AINEWS
The Chinese AI infrastructure company framed its domestic chip partnership as a closed loop of chips, software, and customers.
SenseTime's Galaxy Project is not a chip story yet. It is a forecast with a partner list.
The Chinese AI infrastructure company launched Galaxy with nearly 20 partners at a keynote this week, framing the program as a chip, ecosystem and deployment closed loop. Yang Fan, who leads SenseTime's Large Device Business Group, called the announcement "Intelligent Transformation and Symbiosis." The headline figure is large: SenseTime says its large-scale device platform already processes 2.42 trillion tokens per day. Tokens are units of text that an AI model reads or writes, and 2.42 trillion of them is a real workload. The company projects that figure will scale 25-fold to 10 trillion tokens per day by Q4 2026.
That 10-trillion figure is a company projection, not a measured result. The 2.42 trillion is self-reported; the 10 trillion is a forecast SenseTime will have to hit for the announcement to mean what it sounds like.
Three other numbers share the same source. SenseTime claims its heterogeneous hybrid inference delivers an 85 to 152 percent lift in Model FLOPs Utilisation (MFU), a metric that measures how much of a chip's theoretical compute an inference setup actually uses, on mainstream domestic chips. It claims 1.25x cost-effectiveness against Nvidia's H-series parts. It claims 2.5x token output versus domestic homogeneous inference setups at equivalent cost. None of the three is independently benchmarked in the announcement itself.
The distinction matters because a forecast and a benchmark are different shapes of claim, even when an announcement packages them as one. A 10-trillion-token target can be reported, revised or missed without the underlying inference stack changing. A 1.25x-Nvidia-H cost claim is the kind of figure that only resolves in customer production, where throughput, latency, power and total cost of ownership are visible together. SenseTime did not put either kind of evidence on the table at the keynote.
The partnerships are the part of Galaxy that does not depend on benchmarked performance. SenseTime signed a space computing agreement with Guoxing Aerospace, a Chinese satellite manufacturer, and announced a research partnership with five institutions including the Shanghai Artificial Intelligence Laboratory, a government-backed AI research institute, focused on scientific computing. Neither deal is, on its face, a production deployment claim, and neither is the kind of relationship a reader can price. They are commitments to integrate SenseTime's inference platform with downstream customers that have specific compute profiles: orbital, laboratory, eventually commercial.
For the cost claim to graduate from vendor-stated to measured, three things would have to land. A customer or auditor would need to disclose production-side throughput at a published total cost of ownership. A third party would need to benchmark the 1.25x-Nvidia-H figure under matched precision, batch size and traffic shape, because inference cost is sensitive to all three. SenseTime would have to report Q4 2026 token throughput against the 10-trillion target in a form that external observers can verify, rather than as a slide number from a keynote.
The source article flags the gap itself: vendor-optimised test clusters and customer production environments diverge, and the announcement does not address that gap. Type0 is not in a position to call Galaxy a win, a miss, or a load-bearing moment for China's domestic AI compute stack. The Galaxy Project is a roster of partners and a forecast with a date on it. The next readable signal is whether the 10-trillion-token figure arrives as audited disclosure or as another keynote slide.
SenseTime's investor disclosures page lists the company's filings and corporate governance documents for any reader who wants to compare the announcement cadence against quarterly reporting.