TF Securities said in a recent analysis that open-source AI models, led in part by China-based efforts, are breaking down the dominance that closed-source models once held. Rather than hurting the AI industry as a whole, the firm argued, the shift could create a better opening for cloud providers to regain pricing power and move up the stack as compute platforms. In that process, profit pools would move away from model companies and toward cloud operators with routing control and vertical integration capabilities.
Cloud providers built the roads but did not control the tolls
According to the report, cloud providers have spent the past few years taking on the heavy capital burden of AI infrastructure, including GPU purchases and data center construction. But because most AI workloads still had to call a small number of closed-source models, including those from OpenAI, the real point of demand capture and pricing authority remained at the model layer.
TF Securities described that arrangement as a core contradiction in the AI business model. Cloud companies carried the risk of capital-intensive investment, while the model layer captured a larger share of incremental profit. In the firm’s framing, closed models functioned like toll booths, collecting fees on roads that cloud providers spent time and capital to build. That imbalance, the report said, has been a major reason cloud valuations remained under pressure.
Open-source models are changing the profit split
TF Securities said that structure is starting to shift. Open-source models represented by Meta’s Llama series and DeepSeek have continued to improve in performance while driving down inference costs, making top closed-source systems less necessary for every task.
The report did not argue that closed models are losing all of their advantages. It said they still retain pricing premiums in complex reasoning and high-value use cases. Even so, their monopoly position is being eroded. As models become more interchangeable, model companies lose bargaining power, while cloud providers gain more room to control customer relationships and make routing decisions.
In that view, open-source AI is not destroying industry profits. It is reallocating them.
Tiered routing becomes the key mechanism
The report said AI workloads are likely to move away from the idea of using one model for every job and toward a tiered routing structure. Under that setup, the most difficult tasks would still be assigned to top closed-source models. More routine work would be routed to cheaper small models or open-source models. Stable, high-volume workloads would be deployed by cloud providers on their own infrastructure, with in-house ASICs used to push down inference cost per token even more.
That structure, TF Securities said, gives cloud operators three ways to monetize: renting out GPU compute, self-hosting open-source models, and selling tokens directly to customers. As a result, the share of profit flowing to outside model suppliers would shrink. In the report’s framing, models would no longer sit above the stack as toll booths. They would become compute resources that cloud platforms can dispatch more freely.
Hardware demand still depends on one core variable
TF Securities also said the hardware outlook is not a one-way bullish story. On one side, fiercer open-source competition could bring more participants into model training, while lower token prices could stimulate usage growth. Both would support hardware demand.
On the other side, progress in model compression, repeated gains in inference optimization algorithms, and large-scale deployment of in-house ASICs by cloud providers are all reducing the amount of general-purpose GPU compute needed for each token. That is a meaningful headwind.
The final direction of hardware demand, the report said, comes down to one question: can growth in token volume and compute consumed per task keep outpacing improvements in algorithm efficiency and chip efficiency? If the answer is yes, total compute demand can keep expanding. If not, cloud providers’ capital expenditure growth could peak earlier than expected even if token calls continue to climb quickly.
Whether the cloud-hardware flywheel turns is the main debate
TF Securities laid out a bullish flywheel built on several linked steps. Open-source models weaken monopoly power at the model layer. Cloud providers then improve compute monetization through routing and vertical integration. Lower inference prices create incremental demand, and demand growth once again exceeds efficiency gains. That allows data centers to sustain high utilization and strong return on invested capital, or ROIC, which in turn supports continued cloud spending on compute infrastructure and backs longer-term hardware demand growth.
The report also made clear that this flywheel is not guaranteed. Whether it can really turn is the central assumption behind the thesis and the biggest point of disagreement in the market. TF Securities’ position was direct: only if demand elasticity is strong enough to outweigh efficiency compression can cloud providers and hardware vendors form a real positive feedback loop.