The null hypothesis is generally done so that you make an assumption that there's "no-difference." When you make the assumption that H0:mean<=6.5, you're injecting into the null hypothesis that the mean kilowatt use is possibly < 6.5. We need more info on what kind of test you're computing, but in general the computations to get your test-statsitic and p-value are done with this "no-difference" assumption and not a composite of "<" involved.
Here is a blog-like article on the null hypothesis: http://udel.edu/~mcdonald/stathyptesting.html
These kinds of problems though are naturally suited in a bayesian environment.
There's a great light bulb example in Bayesian Reliability (http://books.google.com/books/about/Bayesian_Reliability.html?id=GIhbskry6NYC ), where a manufacturer makes a claim about light bulb lifetimes. The skeptic students grab some light bulbs test them, and combine the manufacturers claim with their new test data to measure the chances the manufacturer was right.
In the same suite, you could measure how likely it is the mean<=6.5 using the consumers agency data, combined with what the manufacturer claims.
By doing that you've gotten rid of the H0,HA approach. However it's much more complex to do.