Anthropic Plans $21.6B Australian Data Centre Buildout as GPT-5.6 Specs Emerge

Anthropic has reportedly launched a $21.6 billion tender for 1.4 GW of training-grade compute in Australia by 2027, signaling a major infrastructure push. OpenAI provided new details on GPT-5.6's three-model lineup (Sol, Terra, Luna) with Cerebras speeds of 750 tokens/sec, though public access remains limited. Google DeepMind delayed Gemini 3.5 Pro to July 17 for additional pre-training.

Written by AI from the sources it cites; it can be wrong — follow any citation to confirm.

8 sources1 min read

Anthropic seeks $21.6B Australian data centre capacity in massive infrastructure play Anthropic has sent a confidential tender to five local data centre operators for 1.4 GW of compute capacity by 2027, backed by up to $15 billion in new capital. The move signals a major bet on training frontier AI in Australia and raises sovereignty questions. 1 2 3

OpenAI unveils GPT-5.6 Sol, Terra, Luna with 750 tokens/sec on Cerebras hardware The three-model family offers a flagship reasoning model (Sol), a balanced option (Terra), and a fast/cheap model (Luna). Sol achieves 750 tokens per second on Cerebras systems, adds native vision, and extends context length, but remains in limited preview due to government security review. 4 5 6

Google DeepMind delays Gemini 3.5 Pro to July 17, scraps base model for retraining DeepMind decided to scrap the 2.5 Pro base model and extend pre-training to improve mathematical reasoning and SVG generation. The delay aims to better compete with GPT-5.6 and Anthropic's Fable 5. 7

GPT-5.6 Sol and Claude Fable 5 split coding benchmarks as access remains restricted Sol leads Terminal-Bench at 88.8% while Fable 5 holds SWE-Bench Pro at 80.3%. Independent verification is blocked because GPT-5.6 is locked in limited preview. 8