There are ready-made isolated repeater chips for the 12Mbps USB transfer rate:
ADuM4160 by Analog Devices or LTM2884 by Linear Technology. Surprisingly to me, both contain inductive coupling = miniature on-chip signal transformers as the coupling elements, interfaced to the outside world by silicon (CMOS?) buffered transceivers. Makes me wonder why the isolation is not optical these days...
Note that 100Base-TX Ethernet, SATA, PCI-e or RS422, all use a balanced pair in either direction, together comprising a 4-wire full-duplex link. Gigabit and 10Gb Ethernet works that way only on fiber optics I guess.
In contrast, USB low/full/high-speed uses a single balanced pair, in half-duplex mode, where the host and device take turns in talking on the bus, and have to tri-state the line driver when they're finished talking, to give a chance to the other party (somewhat similar to RS485, though many electrical and framing details are different).
Any galvanic isolator, including the chips mentioned above, has to respect that half-duplex direction-switching style of communication. A single signal trafo should theoretically work at 12 Mbps, except for the DC biasing resistors, and the framing possibly isn't "free of DC offset on average" either, making it difficult to just use a passive trafo. Attenuation aside.
Perhaps it's precisely this need for the active isolator to "turn the table" fast enough, to detect the end of transmission in the first place, that makes implementation of a "stupid USB repeater" at 480 Mbps impractical, even in today's silicon. There are supposedly some other changes in the electrical interface for high-speed USB 2.0 (constant current signaling) which may be another factor why high-speed USB doesn't lend itself easily to this kind of 485-style RX/TX switching in a dumb repeater.
Note that there's an alternative approach to the "direction switching" problem: rather than detect a high-Z on the line in an analog fashion, which brings along some inherent latency (lag), the isolator would have to understand the USB protocol, just like a USB hub does - so that it would know when to expect an end of the frame being currently received. And possibly, it would buffer whole frames, before relaying them onto the other side - just like a USB hub does. (Or does it?) Effectively the isolator would have to become a USB hub, with an isolation gap somewhere in there.
It is somewhat surprising to me, that there are no hub-style isolated repeaters. Possibly because ATMEL and friends make hubs, and Analog or Linear (or Avago?) make isolators, but the two gangs don't mix...
The problem of transporting the high bitrate over an isolation gap should not be all that difficult - yet even this area seems surprisingly "underdeveloped", or seems to suffer a gap of some sort. 10Gb Ethernet over fiber has been there for years, with bitwise base-band SERDES (bitstream), transmitted by a "laser" (at least a VCSEL) and received by a photodiode. Yet the DIL-packaged opto-couplers have barely reached 50 Mbps or so. Where does the gap come from? Well it seems to me that the guys making the DIL opto-couplers rely on relatively slow LED sources and photo-transistor receivers. While the guys making fiber stuff make their VCSELS and photodiodes suitable for coupling to a fiber - with adjustable bias current, with a local feedback diode strapped onto the VCSEL etc. Apparently noone got the idea to build an electric-to-electric photocoupler out of those high-grade parts. Note that the fiber-coupled gigabit stuff typically uses AC coupling on the electric interfaces, but that shouldn't be a big problem, if someone bothered to adapt the technology to classic DC-coupled TTL style.
Maybe it's just a conservative old-school view of the industry, on my part. Maybe the gigabit high-bandwidth tech has already moved to a new era, where you can only play in terms of standardized busses and interfaces, and there's no point in making discrete components capable of transferring a stupid simple logic 1/0 on a single signal. Maybe this is just my dinosaur-style thinking that you can still hack things together like that. The modern GHz era seems to "raise the bar" against casual hackers with a soldering iron. Electronics hacking has become a matter of closed labs with expensive equipment, only available to big industry-leading vendors. It's a closed club. From now on, all you can ever hack is software, or maybe some trivial antenna stuff.
Signal transformers are apparently only good into low hundreds of MHz. 1000Base-TX and especially 10GBase-TX take great pains of cunning modulation to squeeze the data into many "bits per symbol", on full-duplex-per-pair balanced lanes, at the cost of power-hungry DSP processing for all the modulation / local echo cancelation / pre-equalization... just to fit inside maybe 200 MHz of bandwidth available through the "magnetics" (signal transformers). If you're into TV antenna technology, you may have noticed that in the upper range, say 500-800 MHz and above, galvanic isolators are strictly capacitive. No matter what core material you choose, inductive transformers are just no good at those frequencies.
In the end... you know what? USB3 seems to use separate balanced pair transmission lines: one pair for TX, one pair for RX. Feels like coming home.