Google’s Quantum Echoes Beats Supercomputers by 13,000x

Google scientists have developed a novel algorithm capable of solving problems on a quantum processor 13,000 times faster than the world’s speediest supercomputers— a leap they say brings quantum computing closer to real-world use in drug discovery, materials science, and other scientific fields.​ Named Quantum Echoes, the algorithm marks a key breakthrough: it not only achieves “quantum advantage” (outperforming classical systems) but is also the first such algorithm verifiable independently by running it on another quantum computer.​

Its ultra-fast performance was demonstrated in benchmark tests on Google’s Willow quantum processing unit (QPU). The researchers detailed the algorithm’s mechanics in a study published October 22 in the journal Nature.​ “Quantum algorithms guide quantum computers to solve problems efficiently, much like software drives classical computing,” Xiao Mi, a Google Quantum AI research scientist who led the work, told Live Science via email. “For either classical or quantum computing to solve future problems, both software and hardware must exist and work in tandem.”​

While the first study proved Quantum Echoes’ quantum advantage, the team also sought to show its practical utility. In a second study—published the same day (October 22) in the arXiv preprint database—they designed a quantum circuit to replicate molecular dynamics as observed in nuclear magnetic resonance (NMR) spectroscopy labs.​ Through this, they uncovered previously unknown details about the atomic spacing and structures of two molecules: [4-¹³C]-toluene (15 atoms) and [1-¹³C]-3′,5′-dimethylbiphenyl (DMBP, 28 atoms).​

The experiment used a small system (15 qubits), but the team noted in the study that future work will enable simulating molecules four times larger—a scale beyond the reach of classical computer simulations.

Allocate compute cost at the same level as revenue

AI products often report one blended cloud bill while monetization happens by feature, customer or modality. That makes it difficult to see which usage is profitable. Tag inference and generation cost by product path so revenue can be compared with the resources consumed to earn it. Text, voice, images and video have very different cost curves, and even two text workloads can differ dramatically because of context length and output volume.

Use successful outcomes as the denominator

Cost per token is an infrastructure metric, not always a business metric. For agent workflows, the more useful measure may be cost per completed task; for image products, cost per accepted generation; for support, cost per resolved conversation. Retries, abandoned generations and moderation failures belong in the numerator because they consume compute without producing the intended user value.

Model utilization and pricing together

Reserved capacity can lower unit cost at high utilization but create expensive idle time when demand is uncertain. Serverless or on-demand infrastructure may have a higher nominal rate but lower commitment risk. Forecasts should compare realistic utilization scenarios rather than list price alone.

Forecast cost variance by workload driver

Cloud cost moves when model mix, context size, output length, concurrency or media usage changes. A forecast that grows total spend only with user count will miss shifts in product behavior. Build driver-based scenarios so finance and engineering can explain whether a variance came from more users, heavier usage, a new model or an efficiency regression.

Separate free credits from sustainable unit economics

Cloud credits reduce cash burn temporarily but do not change the production cost of the product after the subsidy expires. Track both cash spend and normalized infrastructure cost so runway does not look artificially strong during a promotional period. This is especially important before fundraising or pricing decisions because the post-credit margin may be very different.

Use cost budgets as product constraints

Teams can assign target cost per active user, task or premium feature and monitor when a launch moves outside the budget. This makes infrastructure efficiency visible early instead of waiting for the monthly cloud invoice.

Separate fixed capacity from variable consumption

Some AI costs behave like committed capacity while others scale nearly linearly with requests. Finance models should distinguish the two because gross margin responds differently to growth and demand volatility. Committed GPUs may create strong margins at high utilization but poor margins during slow periods, while API-based inference shifts more cost directly with usage.

Track quality-adjusted efficiency

A cheaper model is not more efficient if it increases retries, support tickets or user abandonment. Compare cost changes with quality and completion metrics so optimization preserves the product outcome. Routing can often produce better economics than a single-model strategy by using expensive models only where they materially improve results.

Include support and moderation in variable cost where appropriate

Some AI use cases create human review or support burden that scales with activity. If those costs are excluded from unit economics, a product can appear profitable while operational workload grows faster than revenue.