EXCLUSIVE: AI Data Center Boom Is ‘Overkill,’ Defense AI CEO Warns Compute Demand Will Drop

Artificial intelligence companies may be building far more compute than it ultimately needs, according to Tyler Saltsman, CEO of EdgeRunner AI. The compute being built is “a little overkill,” Saltsman told Benzinga. “I think what will happen is AI will become too expensive, and all of this data center demand will start to drop.” Saltsman expects AI’s future to consist of specialized models working together rather than a single enormous system trying to do everything. ‘The World Doesn’t Need AGI’ “Future intelligence isn’t going to be one big mega-model,” he added. “The world doesn’t need AGI, what it needs is domain-specific GPT-4 level models.” Saltsman, whose company sells specialized on-device AI models, said one EdgeRunner military assistant can run at roughly 4 billion parameters using 8GB to 16GB of video memory on consumer-grade hardware. After training in a data center, EdgeRunner compresses its models so they can run directly on a device without repeatedly sending requests to the cloud. Saltsman said that approach could “completely undercut” the large AI companies’ bottom line by removing recurring per-token charges. He also argued that competition from open-source models helps explain recent calls from Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman to slow frontier AI development. “Open source is catching up, and it threatens their business model,” Saltsman said. Why EdgeRunner Says Smaller Military AI Works Better EdgeRunner builds smaller, military-specific AI models designed to run directly on soldiers’ devices without an internet connection. Saltsman says training the models on military doctrine, terminology and use cases can make them more reliable on battlefield tasks than general-purpose LLMs. He pointed to a recent incident in which a faulty AI-assisted intelligence report wrongly identified a Chinese ship as carrying nuclear-related components, nearly triggering a U.S. military operation. EdgeRunner reduced error rates by 37% in one training run using data from the Army’s Artificial Intelligence Integration Center, Saltsman said. Nvidia Still Wins, but AI Compute Could Shift Saltsman stopped short of predicting weaker demand for Nvidia Corp. (NASDAQ:NVDA), praising its chips and CUDA software platform. Instead, he expects computing power to shift as companies match models more closely to individual tasks. He compared using massive frontier models for military mission planning to firing a $2 million Patriot missile at a $20,000 drone. “If you’re just using these powerful mega-models for mission planning, it can be overkill,” Saltsman said. Prediction-market traders remain bullish on Nvidia. Polymarket gives the chipmaker a 78% chance of ending 2026 as the world’s largest company, compared with Apple (15%), on roughly $7.5 million in trading volume. Benzinga reported Wednesday that McKinsey found the same task can cost up to 30 times more depending on the AI agent used. Read Also: Small AI Model Matches GPT-4.1 at Forecasting as Olas Founding Member Says Frontier Models Face ‘Increasing Economic Pressure’ Image: Shutterstock