A six-qubit ESA testbed, a 150-qubit EuroHPC purchase, a neutral-atom roadmap in Illinois and an NSF-funded integration platform share a location strategy, not a common level of readiness.
Between July 8 and July 29, 2026, five organizations announced on-site quantum initiatives across Europe and North America. Taken together, they look like a procurement wave. Read closely, they are five different instruments with five different definitions of success.
Project | Public commitment | Public timing | What is disclosed |
|---|---|---|---|
CSC selected IQM to supply a system for the LUMI AI Factory | Autumn 2027 | A 150-physical-qubit Halocene H4 followed by phased upgrades | |
Infleqtion at the Illinois Quantum and Microelectronics Park | Vendor-announced deployment | 2027 | Designed to demonstrate more than 50 logical qubits, with an architecture intended to scale beyond 1,000 physical qubits |
A $5 million NSF-funded hybrid testbed with HPE and Rigetti | Full operations expected in 2027 | An integration and research platform, not a standalone quantum computer | |
An installed on-premises quantum system and research pilot | Pilot planned by the end of 2026 | Six silicon-spin qubits in a rack-mounted system | |
An on-site quantum-computing investment | Not disclosed | No public vendor, modality, qubit count or delivery date |
Three of the five projects have public 2027 milestones. ESA's machine is already installed. EY has not published a delivery year. The common thread is therefore not one delivery date, one contract model or one level of technical maturity. It is the decision to bring quantum hardware—or the infrastructure around it—under local operational control.
That does not contradict Gartner's forecast that enterprise AI workloads at scale will not run on quantum hardware through 2028. These buyers are not purchasing proven production AI acceleration. They are purchasing testbeds, integration experience, workforce development, data control and upgrade paths.
Bell-1 is the clearest example because ESA is unusually direct about what the system can and cannot do.
The Equal1 machine has six silicon-spin qubits, operates at about 0.3 kelvin and draws approximately 1.6 kilowatts. Its closed-cycle cooler and control system fit inside a rack-mounted chassis, and its qubits are fabricated using a CMOS manufacturing process similar to conventional chipmaking. That fabrication route is where silicon spin has gathered its strongest manufacturing evidence while its demonstrated qubit counts stayed small. ESA has installed it inside the ESRIN data center in Frascati and integrated it with the agency's high-performance computing environment.
ESA does not present Bell-1 as a production accelerator. Its Phi-lab describes current systems as small and hardware-limited and says that demonstrating quantum advantage for Earth-observation workloads remains a significant scientific challenge. The plan is to use Bell-1 for a year of internal research after commissioning, with a pilot by the end of 2026. Named workloads include hybrid quantum neural networks for land-use and land-cover classification and satellite mission planning.
That makes Bell-1 less ambitious than the larger 2027 systems and more legible as a purchase. ESA is buying a physical development environment: a real device, local access, known operating constraints and a short path from algorithm work to hardware tests. The value is not six qubits in isolation. It is the ability to discover which parts of a hybrid workflow fail when the simulator is replaced by an actual machine.
The public materials reviewed for this article do not include an executed LUMI-IQ contract. The most detailed public EuroHPC instrument is instead the tender package for EuroSSQ-HPC, a separate semiconductor-spin system for SURF in the Netherlands. It should not be treated as the Finnish contract, but it shows how EuroHPC can structure a quantum-hardware acquisition.
The procurement already sets baseline technical requirements for the first-generation system: at least 16 physical qubits, average two-qubit gate fidelity of 99.0%, and a relaxation time of at least 100 microseconds. Its draft supply contract also contemplates several commercially flexible terms:
The contract also disapplies its general price-reduction remedy for low-quality delivery. That does not leave the buyer without recourse. The liquidated-damages cap applies to that remedy for the relevant delivery; the contract separately preserves potential liability and termination rights.
The more important correction is on acceptance. Annex III of the draft contract contains placeholders for the detailed acceptance procedure, and Annex IV leaves the final milestone schedule to the winning offer. But the separate performance-benchmark annex already defines the acceptance architecture.
The listed benchmarks, together with additional benchmarks proposed by the bidder, become an integral part of acceptance. They must be run on the fully installed system under the hosting entity's supervision. The resulting protocol must demonstrate compliance with the performance values and properties committed in the winning offer. If the system does not comply, the supplier is expected to modify it until the tests prove the committed performance. Acceptance follows only after that evidence is recorded and signed.
The required tests include addressable-qubit count; single- and two-qubit gate fidelity and speed; readout fidelity and speed; and coherence measurements. Application-level tests include GHZ-state creation, a variational MaxCut workload and the Bernstein-Vazirani algorithm. At least one threshold is explicit: the GHZ-state benchmark requires fidelity greater than 0.5.
The procurement design is therefore more disciplined than either extreme suggests. EuroHPC does not prescribe every numerical value in advance, and it does not leave acceptance unwritten. The buyer defines the test framework. The bidder commits many of the numerical values. Those commitments then become acceptance obligations.
That is a rational way to buy a rapidly changing system. It allows competing modalities to make different, testable offers without pretending that one universal figure of merit already exists.
The harder problem is the headline number that sounds closest to useful computation: logical qubits. The term still means something different on every vendor roadmap, which is why the Department of Energy's own 2028 fault-tolerance target has required a separate specifications effort to establish which definition counts.
Infleqtion's Illinois announcement says its Sqale system is designed to demonstrate more than 50 logical qubits on a path toward a 100-logical-qubit target, using an architecture intended to scale beyond 1,000 physical qubits. The company's 2025 Form 10-K puts that statement in its proper category. As of December 2025, Infleqtion reported arrays of up to 1,600 trapped atoms, 99.73% two-qubit CZ-gate fidelity, and 12 demonstrated logical qubits. It targeted 30 logical qubits in 2026 and 100 by 2028, while warning that it might not reach the 2028 threshold on the expected schedule.
That does not invalidate the Illinois roadmap. It distinguishes a planned system capability from a demonstrated one.
A logical-qubit count is not a portable performance unit by itself. Two systems can report the same number while using different error-correcting codes, physical-qubit overheads, decoders, logical error rates, gate sets and operating depths. The code landscape has been expanding well beyond surface codes, which widens the distance between two identical-looking counts. A procurement that relies on logical qubits therefore needs to state at least:
Without those fields, a logical-qubit target is not a cross-platform acceptance unit.
IQM's public numbers are easier to test because the initial Halocene release is specified as a 150-physical-qubit system targeting 99.7% physical two-qubit gate fidelity. IQM says the platform supports research on up to five logical qubits and the implementation of Clifford gates. CSC has announced the 150-qubit system and phased upgrades, but the public record does not disclose the exact fidelity thresholds or logical-qubit conditions written into LUMI-IQ's acceptance terms.
Physical-qubit count is directly measurable, but it is not a sufficient measure of useful performance. Logical-qubit count is closer to the end goal, but it is not comparable without the rest of the error-correction specification. Serious procurement has to carry both layers.
The strongest case for on-premises quantum is not that today's QPUs will replace supercomputers. It is that useful quantum systems will have to become components of larger computing environments.
NVIDIA NVQLink specifies up to 400 Gb/s of GPU-QPU throughput, less than four microseconds of FPGA-to-GPU-to-FPGA latency, and 40 petaflops of FP4 classical compute. The point is not simply faster data movement. Calibration, decoding and feed-forward control can make the classical system part of the quantum machine's operating loop.
In one early demonstration, NVIDIA and Quantinuum used a GH200 Grace Hopper system to decode a qLDPC code that encoded eight logical qubits in 30 physical qubits. NVIDIA reported a median decoding time of 67 microseconds and a 5.4-fold reduction in error rate after three rounds of correction. That result is specific to the demonstrated hardware, code, and control loop; it is not a universal timing threshold. It nevertheless shows why tightly coupled classical compute is becoming part of QPU architecture rather than an external service. That is also what NVIDIA's Ising pitch was really about: the classical control plane, not the qubits, carries much of the near-term engineering.
Physical co-location is not mandatory for every hybrid workload. The relevant requirements are latency, determinism, bandwidth, security, and control-system design. For some error-correction and calibration loops, local integration can be decisive. For batch-style algorithm development, remote access may be entirely adequate.
Pittsburgh's TangleLab makes the institutional argument more clearly than the hardware argument. PSC says it is "not deploying a standalone quantum computer." The project is intended to be an open prototype for integrating quantum information processing with advanced cyberinfrastructure, supported by user onboarding, training, workshops and research consulting.
Simulators remain useful for developing algorithms and orchestration software. They do not reproduce every calibration failure, queueing constraint, controller behavior, maintenance event, or facility dependency of a physical system. A center that expects to operate quantum accelerators eventually has to learn those failure modes somewhere.
EY Canada's announcement is technically underspecified but strategically revealing. EY says the on-site system is intended to support sensitive workloads within Canadian borders and give users greater control over data location, management and access. No vendor, modality, qubit count or delivery date has been published, so the machine's capability cannot yet be evaluated. The buyer rationale can: ownership and locality are part of the product.
CSC presents LUMI-IQ as a 150-qubit starting point followed by upgrades that increase qubit count and error resilience. That makes it a platform purchase rather than a single frozen appliance. The durable assets are not only the first QPU. They include the integration layer, operating procedures, trained staff, application pipeline, and ability to absorb later hardware.
That is also why the five July announcements should not be judged against one near-term benchmark. ESA is buying a compact research instrument. PSC is building an integration environment. EY is emphasizing control and data residency. Infleqtion is committing to a roadmap. CSC is procuring a staged quantum-HPC platform.
The immediate test for this market is not quantum advantage. It is whether vendors can manufacture, install, and obtain customer acceptance for larger systems on schedule.
IQM's first-half 2026 filing said the company had sold 26 full-stack systems and delivered 17 since its founding. The Finnish customer agreement added €33 million to backlog, taking the total above €102.1 million as of 3 August. IQM is investing more than €40 million to double its Espoo cleanroom capacity and establish the ability to produce up to 30 full-stack systems annually. That is planned capacity, not current output.
The same filing says IQM's 2026 revenue outlook depends partly on the scheduled delivery and customer acceptance of its first 150-qubit system. It does not identify the customer.
The nearest public milestone is Euro-Q-Exa at the Leibniz Supercomputing Centre. EuroHPC announced in 2024 that the project would include a 54-qubit system in the second half of 2025 and a 150-qubit system by the end of 2026. The 54-qubit Radiance system was inaugurated in February 2026. The public difference between the announced window and the inauguration date does not establish a contractual delay because the delivery and acceptance dates have not been disclosed.
The 150-qubit milestone remains the cleaner test. It is not the Finnish machine, and its specifications and acceptance terms should not be assumed to match LUMI-IQ's. But a customer-accepted 150-qubit deployment by the end of 2026 would provide the first public evidence that IQM has crossed the scale on which its 2027 Finnish commitment depends.
The on-premises quantum wave is real. What it represents is narrower—and more useful—than a rush to production quantum advantage. Buyers are turning QPUs into managed infrastructure, forcing vendors to expose integration requirements, manufacturing capacity and acceptance metrics.
The first serious benchmark for this market is not whether a quantum computer beats a supercomputer. It is whether the supplier delivers the system it bid, proves the performance it committed and leaves the buyer with an environment that can improve after acceptance.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
