10 Startups Making Waves in the Quantum Computing Industry

Quantum computing is sprinting out of the lab and into boardrooms, data centers, and national strategies. If you’re a technology leader, researcher, investor, or policy planner trying to separate real momentum from hype, understanding the startups shaping the field is your best shortcut. In this guide, we spotlight 10 startups making waves in the quantum computing industry, explain why each matters, and show you how to engage them effectively—from prerequisites and step-by-step pilot plans to metrics, pitfalls, and a mini-plan you can run this quarter.

Key takeaways

• Hardware and software are maturing in parallel. Neutral-atom, photonic, and trapped-ion startups are scaling while control, error suppression, and orchestration vendors help you get value sooner.
• You can pilot now—safely and affordably. Access systems via cloud marketplaces and vendor sandboxes, then measure progress with concrete KPIs like circuit depth, fidelity, and total cost per solved instance.
• Start small, design for learning. Use 2–4 week proofs-of-concept, clear baselines against classical methods, and “graduation criteria” to decide whether to scale, pivot, or pause.
• Avoid common traps. Don’t over-index on qubit counts; optimize for use-case fit, error rates, queue time, and the strength of the vendor’s ecosystem and roadmap.
• A portfolio approach wins. Combine one hardware pilot with one performance stack and one design-automation stack to balance risk and learning speed.

PsiQuantum (photonic, fault tolerance first)

What it is & core benefits
PsiQuantum pursues fault-tolerant photonic quantum computing, betting on silicon photonics and large-scale error correction to reach “utility-scale” performance. A landmark public-private partnership will build a facility near Brisbane Airport with a government package of roughly AUD $940M and an aggressive operational target date. That scale—and the manufacturing approach—aims at millions of physical qubits and error-corrected logical qubits suitable for chemistry, materials, optimization, and energy problems.

Requirements & low-cost alternatives

• Prereqs: Cloud access (for program design and simulation), basic quantum programming proficiency (Python), and budget for vendor collaboration or applied research projects.
• Low-cost alternatives: If you’re not ready for a facility-scale engagement, prototype your algorithms on photonic simulators and hardware-agnostic SDKs; validate noise sensitivity and scaling assumptions before moving to hardware.

Step-by-step for a beginner pilot

1. Define a materials or optimization use case where noise and connectivity matter (e.g., variational chemistry or constrained optimization).
2. Build and verify a classical baseline (exact solver or heuristic).
3. Implement a photonics-friendly algorithmic variant (Gaussian boson, linear optics mapping, or error-aware ansatz) in a vendor-agnostic framework.
4. Co-design with PsiQuantum’s team (or partners) to stress-test error-correction assumptions and resource estimates.

Beginner modifications & progressions

• Simplify: Start with small-mode photonic circuits and short depth; evaluate classical post-processing boosts.
• Scale up: Incorporate error-correction resource modeling; move to larger circuits and hybrid workflows.

Recommended frequency/metrics

• Monthly progress reviews.
• KPIs: resource estimates (photons, modes), error budget, expected logical qubits, speedup projections vs. classical.

Safety, caveats & mistakes

• Don’t treat a long-term fault-tolerance roadmap as a near-term SLA.
• Validate claims with explicit logical-qubit and error-correction overhead models.

Mini-plan (example)

• Week 1: Lock a chemistry or grid-optimization target with classical baselines.
• Week 2: Prototype a photonics-compatible circuit and estimate error-correction overhead with vendor input.

Atom Computing (neutral atoms, rapid scaling)

What it is & core benefits
Atom Computing builds gate-based neutral-atom systems using optically trapped atoms. Its next-gen platform has demonstrated a 1,225-site array populated with ~1,180 qubits—an early signal of neutral-atom scalability and dense connectivity.

Requirements & low-cost alternatives

• Prereqs: AWS/Azure credits or vendor access; Python stack (Jupyter, standard quantum SDKs); basic understanding of neutral-atom topology and pulse-level constraints.
• Low-cost alternatives: Start on public datasets and emulator tiers; borrow ready-made ansätze for QAOA/VQE.

Step-by-step for a beginner pilot

1. Choose a problem family that benefits from all-to-all or reconfigurable connectivity (MaxCut variants, routing, clustering).
2. Run on a simulator to find parameter regimes (depth, mixers).
3. Submit small instances to a neutral-atom backend (or queue) and capture real device metrics (success probability, two-qubit error, queue/wait).
4. Iterate with error-aware recompilation.

Beginner modifications & progressions

• Simplify: Low-depth circuits and small graphs.
• Scale: Add mid-circuit measurement (if available), constrained optimization with penalties, and layer-wise learning.

Recommended frequency/metrics

• Weekly runs; track depth vs. fidelity, throughput (jobs/week), and cost per successful solve.

Safety, caveats & mistakes

• Don’t equate “more qubits” with “more useful” without fidelity and depth analysis.
• Confirm device calibration recency before large jobs.

Mini-plan (example)

• Day 1–3: Baseline MaxCut 50–100 nodes on classical heuristics.
• Day 4–7: Port to QAOA, tune on simulator, then submit to neutral-atom hardware for 10–30 node graphs.

QuEra (neutral atoms, analog & digital on the cloud)

What it is & core benefits
QuEra operates Aquila, a 256-qubit neutral-atom machine offering analog Hamiltonian simulation and supporting digital modes—publicly accessible via a major cloud marketplace. The analog mode is valuable for simulating spin models and combinatorial problems; the SDK roadmap includes distinct paths for analog and digital.

Requirements & low-cost alternatives

• Prereqs: AWS account if using the cloud channel, Python, and basic analog-simulation concepts (Ising, Rydberg blockade).
• Low-cost alternatives: Use QuEra’s emulation in their SDK to validate programs before submitting jobs.

Step-by-step for a beginner pilot

1. Map your problem to an Ising-type Hamiltonian (e.g., MIS, MaxCut, lattice models).
2. Prototype with the SDK’s analog builder; validate parameters (detuning, Rabi rate, schedule).
3. Execute small instances on Aquila; compare outcome distributions to classical samplers.
4. Iterate with schedule shaping and constraint encoding.

Beginner modifications & progressions

• Simplify: Small lattices and shallow schedules.
• Scale: Move to programmable geometries and hybrid analog-digital approaches.

Recommended frequency/metrics

• Run biweekly; monitor empirical approximation ratios, sampling quality (TV distance), and queue latency.

Safety, caveats & mistakes

• Analog ≠ “approximate by default”; define acceptance criteria before sampling.
• Carefully encode constraints; many mistakes stem from Hamiltonian mis-specification.

Mini-plan (example)

• Sprint 1: 2D lattice Max Independent Set (20–40 sites) with schedule sweeps.
• Sprint 2: Add problem-specific penalties; compare to simulated annealing.

Pasqal (neutral atoms, European HPC integration)

What it is & core benefits
Pasqal builds neutral-atom systems and is integrating with European HPC initiatives. Under a EuroHPC contract, Pasqal will deliver a neutral-atom system in Italy starting with at least 140 qubits in analog mode, with an upgrade path to hybrid analog/digital by 2027—key for more sophisticated algorithms and HPC coupling. PasqalData Center Dynamics

Requirements & low-cost alternatives

• Prereqs: Access to European HPC centers or cloud partners; team familiar with physics-inspired algorithms.
• Low-cost alternatives: Begin with Pasqal emulators and academic examples to validate Hamiltonian mappings.

Step-by-step for a beginner pilot

1. Choose a simulation or optimization problem appropriate for analog mode (e.g., lattice models, logistics networks).
2. Validate an energy landscape classically (small instances).
3. Run analog schedules with parameter sweeps; record distributions and energies.
4. Plan migration to hybrid analog/digital when available.

Beginner modifications & progressions

• Simplify: Low-dimensional lattices and short schedules.
• Scale: Plug into HPC pipelines (pre/post-processing, ML-in-the-loop).

Recommended frequency/metrics

• Monthly cycles; track solution quality vs. classical heuristics, time-to-result, and cost per accepted sample.

Safety, caveats & mistakes

• Don’t assume analog sampling solves NP-hard problems reliably; combine with problem structure and heuristics.
• Manage queue time around facility maintenance windows.

Mini-plan (example)

• Month 1: Validate two analog schedules on a 30-node graph.
• Month 2: Integrate HPC preprocessing (graph sparsification), compare performance.

Xanadu (photonic hardware + PennyLane software)

What it is & core benefits
Xanadu advances photonic quantum hardware and develops PennyLane, a popular open-source library for quantum ML and hybrid optimization. The company has publicized fundraising goals and reported new photonic system advances aimed at networking and scaling; earlier work demonstrated programmable photonic advantage claims with Borealis available via the cloud. BetaKitThe Quantum Insiderxanadu.ai

Requirements & low-cost alternatives

• Prereqs: Python/NumPy; familiarity with variational circuits and autodiff.
• Low-cost alternatives: Use PennyLane with classical simulators and plug-ins for different backends before touching hardware.

Step-by-step for a beginner pilot

1. Pick a hybrid ML or optimization task (e.g., feature embedding, VQE, VQCs).
2. Build in PennyLane with a differentiable pipeline; validate on synthetic data.
3. Swap simulator with a photonic backend when available; measure loss vs. depth.
4. Optimize with gradient-free methods if noise dominates.

Beginner modifications & progressions

• Simplify: Shallow ansätze and small datasets.
• Scale: Structured encodings, batching, and error-aware training loops.

Recommended frequency/metrics

• Weekly experiments; track validation loss, gradient stability, and wall-clock time.

Safety, caveats & mistakes

• Beware barren plateaus; monitor gradient norms and try layer-wise training.
• Keep a classical baseline to avoid over-crediting quantum effects.

Mini-plan (example)

• Week 1: Reproduce a PennyLane tutorial for hybrid classification.
• Week 2: Swap in a photonic simulator/backend; compare accuracy vs. classical SVM.

ORCA Computing (modular photonics, on-prem testbeds)

What it is & core benefits
ORCA builds modular, fiber-connected photonic systems emphasizing near-term utility and on-prem deployment. In 2025, ORCA delivered a photonic quantum system to the UK’s National Quantum Computing Centre (NQCC), fulfilling a public milestone and underlining its focus on hybrid quantum-classical workflows and rapid installation. Quantum Computing Report

Requirements & low-cost alternatives

• Prereqs: On-prem space (rack-level), photonics-aware IT support, and integration capacity for hybrid workflows.
• Low-cost alternatives: Use remote access or evaluation programs before committing to hardware delivery.

Step-by-step for a beginner pilot

1. Identify a near-term use case (network design, routing, anomaly detection) suited to sampling/optimization.
2. Prototype workflow with synthetic data; define SLAs (uptime, latency).
3. Run workloads on ORCA’s platform (remote or on-prem); measure time-to-first-solution and variance.
4. Integrate with your classical stack for pre/post-processing.

Beginner modifications & progressions

• Simplify: Start with a single photonic source and a narrow workload.
• Scale: Add additional sources/modules and hybrid ML loops.

Recommended frequency/metrics

• Monthly service-level reviews; track install time, usable device hours, and solution quality vs. classical heuristics.

Safety, caveats & mistakes

• Underestimating facility and integration needs.
• Assuming photonic systems require cryogenics (ORCA’s designs target practical operation).

Mini-plan (example)

• Phase 1: 2-week remote pilot on a routing dataset.
• Phase 2: On-prem trial in a test rack with hybrid pipeline.

Universal Quantum (trapped-ion silicon chips, modularity at scale)

What it is & core benefits
Universal Quantum, a University of Sussex spin-out, pursues trapped-ion processors fabricated with silicon microelectronics and modular interconnects. The company unveiled a commercial ASIC for integration into its iQPU, a step toward million-qubit architectures; earlier they secured a large European contract underscoring government confidence in the approach. The Quantum InsiderUniversity of Sussex

Requirements & low-cost alternatives

• Prereqs: Interest in long-horizon, error-corrected computing; algorithm and resource-estimation expertise.
• Low-cost alternatives: Focus on logical-qubit simulations and circuit compilers; validate error-correction resource counts for your use case.

Step-by-step for a beginner pilot

1. Build a logical-qubit resource model (code family, target logical error).
2. Map your application (chemistry/cryptography/optimization) into fault-tolerance costs.
3. Stress-test sensitivity to gate times and error rates; iterate with vendor roadmaps.

Beginner modifications & progressions

• Simplify: Small logical patches and toy workloads.
• Scale: Add surface-code patches and syndrome-extraction modeling.

Recommended frequency/metrics

• Quarterly reviews; track logical error rate, overhead (physical qubits/logical qubit), and projected time-to-solution.

Safety, caveats & mistakes

• Confusing near-term NISQ metrics with fault-tolerant planning.
• Not modeling I/O and interconnect bottlenecks alongside gate fidelity.

Mini-plan (example)

• Quarter 1: Complete an error-correction budget for one application.
• Quarter 2: Validate with updated device parameters and ASIC capabilities.

Classiq (high-level design, synthesis, and OS-like platform)

What it is & core benefits
Classiq provides a high-level, model-first development platform—including a quantum programming language (Qmod) and automated synthesis—that compiles intent into hardware-aware circuits. In May 2025, Classiq announced a $110M Series C (the largest reported round for a quantum-software firm to date), signaling the market’s appetite for design automation tools that scale with hardware progress. Classiq

Requirements & low-cost alternatives

• Prereqs: Python developers; target use cases with complex constraints (scheduling, finance risk, materials).
• Low-cost alternatives: Start with the free tier/evals, export circuits to your preferred backend, and compare to hand-written baselines.

Step-by-step for a beginner pilot

1. Specify the problem at a high level (constraints, objectives).
2. Use Classiq’s modeling to generate candidate circuits; review cost/fidelity estimates.
3. Run on simulators and then on your selected hardware via cloud connectors.
4. Iterate with synthesis options (depth reduction, compression).

Beginner modifications & progressions

• Simplify: Start with small DAGs and single-objective optimizations.
• Scale: Multi-objective formulations and dynamic constraint handling.

Recommended frequency/metrics

• Weekly synthesis iterations; track circuit depth/width, gate counts, and achieved fidelity.

Safety, caveats & mistakes

• Trust but verify: compare generated circuits against hand-tooled baselines.
• Don’t ignore hardware idiosyncrasies—ensure hardware-aware compilation is on.

Mini-plan (example)

• Week 1: Port an existing optimization model into Qmod.
• Week 2: Compile to two backends; pick the more stable for extended tests.

Q-CTRL (error suppression, performance management, education)

What it is & core benefits
Q-CTRL builds “quantum infrastructure software” that suppresses errors and improves circuit reliability. Fire Opal integrates with multiple hardware platforms (including IBM’s ecosystem and IonQ via a major cloud) and, as of June 2025, is available on Rigetti’s latest 84-qubit system—letting users improve results with minimal code changes. q-ctrl.comHPCwire

Requirements & low-cost alternatives

• Prereqs: Existing workloads on IBM/Rigetti/IonQ or analog systems; Python SDK familiarity.
• Low-cost alternatives: Try free or sandboxed tiers, apply built-in optimization recipes to benchmark circuits first.

Step-by-step for a beginner pilot

1. Select two to three representative circuits (e.g., QFT, QAOA, BV).
2. Run “vanilla” jobs on your target device; log accuracy and cost.
3. Re-run with Fire Opal’s error suppression; compare fidelity and time-to-result.
4. Integrate the best setting into your production notebooks.

Beginner modifications & progressions

• Simplify: Start with short-depth circuits.
• Scale: Add layer-wise learning, advanced decoupling, or AI-assisted gate tuning.

Recommended frequency/metrics

• Weekly regression runs; track fidelity gains, success probability, and $/correct answer.

Safety, caveats & mistakes

• Don’t generalize from a single benchmark; test across circuit families.
• Re-baseline after hardware upgrades or calibration changes.

Mini-plan (example)

• Day 1: Baseline three circuits on hardware of choice.
• Day 2–3: Enable Fire Opal and report fidelity, success-rate, and cost deltas.

Quantum Machines (control stack, orchestration, hybrid HPC)

What it is & core benefits
Quantum Machines (QM) provides the OPX control platforms and QUA language, plus tight integrations for hybrid quantum-classical workflows—including early access to NVIDIA DGX Quantum and new facilities that support advanced experiments. In February 2025, QM announced a $170M Series C, reflecting broad adoption across hardware makers. Quantum Machines

Requirements & low-cost alternatives

• Prereqs: Lab or on-prem environment; need for pulse-level control, custom experiments, or hybrid acceleration.
• Low-cost alternatives: Use cloud notebooks to prototype QUA programs; partner with a lab or vendor that already runs OPX.

Step-by-step for a beginner pilot

1. Identify a pulse-level or hybrid workflow bottleneck (latency, calibration time).
2. Write a minimal QUA program to control your experiment/simulator.
3. Profile latency and throughput with and without QM orchestration.
4. Integrate DGX-class accelerators for hybrid loops if available.

Beginner modifications & progressions

• Simplify: Use canned QUA examples and pre-tuned sequences.
• Scale: Multi-QPU orchestration and real-time feedback loops.

Recommended frequency/metrics

• Daily lab sprints; track experiment turnaround time, calibration hours saved, and achieved circuit depth at target fidelity.

Safety, caveats & mistakes

• Don’t skip change-management—control stacks are critical infrastructure.
• Profile end-to-end latency (control + cloud + queue), not just kernel time.

Mini-plan (example)

• Week 1: Reproduce a canonical Rabi/Ramsey experiment with OPX.
• Week 2: Add real-time feedback to cut calibration time 20–30%.

Quick-start checklist (use this before you pilot)

• Define one use case with a classical baseline and success criteria.
• Pick a trio: one hardware (Aquila/neutral-atom or photonic), one performance stack (Fire Opal), one design stack (Classiq or PennyLane).
• Secure access (cloud credits, accounts, NDAs if needed).
• Set KPIs: fidelity, circuit depth, queue time, cost per accepted solution, time-to-first-result.
• Plan a 2–4 week sprint with mid-sprint checkpoint and a go/no-go rubric.
• Document everything (seed, device ID, calibration version) to ensure reproducibility.

Troubleshooting & common pitfalls

• “Our quantum results are inconsistent.” Log device ID, calibration timestamp, and queue time. Re-run after enabling error suppression and compare.
• “We ran out of budget.” Cap shots and batch jobs; use simulators and emulators to prune parameter space before hitting hardware.
• “No improvement over classical.” Revisit problem mapping and constraints. Try analog formulations or different ansätze/encodings first. Physical Review
• “The SDK changed and our code broke.” Pin versions, use Docker/conda, and keep examples minimal.
• “We focused on qubits, not quality.” Track fidelity, error bars, and solution quality—not just counts.

How to measure progress or results

• Technical KPIs:
  • Circuit depth vs. achieved fidelity (per device).
  • Logical success probability for a fixed budget (shots).
  • Effective qubit-hours (paid device time × fraction of valid shots).
• Productivity KPIs:
  • Time-to-first-result (from job submit).
  • Queue latency, iteration speed (experiments per week).
  • Calibration overhead (hours/week). HPCwire
• Business KPIs:
  • Cost per accepted solution (vs. classical).
  • “Learning velocity”: number of validated hypotheses per sprint.
  • Readiness score to proceed to Phase 2 (data, model, and infra gates).

A simple 4-week starter plan

Week 1 — Scoping & baselines

• Select a single use case with measurable outcomes (e.g., MaxCut on a network design subset).
• Build classical baselines and define acceptance thresholds (accuracy, cost, time).

Week 2 — First hardware touch

• Implement the problem in one design stack (Classiq or PennyLane).
• Submit to one neutral-atom/photonic device for tiny instances; log fidelity, queue, and cost. PennyLane

Week 3 — Performance management

• Add error suppression (Fire Opal) and compare deltas.
• Run A/B tests across two hardware targets if possible.

Week 4 — Decision & roadmap

• Review KPIs and go/no-go: scale, pivot to a new mapping (analog vs. digital), or pause.
• If “go,” outline a 90-day plan with budget, team, and ecosystems (HPC, cloud).

FAQs

1. Is qubit count the most important metric when choosing a startup?
   No. Depth at usable fidelity, error rates, queue latency, ecosystem tooling, and cost per accepted solution usually matter more than raw qubit numbers. McKinsey & Company
2. What’s the difference between analog and digital (gate-based) modes?
   Analog directly simulates a target Hamiltonian through continuous evolution; digital composes discrete gates. Analog can be efficient for certain physics and optimization tasks, while digital is more general and aligns with fault-tolerant roadmaps.
3. Can I do anything useful today, or is everything years away?
   Yes—especially in benchmarking, optimization heuristics, simulation studies, and building internal capabilities. Performance stacks and design automation help translate small-scale wins into learning you can act on.
4. How do I keep costs under control?
   Use simulators to prune parameters, cap shots, run during off-peak windows, and adopt error suppression to reduce repeats. Track cost per accepted solution as your north-star metric.
5. Which modality should I pick: neutral atoms, photonics, or trapped ions?
   All have credible roadmaps. Neutral atoms show promise in scaling and flexible connectivity; photonics offers room-temperature operation and networking; trapped ions lead on gate fidelity and error-correction pathways. Portfolio piloting is sensible. ORCA Computinguniversalquantum.com
6. Do I need a PhD team to start?
   No. A small team with Python skills, domain knowledge, and patience for experimental iteration can pilot effectively using high-level tools and managed services. GlobeNewswire
7. How do I validate vendor claims?
   Request device IDs, calibration logs, and reproducible notebooks. Re-run benchmarks across weeks and platforms, and insist on classical baselines for every result. Amazon Web Services, Inc.
8. Should I build on-prem or stay in the cloud?
   Start in the cloud for speed and flexibility. Consider on-prem only when you need data locality, guaranteed latency, or specialized integration—some photonic vendors support rapid on-prem delivery. Data Center Dynamics
9. What’s the credible horizon for fault-tolerant systems?
   Timelines vary. Some photonic and trapped-ion roadmaps target late-decade milestones, with government-backed projects underway to accelerate progress; plan pilots that create value regardless of exact FT dates.
10. Is the investment environment healthy for quantum software, or only for hardware?
    Software is attracting record rounds, including large 2025 raises that validate the design-automation and orchestration layers. Wall Street Journal

Conclusion

A few years ago, quantum pilots mostly produced white papers. Today, they produce playbooks. The companies above—spanning hardware, design automation, performance management, and orchestration—let you start small, learn fast, and build a credible roadmap to scale when the hardware does. Treat your first proof-of-concept as an investment in capability, not just an experiment, and measure relentlessly.

Ready to kick off? Pick one hardware, one performance stack, and one design stack—and run a 4-week pilot with clear KPIs and a go/no-go decision at the end.

References

• PsiQuantum to Build World’s First Utility-Scale, Fault-Tolerant Quantum Computer in Australia, PsiQuantum, May 22, 2024, https://www.psiquantum.com/news-import/psiquantum-to-build-worlds-first-utility-scale-fault-tolerant-quantum-computer-in-australia
• PM’s $940m quantum bet on tech that ‘doesn’t work yet’, The Australian, Sep 10, 2024, https://www.theaustralian.com.au/nation/politics/anthony-albaneses-940m-quantum-bet-on-tech-that-doesnt-work-yet/news-story/3e499edf50249e6288a345e4b2ee57d7
• Brisbane could become computing, quantum physics capital of the world, The Courier-Mail, May 2025, https://www.couriermail.com.au/news/queensland/qld-audit-office-report-on-debt-reveals-quantum-computer-investment/news-story/255b3b70df58874a07f652a38d2396bc
• Quantum Startup Atom Computing First to Exceed 1,000 Qubits, PR Newswire, Oct 24, 2023, https://www.prnewswire.com/news-releases/quantum-startup-atom-computing-first-to-exceed-1-000-qubits-301964712.html
• Atom Computing Wins the Race to 1000 Qubits, HPCwire, Oct 24, 2023, https://www.hpcwire.com/2023/10/24/atom-computing-wins-the-race-to-1000-qubits/
• Atom Computing Reports Advance in Scaling Up Neutral-Atom Qubit Arrays, HPCwire, Feb 15, 2024, https://www.hpcwire.com/2024/02/15/atom-computing-reports-advance-in-scaling-up-neutral-atom-qubit-arrays/
• What to do while pursuing the promise of quantum computing, Brookings Institution, Jul 16, 2025, https://www.brookings.edu/articles/what-to-do-while-pursuing-the-promise-of-quantum-computing/
• Amazon Braket launches Aquila, the first neutral-atom quantum processor from QuEra Computing, AWS, Nov 1, 2022, https://aws.amazon.com/blogs/quantum-computing/amazon-braket-launches-aquila-the-first-neutral-atom-quantum-processor-from-quera-computing/
• Aquila, Our 256-qubit Analog Quantum Computer, QuEra, accessed Aug 13, 2025, https://www.quera.com/aquila
• 2025 — The Neutral Atom SDK (Bloqade), QuEra, Mar 1, 2025, https://bloqade.quera.com/v0.22.1/blog/archive/2025/