On October 21, 2026, IBM CEO Arvind Krishna said on the company’s third-quarter earnings call that if something costs only one-tenth as much but delivers 90% of the results, enterprises will choose it. According to The Register, he said that the inference cost of current open-weight models is only 10% to 25% that of frontier closed-source models, while their effectiveness has reached 90% of the latter.
Benchmark Scores: From a Cliff to a Slope
Krishna described the evolution curve: in 2024, closed-source frontier models scored 70 on broad benchmark evaluations, while open-source alternatives scored only 20 to 30; by the end of 2025, the gap had narrowed to 70 versus 50; by 2026, it had reached “90% effectiveness.” Independent evaluations show that in 2026, leading open-source models such as Qwen3-235B and Llama 3.3 have narrowed the gap with GPT-4o on benchmarks such as MMLU to 5 to 8 percentage points, and some models in specialized tasks such as code generation have approached closed-source levels.
The Hidden Equation of Interest
On the same call, Krishna mentioned that the z17 mainframe brought in $1 billion in additional revenue, Red Hat grew 11%, and IBM’s Spyre accelerator cards will benefit from the trend of enterprises deploying AI on-premises. The conclusion that “open-source models are already good enough” serves IBM’s hardware, on-premises software, and Red Hat container businesses.
Sovereign AI: A Harder Driver Than Cost
Krishna pointed out that enterprises need to “run in a sovereign way, meaning no one can take away the underlying data or underlying logs.” In Europe, the AI Act is pushing the financial and healthcare industries to scrutinize data flows in API calls; in Asia, government procurement contracts often include clauses stating that “data must not leave the country.” For these buyers, local open-source models first satisfy compliance requirements.
The 10% Behind “90% Effectiveness”
Krishna listed low-barrier tasks including summarizing emails, reviewing contracts, answering customer service calls, and drafting email replies. These scenarios have a relatively high tolerance for precision errors. But when code generation must maintain a 99% test pass rate, or legal compliance review requires fine-grained reasoning on edge cases, the 10% gap can create disproportionate risk. A CFO can verify token costs, but it is difficult to audit whether contract review failures stem from a 5-point model gap.
Talent Gap: The Hidden Costs Being Masked
Krishna said: “The question enterprises face is: Where are my kernel engineers? Who is responsible for the underlying network? Who can ensure that the model correctly aggregates across GPUs adjacent in network topology? Who does cache optimization?” The hardware threshold for two or three racks is not high, but specialized knowledge such as model inference optimization, GPU network latency tuning, and quantization-accuracy trade-offs must be absorbed by enterprises themselves. For technology companies with engineering teams, this is not a problem; for traditional enterprises that rely on IT outsourcing, the saved inference costs may be offset by labor costs.
The Real Impact on the Frontier API Business
Krishna’s argument redraws market segmentation. The default path for enterprise AI procurement shifts from “start with GPT-4 and see how it performs” to “default to open-source on-premises deployment, and pay the price premium only when higher precision is needed.” For API vendors such as OpenAI and Anthropic, this means the enterprise customer base will polarize: high-value scenarios such as high-risk decisions, complex reasoning, and multimodality will retain frontier capabilities, while high-frequency, low-complexity tasks will shift to open-source solutions.
Independent Judgment
The improvement in open-source model capabilities and cost advantages are facts. The proposition holds only if enterprises have the ability to operate and maintain local AI infrastructure and their core tasks fall within the range already covered by open-source models. Two years ago, open-source models hovered around 30 points; now they are approaching 90. Enterprises that choose paid frontier APIs today because of the “final 10% capability gap” may reassess in 18 months.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接