A state telco establishes a unit to coordinate AI tokens—the basic chunks of text models read and write—signalling compute is now a product to be managed inside corporate hierarchies, alongside a 360MW AI factory park on Batam island and a Baidu document reading AI that topped…
Compute has stopped being a floating abstraction. It is now reorganising corporate structures, repricing capex bets, and rewarding parameter efficiency, and three signals from Chinese tech outlet 36Kr's evening roundup show each of those moves landing in a single day. State-telco China Mobile confirmed it has created a "Token Office" to coordinate the creation, distribution, and application of AI tokens, the basic chunks of text AI models read and generate, across its business lines. Hours later, an Australian operator backed by Nvidia and a Singapore-based partner disclosed a plan to build a 360MW dedicated AI factory campus on Batam, an island off Singapore, under an eight-year framework worth up to $30B. In parallel, Baidu open-sourced an end-to-end document-reading model that topped four global developer charts inside a day.
China Mobile's Token Office is the most bureaucratically striking of the three signals. The 36Kr evening digest reports the new unit sits alongside the operator's earlier digital-intelligence business unit and "算力办," or computing power office, a layering that reads as deliberate: one office for the AI business, one for raw compute, and now one for the units that compute actually sells. China Mobile did not confirm circulated reports that the office reports directly to core group leadership, so the exact reporting line is an open question. What is confirmed is that a state-run telco with national reach now treats the AI token as an internal product to be coordinated across business lines, not as a side effect of model deployment.
The Batam AI factory park is the capex echo. The same 36Kr roundup carries the news that Nvidia-backed Australian operator Firmus and Singapore-headquartered DayOne will co-develop the 360MW campus using Nvidia's DSX AI factory design, a named reference architecture for dedicated AI compute builds, under an eight-year framework that the roundup frames at up to $30B over six years. The signal is not the dollar figure alone; it is that a hyperscale-class AI campus is being raised on Batam rather than layered onto existing Singapore, Taiwan, or Korean footprints. Compute is being sited for power and latency, not for legacy datacenter geography.
The third signal sits on the demand-elasticity side. Baidu's Unlimited OCR, an end-to-end optical character recognition model tuned for long-document parsing, topped four charts within a day of release: GitHub Daily Trending, GitHub Python, HuggingFace global model trends, and HuggingFace multimodal model trends. The 36Kr write-up cites the model as carrying 3B total parameters, with roughly 570M active at inference, and a composite score of 93.92% on the public OmniDocBench v1.6 benchmark — a figure the roundup frames as a new end-to-end OCR state of the art. The 36Kr evening digest presents both the four-chart-topping achievement and the OmniDocBench v1.6 score; Baidu has not yet published a confirmed model card or HuggingFace repo in the source receipts available, so these benchmark claims are attributed to the 36Kr write-up and require primary-source confirmation before being asserted as established fact. A small active-parameter footprint on a chart-topping model is the open-source reply to compute scarcity: when GPU cycles cost more, parameter efficiency becomes a moat.
The same edition pushes the demand side harder still. A 36Kr-curated recap of Musk's X post on 28 June relays that Grok 4.5 is now in a private beta inside SpaceX and Tesla, with SpaceX intending to ship a "new from-scratch" AI model every month for the rest of the year. English-language coverage in outlets such as NDTV Profit, Free Press Journal, and Prokerala restates the private-beta line. One secondary outlet, Crypto Briefing, layers a "V9 model upgrade" framing onto the rollout that does not appear in the original Musk posts or in 36Kr's recap. Treat the monthly SpaceX cadence as the supported claim; treat the parameter and V9 framing as secondary-outlet editorial until an xAI primary source confirms it.
The non-AI half of the digest is quieter, but each item is worth a watch. The same 36Kr roundup notes that Chinese tea and ice-cream chain Mixue Bingcheng has opened three stores in Bishkek, Kyrgyzstan, its second Central Asian entry after Kazakhstan in April 2025, bringing its overseas footprint to 16 countries. The wire also names a State Council issuance of the 15th Five-Year Plan for Education Development, a 2030 milestone for the country's next planning cycle that the roundup cites but does not link, a US public-pension rebalance estimated at roughly $30B of forced passive selling already on the tape, and a URTOPIA Series B of more than 200M RMB whose lead investor is named in the wire but truncated in the excerpt available and so cannot be confirmed from this digest alone.
What to watch next. The US pension rebalance hits the tape within days, which will show whether $30B of forced selling is absorbed cleanly or prices the AI-infrastructure complex further. Inside China, the Token Office's first public deliverables, not its formation, will tell readers whether compute scarcity is now permanently inside telco org charts or a temporary fix. And the next monthly SpaceX from-scratch model release, telegraphed by Musk in the same digest, will test whether a hyperscaler-class buyer can really keep a monthly training cadence without burning the compute the rest of the same digest says is being rationed.