Vendor and Token Risk in AI: Howard Cohen Explains Shifting Economic Pressure on MSPs

The episode highlights a shift from technology selection to operational risk management in the AI landscape for MSPs. Service providers are being forced to navigate the fast-changing interplay between AI models, the harness software that mediates their deployment, and the financial realities of consumption-based billing. The rapid proliferation of open-source and open-weight AI models, alongside market behaviors from closed vendors and regulatory interventions, is introducing volatility and uncertainty in both cost structures and client offerings. This dynamic creates structural challenges related to margin maintenance, vendor dependency, and responsibility for AI-driven decisions.

The discussion cites the release of GLM 5.2, an open-weight model from Z AI, which now rivals expensive closed models on key benchmarks at a fraction of the cost. At the same time, large-scale investments by commercial AI vendors have yet to deliver returns on expectations, with reports indicating businesses that adopted AI are not seeing projected value. Specific attention is given to operational constraints such as compute scarcity, token consumption variability, and export policy restrictions impacting AI availability. The episode notes that these pressures are driving both vendors and MSPs to reconsider the viability of reliance on expensive, closed offerings versus investigating open alternatives.

Supportive examples include the proliferation of AI “harnesses” (middleware layers like Perplexity, Claude Code, and Cowork) that sit between service providers and underlying AI models, increasing both choice and complexity. Token billing models are highlighted as a source of unpredictability for MSPs, with vendors like Atera and ConnectWise experimenting with different abstractions to shield or pass through token risk to service providers. The potential for on-premises AI deployments using smaller language models is discussed as a cost-mitigation strategy, though this raises further questions about data privacy, infrastructure burden, and long-term vendor roles. Additionally, uncertainty is flagged around sustainability of leading vendors, with projections that at least one major AI player may exit or be acquired within a year due to financial vulnerability.

For MSPs and IT service leaders, these structural and supporting developments translate into increased operational and financial complexity. There is a pressing need to evaluate not just which AI technologies to adopt, but how to architect solutions that can withstand rapid vendor movement, cost swings, and evolving regulatory requirements. Practical safeguards include testing open-source AI models alongside commercial offerings, exercising caution in vendor selection, and closely monitoring evolving consumption billing models. Preparing staff and clients for adaptive, process-oriented approaches—rather than fixed solutions—is positioned as a necessary step to maintain resilience as the AI adoption cycle continues to correct course.

Supported by:
Pax8
CometBackup
Guardz