broo, A., et al. “A quantum computing implementation of nuclear-electronic orbital (NEO) theory.” J. Chem. Phys. 158(21), 214119 (2023) | — |
R7 | Nykänen, A., Miller, A., Talarico, W., et al. “Toward Accurate Post-Born–Oppenheimer Molecular Simulations on Quantum Computers.” J. Chem. Theory Comput. 19(24), 9269–9277 (2023) | — |
R8 | Kovyrshin, A., Skogh, M., Tornberg, L., et al. “Nonadiabatic Nuclear–Electron Dynamics: A Quantum Computing Approach.” J. Phys. Chem. Lett. 14(31), 7065–7072 (2023) | — |
R9 | Kovyrshin, A., Manawadu, H., Altamura, S., Tornberg, L., Broo, A. “Approximate quantum circuit compilation for proton-transfer kinetics on quantum processors.” Phys. Chem. Chem. Phys. 28, 3035–3054 (28 January 2026) | DOI 10.1039/d5cp04097c |
R10 | Huang, B., Chen, Y.-T., Gupt, B., Suchara, M., Tran, A., McArdle, S., Galli, G. (AWS / Univ. Chicago / Braket / Azure Quantum / Argonne). “Evaluating a quantum-classical quantum Monte Carlo algorithm with Matchgate shadows.” Phys. Rev. Research 6, 043063 (24 Oct 2024) | arXiv:2404.18303 |
R13 | Merz, K. and 23 co-authors (Cleveland Clinic / RIKEN / IBM / Michigan State / Univ. Tokyo). “Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing: quantum chemistry of protein-ligand complexes” (1 May 2026). Preprint, not peer reviewed | arXiv:2605.01138 |
A2. IonQ error-correction, benchmarking and energy corpus
Ref | Citation | Identifier |
R14 | Tripier, Chung, Young, Alam, Bjork, Brodutch, Buessen, Coble, Dellaert, Maslov, Roetteler, Tham, Webster, Ye, Gamble, Maksymov, Marceaux, Delfosse (IonQ). “Fault-Tolerant Quantum Computing with Trapped Ions: The Walking Cat Architecture” (21 April 2026) | arXiv:2604.19481v1 |
R15 | Ye, Maksymov, Delfosse (IonQ). “Real-time decoder for a MegaQuOp quantum computer using a single CPU” (25 August 2026) | arXiv:2608.25027v1 |
R16 | Tham, Goldman, Debnath, Patel, Saraladevi, Nguyen, Nielsen, Pisenti, Wright, Gamble, Delfosse (IonQ). “Breakeven demonstration of quantum low-density parity-check codes” (4 June 2026) | arXiv:2606.06455v1 |
R19 | IonQ. Measuring what matters: A scalable framework for application-level quantum benchmarking (announced 14 April 2026). Public code released | arXiv:2604.11781 · github.com/ionq-publications/apps-benchmark |
R20 | Support cited on IonQ Analyst Day slide 34 for the 800-logical-at-10,000-physical claim. The comparator side is Bravyi, Cross, Gambetta, Maslov, Rall and Yoder (IBM), Nature 2024 — bivariate-bicycle [[144,12,12]] | arXiv:2308.07915 (IBM) · arXiv:2503.22071 (Ye and Delfosse, IonQ) |
R21 | Hughes, Srinivas, Löschnauer, Knaack, Matt, Ballance, Malinowski, Harty, Sutherland (Oxford Ionics / Univ. Oxford). “Trapped-ion two-qubit gates with >99.99% fidelity without ground-state cooling” (20 October 2025) | arXiv:2510.17286 · DOI 10.48550/arXiv.2510.17286 |
R22 | Knitter et al. (IonQ · QuantumBasel · Center for Quantum Computing and Quantum Coherence). “Measuring Accuracy and Energy-to-Solution of Quantum Fine-Tuning of Foundational AI Models.” Submitted to IEEE Quantum Week. Preprint, not peer reviewed. Direct electrical measurement on IonQ Forte Enterprise; GPU comparator measured with CodeCarbon on an NVIDIA L4 | arXiv:2605.02798 |
B. Company primary sources, filings and vendor documents
Ref | Citation | Identifier |
R2 | IonQ. “Ecosystem Innovation: IonQ’s Life Sciences Application Workflow Accelerates the Drug Development Process.” ionq.com/blog, 9 June 2025. Carries the 20× derivation and its four qualifiers | — |
R3 | IonQ, Inc. Q2 2025 results, Exhibit 99.1 to Form 8-K, and Q2 2025 earnings call transcript, 6 August 2025 | SEC CIK 1824920 |
R17 | IonQ, Inc. “IonQ Speeds Quantum-Accelerated Drug Development Application With AstraZeneca, AWS, and NVIDIA,” 9 June 2025. Carries the Broo quotation | — |
R18 | IonQ / SkyWater Technology transaction record — definitive agreement 26 January 2026; stockholder approval 8 May 2026; final regulatory approval July 2026; completion 31 July 2026. Both parties’ own releases | SEC CIK 1824920 |
R23 | IBM. IBM lays out clear path to fault-tolerant quantum computing, IBM Quantum blog. Carries the FPGA-or-ASIC real-time decoding statement | ibm.com/quantum/blog/large-scale-ftqc |
R24 | IonQ, Inc. Q2 2026 earnings call, verified line by line against call audio; and IonQ Analyst Day materials. Carries the 1H 2027 commissioning target, the integrated-QPU milestone, the Lightsynq exclusion, the break-even qLDPC result, and the financial figures | Q2 2026 earnings call |
R25 | IBM. IBM Technology Atlas — Quantum roadmap, marked Updated April 2025. Nighthawk, Loon, Kookaburra, Cockatoo, Starling and Blue Jay rows; the 2 MW Blue Jay power figure | ibm.com/roadmaps/quantum.pdf |
R26 | IonQ / University of Cambridge Quantum Innovation Centre announcement, March 2026; and Business Region Göteborg, “Amerikanskt kvantteknikbolag etablerar sig i Göteborg,” 3 December 2024 (IonQ AB, Chalmers Next Labs) | — |
C. Regulatory and programme sources
Ref | Citation | Identifier |
R27 | Covington & Burling LLP (9 September 2024), Arnold & Porter (10 September 2024) and Miller & Chevalier (10 September 2024) advisories on the BIS interim final rule. Independently corroborate ECCN 4A906 / 4D906 / 4E906, the 34-qubit threshold, the 5 November 2024 compliance date, and IEC eligibility for the United Kingdom | — |
R28 | STFC Hartree Centre programme record — HNCDI announced 4 June 2021 as a five-year, £210m STFC–IBM partnership; IBM–Hartree collaboration dating to 2013; Hartree Centre Strategy 2024 to 2029 naming HNCDI as flagship programme | — |
Sourcing limitations, stated rather than implied. No archiving has been performed on URL-borne sources; entries R2, R17, R23, R24, R25, R26 and R28 are the most exposed to quiet editing of the underlying pages. R6, R7 and R8 rest on the bibliographic record rather than a direct read. R13 was read at abstract and press-record level only, which is why the 210× is tagged UNDISC rather than disputed. The third-party cryogenic power figures in §8 are consolidated from technical literature identified in the source research and are not attributed here to a single named study; a reader wanting to check them should treat that paragraph as the weakest sourcing in this essay.
20. DISCLOSURE OF INTERESTS
Position. The author holds long equity positions in IonQ, Inc. (NYSE: IONQ).
Not investment advice. Nothing here is a recommendation to buy, sell or hold any security.
No comments:
Post a Comment