Dear authors,

I found this paper to be very well written and deals with an important topic that has a lot of uncertainty surrounding it. I think this is an impressive approach to modeling C3/C4 species abundance in response to climate change. With that said, I appreciate the acknowledgment and discussion of the many caveats to your approach.

In general, I think the details provided on modeling methods are thorough and clear, but there are a few things I found missing. In particular, I was not able to find the details on the model selection strategy in the supplemental methods. How did you find your "best" models among the 2^90 candidates? I see details on the Koshkina et al. method for integrated SDMs, but no info on model selection.

Below are specific line-by-line comments:
47-50: might also suggest mentioning a state shift from perennial-to-annual dominance, which you can find references for and consequences of in California grasslands (e.g., Koteen, Baldocchi, and Harte, 2009) among other areas.

63:68: my understanding is that C4 species tend to be most abundant in the tropics, thus warm, humid environments. I could be mistaken, but saying they are competitively dominant in hot, dry environments seems to contradict that. The references in the last sentence of this paragraph seem to support my case.

107: Is there supposed to be a link in the parentheses for the SEINet data portal?

125: I suggest checking and reporting variance inflation factors for the models you fit in addition to using r > 0.7 as a cut off for collinearity.

155: Did I miss it in the supplementary methods: where are the details on the model selection strategy?? How did you find a “best” model among the 2^90 candidates?

164-166: This may not matter, but I’m curious why the cross-validation groups were split by random quadrants across the western US rather than four equal-sized groups of random pixels. By “quadrants”, I assume that means all pixels within a group have adjacency, and therefore might have some degree of spatial correlation. But couldn’t you avoid this if you split the data into four random groups with completely random pixels in each group?

Figure 2: Could you maybe include the datapoints of the 1000 lambda values as background visualization below the model prediction lines? Could be helpful to get a sense of how well the models appear to fit the data. Maybe semi-transparent to avoid it looking cluttered.

Figure S12: It doesn’t look like there is very much variability in soil depth across the region (most of the area looks like nearly 200 cm). Do you trust that to be the case? Are there other sources of data (USDA gridded soil survey, maybe) that you could get more reliable soil measurements? Also, the title for this figure says “Soil depth (m)” when it should be (cm).

Figure S14: the color direction of this scale is not intuitive. Can you reverse to make temperature increases in shades of orange and decreases in shades of blue?

Discussion: Break up some of the longer paragraphs in the discussion to aid readability.

Table S2 (and relevant throughout): were the predictor variables scaled/standardized prior to running models? Are these standardized coefficients?

383: Also consistent with recent experimental evidence that the C4 grass Aristida oligantha is poised to increase in the PNW with warming (https://doi.org/10.1002/ecy.3464).

The authors made efforts to address most comments; those which were not addressed were relatively minor. The methods are substantially more detailed. The analyses and topic are worthy of publication. The only two minor issues are in line 175 where there seems to be a word missing (sentence doesn't make sense) and tables 2 and 3 seem to have formatting issues that made them difficult to read.