Ah, yes, thanks for setting me straight. Running a software load balancer on your own t1.micro might seem inexpensive as compared to ELB, but it has its caveats.

If you are debating whether to manage your own load balancer (DIY – do it yourself) vs. using ELB then you need to consider:
a) whether you will actually require the elasticity ELB provides – that is, whether your load can be expected to have much static range or dynamic range (to borrow vocabulary from photography)
b) your cost of labor for the operational skills required (troubleshoot problems, fix ‘em, roll out the updates, etc.)
c) the technical debt incurred by using a homegrown (read: error-prone) system as opposed to a more mature offering with a record of reliability
d) the relative cost of the computing resources for each solution.

In most places in the US and EU, item (b) tends to tip the scale overwhelmingly away from DIY.
In most companies where agility is a priority, item (c) also tips the scale away from DIY.

Sounds like material for another article.

To be clear, I meant specifically that you would use t1.micros to host your own load balancer. The application would presumably be hosted on some more powerful instance.

Indeed. Running your application on four t1.micro instances is still cheaper than using a single m1.small. But only certain types of applications are OK on t1.micros because of the CPU throttling.

Then again

a) the bandwidth is cheaper and
b) t1.micro instances definitely have less load capacity than an ELB instance
c) No redundancy at that price.