Seven practices, one engineering team. We design, build, benchmark and operate the compute layer beneath your hardest problems — on-premise, hybrid or sovereign cloud.
A genomics pipeline and a trading engine are not the same machine. We start from your regulatory perimeter, your data gravity and your deadlines — then design backwards.
Quantum Computing
General, useful quantum advantage does not exist today. No vendor roadmap we have read justifies planning a production workload around it. So we build the two things that pay off now — the classical scaffolding and team capability that make you ready, and the post-quantum cryptography migration that is already overdue.
The Constraint
Every stalled quantum programme we have been asked to review failed the same way. It was scoped to a hardware capability that did not arrive. So we scope to what classical machines can do now, and treat the quantum processor as a back-end that may improve.
Fundable · defensible · useful without a QPU
Declined · with the reason in writing
Hybrid Architecture
A variational algorithm spends the overwhelming majority of its wall clock in a classical optimiser, a scheduler queue and data marshalling. That is the part we engineer, because that is the part that decides whether the experiment finishes.
Variational algorithms — VQE for chemistry, QAOA for combinatorial problems, parameterised circuits for machine learning — all share one shape. A short quantum circuit produces expectation values. A classical optimiser reads them and proposes better parameters. Repeat, thousands of times, each iteration carrying shot noise.
Treat that loop as ordinary high-performance computing and it behaves. Treat it as a quantum mystery and it stalls on queue waits, unversioned encodings and results nobody can reproduce. We build it with the same discipline we apply to a Slurm-scheduled CFD campaign, because structurally that is what it is.
How we run the classical halfThe problem is mapped to a cost function or Hamiltonian, then to a parameterised circuit. Encoding choice moves the result far more than hardware choice does. We keep encodings in version control beside the classical code, with the mapping documented well enough that a reviewer can argue with it.
Circuits are compiled to a back-end's native gate set and qubit coupling map. Transpilation is not neutral. Depth and two-qubit gate count change by large factors between optimisation passes, and depth is what decoherence consumes. We record both counts per commit and treat a regression as a build failure.
The same circuit runs on a local state-vector simulator, a tensor-network back-end, or a remote QPU behind a broker queue. The back-end is a configuration value, never a code change. That single decision is what lets a team keep working when a QPU is unavailable, expensive or out of jurisdiction.
Measured expectation values feed COBYLA, SPSA, or a gradient-based method. This is where nearly all the compute time is spent, and it profiles like any other numerical code — vectorisation, parallel evaluation of the parameter batch, and a cache of results so a restart does not repay the whole cost.
Every run is compared with a classical solver on the same data and the same time budget. The baseline usually wins. When it does, that goes in the headline of the report rather than a footnote, because a research programme that cannot record a negative result will keep paying for the same experiment.
Classical Simulation
Exact state-vector simulation is exponential in qubit count. Tensor-network methods reach much further on structured circuits and fail on unstructured ones. Both limits are computable in advance, which makes them honest planning inputs.
A state vector holds one double-precision complex amplitude per basis state: 16 bytes each, 2n of them. Allow a further two to three times that for gate work buffers and distributed transposes. Past a single large-memory node the state is split across an MPI rank set, and a two-qubit gate spanning the split becomes an all-to-all exchange.
At that point the simulation is interconnect-bound, not FLOP-bound. We size these runs the way we size any tightly coupled job — InfiniBand or RDMA over converged Ethernet, rank placement pinned to the topology, and NUMA-aware allocation so the transpose does not cross a socket twice.
GPU back-ends move the ceiling to HBM capacity and the NVLink domain. Multi-GPU distribution then behaves like a strong-scaling problem with a very unforgiving communication pattern, and GPUDirect matters more than clock speed.
| Qubits | Amplitudes | State vector | Practical host |
|---|---|---|---|
| 30 | 1.07 × 109 | 16 GiB | One workstation |
| 34 | 1.72 × 1010 | 256 GiB | One dual-socket node |
| 36 | 6.87 × 1010 | 1 TiB | Large-memory node |
| 40 | 1.10 × 1012 | 16 TiB | 16+ nodes, RDMA fabric |
| 45 | 3.52 × 1013 | 512 TiB | Full cluster, communication-bound |
| 50 | 1.13 × 1015 | 16 PiB | Out of reach by this method |
Matrix-product-state and contraction methods simulate hundreds of qubits when entanglement stays low and circuits stay shallow. Raise the entanglement and the bond dimension grows until the method collapses. A qubit count quoted without the circuit's structure and depth tells you nothing at all.
Simulating a realistic device means density matrices, not state vectors, and the memory requirement squares. A 20-qubit noisy simulation can cost more than a 40-qubit ideal one. We size for the noise model you actually need, and often the answer is a sampled trajectory approach instead.
Development Environments
We stand up environments your researchers can use on Monday — pinned toolchains, reproducible containers, simulator back-ends sized to the questions being asked, and continuous integration that executes circuits rather than merely linting them.
Qiskit, Cirq and PennyLane each have gravity in different research communities, and forcing a single framework on a mixed team wastes months. We containerise all three against a shared simulator layer and a shared result store, so a comparison between them is a configuration change rather than a rewrite.
Simulator jobs are submitted through Slurm alongside the rest of your HPC work, with the same accounting and the same fair-share rules. The notebook and API tier runs on Kubernetes so a researcher gets an interactive session without a ticket. Seeds, shot counts and calibration snapshots are recorded with every run, because a quantum result without its metadata cannot be reproduced.
Every commit runs the circuit against a small simulator and asserts on more than correctness. Transpiled depth, two-qubit gate count and estimated shot cost are recorded as build outputs, and a regression in any of them fails the pipeline. Numerical results are checked against analytic values where they exist.
QPU time is metered and queues are unpredictable. Each submission carries a shot budget and an estimated spend, checked against the project allocation before the job leaves the building. Jobs that would exceed it are refused with a message, not discovered on an invoice a month later.
We do not own quantum hardware and will not pretend to. Access runs through commercial brokers to superconducting, trapped-ion and neutral-atom devices via one submission layer, with queue time, calibration state and cost logged per run so the comparison between back-ends is evidence rather than anecdote.
The deliverable includes your team running the environment without us. Paired sessions on encoding and transpilation, a runbook for the simulator tier, and a written assessment of which of your problems are worth revisiting when hardware improves — and which are not, with the reasoning kept.
Problem Classes
These are the categories that appear in every quantum business case brought to us. The assessment column is our current engineering view. We will change it when the evidence changes, and we will tell you when we do.
| Problem class | The usual claim | Our assessment | What we would do this year |
|---|---|---|---|
| Combinatorial optimisation | QAOA and annealing beat classical solvers on scheduling, routing and portfolio problems. | No. On every benchmark we have reproduced, a tuned classical solver — CP-SAT, a good MIP formulation, or simulated annealing — wins on both wall clock and solution quality. | Fix the classical formulation. That is usually where the order-of-magnitude sits. Keep a QAOA branch as a research track with a fixed, small budget. |
| Quantum chemistry | VQE will give ground-state energies beyond the reach of coupled cluster. | The most scientifically credible class, and still not useful at industrial system sizes. Measurement overhead and noise dominate past a handful of active orbitals. | Build the DFT and CCSD(T) pipeline properly first. Use simulators to develop ansatz design and active-space selection so the science is ready when devices are. |
| Sampling | Quantum devices sample from distributions no classical machine can reproduce. | The strongest evidence in the field, attached to the least useful problems. Demonstrations use circuits chosen because they are hard to simulate, not because anyone needs the output. | Treat as science worth following, not a budget line. Revisit if a sampling task with a real application appears. |
| Quantum machine learning | Quantum kernels and parameterised circuit models outperform classical models. | No evidence at useful scale. Reported advantages have generally not survived comparison with a properly tuned classical model on the same data and the same budget. | Spend the money on data quality, feature engineering and classical training capacity. The return is larger and it is measurable. |
| Cryptanalysis | Shor's algorithm will break RSA and elliptic-curve cryptography. | Correct in principle, and the reason for the next section. It needs millions of physical qubits and error correction nobody has built. The timing is unknown; the eventual outcome is not. | Migrate now. This is the one item on the page with a deadline, and the deadline is set by how long your data must stay confidential. |
Post-Quantum Cryptography
A quantum computer does not need to exist for this risk to be live. An adversary records your encrypted traffic today and decrypts it when the hardware arrives. For anything with a twenty-year confidentiality requirement, the decision has already been made on your behalf.
Harvest now, decrypt later is not a future threat. The harvesting is in the present tense. Only the decryption is in the future.
The migration is a multi-year programme for any estate of consequence, and the hard part is not the algorithms. It is discovery. Most organisations cannot list where they perform asymmetric cryptography, which means they cannot scope the work, cost it, or tell a regulator when it will be finished.
So we start with an inventory and an exposure ranking, not a product. The register that comes out of it is useful even if the timeline slips, because it is the same artefact you need for certificate lifecycle management, for Essential Eight patching evidence, and for any credible audit of key custody.
You cannot migrate what you cannot find. We inventory every place the estate performs asymmetric cryptography: TLS terminators and load balancers, VPN concentrators, code-signing chains, HSM key ceremonies, database wrapping keys, S/MIME, MQTT and industrial control certificates, firmware with keys compiled in, and the appliance nobody holds credentials for.
Output — register with owner, key type, lifetime, renewal path
Rank each data flow by how long its confidentiality must hold. Traffic whose value expires in a week is not urgent. Material with a twenty-five-year classification, health and genomic records, source code for long-lived platforms, and legal holds are urgent today, because a copy taken this year stays readable to whoever eventually acquires the capability.
Output — ranked flows, with the urgent set named
Key establishment moves to ML-KEM, signatures to ML-DSA, with SLH-DSA where a conservative hash-based signature is warranted. In practice we deploy hybrid key exchange first, pairing X25519 with ML-KEM, so a weakness in either component does not cost you the session. Signature migration follows, and it is the slower half because trust anchors move at the speed of their vendors.
Output — sequenced migration plan with rollback at each step
The purpose of the exercise is to be able to do it again. Algorithms behind an interface, key types as configuration, certificate lifetimes short enough that rotation is routine rather than an incident. Post-quantum keys and signatures are substantially larger, which changes handshake sizes, packet fragmentation and embedded flash budgets. Test that on the constrained devices first.
Output — agility standard, plus the size and latency evidence
Engagement
Each is fixed in scope and ends in a written document you can take to a board, a research committee or a regulator — including the parts where our answer is that you should not proceed.
Cryptographic inventory, exposure ranking against confidentiality lifetimes, and a sequenced migration plan with owners. The one item on this page we would start this quarter regardless of your view on quantum computing.
Start here
A sized simulator tier on your HPC estate or ours, the three major frameworks containerised against it, Slurm and Kubernetes integration, circuit CI, and a runbook your team owns at handover.
Typically 6–10 weeks
One candidate problem, encoded properly, simulated at the largest honest scale, and benchmarked against the best classical solver we can build for it. The report says whether to continue. Most of ours say not yet.
Fixed scope, fixed price
Almost certainly not. Ownership brings a cryogenic or vacuum facility, a calibration team and a device that will be obsolete before it is useful. The exception is an organisation whose actual business is quantum hardware research, and those clients are not asking us this question.
The cryptographic inventory. It is the only part of this page with a real deadline, the work is entirely classical, and the register it produces pays for itself in certificate management alone. If budget allows one thing, it is this.
The question has no answer in that form. A count without circuit depth, two-qubit gate fidelity, connectivity and measurement error tells you nothing. Fifty noisy physical qubits with poor connectivity are worth less than twenty well-connected ones, and both are worth less than a good classical solver.
Through a broker, as a research exercise, yes. Not for a production result, and not for anything that cannot lawfully leave the country. Expect queue waits measured in hours, results dominated by noise, and a classical baseline that finishes first. That outcome is still worth documenting.
Then the work still holds. Your durable assets are the problem encoding, the classical baseline, the CI pipeline and the people who understand all three. Swapping in a better back-end is a configuration change, which is the whole reason we insist on that boundary from the first commit.
No. We hold no quantum hardware reseller agreement and take no margin on brokered QPU time, which is why we can tell you that the simulator is the better purchase. Where a broker subscription makes sense, you contract it directly and we integrate against it.
Let's Talk
Send the problem you were told needs a quantum computer. We will encode it, simulate it at the largest scale the arithmetic allows, benchmark it against the best classical solver we can build, and tell you which one wins. If it is the classical one, you have saved a programme budget.
Cleared personnel available for classified discussions. Simulation stays in-country by default.