The morning digest has one number that made me look twice. SenseTime processed an average of 2.4 trillion tokens per day as of July. But that company’s invoice carries no unit called “tokens.” Instead, the unit written on it is the completed tasks its agents finished. It looks like a small change, the kind where a meter’s dial moves one tick. But I see this as the freshest signal in the AI market today. The 18 news items selected today are mostly stories about “capacity.” Building factories, pulling in power, stacking chips. Yet in the middle of all that, the one story where the “unit” changed shows the bigger picture. Let’s read the direction this signal points, together with the infrastructure competition news from Korea and around the world.

An image visualizing the concept behind this post: the invoice unit changes from tokens to completed tasks A visual of the core concept of this post.

Why the income statement is not about “tokens”

Let’s start with the numbers. In the first half of 2026, SenseTime recorded net profit of 617.3 million yuan, about 1.23 trillion won, on an IFRS basis. It is the company’s first half-year profit since its 2021 IPO in Hong Kong. Revenue was 2.91 billion yuan, up 23.4% year over year. The interesting part is not the scale but the composition. Generative AI revenue was 2.33 billion yuan, about 80% of the total. Recurring revenue from subscriptions and services was 1.14 billion yuan, up 124.4% year over year, making up about 40% of revenue. The gross margin held in the 41% range.

Placed side by side with competitors, the difference is clearer. MiniMax posted a net loss of 358 million dollars, and Zhipu AI a net loss of 2.07 billion yuan. Even with triple-digit revenue growth, the losses stayed intact. The “grow but stay in the red” structure was already solid in China’s AI industry. SenseTime’s first profit is the event that proved, on the income statement, that a billing model beyond token price competition actually exists.

From the company’s own announcements, the outline of the shift is clearer. SenseTime moved away from a structure dependent on centralized government agencies and large enterprise IT procurement, and expanded its channels into AI agent workflows and solo entrepreneurs, the solo-preneur segment. Overseas revenue grew 127% year over year. CEO Su Li said the company will stop wearing itself out in the race to grow parameters and concentrate R&D on breakthroughs in the intelligence of the models themselves. CFO Wang Zheng expects to maintain adjusted EBITDA profitability and turn annual operating cash flow positive. Reorganizing capabilities into three axes, one model system, one token factory, and one agent harness, fits the same context. Each axis is meant to form an independent commercial closed loop. The company is also pushing to build a 40,000 petaflops-scale AI computing center in Hong Kong by 2030. A company that sells tasks is at the same time building inference capacity.

Why the meter no longer measures fuel

To understand the change, think of a meter. AI has been sold like electricity or fuel. Use more tokens and you pay more; use fewer and you save. What you buy is input, that is, fuel. Whether that fuel actually finished the job was a separate matter. The structure settled only on how much was burned.

The weakness of that model is that the buyer has a hard time knowing “did the job end well.” If lots of tokens were burned but the output fell short of expectations, cost and value drift apart. Suppliers could only lower the fuel unit price. So the industry slid into a race to cut token prices, and the deeper the cuts, the bigger the losses. The MiniMax and Zhipu AI losses seen above are the direct product of that structure. The fuel got cheaper, but no one’s books kept a profit.

SenseTime flipped the meter’s direction. Instead of settling on how much fuel was burned, it settles on whether the task was completed and delivered. What is bought changes from input to output. The 41%-range gross margin and the 124%-range recurring revenue growth are evidence that even when you are paid in tasks, the business leaves something on the income statement. In a structure that is paid for completion rather than in a fuel race, profit has a place to sit. This change is not a price adjustment. It is a move of the business model itself.

Capacity keeps piling up while the unit on the invoice changes

While the meter’s dial changes, Korea and the world are busy building capacity. In the second quarter of 2026, global DRAM revenue was 154.73 billion dollars, about 208.3439 trillion won, up 59.5% quarter over quarter. A quarterly record. Samsung Electronics holds a 39.4% share, SK Hynix 24.9%, and the two Korean firms combined record 64.3%. However, the price increase for general-purpose DRAM contracts in the third quarter is expected to slow to 13-18%. That is a signal of moving into a phase where the pace of increases bends but the level holds.

Brookfield, Naver, and NVIDIA have formalized a 10 billion dollar investment in Korea’s sovereign AI factory. Naver’s Gaj Sejong site expands from 55MW to 200MW by 2028, and NVIDIA puts in 1 billion dollars directly. Brookfield is pushing up to 9 billion dollars of project financing at the non-binding term sheet stage, and a 1GW-scale follow-on project is also on the table.

The move of building capacity and pushing it out beyond the gate is happening alongside. Equinix, together with NVIDIA and Together AI, announced “Inference Exchange,” a distributed AI inference platform supporting more than 200 open models. On a mesh network spanning 36 countries and 77 markets, it aims at metro edge inference and sovereign AI. First-quarter 2027 operation is the goal. It is the scene of a data center operator reaching from “capacity” into “inference.”

Moody’s analysis points to a tighter fact. US hyperscalers spend 785 billion dollars on AI investment this year, while China’s big tech spends 140 billion dollars, about one fifth. But “computing capacity per dollar” is closing fast. Under the national program “Eastern Data, Western Compute,” China is moving training workloads inland, where the power grid is rich and the climate is cool, and stacks on top free land from local governments, power subsidies, and tax relief to lower build costs. By IEA statistics, installed capacity at the end of 2025 was 52GW in the US and 28GW in China; China is projected to reach 67GW by 2030 on 19% annual growth. Huawei’s CloudMatrix 384 rack delivers about 2x the performance of NVIDIA’s GB200, but its power consumption reaches 4.1x. The benefit of cheap power is offset in large part by power overconsumption.

Even so, the construction does not stop. NVIDIA’s data center business earned 89 billion dollars, about 92% of the 96.2 billion dollars in revenue for the recent quarter, and Jensen Huang said more than 100,000 Grace Blackwell NVL72 systems went into the Astra training, and previewed an additional 400,000 GPUs to come online. Separate from the safety debate, electricity keeps flowing into the factories that convert it into tokens.

Capacity keeps piling up. But the fresh signal today is that the unit of value is moving from “the amount of capacity and tokens” up to “the completed task.” While we are debating who will build the bigger factory, the buyer’s question was quietly changing. What does the factory in the end complete and deliver? That is the question of the day.

When the unit becomes the “task,” what companies ask first

The moment settlement happens in task units, the question a company throws changes. It moves from “how much is the token?” to “can I trust that task?” Two news items in today’s digest point the same direction.

The first is a warning from inside OpenAI. Four days after OpenAI unveiled GPT-6 Astra, which it designated as its first general-purpose AI, its chief scientist Jakub Pachocki is publicly calling for a voluntary slowdown in development. The context is that AI has become hard to understand and control for humans, and that recursive self-improvement, where AI develops itself without human intervention, will arrive soon. It is the warning that “no one is prepared for the consequences of rapid development.”

The second is evidence that people actually react to the output. According to a survey treated as a watch point for the 2026 national audit, 94.4% of children and adolescents have experience using an AI chatbot, and 75.3% of them act on the answers or seriously consider doing so. Separately, the government has set budgets for 741 AI projects across 41 ministries at 9.9 trillion won. Alongside it, an estimate that running 260,000 high-performance GPUs requires up to 1GW of constant power was also treated. In Korea, NIA announced it is creating an “Agentic AI Innovation Team” through an organizational restructuring and will work to build an ecosystem where citizens use agents in daily life. If money and action are truly being priced onto the output, who performed it correctly and how is a matter of accountability, not a research topic.

So a company buying in task units demands three things. First, verification that the task was properly completed. Second, the boundary that lets agents move autonomously without crossing into risk, that is, governance. Third, the ledger that answers how much this task cost and where it went, that is, cost attribution. These three are also questions that a world settling only in tokens could never answer.

Lens: where tokens get turned into tasks

Let me add one lens here. The company pains today’s news exposed, certainty about tasks and safe execution, cost attribution, and sovereignty, are exactly where ThakiCloud’s Agent-Native Cloud Paxis already stands as a formal product, v1.1 GA. The moment the unit on the invoice becomes the “task,” the value of the execution environment that sits between tokens and tasks rises.

Paxis is an execution environment that treats Skills and Tools, Policies and Audit Logs as first-class resources, not as attachments. What answers the demand for verification is the policy gate and the audit logs. They leave a record that can be pulled out afterward to show what the agent did. The boundary for safe execution is made by isolated sandbox runs and governance over autonomy from L0 to L3. What answers the cost attribution question is the CostRouter. It picks the model per job and writes into the ledger how much this task cost. Placing the execution environment on sovereign or on-prem K8s, on ai-platform, responds to the sovereignty demand.

If a company wants to buy and sell in completed tasks, the layer that makes that settlement trustworthy is Paxis. As SenseTime’s income statement showed, the 41%-range gross margin and 124%-range recurring revenue sit only on top of a structure that can be paid in task units.

The meter has already changed. What is needed now is to build the workbench that turns tokens into tasks people can trust.

SenseTime’s profit is not the business of one Chinese company. It is a statement for the whole industry that the structure of selling tasks holds. Capacity keeps piling up and electricity keeps being turned into tokens. But the unit on the invoice changes. I expect more companies next quarter will price not by “how much per token” but by “how much per task.” And the execution environment that can read that price tag will, in the end, be the variable that decides this game.

References

This post was written by synthesizing the following news items.

Tags: agent-governance, agentic-ai, ai-infrastructure, korea-ai, outcome-based-pricing, sense-time, task-based-billing

Categories:

Updated:

You may also enjoy