"It felt like it completely lost personality," remarked Hiten Shah on a recent Forward Future Live session, dissecting the perplexing user reactions to OpenAI's GPT-5 launch. This sentiment, echoed widely across tech circles, highlights an unexpected dimension in AI development: the deeply human attachment users form to AI models, an attachment seemingly disrupted by performance-driven upgrades. Beyond raw benchmarks, the model's subjective "feel" proved a critical, yet perhaps unquantifiable, success metric.
Matthew Berman and co-host Hiten Shah recently convened Jordan Nanos of SemiAnalysis on 'Forward Future Live' to dissect the tumultuous week in artificial intelligence, focusing on the mixed reception of GPT-5, geopolitical shifts in the semiconductor industry, and the evolving landscape of AI compute. The swift user backlash over GPT-5's perceived shift in character, despite its domination of traditional benchmarks, underscores a growing disconnect between technical advancement and user experience. This rapid evolution also saturates traditional performance metrics, rendering simple benchmark comparisons less meaningful as models constantly adapt and improve.
The conversation then pivoted to the high-stakes game of semiconductor geopolitics, ignited by the Trump administration's unusual deal with Nvidia and AMD. These chipmakers pledged 15% of their advanced chip revenue from sales in China to the U.S. government for export licenses. Jordan Nanos characterized this as "classic deal-making," forcing US companies to choose between a substantial market share and compliance, while simultaneously prompting China's government to instruct its tech companies to cease purchasing US chips, citing security concerns. This move by China is less about genuine security and more about accelerating their domestic chip architecture and manufacturing capabilities.
The battle for AI compute infrastructure is intensifying, extending beyond traditional hyperscalers like Google Cloud and Azure. A new breed of "NeoClouds," including CoreWeave, Crusoe, Nebius, and Lambda, are emerging, often converting former crypto-mining facilities into high-density GPU data centers. These players offer aggressive pricing, with some quoting H100 access at under two dollars per hour, and a more hands-on, user-centric support model compared to the larger cloud providers. The sheer scale of investment in this "build-out" is staggering, measured in gigawatts and billions of dollars.
The race for custom silicon is also heating up, with companies like AWS (Trainium), Google (TPUs), Meta, XAI, Cerebras, and SambaNova all developing proprietary chips. While raw performance is a driving factor, the ability to provide seamless integration, robust software stacks, and responsive user support for these specialized architectures is becoming a critical differentiator. This complex, multi-faceted landscape is still in its nascent stages, with the ultimate winners yet to be determined by both technological prowess and market adaptability.
