You're looking for a GBW of 400MHz, which isn't going to be easy to achieve with 2N3904s (fT ~ 300MHz -- note, by the way, that fT is GBW in terms of current, not voltage; it's still possible to get real voltage or power gain at frequencies above fT!). A single stage certainly won't, but several rubbed together might get the job done, or you can look at higher frequency devices.
One thing is true: you can always go for very marginal gain per stage, maximizing bandwidth as much as possible; then, it's only a matter of chaining enough stages to get the gain required (which goes exponential in the number of stages, after all!). Obviously, a gain of 1 takes infinite stages, so you can't go too low; supposedly, the ideal gain per stage is something bizarre like sqrt(pi) (~1.77) or the like. Practically speaking, you don't want to go to so much bother, copy-and-pasting stages together, so you'd shoot for a modest amount, like 5-10 per stage. This also helps with noise, which is additive per stage, so if your gain is very small, noise will quickly dominate (the noise, accumulated from the first couple stages, only grows exponentially (along with gain) through later stages!). A good, low noise, modest gain, front end helps magnify your signal without adding too much noise, allowing subsequent stages to be fairly noisy with little consequence.
Here's an example:
http://seventransistorlabs.com/Images/WidebandAmp.pngMMBTH10 is something like a shrunken, peppy 2N3904, with fT a bit over 800MHz, or something like that. The V/I ratings are less than half a 2N3904's, so the scaling analogy probably isn't far off.
This amplifier is intended for 50 ohm service, so the input and output impedances are fairly low. Hence, a common base input stage was chosen. This still isn't very good, because shunt feedback to the base (necessary to set gain, which increases bandwidth and decreases distortion) increases that impedance, I think into the 1kohm range (which looks capacitive at high frequencies where gain starts to roll off, so it momentarily becomes a good match, then it goes bad again).
To address stray capacitance -- not so much of the stage itself (that would be hard), but of the things attached to it (like the feedback caps and the next stage), a gyrator (tube enthusiasts will recognize it as a "mu follower") load is used. This has an emitter follower output characteristic (when taken from the emitter, obviously..), nice and low impedance, while somewhat isolating the load from the active gain node (the bottom transistor collector). The active top load delivers about double the bias current of the first transistor, excess being sunk through the 680 ohm resistor. I think... this was necessary to get the voltages to line up correctly, while also maintaining linearity (you don't want the transistors swinging into cutoff half the time) and extra pull-down oomph.
The second stage is the same thing, except the 680 pull-down goes through two diode drops and a small resistor (huh, it's not labeled -- seems I used 27 ohms in the real thing), which provides DC bias for the emitter follower output (which could've been more MMBTH10/81 action, but I figured the output impedance of the emitter follower was good enough to drive somewhat 'fatter' transistors, better suited to driving a low impedance; I don't think I noted much difference in the simulation*).
*But then, the simulation was claiming a 300MHz bandwidth, which is optimistic at best. I forget how I measured 100MHz, think it was step response (against a 350MHz scope).
Anyway, the second stage is arranged as a shunt feedback amplifier, much as you'd build an inverting op-amp stage: the 470 ohm coming from the previous stage, into the transistor base, is balanced by the 2.2k coming from the output, for a gain of about 4 on this stage. (Similarly, the 6.2k and 2.2k resistors on the base of the first stage set noninverting gain around 4. Total gain is around 16, but between source padding and load resistance, it comes down to about 10 total, or 20dB.)
Whether this type of design is applicable to your present problem, I don't know. For example, the "raising input impedance" trick might come in handy. There are ways of bootstrapping the input impedance away to infinity (and back to negative impedances, for that matter), handy for very weak signals. Or, if you don't need wideband, but only narrowband use (RF amps), transformers and inductors can do an order of magnitude better than my poorly optimized circuit. A 2N3904 might even achieve the requisite 26dB at 20MHz, if it's tuned to a limited bandwidth (say +/-10%).
Tim