<p>We study a truncated kernel ridge regression (T-KRR) estimator for nonparametric regression. The approach is based on substituting the full <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(n\times n\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>×</mo> <mi>n</mi> </mrow> </math></EquationSource> </InlineEquation> random kernel matrix by its first <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(n\times N\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>×</mo> <mi>N</mi> </mrow> </math></EquationSource> </InlineEquation> block, with <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(N \ll n.\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>N</mi> <mo>≪</mo> <mi>n</mi> <mo>.</mo> </mrow> </math></EquationSource> </InlineEquation> The study of the T-KRR convergence rate is based on two varying probability measures. The first measure is associated with the observed data with an unknown pdf. The second measure has a known pdf and it is associated with the kernel <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\({\mathbb {K}}\)</EquationSource> <EquationSource Format="MATHML"><math> <mi mathvariant="double-struck">K</mi> </math></EquationSource> </InlineEquation>. We develop rules for the choice of the optimal values of the tuning parameters of the T-KRR. This latter requires much smaller computational load than the full KRR and has the same optimal convergence rate as this latter. Also, we provide numerical simulations to illustrate the results of this work and to compare the T-KRR with two others competing scalable KRR.</p>