11/16 2026
488
Text and Images | Tang Jie
At the 2026 Baidu World Conference, the release that caught me off guard—and genuinely impressed me—was the Token Exchange.
The Token Exchange is a large model trading platform launched by Baidu. Robin Li called it the exchange for large models at the conference. It brings together quality models and model providers, enabling enterprises and even individuals to conveniently find, purchase, and use large model tokens (Token), making "trustworthiness" a prerequisite for model trading.
The platform's operation can be summarized in four stages: "admission—monitoring—rating—trading." Models and suppliers undergo verification before entering the platform; after entry, they are subject to continuous probing and invocation monitoring; their performance is converted into credit records visible to users; and users make selections and settlements under comparable price and credit information. Supporting these four stages are the platform's two core capabilities: "authenticity verification" and "credit assessment."
First, where the surprise lies. From publicly available information, among leading cloud vendors, Baidu is the first to build open onboarding, authenticity verification, and credit rating into an independent platform. In the past, those selling multiple models together were mainly relay stations reselling model APIs and independent aggregation platforms.
Where the surprise lies: A new platform typically takes time to build up both supply and demand, yet the Token Exchange accumulated a considerable volume within a month of trial operation. According to Baidu, within one month of launch, onboarded supply resources exceeded 100, daily average invocations surpassed 50 million, and daily average Token invocations exceeded 700 billion. The rapid volume growth indicates that the market had actually been waiting for such a product to emerge.
In my view, the true value of the Token Exchange lies in answering two questions that AI cannot avoid as it moves toward industry: Can Token become a true commodity, and will enterprises dare to put AI spending into their daily operations?
01 Intelligence Is Not Yet a Commodity
For intelligence to truly become an industry, its carrier—Token—must first become a commodity. Yet in the current trading process, it still faces three major uncertainties.
First, the same model listed with different suppliers may have the same name and similar pricing, but the actual service delivered differs, making direct comparison difficult. A monitoring report by the China Academy of Information and Communications Technology in October 2026 showed that for the same model across different platforms, the maximum difference in TPS (tokens generated per second) was nearly twofold—something invisible on a price sheet.
Second, speed differences are still a minor issue. More fundamentally, what is delivered may not be the model claimed—for example, passing off a lower-tier model as a premium one, or applying quantization (compressing the model to reduce costs) to gain throughput—and buyers have no way to verify. In March 2026, the CISPA Helmholtz Center for Information Security in Germany tested 24 API interfaces reselling large models and compared them using model fingerprints (unique features in model outputs that can identify which model it actually is); 45.83% failed verification.
Finally, a model may be qualified at the time of purchase, but no one guarantees it won't change afterward—and it's not necessarily intentional fraud. Since August 2025, Anthropic's Claude experienced degraded response quality for over a month due to infrastructure failures. Its own internal evaluations did not detect this; it was only discovered through user feedback.
These three uncertainties combined leave buyers unable to distinguish good from bad and only able to compare prices; they also leave conscientious suppliers with high costs and high prices that buyers cannot justify, ultimately preventing them from commanding fair prices.
This industry status quo is precisely what the Token Exchange aims to break through: making Token comparable in price, verifiable, and accountable.
The first thing the Token Exchange does is unify testing conditions, metrics, and versions—first determining whether different providers' services are comparable under the same conditions, then displaying comparisons so that pricing for the same model can be compared. On November 12, on the Token Exchange's homepage, the popular model glm-5.3-flash had 12 suppliers listed simultaneously, allowing users to compare prices directly.
The second thing is verifying whether what suppliers sell is actually the model they claim. The Token Exchange first uses model fingerprint comparison to identify the model's identity, then uses capability tests and performance tests to verify the claimed capabilities. Individual buyers cannot do this: Moonshot's test suite for testing third-party channels requires two 8-GPU servers and about 15 hours per run, and must be re-run whenever the model version changes.
Services change, and one-time verification is insufficient, so the third thing is continuous monitoring. The Token Exchange conducts regular probing (periodically dispatching test requests for spot checks) while monitoring real invocations, converting each supplier's performance into a credit score by measuring the gap between its claims and actual measurements, and alerting users when ratings change. One-time matchmaking thus becomes an accumulable, queryable performance history.
These three things change the way intelligence is measured, verified, and priced, transforming Token from a project-based business requiring fresh investigation and negotiation for every order into a commodity that can be bought and sold with confidence. From the Token Exchange's logic, it's not hard to see that this is actually the necessary path for any factor of production to become a commodity.
Electricity first had kWh metering, grid dispatch, and power quality assessment; telecommunications first had call records, settlement prices, and service quality metrics. In all cases, metering was unified first so that prices had a basis, then third-party quality inspection and performance records followed so that buyers and sellers dared to engage in long-term cooperation.
Intelligence has only reached the first step—Token-based billing. Quality inspection and performance records still rely on each buyer fumbling through contracts individually, with varying standards. By June this year, the national daily average Token invocation volume had exceeded 500 trillion. The larger the volume, the more this duplicated investment grows, and transaction costs rise accordingly. What the Token Exchange supplements is precisely these latter two steps, while also aligning the metering standards.
Policy is also pushing this forward. In March this year, Liu Liehong, Director of the National Data Administration, called Token a "value anchor of the intelligent era." In August, the General Office of the Ministry of Industry and Information Technology issued a document proposing increased procurement of Token and other services. In September, Beijing included establishing a Token quality evaluation system and building a Token distribution platform in its action plan.
Liu Peng, Professor at the School of Public Administration and Director of the Center for Evidence-Based Governance and Public Performance at Renmin University of China, believes that trustworthy verification and authenticity governance of large models have become important issues. The Token Exchange is "an exploratory innovative practice" that provides a professionalized carrier for trading related factors. Advancing through market mechanisms helps prevent redundant construction of large models and can provide reference for the subsequent construction of national-level standards systems.
Only when good products can be recognized can they command fair prices, and competition will shift from who is cheaper to who is more reliable. For intelligence to become an industry, it must first become a commodity.
02 Bringing Intelligence into Budgets
Imagine an enterprise undergoing intelligent transformation that wants to procure large model services from the market. Before the money is spent, it must pass through several internal gates—procurement, business and IT, and finance—each with its own concerns.
What the procurement department sees on the bill is only the Token unit price. What's invisible is due diligence, price comparison, trials, and disputes after failures. When connecting several suppliers simultaneously, evaluations must be re-run, price inquiries restarted, and responsibilities redrawn. These costs don't appear on the bill but genuinely consume decision-making time and energy.
What makes business and IT departments even more hesitant is the situation after something goes wrong. When integrating AI into core systems, the biggest fear is often not insufficient model capability but that no one takes responsibility when problems arise: the model vendor blames the prompt, the platform blames model behavior, the integrator blames data quality—each passing the buck, ultimately causing problems to be shelved and left unresolved. The Token Exchange's credit records at least provide documented evidence of whether the model service component has met standards.
Ultimately converging on the finance department, expenditures with unclear costs and undefined responsibilities can only go through special budgets, negotiated annually and settled upon acceptance. To enter the annual operating budget, billing must be usage-based, auditable, and predictable—and the Token Exchange's metering, authenticity verification, and credit assessment correspond precisely to these three prerequisites.
Enterprises that can maintain their own evaluation teams and handle all of this internally are few. Most are three types: SMEs with limited budgets that must use AI, ISVs (Independent Software Vendors) that profit by reselling model capabilities, and local state-owned enterprises and industry clients constrained by strict procurement processes. Their price sensitivity varies, but all can only rely on externally verifiable references.
For them, comparable prices and third-party performance records provide a basis for choosing suppliers and make upward reporting clear. A unified entry point and switchable supply pool lower the usage threshold, and only then will the application layer's explosive growth truly accelerate.
This is the value the Token Exchange brings to enterprises seeking to transform their businesses with AI: leveling the coordinate system for Token and making AI implementation a reality.
Few vendors in China possess chip, cloud, model, and agent capabilities simultaneously—and whether this reference framework is trusted depends precisely on who does the evaluating: authenticity verification must be accurate, and credit assessment must be sustained over time. Baidu's confidence in launching the Token Exchange comes from its full-stack technology and long-term accumulation.
To evaluate models accurately, one must first understand models. Judging "is this really the model it claims to be" and "do claimed capabilities match actual capabilities" requires the internal perspective Baidu has accumulated through years of model training and evaluation—something external institutions find hard to replicate simultaneously.
To measure over time, one must be able to sustain operations. Probing and full-link real-time monitoring must run continuously—a ongoing expense that requires sufficient compute and inference engineering capability to bear.
Beyond evaluating and measuring, the platform must also handle multiple models and multiple suppliers. Baidu's Qianfan platform has long integrated Wenxin and third-party models uniformly, and the Token Exchange is an extension of this engineering capability. For credit ratings to truly enter enterprise procurement processes, government-enterprise delivery, compliance, and operations systems are also needed.
Beyond technology, the core competitiveness of an exchange is matchmaking understanding—that is, understanding different buyers' needs. Large model buyers include some of the most demanding large clients with stringent requirements for availability, compliance, and accountability, as well as SMEs and even individuals.
According to Baidu, Baidu AI Cloud has served 80% of central state-owned enterprise clients and ranked first in large model market share among central SOEs in 2025, making it more familiar with large clients' compliance and accountability requirements. At the same time, Baidu also has experience serving SMEs and OPCs (one-person companies), enabling it to understand different buyers' needs across multiple dimensions.
Whether a rating system carries weight depends on two things: whether the evaluator knows the industry, and whether buyers actually use it to make procurement decisions. Baidu has both.
03 Conclusion
Once ratings are used by buyers, the beneficiaries are not just buyers.
The compute market is currently unevenly heated. Media investigations show that top internet companies' self-built compute clusters are in short supply, while monitoring data from Inspur AI Research Institute indicates that in the first quarter of this year, the average GPU utilization rate of national intelligent computing centers was only about 30%.
Top-tier clusters being insufficient shows that demand exists and the gap needs to be filled by compute elsewhere. But on the other side, many compute centers have GPUs and can run model services, yet cannot prove their services meet standards or find customers. What's missing between the two sides is a place buyers can trust and find.
What the Token Exchange connects is precisely this link, opening a monetization outlet for existing compute and improving the compute efficiency of the entire industry.
And this layer's value already has valuation benchmarks overseas. In August 2026, Stripe announced the acquisition of model aggregation platform OpenRouter, with media-reported prices between $7 billion and $8 billion. This layer is valuable because it simultaneously brings several things: the industry gains unified trading standards, enterprises dare to write AI into annual budgets, and idle compute finds buyers.
Models will keep evolving with each new generation, but once standards are set, they will serve the entire market for the long haul. What the Token Exchange establishes is precisely this set of standards—and in the intelligence industry, its significance is on par with any large model.
Disclaimer: This article is intended solely for learning and exchange purposes and does not constitute investment advice.
We welcome your likes, views, and shares. Your support drives our continued updates!