NEC Exits Hardware, Leaving Behind Quantum Annealing Accounts

Let's look at some data from the reports first. In March this year, NEC halted R&D on actual quantum computer hardware; the company is still exploring quantum computing technologies and services; the focus has shifted to quantum annealing; relevant researchers have moved to Fujitsu; stock prices dropped. Putting these facts together, it looks more like a budget reallocation.

According to Nikkei Asia and other reports, NEC has stopped R&D on physical quantum computers but will continue to explore quantum computing technologies and services.

I work in computer vision and often tell students: "Is this direction good for publishing papers?" When funding tightens, the question becomes: "Can we survive the mid-term review?" NEC's exit from general-purpose quantum hardware has a very simple reason: the return cycle is too long. The academic world fears this kind of cycle too. If a project takes three to five years just to produce reproducible results, with equipment purchases, cryogenic maintenance, hiring, paper writing, and patent filings in between, many groups get drained dry.

It helps to view quantum computing as two distinct paths. General-purpose quantum hardware pursues error correction, scalability, and the ability to run arbitrary algorithms.

Quantum annealing is more like a specialized service, solving specific optimization problems to deliver an acceptable result to clients. Both are called quantum computing, but their experimental designs are completely different.

General hardware is like building new foundation models in CV. Reviewers ask: Are the baselines sufficient? How is noise calculated? Is the scale reproducible? Is there a practical advantage? I raise these points often in my reviews. Quantum hardware is even trickier, involving materials, superconductivity, cryogenics, control electronics, and error correction—software tuning is just one layer. Farhi et al.'s 2001 paper on quantum annealing provided a physics-inspired path, while Preskill's 2018 proposal of NISQ acknowledged that devices are noisy and limited in scale. NISQ left a window for research but also a question mark for business.

The annealing route looks more like a deliverable optimization service. Reports say NEC continues to develop and apply quantum annealing, having previously attempted to use vector annealing to improve IT equipment maintenance parts delivery. The entry point is small but critical. It doesn't need to answer when the universal quantum computer will arrive; it only needs to answer whether a specific scheduling, combinatorial optimization, or maintenance routing problem can be solved with lower cost and time. For units with tight budgets, the problem boundaries are friendlier.

NEC's adjustment was foreshadowed. Reports mention they've been doing quantum computing since the 1990s and have published progress on superconducting tech. They understand hardware and know how hard it is. A common phenomenon in academia is that the closer you get to the bottom layer, the easier it is to publish papers, but the harder it is to engineer solutions. Long-term hardware pursuit often yields results in experimental conditions, device metrics, and system demos; customers care more about whether results are reusable, if services can go live, and if costs can be calculated clearly.

Talent flow also tells the story. Japanese media reported that NEC's quantum computer researchers are moving to Fujitsu. This resembles lab recruitment: when funding is low, scattered directions are the biggest risk. Losing one person might mean losing a whole line. By absorbing these talents, Fujitsu is, in the short term, concentrating equipment, algorithms, and customer validation. In scientific research, this is called resource restructuring; in business, it's called cutting losses.

I wrote an article recently about computing power procurement, with the core idea being: if you're broke, owe the most troublesome debt first. Quantum hardware is similar; the chip is just the surface, the real pressure is long-cycle cash flow. General hardware is like a road without an end: today solve qubit count, tomorrow coherence time, the day after error correction and compilation. Annealing is more like a project-based model, delivering per problem. For listed companies, budgets are easier to approve; for labs, students can have writable papers and runnable experiments.

However, annealing isn't guaranteed to win. Whether it brings stable advantages over classical algorithms still depends on reliable experimental design. Good experiments can't just show pretty cases; they need public benchmarks, control methods, randomness handling, and cost tables. Reviewers often ask if it's cherry-picking. Specialized quantum services fear this too. Proving superiority with just a few small instances might get a paper published, but business may not hold up.

My judgment is that NEC has swapped its "hardware dream" for an "annealing ledger." It retains technical capabilities but no longer ties the company's fate to general hardware. For enterprises, this is financial discipline. For research, it's a reminder: frontier tech should be viewed in layers; bottom-layer breakthroughs deserve respect, but most organizations are better suited for middle and application layers. CV has followed this pattern too; those who changed the industry were the people doing data, evaluation, engineering, and scenarios who brought models down to earth.

When choosing AI, quantum, LLM, or research directions, don't first ask "Is this direction good for publishing papers?" Look first at "What reproducible results can be delivered next quarter?" Break the roadmap into three milestones: a publicly runnable benchmark, a reproducible experiment, and a cost table. If the hardware route can't pass these three gates, pivot to services, tools, or data. When funding is tight, a narrow verifiable path is worth more than a grand broad road.