IBM's Krishna Downplays AI Threat as AI Chip Race Heats Up

IBM Moves to Calm Investor Nerves

IBM chief executive Arvind Krishna has moved to reassure investors that artificial intelligence will not undercut the company's core software business, arguing that demand for the company's mainframe hardware is actually accelerating and that any lag in software growth should even out within roughly a year 1. The comments come amid broader anxiety on Wall Street that generative AI tools could erode demand for traditional enterprise software by automating tasks that once required licensed applications or by compressing the services layer that has long supported IBM's margins 1. Krishna's framing suggests IBM sees AI less as a threat to its legacy business and more as a catalyst pushing customers toward new mainframe capacity, buying the company room to argue that its hybrid cloud and software strategy remains intact even as the industry's center of gravity shifts toward AI infrastructure 1.

The Infrastructure Race Behind the Reassurance

Krishna's comments arrive against the backdrop of a fast-moving contest to build the chips and systems that power AI workloads, a race in which IBM is not a headline player but whose outcome will shape the competitive pressures on every enterprise technology vendor. Advanced Micro Devices has been positioning itself as the most credible challenger to Nvidia's dominance in AI accelerators, unveiling a new generation of AI infrastructure meant to compete directly with Nvidia's chips and systems at a San Francisco event 2. That push has been reinforced by a partnership between AMD and Cerebras, which are combining AMD's EPYC processors and Helios rack-scale architecture with Cerebras's wafer-scale chip technology to target low-latency, high-throughput AI inference — the stage of AI deployment where models actually respond to user queries rather than being trained 4. Inference workloads have become an increasingly important battleground because they run continuously in production and drive the ongoing hardware costs that companies incur once an AI model is deployed at scale, in contrast to the one-time capital costs of training.

Chipmakers Widen the Field

The competitive landscape extends beyond AMD and Nvidia. Google and MediaTek have partnered on a next-generation chip, internally referred to as "Triggerfish" and tied to Google's TPUv9 line, designed to handle both training and inference for what the companies describe as agentic AI systems — models capable of taking autonomous, multi-step actions rather than simply answering prompts 5. That effort signals that major cloud providers are continuing to invest in custom silicon rather than relying solely on merchant chips from Nvidia or AMD, a strategy aimed at controlling costs and performance for their own inference-heavy workloads as agentic applications proliferate 5. Taken together, the AMD-Cerebras collaboration and the Google-MediaTek chip effort illustrate how the industry is fragmenting into multiple approaches to the same underlying problem: delivering AI inference at a cost and speed that makes large-scale deployment economically viable 45.

Hardware's Growing Reach

The push into new silicon is matched by a broader movement of AI out of research labs and into physical, tangible systems. At the VivaTech conference in Paris, demonstrations included robots controlled through direct human-mind interfaces, an example of how AI is increasingly being embedded into hardware that interacts with the physical world rather than remaining confined to software interfaces and cloud services 3. While this kind of brain-computer robotics work sits at a different point on the spectrum than enterprise mainframes or data-center accelerators, it underscores the same overarching theme running through this period of AI development: computing power and specialized hardware are becoming the central currency of progress, whether the goal is running large language models in a data center or letting a person move a robotic limb with their thoughts 3.

Why It Matters

IBM's attempt to reassure investors reflects the tension facing established technology companies that built their businesses on software licensing and services now that hardware-driven AI capability is advancing so quickly 1. Meanwhile, the intensifying competition among AMD, Nvidia, Google, MediaTek, and Cerebras over inference-optimized chips suggests that the cost of running AI models — not just training them — is becoming the primary battleground for the industry 245. As inference costs and hardware efficiency increasingly determine which companies can profitably scale AI products, the pressure on legacy software vendors like IBM to prove their platforms remain relevant is likely to persist well beyond Krishna's one-year timeline 1.