Two developments this week have reset the terms of the global AI race: xAI’s Grok 4.5 has quietly entered private beta testing inside SpaceX and Tesla, while a Chinese lab has open-sourced a 1.6-trillion-parameter model trained entirely on domestic chips, released under a permissive MIT licence. Together, the moves underline how quickly frontier AI capability is diffusing beyond the handful of US labs that dominated the conversation just two years ago.
Grok 4.5 Stays Inside the Family — For Now
xAI has not made Grok 4.5 publicly available, instead routing early access through Elon Musk’s other companies, where it is reportedly being tested against internal engineering, manufacturing-optimisation, and customer-support workloads. Keeping the rollout inside SpaceX and Tesla gives xAI a controlled environment to stress-test the model on messy, high-stakes internal data before a wider release, a pattern increasingly common among frontier labs wary of unveiling flagship models directly to the public.
The private beta approach also lets xAI gather real-world performance signals without competitors easily benchmarking the model, at a time when leapfrogging rivals on public leaderboards has become less central to the competitive narrative than infrastructure, deployment speed, and enterprise integration.
China’s Open-Source Gambit
The more disruptive news may be the release of a 1.6-trillion-parameter model trained entirely on domestic Chinese chips — a detail as significant as the model’s scale. It signals that Chinese labs are no longer solely reliant on Nvidia hardware to train frontier-scale systems, undercutting the effectiveness of years of US export controls aimed at slowing China’s AI progress by restricting access to advanced GPUs.
Releasing the model under an MIT licence, rather than a more restrictive research-only terms of use, is also a strategic choice: it invites global developers to build on top of the model freely, in the same way Meta’s Llama family and other Chinese releases such as DeepSeek have done, seeding adoption far beyond China’s borders and creating soft-power leverage in the broader AI ecosystem.
Infrastructure Becomes the New Battleground
These releases arrive alongside a broader shift in how AI competition is playing out. Meta is reportedly preparing to rent out its own AI compute capacity, Google’s data centres drove a record 37% jump in electricity use, and multiple frontier labs are racing to secure custom silicon rather than compete purely on model architecture. Anthropic has entered preliminary talks with Samsung Electronics to manufacture a custom AI accelerator using Samsung’s 2nm process, according to people familiar with the discussions — a move that would reduce reliance on Nvidia and TSMC in the same way OpenAI, Google and Amazon have pursued their own custom chip programmes.
The convergence of these threads — private beta testing, open-source releases trained on non-Nvidia silicon, and a scramble for proprietary chips — suggests the next phase of AI competition will be decided as much by who controls compute and distribution as by who has the single best-performing model.
Industry Reaction
AI researchers have reacted to the Chinese open-source release with a mix of technical admiration and geopolitical concern, noting that a fully domestic-chip-trained model at this scale removes a key assumption underpinning years of US chip-export policy. Enterprise buyers, meanwhile, are watching Grok 4.5’s internal deployment at Tesla and SpaceX for early signals of how xAI intends to commercialise the model once it exits private beta — likely through the same enterprise and developer API channels that have become the industry’s default distribution model.
What to Watch Next
xAI has given no public timeline for Grok 4.5’s wider release, and the Chinese model’s real-world performance against Western frontier systems will only become clear as independent developers begin fine-tuning and benchmarking it over the coming weeks. Both developments reinforce a trend already visible through 2026: the gap between the best closed models and the best open ones continues to narrow, even as the compute and chip infrastructure required to train them becomes the more consequential — and more contested — resource.
Leave a comment