We introduce Acacia, a graph foundation model trained on the web graph. Acacia can handle arbitrary feature dimensions and semantics without extra training, and it directly supports tasks such as node classification, link prediction, node clustering, and graph generation. It possesses in‑context learning abilities and does not rely on any pretrained large language model.\ \ Unlike existing graph foundation models that often require additional classification heads or feature projectors for new graphs or labels, Acacia works out of the box. Many current models gain capabilities by being stitched together with pretrained LLMs, whereas Acacia is trained from scratch using only the Common Crawl web graph. This demonstrates that graph models can acquire emergent capabilities from scratch similar to LLMs.\ \ Review: Acacia highlights the potential of learning graph structures from scratch, opening new avenues for universal graph modeling.