A major shift is reshaping the AI economy, and Canadian tech leaders need to understand it now. The most-used AI models are increasingly not the same models that capture the most revenue. Open weights systems are gaining rapidly in token volume because they are cheap, flexible, and controllable. Meanwhile, closed frontier models from companies such as OpenAI and Anthropic continue to command premium prices for the most difficult work.
For Canadian tech companies, this is not an abstract market trend. It is a strategic decision involving data sovereignty, cloud costs, supplier risk, intellectual property, security, and long-term competitiveness. Businesses across the GTA and the wider Canadian economy are being pushed to choose between renting frontier intelligence through an API and building durable internal capabilities around models they can host, customize, and control.
The result is a two-track AI market. Open weights models appear positioned to dominate usage. Closed models appear positioned to retain a disproportionate share of revenue. The businesses that understand why this split is happening will be better prepared to design AI systems that are economical, secure, and genuinely useful.
Table of Contents
- AI Usage Has Flipped Toward Open Weights Models
- Token Share and Revenue Share Tell Completely Different Stories
- Why the Best AI Models Still Command a Premium
- The Price Gap Is Astonishing, but Price per Token Is Not Enough
- China’s Open Weights Models Are Expanding the Market
- Privacy, Ownership, and the Strategic Value of Control
- Platform Risk Is Becoming a Board-Level AI Issue
- Open Weights Models Also Drive Infrastructure Demand
- The AI Economy May Split Into Three Layers
- The Geopolitical Risk Canadian Tech Cannot Ignore
- How Canadian Businesses Should Build an AI Model Strategy
- The Bottom Line: AI Volume Will Be Open, AI Revenue Will Be Premium
AI Usage Has Flipped Toward Open Weights Models
Data from Vercel, a platform used for web and AI application hosting, points to a sharp change in the composition of model usage. Over a period spanning June through August, the share of tokens processed through open weights models rose while the share processed through closed models declined.
That distinction matters. Closed weights models are proprietary systems that users access through a provider’s product or API. OpenAI’s models and Anthropic’s Claude models fit this category. The developer uses the system, but does not receive the model weights or gain direct control over its deployment.
Open weights models make their trained weights available for download and use. An organization can deploy these models in its chosen environment, fine tune them for a specific workflow, and decide how data is handled. They are often described casually as open source, although open weights and open source are not always identical concepts. The crucial business point is that open weights provide significantly more operational control.
For Canadian tech decision-makers, the increase in open weights usage signals that AI is moving beyond a narrow set of dominant cloud APIs. More organizations can now run capable models in private infrastructure, through third-party inference providers, or within controlled enterprise environments.
Token Share and Revenue Share Tell Completely Different Stories
The headline numbers reveal the complexity of this market. DeepSeek reportedly surpassed Anthropic in token share on the referenced platform, with 25.2% compared with Anthropic’s 24.5%. In other words, more AI output and processing activity was flowing through DeepSeek.
Yet token volume is not revenue. The spending data is dramatically different. DeepSeek accounted for only 2.8% of total model spending, while Anthropic represented 64.6%. Anthropic’s spending share was roughly 23 times greater, despite DeepSeek’s slightly higher token share.
This is the defining contradiction of the current AI market:
- Open weights models can serve a huge amount of total usage.
- Frontier closed models can capture much more economic value per task.
- Organizations are willing to pay substantial premiums when a small improvement in model capability changes the outcome.
The same pattern holds at a broader level. The top models account for close to half of token volume, but close to 90% of total revenue. That concentration is a warning and an opportunity for Canadian tech. Cost-efficient models can expand access to AI, while frontier providers retain enormous pricing power in high-stakes use cases.
Why the Best AI Models Still Command a Premium
AI benchmark gaps are often measured in only a few percentage points. On paper, a model that scores 95 instead of 98 may seem nearly interchangeable. In practice, the final few points can be enormously valuable.
Consider a use case such as high-frequency trading. A slight edge in accuracy, reasoning, speed, or decision quality can be worth billions of dollars over time. In these circumstances, an enterprise does not choose a model based only on the cheapest cost per token. It chooses the model most likely to deliver the correct answer under pressure.
This logic applies far beyond finance. Complex legal analysis, mission-critical software development, sophisticated research, high-value corporate transactions, and specialized enterprise planning can all justify higher spending when incorrect output creates significant business risk.
That is why leading closed-model providers can retain strong revenue even as less expensive alternatives improve. Canadian tech firms should recognize that the market is not moving toward one universal winner. It is fragmenting based on the value and risk of each task.
Low-cost AI is powerful when “good enough” creates value. Premium AI remains powerful when the cost of being wrong is much higher than the cost of inference.
The Price Gap Is Astonishing, but Price per Token Is Not Enough
The gap between frontier and open weights pricing can be extreme. One example compares an Anthropic frontier model priced at US$50 per million output tokens with DeepSeek V4 Flash at US$0.18 per million output tokens. The models are not in the same class, but the contrast makes a crucial point: for many everyday tasks, lower-cost systems can be overwhelmingly more economical.
If an inexpensive model can support internal knowledge searches, summarize documents, classify content, draft routine communications, or power basic software workflows, few organizations will want to spend premium rates unnecessarily. This is especially significant for startups, mid-market firms, and Canadian tech teams operating under constrained cloud and AI budgets.
However, the price shown on a model card is not the full cost of doing work. A better measure is cost per completed task.
A lower-priced model may use more tokens, require more retries, or need more agent steps to reach the same conclusion. One example compared Kimi K3 with GPT 5.6 Sol. Kimi K3 had a lower per-token price, but the total cost of completing a task was close: US$0.84 versus US$0.96. The difference came from token efficiency. The cheaper model needed more tokens to finish the same work.
Canadian tech leaders evaluating AI should therefore avoid simplistic procurement comparisons. The right assessment includes:
- Cost per successfully completed workflow
- Accuracy on company-specific tasks
- Latency and reliability
- Human review time required
- Security and privacy obligations
- Deployment, maintenance, and infrastructure costs
- Long-term dependence on a single model provider
China’s Open Weights Models Are Expanding the Market
One of the most consequential developments in AI is the strong performance of open weights models released by Chinese companies. DeepSeek, Kimi, Qwen, and GLM are increasingly central to the global model ecosystem. Their systems are widely used because they can offer high capability at a much lower cost than premium closed alternatives.
Artificial Analysis rankings illustrate how close the competition has become. Claude Opus 5 Max sits at the leading edge, with GPT 5.6 also near the top. But open weights systems are close behind. Kimi K3 Max, GLM 5.3, and Qwen 3.8 appear among high-ranking models, alongside Meta’s Muse Spark.
This matters because most enterprise workloads do not require absolute frontier performance. Many operational tasks are repeatable, bounded, and highly specific. A capable open weights model, adapted using internal data and evaluations, may outperform a general-purpose frontier model on a company’s actual workflow.
For Canadian tech, the immediate benefit is clear: competitive model supply lowers costs and broadens access. More businesses can experiment with AI without accepting a premium closed-model bill for every request. This development can make AI adoption more practical for organizations outside the largest banks, telecom companies, and global platforms.
Privacy, Ownership, and the Strategic Value of Control
Cost is only one reason enterprises are turning toward open weights models. Privacy and ownership can be equally important, especially in sectors that handle confidential records, proprietary research, legal materials, or sensitive customer information.
When a business runs an open weights model in a controlled environment, it can determine where prompts, documents, outputs, and fine-tuning data reside. It can choose its hosting arrangement, govern access, and limit exposure to external platforms. That is a powerful proposition for Canadian tech organizations considering sensitive deployments.
Several notable companies have already used open weights models in their products and services. Examples include Thomson Reuters using Qwen, Harvey using Kimi models for legal AI, Airbnb using Qwen, Cursor using Kimi K2.5, and Perplexity using DeepSeek.
Harvey provides a useful illustration of the opportunity. The legal AI platform fine tuned Kimi K3 to create a model adapted to legal work. Its customized system performed strongly across legal benchmarks, including legal agent, contracts, and corporate law evaluations. The broader lesson is not that one model automatically solves legal work. The lesson is that a strong base model can become much more valuable when it is adapted to a specialized domain.
That capability changes the business equation. Rather than sending every task to a general-purpose external system, an organization can build an AI layer that reflects its own terminology, workflows, standards, and institutional knowledge.
From Renting Intelligence to Building Enterprise Equity
A useful analogy compares closed-model usage to renting a home. Renting can be convenient. The property is available immediately, and someone else handles much of the underlying maintenance. But the renter is not building equity in an asset that compounds over time.
Canadian tech companies that rely exclusively on external AI APIs may receive powerful capabilities quickly, but they may not be building a lasting intelligence asset. Their prompts, workflow designs, context structures, and operating patterns can become deeply tied to another company’s platform.
By contrast, an organization that fine tunes an open weights model, develops proprietary benchmarks, organizes high-quality internal data, and integrates the model into its operating processes is building an asset that can improve over time. The model itself may change, but the enterprise knowledge, evaluation framework, and implementation discipline remain valuable.
Platform Risk Is Becoming a Board-Level AI Issue
Platform risk arises when an organization builds a critical part of its business on another company’s service and loses meaningful control over costs, availability, product direction, or competitive exposure.
In AI, this risk is particularly acute. A company that depends entirely on one closed-model provider can face several challenges:
- A provider could raise prices or alter usage terms.
- A model change could affect performance in production workflows.
- Access could be restricted or interrupted.
- The provider could launch products that overlap with a customer’s offering.
- Enterprise context and usage patterns could create uncomfortable strategic dependencies.
Open weights models do not eliminate every risk. They require technical expertise, hosting capacity, evaluation, governance, and ongoing maintenance. But they create more options. A business can self-host, move among inference providers, or use a range of specialized deployment partners. Competition among those providers can improve pricing and reduce lock-in.
For Canadian tech executives, the strongest position may be a deliberate multi-model strategy. A business can preserve access to frontier closed models for demanding tasks while developing open weights capabilities for private, high-volume, or specialized workloads.
Open Weights Models Also Drive Infrastructure Demand
The rise of open weights AI is not necessarily bad for infrastructure providers. In fact, increased model availability can expand demand for compute. When more developers, startups, and enterprises can deploy AI models, total token generation rises. More tokens require more chips, servers, hosting capacity, and inference infrastructure.
Each token still represents real computational work. An open weights token does not become free to generate simply because the underlying model is downloadable. The cost structure differs because of the surrounding ecosystem, pricing strategy, deployment choices, and reduced proprietary premium.
This is why greater open weights adoption can be positive for chip and data-centre demand. It also creates a major opportunity for Canadian tech infrastructure providers and cloud operators. Organizations that want more control over model execution will need reliable environments in which to run them.
The likely outcome is not the end of centralized AI infrastructure. It is a wider distribution of AI deployment. Some workloads will remain with the largest frontier labs. Others will move to specialized inference providers, private clouds, and enterprise-controlled systems.
The AI Economy May Split Into Three Layers
A useful framework divides future AI spending into three distinct segments. This structure helps Canadian tech leaders avoid treating every model choice as a binary decision between “open” and “closed.”
- Cheap generalist models: Commodity open weights systems likely handle the greatest volume of routine work while accounting for a smaller share of total spending.
- State-of-the-art specialists: Enterprises can use open weights models with proprietary context and fine tuning to create specialized systems. This segment can capture substantial business value because it is tailored to particular industries and workflows.
- Absolute frontier generalists: Closed labs will continue to provide the highest-performing broad models. These models may serve a smaller portion of total volume, but capture significant revenue because of their premium value in hard tasks.
This three-layer model is more realistic than claims that open weights AI will completely replace closed models or that proprietary labs will permanently control every important application. Different tasks demand different economics and different levels of control.
The Geopolitical Risk Canadian Tech Cannot Ignore
The rise of Chinese open weights models creates a longer-term strategic concern. Today, many of these models are trained and served using Nvidia hardware. But China is also developing domestic chip capabilities. If models and chips become closely co-designed, organizations that build deeply on those models could eventually become dependent on a particular technology stack and supply chain.
This issue has direct relevance for Canadian tech, even though much of the debate is framed around the United States. Canada operates in a North American technology and security environment, while also seeking access to global innovation. Businesses must balance the attraction of lower-cost and high-performing models with a clear understanding of vendor origin, infrastructure dependencies, data handling, and future portability.
The essential concern is not that organizations should reject open weights models. It is that model selection should be part of a broader resilience strategy. A company should know whether its applications can move across models, hosting providers, chips, and cloud environments if conditions change.
How Canadian Businesses Should Build an AI Model Strategy
The race between open weights and closed frontier models is moving quickly. Canadian tech leaders should not wait for a single standard to emerge. Instead, they should build the internal capability to test, compare, govern, and deploy multiple model types.
A practical strategy includes the following actions:
- Classify workloads by risk and value. Reserve premium frontier models for problems where the best possible answer creates meaningful economic value.
- Use lower-cost models for scale. Evaluate open weights models for routine, high-volume, and cost-sensitive tasks.
- Create internal benchmarks. Generic public benchmarks are informative, but a company’s own data and workflows determine which model is actually useful.
- Measure completed-task economics. Include retries, token use, human review, latency, and operational overhead.
- Protect sensitive information. Establish clear rules for what data can enter external APIs and what should remain in controlled environments.
- Preserve portability. Avoid designing critical workflows around a single provider without a realistic exit path.
- Develop fine-tuning expertise. Specialized capability can create a meaningful competitive advantage over generic AI use.
The most important capability is not merely choosing a model. It is building organizational fluency around AI deployment. Canadian tech firms that learn to operate open weights models, manage private data, and compare systems rigorously will have far more leverage as prices fall and competition intensifies.
The Bottom Line: AI Volume Will Be Open, AI Revenue Will Be Premium
The AI market is becoming more competitive, more decentralized, and more strategically complicated. Open weights models are likely to win significant token volume because they are inexpensive, customizable, and easier to control. They can expand access to advanced AI across the Canadian economy and enable companies to build specialized systems around their own knowledge.
At the same time, OpenAI and Anthropic are likely to retain major revenue opportunities. Their frontier systems deliver a level of capability that remains highly valuable when the consequences of error are large. Recent aggressive price reductions by OpenAI also show that the frontier labs are responding to competitive pressure rather than standing still.
For Canadian tech leaders, the winning approach is neither blind loyalty to closed APIs nor automatic enthusiasm for every open weights release. It is a portfolio mindset. Use premium intelligence where it earns its premium. Build controlled, customized capability where privacy, cost, and enterprise ownership matter most.
The question facing every Canadian technology organization is urgent: is its AI strategy creating a short-term convenience layer, or a long-term intelligence asset that compounds with the business?
Frequently Asked Questions About Canadian Tech and Open Weights AI
What is the difference between open weights and closed AI models?
Closed models are accessed through a provider’s platform or API, while open weights models can be downloaded, deployed, and customized by an organization. Open weights models generally offer greater control over hosting, data handling, and fine tuning.
Why are open weights models gaining token volume?
Open weights models are often substantially less expensive and can be run in a variety of environments. Their flexibility, lower cost, and customization potential make them suitable for a large volume of routine and specialized enterprise tasks.
Why do closed AI providers still capture more revenue?
Closed frontier models can provide better performance on the most difficult tasks. When a small gain in accuracy or reasoning has major financial consequences, organizations may be willing to pay a substantial premium for the best available model.
What should Canadian tech companies measure when selecting an AI model?
Companies should measure the cost per successfully completed task, not only the cost per token. Evaluations should include model accuracy, token efficiency, latency, privacy requirements, human review time, infrastructure costs, and provider dependency.
Can open weights models reduce platform risk?
They can reduce dependence on a single model provider because an organization may self-host a model or choose among multiple inference providers. However, open weights deployments still require strong governance, technical expertise, infrastructure planning, and security controls.
