At WAIC 2026, SenseTime offered a counterintuitive answer to the compute crunch: stop trying to build a stronger chip — split the race into legs and let different chips run a relay.
This is not a slide-deck theory. SenseTime has the system running in production, and — for the first time — a hybrid inference cluster with domestic Chinese chips participating is profitable in real customer deployments. Alongside the announcement, SenseTime set a year-end target of 10 trillion tokens per day, a 25× increase since the start of the year.
Agents Are Pushing Compute Demand to the Limit
The past year brought a fundamental switch: large models went from “keeping people company” to “doing work.” Latest estimates from the University of Michigan and Stanford show that on real software-engineering tasks, a coding agent processes roughly 154 input tokens for every 1 output token — and these agents do not eat once a day; they run 24/7.
Industry consensus puts this year’s inference compute demand at 5 to 10 times last year’s. Meanwhile, supply of NVIDIA’s high-end cards has barely grown. Demand exploding against flat supply — the whole industry is asking the same question: now what?
The Triathlon Insight: Split Prefill From Decode
SenseTime’s idea comes from a plain analogy — the triathlon. Rather than one athlete swimming the entire course, three specialists each run their strongest leg. Every large-model inference divides into two operations:
-
Prefill (reading the question): ingest the user’s documents, context, history — a compute-throughput job.
-
Decode (writing the answer): emit tokens one by one — a memory-bandwidth job.
Traditionally, one card does both start to finish. SenseTime splits the workflow: domestic general-purpose cards run Prefill — their compute throughput is competitive, so the job suits them — while NVIDIA’s high-end cards run Decode, their absolute strength. This is heterogeneous hybrid inference: not asking domestic chips to fight NVIDIA head-on, but assigning every card the work it does best.
How Hard Is Deployment? 30 Prefill Nodes per Decode Node
Simple to say, brutal to build. A pure-NVIDIA setup needs 4 Prefill nodes per Decode node; with domestic chips the ratio becomes 30:1 — network complexity, scheduling complexity, and fault-tolerance complexity all spike at once. SenseTime chief scientist Zhang Xingcheng acknowledged the years of unglamorous work underneath: compilers, operators, networking, caching, scheduling, fault tolerance, layer by layer. Only after system-level tuning does this heterogeneous system — far more intricate than any pure-NVIDIA equivalent — sustain long-run stability at 99.9% reliability. Adapted mainstream models include GLM 5.2, DeepSeek V4, Kimi K2.7, MiniMax M3, and SenseNova V6.7.
The Breakthrough: Domestic Compute Finally Gets a Seat at the Table
The old domestic-compute storyline ran: self-sufficiency matters → our chips can run → and then nothing. Runnable but uneconomical; usable but unsold. The surprise this time: a hybrid cluster with domestic chips participating is actually profitable. SenseTime co-founder Yang Fan’s definition is unambiguous: revenue covering depreciation, amortization, rack rental, electricity, and service staffing, with margin to spare. Real money.
Three steps made it work:
-
Squeeze the domestic silicon: targeted optimization across three different domestic vendors lifted single-card MFU (effective utilization) roughly 2× versus factory defaults — the best card gained 152%.
-
More output per dollar: against same-cost all-domestic inference, the heterogeneous hybrid produces 2.5× the tokens.
-
Scale it up: 400 billion tokens/day at the start of the year → a projected 2.42 trillion/day by late July → the 10 trillion/day year-end target. A 25× year.
For the first time, domestic compute entered the real commercial market without waiting for any single card to fully match NVIDIA — system-level division of labor got it there early.
The Token Factory: Running Compute Like a Plant
Why does SenseTime profit from the same domestic cards where others lose money? Yang Fan is blunt: many self-described compute providers are essentially holding machines for the big labs on long leases — a financial service, not an operation. Imagine a factory where the building, machines, assembly line, warehouse, and sales each belong to a different company; forget optimization, the meetings alone eat the margin. SenseTime runs end-to-end control: from facility physics to servers, network architecture, compilers, operators, frameworks, the inference engine, scheduling, and finally customer token traffic, one organization owns the whole chain. No single link is necessarily the longest plank — assembled, they create system-level optimization space. The token-factory logic maps cleanly: chips are the machines, the network is the assembly line, KV cache is the warehouse, scheduling is the plant manager, electricity is the raw material, and tokens are the product.
Another Dimension of Coordination: Electricity
Once chips, network, and business are connected end to end, optimization reaches down to the power bill. At the Lingang datacenter, SenseTime deployed energy storage with CATL, doing elastic scheduling against time-of-use pricing and compute forecasting: token output per unit of electricity up about 80%, power costs roughly 10% below regional IDC peers. On July 10, when Shanghai’s grid strained, the state utility issued a demand response and SenseTime’s Lingang facility released 75% of its capacity within two hours, saving 46,000 kWh. Large AI datacenters are becoming the grid’s virtual batteries — shedding load at peaks, running flat-out on inference in the troughs.
From One Factory to a Network: the Galaxy Project and Space Compute
The token factory works, but one company’s end-to-end excellence cannot carry an ecosystem. At WAIC, SenseTime launched the Galaxy Project, uniting domestic chip vendors — Cambricon, MetaX, Huawei Ascend, Moore Threads — and infrastructure providers including SiliconFlow, with four targets: 1 token factory, 5 ten-thousand-card domestic clusters, 10 co-created technical directions, and 200+ AI startup organizations. Further out, three frontier paths are in the works: optical computing to cut data-movement energy, quantum computing for high-dimensional optimization beyond classical reach, and space-based compute — a signed agreement with CAS Space to launch four “SenseTime” compute satellites in Q4 2026, a commercial hybrid satellite-ground backbone by 2028, and a thousand-satellite integrated AI constellation as the 2030 goal. Yang Fan’s reasoning on space: China’s play is not that ground power is too expensive — it is that huge swaths of the world have no 3G or 4G coverage, and satellite-hosted compute is the most effective path to reach them.
The Winner’s Logic Changes in the Agent Era
For years, AI infrastructure competition looked like an arms race: most GPUs, highest peak FLOPS, most spectacular datacenter. In the agent era that playbook is failing. The model does not care which card it runs on; the user does not care how many P of compute sit in your basement. They care about three things: are tokens fast, are they cheap, is the service stable. Tomorrow’s winner is not necessarily the company with the most GPUs — it is the one that best organizes different chips, different electricity prices, different models, and different tasks into one efficient production line. What SenseTime showed at WAIC 2026 is more than a domestic heterogeneous inference scheme. It is an early answer to a bigger question: when AI truly enters mass production, what should the next-generation factory look like?
References: Synced coverage, “Agents blow up compute? A new species from China: 10 trillion tokens a day, and profitable” · SenseTime’s official WAIC 2026 announcements
