AI companies often fear that Nvidia might swiftly render their efforts obsolete, but Tuhin Srivastava, cofounder of AI inference platform Baseten, remained calm when faced with this challenge. “This is the thing about AI — you gotta burn the boats, ” Srivastava told Business Insider, indicating a commitment to move forward despite uncertainty. While he hasn't fully abandoned his current approach, he has prepared for change. Earlier this year, DeepSeek's breakthrough AI model challenged Srivastava and his team who struggled to scale inference—the computational process generating AI outputs—due to complex bottlenecks slowing down model responses. Baseten’s customers demanded faster access. Although Baseten had Nvidia’s cutting-edge H200 chips capable of running the advanced model, Nvidia’s Triton Inference Server software was glitching under the heavy inference load. To overcome this, Baseten developed its own inference platform, which they continue to use. However, in March, Nvidia unveiled Dynamo, a new open-source inference platform at its GTC conference designed to handle large-scale reasoning models efficiently. Nvidia CEO Jensen Huang described Dynamo as “essentially the operating system of an AI factory. ” Srivastava acknowledged Nvidia’s move was inevitable and expects that once Nvidia's platform surpasses Baseten’s, his team will quickly pivot and adopt it—an adjustment he anticipates will take only a few months. The AI landscape is highly dynamic: models constantly grow more complex, requiring immense computing power and engineering expertise, but later often shrink as efficiencies and mathematical improvements emerge. Developers must balance cost, speed, accuracy, and hardware demands, continually reshuffling strategies.

Karl Mozurkewich, principal architect at cloud firm Valdi, cautions against becoming attached to any single framework or method. Similarly, developer and YouTuber Theo Brown, whose company Ping builds AI tools, values AI’s tendency to make previously “holy” industry practices cheap and disposable. Early in his career, Brown learned to proactively rebuild projects quickly rather than getting stuck in sunk cost fallacy or waiting for permission. This agile, relentless approach is typical among leading AI innovators and distinguishes startups from larger corporations. Quinn Slack, CEO of AI coding platform Sourcegraph, finds that about 80% of Fortune 500 companies can be convinced in a single meeting to abandon fragile early AI systems and adopt better solutions. Meanwhile, Ben Miller, CEO of Fundrise, a real estate investment platform integrating AI, takes a more cautious stance. He focuses on reliable models that meet product needs rather than chasing cutting-edge innovations that may not justify engineering time—especially for larger organizations and ventures built higher up the software stack. The AI stack layers from Nvidia GPUs at the hardware base, through platforms like Baseten, up to AI models such as R1 and GPT-4o, and finally to consumer-facing applications like Miller’s. Mozurkewich notes that advancing “bleeding-edge” features does not guarantee growth in customers or revenue; at the end-user level, rapid change and disruption offer diminishing returns. In sum, AI development demands a willingness to pivot swiftly in response to technological shifts, balanced by pragmatism about timing and scope of adopting new advances. This dynamic environment fosters continual evolution, propelled both by giants like Nvidia and agile startups willing to burn their boats and rebuild.