Jane Street's key insight is that a token factory is not a fundraising story, but an operating system.

Author and source: Hongyi Jiaofu Kaige, Ye Kai Wen

Many people’s first reaction upon seeing Jane Street’s Texas AI data center is envy—envy of its wealth, able to afford GPUs, build liquid-cooled server rooms, and dare to deploy thousands of high-end chips into a single trading system.

This reaction is normal, but it's easy to misinterpret.

(ChatGPT-generated image)

What makes Jane Street truly worth studying is not that it has 4,032 GPUs, nor that it transformed its traditional data center into a liquid-cooled AI facility; the key point is that it has reimagined computing power as a frontline trading productivity asset. In other words, it is not building a "data center"—it is building a Token factory that continuously produces trading decisions, risk signals, model iterations, and market insights.

This change is important.

To traditional financial institutions, data centers are cost centers—servers, networks, cooling, and power are all expenses. But at Jane Street, the data center becomes a production line: electricity flows into the facility, GPUs perform training, models participate in trading, and trades generate revenue. What flows through this system is not steel or square footage, but countless model invocations, training tasks, and inference computations. In today’s terms, this is tokenized intelligent production capacity.

The lesson for China is direct. China is discussing tokenization overseas, talking about compute token factories, and considering how Hong Kong-listed companies shift from real estate, industrial parks, and manufacturing toward AI infrastructure. Jane Street offers a concrete example: a token factory isn’t about concepts, issuing a coin, or packaging compute as a financial product—it’s about fully integrating electricity, liquid cooling, GPUs, networking, models, scheduling, and internal settlement mechanisms to make compute a true engine of corporate revenue and valuation.

Jane Street is not a trading firm with AI, but an AI factory with trading desks.

Jane Street has long been regarded as one of Wall Street’s most secretive quantitative trading giants. It is not a traditional investment bank, nor an asset management firm that relies on management fees; instead, it profits by using its own capital, trading models, low-latency systems, and global market liquidity to capture spreads and risk premiums.

Public records show that Jane Street trades multiple asset classes across 45 countries and more than 200 trading venues, with a workforce of approximately 3,000 employees. It has a strong technology culture, having long relied on the OCaml functional programming language, and emphasizes proprietary software development, low-latency systems, automated risk management, and highly reliable trading infrastructure. Jane Street’s official website describes its machine learning team candidly, characterizing it as “a research lab connected to a trading desk,” and notes that financial market data flows in like a flood—most of it noise—requiring the machine learning team to extract tradable signals from it.

This statement is crucial.

Most companies use AI to improve office efficiency. Jane Street uses AI to accelerate market understanding. The former reduces costs; the latter generates revenue. The former treats AI as a tool; the latter treats AI as a means of production.

Therefore, Jane Street’s AI data center transformation cannot be understood as a typical corporate digital upgrade—it’s more like a trading firm disassembling and rebuilding its core engine. In the past, a trading firm’s core consisted of traders, mathematical models, market data, and execution systems; today, its core is data, computing power, models, networks, and an internal resources market.

This is the first meaning of the Token Factory: intelligence does not emerge out of nowhere—it is produced.

From six Dell servers to 4,032 GPUs, what has truly changed is the mode of production.

Jane Street’s website mentions that twenty years ago, its “cluster” consisted of just six Dell servers piled on the floor of its office; today, its new data center in Texas houses 4,032 GPUs and employs a liquid-cooling architecture. This transformation is vividly striking—it is not merely a hardware upgrade, but a fundamental shift in the enterprise’s production relationships.

Early trading systems were closely located to traders. Machines were placed in offices, so issues could be addressed directly—even by unplugging the power. At that time, computing power served primarily as a supporting tool for trading teams.

Later, the trading system entered the engineering phase. Networks, risk control, backtesting, automated execution, and monitoring systems gradually became platformized, with computing power becoming a shared foundational resource across multiple teams.

We are now in the third phase. GPUs, liquid cooling, power, fiber optics, storage, scheduling systems, and model training have been integrated into an AI trading factory. Data centers are no longer part of the support function but are directly embedded into trading capabilities.

This is particularly insightful for Chinese companies.

Many companies talk about AI transformation but remain stuck at the stage of “buying models,” “deploying systems,” and “integrating APIs.” Truly competitive enterprises are turning AI infrastructure into their own production systems. Those who can acquire electricity more cheaply, run GPUs more stably, train models faster, and allocate computing power more efficiently will gain new cost and speed advantages in the AI era.

This is not the responsibility of the technical department.

This is a matter for the board.

Liquid cooling is not an engineering detail—it's the foundation of the Token factory.

One of the most notable aspects of the Jane Street Texas data center is its high-density liquid cooling retrofit. Public materials and accompanying documentation highlight that this data center supports high-density GPU racks rated at GB300 levels, with individual rack power consumption reaching approximately 140 kW—a density that traditional air cooling can no longer accommodate.

According to NVIDIA's official specifications, the GB300 NVL72 features a fully liquid-cooled, rack-scale architecture, integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs for AI inference, AI reasoning, and large model training. HPE’s disclosed information on the GB300 NVL72 also confirms that such systems are liquid-cooled, rack-scale solutions designed for training, fine-tuning, and inference of models exceeding one trillion parameters.

This indicates a trend: the future competition among AI factories will not just be about chips, but about systems engineering.

Even the most powerful single GPU cannot deliver its full value if the data center cannot support high power density, if the cooling system is unstable, if power distribution is imprecise, or if networking and storage become bottlenecks. The bottleneck in AI computing is expanding beyond the chip itself to include power, liquid cooling, data centers, networking, and scheduling.

This is very important for the transformation of China's token factories.

China has numerous industrial parks, old factories, data centers, cloud computing campuses, and local computing projects. Simply moving servers into these facilities does not constitute a Token Factory. A true Token Factory must possess high power density capacity, liquid cooling retrofitting capability, stable power supply, energy management, and task scheduling capabilities.

Without these underlying capabilities, computing power assets can easily become paper assets—appearing to have purchased equipment, but with low actual utilization, unstable customers, high energy costs, and significant depreciation pressure, ultimately turning into a new burden of heavy assets.

So, the first hard lesson from the Jane Street case is that the foundation of a token factory is not the whitepaper, but electricity and liquid cooling.

Hashrate must have a price; otherwise, GPUs would be like a public cafeteria.

What Jane Street most deserves for Chinese companies to learn from is not liquid cooling, but its internal computing power pricing mechanism.

Internal analysis mentions that Jane Street designed an internal computational currency called "Hive Bucks," allowing different teams to compete for GPU resources through an auction-like process. Public reports also indicate that Jane Street uses this internal currency to auction GPU compute time, helping teams allocate computational power based on the value of their tasks.

This mechanism is very important.

After many companies purchase GPUs, the biggest issue isn’t lack of demand, but the inability to prioritize it. Every team claims their model is critical, every project wants to run first, and every manager wants to claim more resources. As a result, GPUs end up being tied up by low-value tasks for long periods, while high-value tasks can’t get queued. Compute power, though technically a company asset, effectively becomes an internal cafeteria.

Jane Street’s approach more closely resembles a market mechanism. Teams that believe their tasks are more valuable bid higher using internal budgets. GPU time is no longer a free public resource but a productive asset with an opportunity cost. As a result, compute allocation shifts from administrative approval to an internal market.

This has direct relevance for the "Hashpower Token Factory."

True computational power tokens should not primarily be understood as tokens for external trading, but rather as internal units for measuring and settling resources. GPU hours, model invocations, task priorities, power capacity, cooling resources, and customer orders can all be priced and scheduled through a unified internal measurement system. Only by first fully understanding and optimizing computational power usage within the enterprise can computational power be productized, financialized, and assetized in the future.

If internal computing power has no price, discussing tokenization is likely just conceptual packaging.

Building your own infrastructure and moving to the cloud are not mutually exclusive—they represent a layered approach to core capabilities and elastic capabilities.

Many people ask, since Jane Street already has its own data centers, why does it still sign a large-scale AI cloud agreement with CoreWeave?

According to CoreWeave’s official announcement, Jane Street has committed approximately $6 billion to use CoreWeave’s AI cloud platform and invested $1 billion to acquire equity in CoreWeave, supporting large-scale machine learning and AI capabilities for trading. The announcement also stated that CoreWeave will provide computing resources, including next-generation NVIDIA Vera Rubin technology. Reuters also reported that, through this transaction, Jane Street has become one of CoreWeave’s key shareholders while gaining access to large-scale AI cloud capabilities.

This shows that Jane Street did not blindly build its own infrastructure or simply migrate to the cloud—it adopted a hybrid architecture.

Core, sensitive, low-latency, and highly customized workloads are suitable for self-hosting. Elastic, cutting-edge chip resources, cross-regional scaling, and intermittent burst demands can be entrusted to AI cloud providers like CoreWeave.

This approach is highly relevant for Chinese H-share listed companies.

Many traditional publicly listed companies transitioning into computing power often fall into two extremes: one is fully building in-house infrastructure, requiring massive capital investment, but failing to attract sufficient customers, leading to severe cash flow pressure; the other is relying entirely on external cloud services, lacking core infrastructure and asset accumulation, ultimately being limited to light consulting and light integration services, making it difficult to achieve higher valuations.

A more reasonable approach is layering.

The company can retain control over its core campus, power resources, liquid-cooled data centers, and critical customer loads, while partnering with cloud providers, GPU service providers, and model companies to access elastic computing power and technical capabilities. This approach ensures asset accumulation and operational flexibility, avoiding lock-in to a single model.

Trading revenue proves one thing: AI infrastructure can directly enter the profit system.

Why is Jane Street willing to invest so heavily? Because its AI infrastructure is not just for show—it integrates directly into its profit system.

Reuters reported that Jane Street's net trading revenue in 2025 reached $39.6 billion, surpassing multiple major competitors such as Citadel Securities and Hudson River Trading, as well as the trading revenues of several large investment banks. The report also noted that Jane Street's performance was aided by market volatility, its algorithmic trading capabilities, and gains from AI-related investments. The Financial Times of the UK also reported that Jane Street's 2025 revenue nearly doubled to $39.6 billion, highlighting its investments in AI-related companies such as CoreWeave, Anthropic, and Thinking Machines Lab.

This data shows that Jane Street’s AI data center cannot be evaluated solely based on “cloud cost savings.” For them, the value of compute power may be reflected in faster model training, deeper backtesting, more stable execution, enhanced risk identification, and greater trading capacity.

Traditional companies often measure AI’s ROI by how much labor they save. Jane Street’s approach is more aggressive: computing power isn’t about saving money—it’s about making money.

This is also something China’s Token Factory should learn. If a computing power center can only say, “I have X petahashes of computing power,” its value is incomplete. It must address more business-oriented questions: Who are these computing services for? Whose revenue are they helping to increase? What costs are they reducing? What cycles are they shortening? What customer loyalty are they building? And ultimately, can they generate sustainable cash flow?

Computing power only holds real value when integrated into a customer's business system.

The insight for Chinese tokens going global lies in transforming electricity into intelligent services.

China is discussing tokenization going global, but it最容易陷入模型叙事. DeepSeek, Qwen, Zhipu, Kimi, MiniMax, ByteDance’s video models—these are certainly important. But the Jane Street case reminds us that models are merely an intermediate layer. What truly creates industrial advantage is the complete closed loop spanning from electricity to computing power, from computing power to tokens, from tokens to applications, and from applications to revenue.

China's advantage lies precisely within this closed loop.

China possesses green power resources, power infrastructure, data center construction capabilities, engineering delivery capabilities, and a large model ecosystem, along with high-frequency application scenarios such as short videos, foreign trade, customer service, gaming, education, and finance. If these resources can be integrated into a Token factory, tokenization overseas will not merely mean the export of model APIs, but the export of China’s digital infrastructure capabilities.

Jane Street applies AI data center services to trading, while China’s token factories can serve a broader range of scenarios: foreign trade companies can use agents for product selection, customer service, translation, and marketing; short-form video companies can use AI for translation, dubbing, editing, and distribution; manufacturers can use AI for quoting, production scheduling, supply chain forecasting, and after-sales service; financial institutions can use AI for risk control, investment research, and trading support.

These applications consume tokens.

The greater the token consumption, the more valuable the computing power factory. The more mature the computing power factory, the greater the cost advantage for Chinese AI services expanding overseas. Once a cost advantage reaches scale, it becomes an industrial advantage.

The lesson for Hong Kong-listed companies is not to chase AI trends, but to restructure their balance sheets.

Hong Kong-listed companies, particularly those in real estate, industrial parks, construction, property management, energy, and manufacturing, should recognize a deeper direction from the Jane Street case.

Not every company needs to develop large models, nor does every company need to become an AI cloud provider. But many companies can repurpose their existing assets as part of AI infrastructure.

Old factories can be retrofitted into liquid-cooled data centers. Industrial parks can connect to computing nodes. Energy assets can be linked to data centers. Property management can integrate intelligent agents. Listed platforms can acquire AI computing assets through mergers, equity offerings, or strategic partnerships. The key is that the transformation must enter the revenue structure, not remain confined to announcements.

If a Hong Kong-listed company wants to talk about a token factory, it must at least answer several questions:

Is there stable power supply? Is there suitable space for renovation? Does it have liquid cooling and high-density data center capabilities? Are there real customers? Are there model or application partners? Is there an internal pricing mechanism for computing power? Is there a pathway to include computing power revenue in financial statements?

If these questions cannot be answered, AI computing power is merely a market capitalization management talking point. If they can be clearly answered, traditional listed companies have the opportunity to shift from valuing legacy assets to valuing digital infrastructure.

Jane Street's key insight is that a token factory is not a fundraising story, but an operating system. It requires technology and financial discipline, engineering capability and customer orders, computational power and an internal resource market.

China's version of the Token Factory shouldn't just focus on hardware—it must also learn about organizational structures.

If you only learn about Jane Street buying GPUs, implementing liquid cooling, and using CoreWeave, you're still missing the deeper picture.

What you should focus on learning is the organizational structure.

Jane Street integrates trading, research, engineering, computing power, and capital into a single system. Models are not academic papers by researchers, but integral components of the trading system; computing power is not an IT department cost, but the fuel for strategy iteration; internal currency is not a gimmick, but a resource allocation mechanism; external cloud is not a replacement for in-house infrastructure, but a means of elastic scaling; capital investment is not financial portfolio management, but a strategic positioning in the AI infrastructure ecosystem.

Many Chinese companies are undergoing AI transformation, but the issue is often not technology—it’s organizational silos. Business units don’t understand computing power, IT teams don’t understand revenue, finance teams focus only on costs, and the board only sees concepts, making it hard to close the loop.

For the Token Factory to succeed, its organizational structure must change. Hash power needs an owner, electricity needs a cost account, models need use cases, customers need usage metrics, tokens need an internal price, and revenue must be attributable. Otherwise, no matter how much hash power there is, it will only amount to scattered resources.

Conclusion: The financial giants of the future may begin as AI factories.

The Jane Street case illustrates one thing: the next generation of financial giants may appear to be trading firms, but beneath the surface, they could be AI factories.

It converts electricity into computing power, computing power into models, models into trading decisions, and trading decisions into profits. Once this chain is fully operational, an AI data center becomes not just a cost center, but a part of the profit system.

For China, the value of this case does not lie in copying Jane Street. China does not need every company to engage in quantitative trading, nor does it need every Hong Kong-listed company to build data centers with 4,032 GPUs. What truly needs to be learned is its underlying methodology: treat the token factory as a production system, not a conceptual one; treat compute power as an asset that can be priced, scheduled, audited, and financed—not merely a collection of servers; and treat token internationalization as a complete closed loop encompassing electricity, compute, models, applications, and settlement—not just a simple API export.

In the future, whoever controls the Token factory controls intelligent production capacity. Those who can transform old factories, old industrial parks, old data centers, and old listing platforms into token production lines will have the opportunity to achieve new valuations in the AI era.

Wall Street has already provided the sample.

Jane Street did not build a data center; it built an intelligent engine that transforms noisy data into trading profits. What China aims to do is integrate this engine into broader industrial scenarios, enabling manufacturing, foreign trade, content, finance, and the Hong Kong stock market to connect their own token factories.

In the old era, people looked at land and buildings.

In the new era, observe electricity, computing power, and tokens.

This is precisely what makes the Jane Street case truly worth analyzing.