Above 10 GHz, keep nickel off the RF trace. Immersion silver or OSP on the trace side, ENIG or ENEPIG only where a wire bond or a connector needs it. Below 3 GHz the finish is a solderability and shelf life decision and any of them will do; from 10 GHz upward it is a loss decision, and ENIG costs about 0.4 dB per inch on a 50 ohm microstrip that would otherwise sit at 1.2 dB per inch.
That number surprises people who treat the finish as a coating. It is not the gold. It is the nickel under the gold.
Where the current actually flows
At high frequency the current crowds into a thin layer at the conductor surface. The depth of that layer falls with the square root of frequency: in copper it is about 2.1 micrometres at 1 GHz, 1.2 at 3 GHz, 0.65 at 10 GHz, 0.38 at 30 GHz and 0.24 at 77 GHz.
ENIG as we publish it is 100 to 200 microinches of nickel, which is 2.54 to 5.08 micrometres, under 2 to 3 microinches of gold, which is 0.05 to 0.08 micrometres. Compare those with the skin depth at 10 GHz. The nickel layer is four to eight times thicker than the depth the current needs, so the current flows in nickel. The gold is about one tenth of a skin depth, so the gold itself contributes almost nothing either way. The nickel is the whole story.
Nickel is a poor conductor next to copper: bulk resistivity 6.99 x 10^-8 ohm metre against 1.68 x 10^-8 for copper, a factor of 4.2. On a 10 GHz surface it gives a surface resistance of 0.052 ohm per square against 0.026 for copper, a factor of two even before the magnetic permeability of nickel is counted. Nickel is ferromagnetic, and a permeability above one shrinks the skin depth further and raises the surface resistance again. Two is the conservative case; with a permeability of ten the ratio is closer to six.
The loss, in decibels
Put that on a real line. A 50 ohm microstrip on 0.20 mm of FR-4 class dielectric with a dielectric constant of 4.2 needs a 0.298 mm wide trace, so the conductor loss at 10 GHz is 0.38 dB per inch in copper and 0.78 dB per inch when the same trace is covered by 2.5 micrometres of nickel. The dielectric under the same line contributes about 0.82 dB per inch, so the total goes from roughly 1.20 to 1.60 dB per inch, a third more loss for a finish that looked cosmetic.
At 30 GHz the same comparison runs 0.66 dB per inch in copper against 1.35 in nickel, on a dielectric that already costs 2.45. The nickel penalty grows with frequency because the skin depth keeps falling while the nickel thickness does not.
Immersion silver behaves like copper: silver is slightly more conductive and the layer is a fraction of a micrometre. OSP is an organic film a few tenths of a micrometre thick, so the current stays in copper. Neither adds a magnetic layer, which is the point.
When ENIG is still the right answer
ENIG is not a mistake. It is flat, it survives storage, it solders well after multiple reflows, and it is the finish that lets a 0.5 mm pitch BGA sit down without paste wicking. Below 3 GHz its 0.1 to 0.2 dB per inch penalty disappears under the dielectric loss of ordinary FR-4. It is also the base for ENEPIG, where a palladium layer sits between nickel and gold to stop nickel diffusing into the gold and to make gold wire bonding reliable.
So the decision is not "which finish is best" but "which nets need which finish". On a mixed board the RF side wants silver or OSP, the digital side and the connectors are happy with ENIG, and the wire bond pads need ENIG, ENEPIG or hard gold. Selective finishing exists for exactly this case.
Three things that go wrong
One finish for the whole board. A 24 GHz radar board finished with ENIG everywhere, because the BGA side asked for it, and then a link budget that comes in short. The fix is to name the nets that get silver and the nets that get ENIG on the fabrication drawing, and to state the sequence, because two finishes on one copper layer means two plating runs and a masking step between them.
Buying thicker gold to fix loss. At 10 GHz a gold layer would need to be several micrometres thick before it carried the current, and the 0.05 to 0.08 micrometres on a standard ENIG board does not. Hard gold on an edge connector is a wear coating, and it still sits on nickel. Thicker gold buys wear life and wire bond reliability, not insertion loss.
HASL on fine pitch or on an RF edge. Hot air levelling leaves a thickness that varies across the panel, and at 0.4 mm pitch the air knife can bridge pads. It also puts the board through a second 245 degrees Celsius excursion for the sake of the finish. Where solderability is the only requirement and the pitch is coarse, HASL is fine and cheap; where either frequency or pitch is tight, it is the wrong tool.
What SUPERB PCB controls on an RF build
Published finishes: OSP, ENIG with 100 to 200 microinches of nickel and 2 to 3 microinches of gold, selective ENIG, ENEPIG, immersion tin, immersion silver, lead-free HASL, hard gold, soft gold and flash gold. ENIG pad spacing is 5 mil minimum, RF line width and spacing go down to 0.05 mm, and impedance is held to plus or minus 5%, with plus or minus 2% on critical nets. A TDR coupon goes on every panel, on the same layer pair as the product.
For a board above 10 GHz the finish is agreed with the stack, because the same coupon that proves the impedance also shows whether the finish moved the loss. That work sits on our RF and microwave PCB page and the build side on RF PCB. SUPERB Tech quotes the finish by net, not by board, once the drawing names which side is RF. The general trade-offs between ENIG, HASL and immersion silver, without the frequency argument, are in choosing ENIG, HASL or immersion silver.
One limit worth stating: we do not measure RF insertion loss ourselves. DC, functional and customer fixture tests run on the line; S-parameters are measured by the customer laboratory or by a partner RF laboratory. What we control is the geometry, the copper, the finish thickness and the coupon, which is what sets the loss before anyone connects a network analyser.
What to send with the order
Frequency band, which side or which nets carry RF, the loss budget per inch if there is one, whether anything is wire bonded, the connector type, how many reflows the board will see, and how long it will sit in storage before assembly. Those seven lines decide the finish, and they decide it before the panel is etched rather than after the first prototype fails its link budget.