Explaining Why D-Wave’s Quantum AI Play Is a Dead End
Explaining Why D-Wave’s Quantum AI Play Is a Dead End
D-Wave has spent years positioning itself as a Quantum AI company. CEO Alan Baratz regularly tells investors that quantum annealing can train AI models faster, more accurately, and with lower energy consumption than classical computers. On earnings calls, he speaks confidently about “exploring generative AI architectures” and “the synergies and benefits of these complementary technologies.” The pitch sounds compelling: while competitors promise quantum breakthroughs sometime in the next decade, D-Wave claims to be delivering AI value today. But the technical reality tells a different story. D-Wave’s only AI project is a single proof-of-concept completed nearly two years ago that never progressed beyond the experimental phase. The underlying technology targets a legacy AI architecture the industry abandoned years ago. And even if the approach worked exactly as advertised, it addresses such a small piece of the AI pipeline that there is no meaningful revenue opportunity.
Japan Tobacco: Their Only AI Project
When D-Wave discusses its Quantum AI vision, it can only point to a single successful project: a drug discovery collaboration with Japan Tobacco International completed nearly two years ago. This is not one success among many. It is the only AI project D-Wave can credibly cite, and there has been no follow-on work, no expansion, and no second customer.
Note: D-Wave does discuss a spin glass project, but this project was never considered successful by insiders associated with the project.
The project itself was a proof-of-concept for generating novel small molecules that could eventually become pharmaceuticals. The AI system was a hybrid architecture combining three components: a Discrete Variational Autoencoder (DVAE) to represent molecular structures, a Transformer to handle the actual generation of new molecules, and a Graph-Restricted Boltzmann Machine (GRBM) embedded in the middle. D-Wave’s contribution was using their quantum processor to train just the GRBM component during one phase of the process. Everything else, including the DVAE, the Transformer, and the overall generative logic, ran on classical hardware (GPUs and silicon-based compute).
Japan Tobacco declared the project a success… but without any evidence. No quantitative benchmarks exist demonstrating that the quantum processor actually outperformed classical sampling methods on the same task. D-Wave cannot independently verify or reproduce the claimed results – an exercise that the founder of Coherence Report himself wrestled with while trying to market the approach. The company simply has a client who says it worked, with no documented basis for that assessment that anyone outside the project can evaluate. And it’s clear that no one at D-Wave pushed for empirical validation of the results – which is telling in itself.
Moreover, there has been no production deployment at Japan Tobacco, and no expansion of scope. No other AI customers have followed, despite a frenzy of demand for other quantum AI providers (using gate logic technology). And yet, CEO Alan Baratz continues to reference this single, two-year-old proof-of-concept in nearly every earnings call and public appearance as evidence that D-Wave’s Quantum AI strategy is working. When your flagship AI success story is a single PoC from 2024 that never went anywhere, and you are still talking about it in 2026, that is not momentum. That is a company with nothing new to say.
Why This Approach Will Not Scale
The problem is that there are significant technical limitations. In order to use the QPU, D-Wave and Japan Tobacco had to implement an older model called an RBM – Restricted Boltzmann Machine. These were popular at one time, but have been replaced by Transformers and Variational Autoencoders because they struggle with the massive, multi-dimensional datasets modern AI requires. D-Wave’s hardware only works with binary variables, meaning each input must be reduced to a yes-or-no, 0-or-1 decision. Modern neural networks operate on continuous gradients, making tiny adjustments across billions of parameters during training. There is no natural way to map large language model training onto binary energy minimization. Meanwhile, classical computing already handles discrete variables efficiently. Techniques like Vector Quantization (used in the original DALL-E) and mathematical shortcuts like the Gumbel-Softmax trick allow standard GPUs to do this work without quantum hardware. The industry leaders in AI-driven drug discovery, such as DeepMind’s AlphaFold and NVIDIA’s BioNeMo, skip RBMs entirely, using diffusion models and Transformers scaled across thousands of GPUs. D-Wave’s quantum annealer will never be able to integrate into these technologies.
The Commercial Dead End
Even setting aside the technical limitations, the Japan Tobacco project reveals a fundamental problem with D-Wave’s AI revenue story. The quantum component was an insignificant piece of a much larger project. The Transformer did the generation, and the DVAE handled the molecular representation. Classical hardware ran the show.
The work done by the QPU is easily skippable. Swapping in a quantum processor for this one sub-task does not change the fundamental economics of the pipeline, nor does it offer D-Wave a meaningful cloud revenue opportunity. Training an RBM is a one-time or infrequent task, not an ongoing computational workload. It does not generate the continuous cloud usage that D-Wave’s business model requires. A client might buy a few hours of quantum processing time for one research project, but there is no path to the high-margin recurring revenue D-Wave has promised investors. The company is not building the future of AI. It is attaching a quantum label to a technique the industry abandoned years ago, and hoping no one looks too closely at the only use case they have to show for it.