IBM's Arvind Krishna: Open AI Models Hit 90% Effectiveness at Just 10% of Frontier Model Costs

Open-Weight AI Models Close the Performance Gap

IBM CEO Arvind Krishna made a compelling case for open AI models during the company's third-quarter 2026 earnings call, revealing that the performance gap between frontier models and open-weight alternatives has dramatically narrowed1

1

3

3

. In late 2024, when closed frontier models scored around 70 on broad-based benchmarks, open models lagged significantly at 20 to 30. By late 2025, that gap had shrunk to 70 versus 50, and open-weight AI models now deliver approximately 90% of the effectiveness of top-tier frontier models1

1

.

What makes this shift remarkable is the economics. Arvind Krishna emphasized that open AI models currently run at 10% to 25% of the token cost of frontier models1

1

3

3

. Krishna posed a direct question to enterprises: "Would you like something that's 10% of the cost if it's 90% as good as the best, or do you want to pay that 10 times higher cost?"3

3

. For most enterprise tasks like summarizing emails, examining contracts, answering customer calls, and composing emails, these cheaper alternatives prove sufficient1

1

.

On-Premises AI Infrastructure Demand Surges

The improved AI model economics are fueling explosive interest in on-premises AI infrastructure. Krishna noted that enterprises can deploy effective AI infrastructure with just two or three racks, making the barrier to entry surprisingly low3

3

. However, implementation challenges remain significant. Enterprises struggle to find kernel engineers, network specialists who can cluster models on GPUs in close network proximity, and experts who can optimize caching for faster response times1

1

.

IBM is positioning itself to address these deployment hurdles through software solutions and hardware offerings. Krishna predicted that demand for on-prem AI will drive sales of IBM's Spyre mainframe accelerator cards1

1

. The company's current-generation z17 mainframes have already generated $1 billion more revenue than any previous model, while the new rackable models delivered "the strongest midrange launch in IBM Z in nearly two decades"1

1

.

Enterprise AI Adoption Accelerates Beyond Experimentation

Arvind Krishna told financial analysts that enterprise AI is transitioning from experimentation into large-scale production, with IBM's hybrid cloud, software, data, automation, and security products forming the foundation for this shift2

2

. Red Hat bookings accelerated back to the high teens, while Red Hat OpenShift adoption strengthened as organizations scaled AI applications and agents2

2

.

Krishna emphasized that deploying more models or building more agents won't be enough. "The advantage will belong to a company that can operate AI securely, efficiently and sustainably across the enterprise," he stated2

2

. IBM has assembled its approach around four platforms: Confluent for real-time business context, watsonx including Bob and Orchestrate for software development and agent orchestration, HashiCorp and Apptio for identity, policy, costs and performance, and Red Hat for open hybrid deployment across clouds, on-premises systems, sovereign environments, and the edge2

2

.

Sovereign AI and Data Control Drive Strategic Shift

The economics of open-weight AI models are converging with another critical enterprise trend: sovereign AI. Companies increasingly want to keep sensitive data, logs, and model activity under their own control, fueling demand for private AI infrastructure3

3

. IBM's Bob, the company's agentic AI development technology, can now run on-premises, allowing customers in regulated and sovereign environments to maintain code and models on their own infrastructure2

2

.

Data control remains central to IBM's strategy. Krishna revealed that 72% of organizations report inadequate real-time data infrastructure2

2

. Confluent's Streamhouse continuously captures, transforms, and queries real-time data, making governed information available to applications, analytics, and AI adoption. "Agents need the current state of the business," Krishna explained2

2

.

HashiCorp Vault can register, authorize, audit, and revoke agents while securing their credentials. IBM Agent Identity and Vault are partnering with Nvidia's Open Agent Safety Platform to strengthen agent security2

2

. "Governance matters as much as intelligence," Krishna stated. "As agents multiply, enterprises need to know who each agent is, what authority it has, and what it can access"2

2

.

IBM Financial Performance and Market Implications

IBM reported $17.1 billion in total revenue for the third quarter, representing 5% year-over-year growth1

1

. Software revenue grew 9% to reach $7.8 billion for the quarter, with Red Hat as the standout performer, growing 11% and increasing annual recurring revenue by 30%1

1

. Red Hat's virtualization products have secured $800 million in deals with a pipeline of an additional $750 million, a solid result for a product line launched in 20241

1

.

IBM's infrastructure revenue grew 5% year-over-year to $3.7 billion, though this represents slower improvement than rivals Dell, HPE, and Lenovo1

1

. Mainframe revenue fell 20%, which IBM expected given it's now 18 months since the z17 launch1

1

. The stock declined 4.64% in overnight trading and has dropped 10.85% over the last year3

3

.

IBM is also addressing AI economics through innovation. Apptio now connects AI spending with business results, providing visibility into usage, token consumption, and operating costs2

2

. The open-source LLMD project, developed with Red Hat and Google, reuses processed context to help clients run large models at five to ten times lower cost through more efficient use of existing GPU clustering infrastructure2

2

. Watch for how enterprises balance the performance gap against cost savings, whether IBM's deployment software can solve the technical skills shortage, and if the sovereign AI trend accelerates data control requirements that favor on-premises infrastructure over cloud-based frontier models.