diff --git a/pyproject.toml b/pyproject.toml
index 2c23a17cf2..4bf2cde4f7 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -180,6 +180,7 @@ exclude=['pydantic/v1', 'tests/mypy/outputs']
 [tool.ruff.extend-per-file-ignores]
 "docs/**/*.py" = ['T']
 "tests/**/*.py" = ['T']
+"tests/benchmarks/test_fastapi_startup.py" = ['UP006']
 
 [tool.ruff.pydocstyle]
 convention = "google"
diff --git a/tests/benchmarks/test_fastapi_startup.py b/tests/benchmarks/test_fastapi_startup.py
new file mode 100644
index 0000000000..0488b873a6
--- /dev/null
+++ b/tests/benchmarks/test_fastapi_startup.py
@@ -0,0 +1,132 @@
+"""https://github.com/pydantic/pydantic/issues/6768"""
+from __future__ import annotations
+
+from typing import Any, Dict, Generic, List, TypeVar
+
+from typing_extensions import Annotated
+
+from pydantic import BaseModel, TypeAdapter, create_model
+from pydantic.fields import FieldInfo
+
+TYPES_DEFAULTS = {int: 0, str: '', bool: False}  # some dummy basic types with defaults for some fields
+TYPES = [*TYPES_DEFAULTS.keys()]
+# these are set low to minimise test time, they're increased below in the cProfile call
+INNER_DATA_MODEL_COUNT = 5
+OUTER_DATA_MODEL_COUNT = 5
+
+
+def create_data_models() -> list[Any]:
+    # Create varying inner models with different sizes and fields (not actually realistic)
+    models = []
+    for i in range(INNER_DATA_MODEL_COUNT):
+        fields = {}
+        for j in range(i):
+            type_ = TYPES[j % len(TYPES)]
+            type_default = TYPES_DEFAULTS[type_]
+            if j % 4 == 0:
+                type_ = List[type_]
+                type_default = []
+
+            default = ... if j % 2 == 0 else type_default
+            fields[f'f{j}'] = (type_, default)
+        models.append(create_model(f'M1{i}', **fields))
+
+    # Crate varying outer models where some fields use the inner models (not really realistic)
+    models_with_nested = []
+    for i in range(OUTER_DATA_MODEL_COUNT):
+        fields = {}
+        for j in range(i):
+            type_ = models[j % len(models)] if j % 2 == 0 else TYPES[j % len(TYPES)]
+            if j % 4 == 0:
+                type_ = List[type_]
+            fields[f'f{j}'] = (type_, ...)
+        models_with_nested.append(create_model(f'M2{i}', **fields))
+
+    return [*models, *models_with_nested]
+
+
+def test_fastapi_startup_perf(benchmark: Any):
+    data_models = create_data_models()
+    # API models for reading / writing the different data models
+    T = TypeVar('T')
+
+    class GetModel(BaseModel, Generic[T]):
+        res: T
+
+    class GetModel2(GetModel[T], Generic[T]):
+        foo: str
+        bar: str
+
+    class GetManyModel(BaseModel, Generic[T]):
+        res: List[T]
+
+    class GetManyModel2(GetManyModel[T], Generic[T]):
+        foo: str
+        bar: str
+
+    class GetManyModel3(BaseModel, Generic[T]):
+        res: Dict[str, T]
+
+    class GetManyModel4(BaseModel, Generic[T]):
+        res: Dict[str, List[T]]
+
+    class PutModel(BaseModel, Generic[T]):
+        data: T
+
+    class PutModel2(PutModel[T], Generic[T]):
+        foo: str
+        bar: str
+
+    class PutManyModel(BaseModel, Generic[T]):
+        data: List[T]
+
+    class PutManyModel2(PutManyModel[T], Generic[T]):
+        foo: str
+        bar: str
+
+    api_models: List[Any] = [
+        GetModel,
+        GetModel2,
+        GetManyModel,
+        GetManyModel2,
+        GetManyModel3,
+        GetManyModel4,
+        PutModel,
+        PutModel2,
+        PutManyModel,
+        PutManyModel2,
+    ]
+
+    assert len(data_models) == INNER_DATA_MODEL_COUNT + OUTER_DATA_MODEL_COUNT
+
+    def bench():
+        concrete_api_models = []
+        adapters = []
+        for outer_api_model in api_models:
+            for data_model in data_models:
+                concrete_api_model = outer_api_model[
+                    data_model
+                ]  # Would be used eg as request or response body in FastAPI
+                concrete_api_models.append(concrete_api_model)
+
+                # Emulate FastAPI creating its TypeAdapters
+                adapt = TypeAdapter(Annotated[concrete_api_model, FieldInfo(description='foo')])
+                adapters.append(adapt)
+                adapt = TypeAdapter(Annotated[concrete_api_model, FieldInfo(description='bar')])
+                adapters.append(adapt)
+
+        assert len(concrete_api_models) == len(data_models) * len(api_models)
+        assert len(adapters) == len(concrete_api_models) * 2
+
+    benchmark(bench)
+
+
+if __name__ == '__main__':
+    # run with `python tests/benchmarks/test_fastapi_startup.py`
+    import cProfile
+    import sys
+
+    INNER_DATA_MODEL_COUNT = 50
+    OUTER_DATA_MODEL_COUNT = 50
+    print(f'Python version: {sys.version}')
+    cProfile.run('test_fastapi_startup_perf(lambda f: f())', sort='tottime')