Galaxy Plan: Not Just a Story, but a 'Bet' on Domestic Computing Power

When SenseTime announced "5 ten-thousand-card domestic intelligent computing clusters" alongside nearly 20 chip manufacturers, my first reaction wasn't excitement, but pulling up SenseTime's balance sheet. As a family office investment manager, the habit is always to check cash flow and capital expenditure plans first.

A ten-thousand-card cluster implies a single cluster scale of at least 10,000 GPUs; five means 50,000. Based on estimated unit prices for current domestic AI chips (such as Ascend 910B, Cambricon Siyuan 590), hardware investment won't be less than 15 billion RMB, excluding server rooms, electricity, cooling, and operations. SenseTime's 2025 financial report shows cash and equivalents around 8 billion RMB, with operating cash flow persistently negative. If this "Galaxy Plan" truly materializes, SenseTime must introduce external capital or rely on government subsidies and industrial funds.

At WAIC 2026, Yang Fan stated the intention behind the "Galaxy Plan" was to "solve the problem of domestic computing silos," but in my view, this looks more like a "computing bet agreement"—using ecosystem binding to exchange for capital and orders.

Short Term: Execution Risk is the Biggest Variable

The engineering difficulty of ten-thousand-card clusters far exceeds imagination. Top domestic manufacturers began planning ten-thousand-card clusters in 2023, but few are running stably today. Reasons include: gaps in single-card performance, interconnect bandwidth bottlenecks, and immature software stacks. SenseTime has its own Large Device platform and has operated thousands-of-cards-scale clusters, but ten-thousand-card scale means exponential upgrades in network topology, thermal management, and fault rate control.

More critically, this plan aggregates chip makers like Cambricon, MetaX, Hygon, Ascend, Moore Threads, and Biren. Each has incompatible architectures, operator libraries, and communication protocols. What SenseTime needs to do isn't "assemble" a ten-thousand-card cluster, but "unify" scheduling of these heterogeneous chips. It's like asking an orchestra to play five different pieces simultaneously—no matter how skilled the conductor, noise is hard to avoid.

From a capital perspective, such "grand alliance" projects often face the "free-rider" dilemma: everyone hopes others pay and work while they enjoy ecosystem benefits. If SenseTime, as the initiator, doesn't invest enough proportionally, partner enthusiasm will wane. Historically, similar "domestic CPU alliances" mostly ended as signing ceremonies without real implementation.

Long Term: If Successful, SenseTime Gains a Moat as a "Computing Infrastructure Operator"

But looking at it another way, if the "Galaxy Plan" actually works, SenseTime's valuation logic will fundamentally change.

Currently, SenseTime's valuation relies mainly on the "Large Model + AI Application" story, but the market remains skeptical of its commercialization ability—2025 full-year revenue under 6 billion RMB, losses still expanding. If the "Galaxy Plan" succeeds, SenseTime transforms into a "computing infrastructure operator," similar to a "domestic CoreWeave." Valuation logic for such assets isn't PE, but EV/EBITDA multiples, focusing on rental income per unit of compute and utilization rates.

Assuming all 5 ten-thousand-card clusters are built, with single-cluster compute around 5000 PFLOPS (FP16), calculated at current domestic compute market prices (approx. 10-15 RMB/card-hour), plus government subsidies and large contracts from industry clients, annualized revenue potential could exceed 5 billion RMB. This excludes value-added income from AI training services. If SenseTime proves its "rent-collecting" capability, the market will offer higher valuation multiples.

More importantly, this creates an "ecosystem barrier." Once dozens of domestic chip manufacturers rely on SenseTime's scheduling platform and O&M system, switching costs become extremely high. SenseTime becomes the "connector" and "integrator" of the domestic computing supply chain, similar to ARM's role in the mobile chip ecosystem. This position is exactly the "long-term compound asset" most valued by family offices.

But Beware One Trap: The Cost of Diversification Strategy

SenseTime is simultaneously doing "Large Models," "Autonomous Driving," "AI Content Generation," "Smart Cities," and "Computing Clusters." Every item burns cash, and every item faces strong competitors. Family offices fear nothing more in asset allocation than "trying to do everything and doing nothing well." If the "Galaxy Plan" is just another "story" for SenseTime, it will further drag down cash flow, causing bleeding in core businesses.

From an investment angle, I'd watch these key indicators:

• Timeline for landing the first batch of clusters (Can they light up before Q1 2027?)
• Proportion of self-funded capital for the "Galaxy Plan" (Does it exceed 30%?)
• Core client list (Are any internet giants signing long-term contracts?)
• Exclusivity with chip manufacturers (Do they collaborate only with SenseTime, or join multiple similar plans simultaneously?)

If these indicators aren't met within 6 months, the so-called "Galaxy Plan" is just a "domestic computing business card," having almost no impact on SenseTime's valuation.

Action Advice:

If you're watching this sector, treat the "Galaxy Plan" as an "option" rather than an "asset." Don't chase highs in SenseTime stocks or primary market shares due to news hype in the short term. Instead, focus on "selling shovels" companies in the domestic computing upstream/downstream, such as heat dissipation, optical modules, and high-speed interconnect chips—these benefit regardless of who builds the clusters. For SenseTime itself, wait until H2 2026 to see if there are substantive engineering progress updates and order disclosures. Only then, if indicators are met, is it worth heavy positioning.

Original Link: https://www.ithome.com/0/979/115.htm