> There is no indication (not even from the theory side) that the D-Wave approach can achieve any speed-up over a classical computer.

You're wrong about the theory side of things[1, 2], by the way. The discussion around adiabatic quantum computing is weirdly politicized. It's true that D-Wave's computer isn't a universal quantum annealer; but in theory such a thing is equivalent to a gate-model computer.

[1] https://arxiv.org/pdf/1611.04471.pdf

[2] https://iopscience.iop.org/article/10.1088/0256-307X/35/11/1...

An adiabatic quantum computer is equivalent to the gate model. D-Wave do not have an adiabatic quantum computer, so there is indeed no reason, even theoretical, to believe their hardware provides computational advantage. The word "quantum annealer" is just a confusing term used differently by everyone, occasionally meaning "adiabatic quantum computer" and occasionally meaning what D-Wave have.

You sound very confident of that, but the fact that investors and developers keep moving to the platform shows that there's a good chance you're wrong.

From what I can tell, the new hybrid architecture that D-Wave rolled out recently can handle way bigger optimization problems than before, and slice and dice them to fit on the still-fairly-small quantum chip. As long as optimization problems are important (and they seem to increasingly be so), something which can find lower energy states (better solutions) in a shorter time is indeed an advantage.

I for one find it refreshing that they're not always blowing smoke about supremacy the way Google is - instead, they keep pushing forward actual results, like this one published in Nature: https://www.nature.com/articles/s41586-018-0410-x about some spinglass thing I don't understand, but it seems like it's something ideally suited for what they're building.

It remains to be demonstrated that D-Wave has an asymptotic advantage over classical optimization algorithms. Even synthetic spin glass benchmarks specifically designed to give the DW an advantage show no advantage.

Furthermore, real optimization problems have high connectivity, high weight cost function terms, and require higher precision than available. The first two will require an unacceptable overhead for fixed topology solvers. DW is not useful for real world optimization problems.

Google on the other hand has actually done something a classical computer can not do

> Furthermore, real optimization problems have high connectivity, high weight cost function terms, and require higher precision than available.

Have you seen their hybrid system they just released? https://www.zdnet.com/article/this-cloud-service-lets-you-us...

Looks like it's a strong contender for people who have bigger, higher-connectivity problems that want to get ready for when the quantum processors are bigger.

You can't get thousands of variables with all to all connectivity on a quantum annealer. The noise will kill it. I.e. you will never have such a device with reasonable coupler precision.

The techniques for decomposing a problem and solving the sub problems are a big (exponential) overhead because of frustration effects which require you to look at the entire problem at once.

And for context the PUBO -> QUBO mapping (high order to low order) + topology embedding can blow up the number of variables by an order of magnitude or more. The solution space scales exponentially with the number of variables so a naive solver will again deal with a messy overhead.