iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Chinese AI models are gaining ground not necessarily by beating every benchmark, but by delivering acceptable results at lower token costs. Intel CEO Lip-Bu Tan said at a DBS-organized panel in Singapore on October 12, 2026, that models such as DeepSeek, Kimi and GLM are “good enough” for many applications and are being adopted quickly because they are cheaper to use.
What Lip-Bu Tan said about Chinese AI models
Reuters reported that Tan described Chinese frontier models as having lower token costs and very fast adoption. His point was practical rather than absolute: a model does not need to lead every benchmark if it performs reliably enough for a company’s target task and costs substantially less per request.
“I’m a strong believer in open source, that’s the best way to drive efficiency,” Tan said. He added that markets will determine which models benefit particular use cases. That puts the emphasis on matching a model to a job instead of declaring one universal winner.
Why token price can matter more than a benchmark lead
Most commercial AI services bill separately for input and output tokens. Input tokens represent the prompt and supplied context; output tokens represent the generated answer. A lower price per token reduces the cost of every call, so the savings compound when an application handles millions of requests.
#1 Best Overall
Lower inference costs can change a product decision. A company may be able to add an AI feature to every customer account, run longer context windows, or automate more internal work when each interaction costs less. The relevant measure is therefore cost per useful task, including retries and human review, rather than the headline price alone.
Where “good enough” is often sufficient
- Customer-support classification and routing
- Document extraction and summarization
- Drafting routine business correspondence
- Code completion for common patterns
- Internal search and question answering over controlled documents
More demanding work—such as difficult mathematical reasoning, safety-critical decisions, or highly specialized research—may justify paying for a model with stronger verified performance. The cheaper option is not automatically the better option if errors create costly rework.
Rank #2
Which Chinese models were named
| Model | How it appears in the discussion | What a buyer should verify |
|---|---|---|
| DeepSeek | One of the examples Tan cited as “good enough” for many uses | Task accuracy, current API rates, regional availability and data-handling terms |
| Kimi | Named by Tan; its K3 release was associated with a sharp increase in downloads | Latency at your workload, context limits and enterprise support |
| GLM | Named by Tan as another practical, lower-cost option | Language performance, deployment choices and compliance documentation |
| Qwen | Airbnb CEO Brian Chesky described Alibaba’s model as “fast and cheap” in October 2025, according to Yahoo Finance | Whether the available version and terms fit your geography and security requirements |
The descriptions above do not establish that any one model is best overall. They identify candidates for testing against a specific workload.
Free tools Windows power users keep installed
One-click scans. No signup required.
Signs that adoption is accelerating
Associated Press reporting found Chinese models occupying five of the top positions on OpenRouter during the period it examined. OpenRouter rankings are a snapshot, not a complete measure of enterprise usage, but they indicate strong interest among users choosing among hosted models.
Sensor Tower estimated that Kimi exceeded 930,000 downloads in the week after K3’s release, up 200% from the previous week. U.S. downloads were approximately 86,000, a 387% weekly increase. Download figures measure consumer interest and app distribution, not paid business deployment, so they should be read as adoption signals rather than proof of production reliability.
How companies should compare models
Tan’s argument is most useful when turned into a procurement test. Compare alternatives on the same representative workload and record the full operating cost.
- Define the task. Specify the inputs, required output format, acceptable error rate and human-review process.
- Measure useful-task cost. Include input and output tokens, retries, tool calls, storage, monitoring and reviewer time—not just the advertised token rate.
- Test quality on your data. Use a held-out evaluation set that reflects your languages, documents and edge cases. Check factuality, instruction following and refusal behavior.
- Record latency and capacity. Measure time to first token, total response time, throughput and rate-limit behavior at expected peak load.
- Review governance. Confirm where prompts and outputs are processed, whether they are retained or used for training, available contractual controls, and the jurisdictions involved.
- Price switching and deployment. Account for SDK changes, prompt rewrites, fine-tuning, observability, staff training and the cost of running a fallback model.
A model that is 30% cheaper per token can still be more expensive overall if it produces twice as many unusable answers. Conversely, a modest quality gap may be acceptable when the lower-cost model handles routine requests and a more capable model is reserved for exceptions.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Why Intel is discussing model economics
The comments come as Intel reports growing demand tied to data centers and AI. Yahoo Finance, citing Intel’s 2026 earnings release, reported that data-center and AI revenue rose 59% to $6.3 billion in the second quarter. Yahoo Finance also reported, based on Intel’s 2025 annual report, that China accounted for about $12.7 billion of Intel’s $52.9 billion in 2025 revenue by billing location.
Best Value
Those figures are company-reported and provide business context for Tan’s remarks. Cheaper models can increase the number of AI requests companies run, which may expand demand for computing infrastructure even when each individual request generates less revenue for a model provider.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What lower-cost models do not solve
- Accuracy risk: “Good enough” depends on the consequence of an error. Medical, financial, legal and safety-sensitive workflows require stronger validation.
- Data sovereignty: A low token price does not answer where data is processed or which organization can access it.
- Availability: Popular models can face rate limits, regional restrictions or sudden changes in access terms.
- Operational dependence: Switching providers may require prompt, schema and evaluation changes even when models appear API-compatible.
- Benchmark ambiguity: Results vary by test set, language, context length and prompting method; a leaderboard position is not a substitute for your own evaluation.
Bottom line for buyers
Chinese models are becoming credible competitors because practical performance and low token prices can outweigh a small benchmark disadvantage. DeepSeek, Kimi and GLM are the examples Intel’s CEO highlighted, while adoption indicators for Kimi and OpenRouter’s rankings show that users are actively testing the value proposition. Businesses should treat “good enough” as a measurable threshold for each workflow: test quality, latency, governance and total cost together, then use the least expensive model that consistently meets the requirement.
Frequently Asked Questions
Does a lower token price guarantee a cheaper AI system?
No. Total cost also includes retries, human review, infrastructure, monitoring, support and any engineering work needed to switch or maintain the model.
Recommended Free Tools
Are Chinese AI models suitable for every business task?
No. Suitability depends on the task’s error tolerance, data-governance requirements, language needs, latency target and availability in the company’s region.
What did Intel’s CEO mean by “good enough”?
He meant that a model can win adoption by meeting the practical requirements of a use case at a lower cost, even without leading every benchmark.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
