The ATM approach is good in that we know it works and it's easy to integrate with existing legacy equipment like the voice telephone switch. It's bad in that it's an old approach, and it's unfortunate these bleeding-edge infrastructure vendors didn't bring fresher methods to these otherwise rather unconventional networks that they want us to build. The alternative, IP QoS, is based on packet scheduling algorithms like CSZ and HFSC. These have the advantage that, unlike ``virtual circuits,'' the low-latency application can use less or more bandwidth than it reserves. If it uses less bandwidth, then bursty high-latency traffic can still take up all the free space on the link, including space to which the low-latency application has first dibs. If the low-latency application uses more bandwith than it reserved, then the network can still push the data through, but it won't necessarily keep the latency promises.
Telco zealots tend to claim these scheduling algorithms are not ``deterministic,'' but that is just their fancy way of admitting they don't understand how the algorithms work---they only understand the trivial solution of time-division multiplexing. Mathematical proofs validate the HFSC scheduler's desireable characteristics, as does experiment. I think the big problem with IP QoS from the telco perspective is that it isn't expensive enough. For example, since it works on commodity computing machinery with cheap Ethernet interfaces, one can use it without buying an expensive PBX ``switch'' or ridiculous proprietary videophone terminals with synchronous network interfaces. It even works with multiplayer games. At the end of the pre-semiconductor Age, when many of these nay-saying dinosaurs were still alive, telcos replaced their mechanical switches with electronic switches. But the electronic switches work pretty much exactly like the mechanical switches did, and they're still incredibly expensive. Any system that works differently and costs much less must be junk, because ``there's no such thing as a free lunch,'' and the last electronic transition (to the extent that it even was a transition) proved this cliché in their minds.
A reasonable objection to IP QoS is the lack of any large-scale proving ground. The logical place for the Proving would be The Internet, but unfortunately most Internet traffic passes through unusually large core routers that are too strange to be extended with IP QoS through a simple software upgrade and too expensive to replace with nonexistant IP-QoS-friendly replacements. Getting QoS on these core routers is a technical challenge as well as an investment challenge. I think the technical challenge can be overcome, but the investment challenge is much harder, especially when all these telco dinosaurs claim IP QoS will never work because only telco networks are ``deterministic'' (meaning: humans can understand their switching behavior through trivial inspection). The fixed meshes would have been a great large-scale proving ground for IP QoS scheduler research: the network has many hops and heavy traffic, and unlike the general Internet the carrier has end-to-end control of the routing software inside the customer nodes so the QoS queueing could be deployed ubiquitously on the local network. This leaves us with one of the same problems ATM had: the QoS coolness only helps you for (1) VoIP calls that get translated from Internet to telco-land at the mesh carrier's central office, and (2) arbitrary use within the mesh. The difference is that (2), in this case, includes whatever wacky flash-in-the-pan multiplayer game you want to play on your computer against other on-the-mesh friends, rather than ATM's (2) of single-purpose thousands-of-dollars sealed boxes with OC-12 interfaces.