At Rise of AI 2025 in Berlin, Thomas Schmidt, co-founder of Amber AI, an Nvidia elite partner founded in 2006, argues that the industry must move from AI hype to real, useful impact. He traces the field from AlexNet in 2012 through Nvidia's 2016 shift into AI-focused hardware and the 2022 ChatGPT moment, citing figures that generative AI returns 3.5 US dollars per dollar spent on average, and up to 8 times for the best-performing companies. He argues much enterprise infrastructure is still outdated CPU-based hardware, and that moving to current-generation GPUs, especially Nvidia's Blackwell architecture, cuts power use and cost while shifting workloads from training toward inference. Schmidt presents a human-centric design framework for making AI actually get used inside companies, emphasizing starting from real business pain points rather than technology demos, and describes three concrete use cases: accelerating patent-lawyer document review, optimizing logistics and cargo loading, and using computer vision to identify chick sex in eggs before hatching to stop unnecessary culling. He closes with a call to prioritize trust, impact and workflow integration over model sophistication.
Key takeaways
- Schmidt cites generative AI ROI of 3.5x on average, and up to 8x for leading companies, per dollar spent.
- He argues moving legacy CPU workloads to current-generation GPUs cuts both cost and energy use.
- A patent-review use case cut document analysis time from one to two weeks down to minutes or hours.
- A computer-vision use case determines chick sex from eggs before hatching to eliminate unnecessary culling.
- He warns companies most often fail by over-engineering solutions and running endless pilots instead of fixing one real pain point.
Detailed notes
Amber AI and Nvidia's early GPU-for-AI bet
Schmidt introduces Amber, an Nvidia elite partner founded in 2006, when one of his co-founders, then at the Technical University of Munich, was among the first to code on GPUs and worked directly with Nvidia's Jensen Huang on early projects with BMW. Amber, then called Fluidyna, was acquired by the US company Altair; in 2022 the AI team spun back out as Amber to chase Nvidia and AI as the next big opportunity, while Altair itself was later acquired by Siemens. He notes Amber was Nvidia's 2024 Rising Star partner and won 2025 Nvidia Partner of the Year for Central Europe.
From AlexNet to ChatGPT: the state of AI and its ROI
Schmidt sketches the timeline from AlexNet in 2012, through Nvidia's 2016 decision to build dedicated AI hardware after Dell and HPE had not yet seen the opportunity, to transformers and the 2022 ChatGPT moment that brought AI into mainstream attention. He cites figures that generative AI delivers an average return of 3.5 US dollars per dollar invested, with leading companies reaching up to 8 times return, that an estimated one-third of companies expect to double monetization using generative AI, and that by 2028 much more coding and software testing will be done through generative AI. He stresses generative AI is only part of what AI can do.
Infrastructure economics: from legacy CPUs to Blackwell
He argues many customers, from small firms to large corporations, still run legacy CPU infrastructure that could be GPU-accelerated, and that moving workloads to GPUs would cut cost, energy use and heat generation even as AI overall consumes more energy. He contrasts a decade-old Titan-based system worth roughly 130 million US dollars with a DGX/HGX-class system from two years ago worth about 400,000 US dollars, delivering similar power at a fraction of the cost and energy. He notes most current deployments still use Nvidia's Hopper architecture, but that production is shifting to Blackwell, which he says Jensen Huang himself describes as leaving no real reason to keep buying Hopper, particularly because Blackwell's efficiency gains are larger for inference than for training, and the industry overall is shifting from training toward inference workloads.
A human-centric framework for making AI useful
Schmidt argues that past the peak of AI hype, many pilot projects went nowhere because they were built as tech demos rather than around real customer pain points. He proposes a human-centric design framework: understand the specific problem, engage the humans who will use the system, and iterate based on their feedback rather than aiming for 100% accuracy. As an example, he describes a customer-service chatbot now in its 53rd iteration, continually improved based on user complaints, until it reaches a point where users prefer it to alternatives. He adds that in Europe, AI also needs to be explainable rather than a black box, and should be accessible to every relevant worker rather than a narrow team.
Three use cases: legal, logistics, and hatchery ethics
Schmidt describes working with patent lawyers, where customized models for text and drawings, combined with human validation, cut document-review time on some projects from one to two weeks down to minutes or hours. In logistics, Amber helped a customer optimize the packing of large concrete blocks onto lorries and used computer vision to inspect cargo safety, reducing fuel and vehicle needs. A third example addresses the culling of male chicks in the poultry industry: using deep learning and computer vision to determine chick sex from eggs before hatching, eliminating the need to hatch and then cull males at all.
Q&A highlights
Asked about data storage and networking as a bottleneck, Schmidt says customers often underinvest here relative to compute, comparing it to buying a Ferrari and fitting it with tiny wheels, and names Dell, DDN and NetApp as partners working to keep pace with Nvidia's networking speed. Asked about the biggest pitfall in shifting from 'clever' to genuinely useful AI, he says companies commonly over-engineer solutions instead of simplifying, should prioritize quality data over quantity, and often get stuck running pilot after pilot without management involvement to fix real pain points and move projects into production.
Quotes
AI should have a focus of human-centric design framework.
Thomas Schmidt
You're not buying the Ferrari and then put tiny wheels on it and not get the power on the street.
Thomas Schmidt
A lot of companies over-engineer. Sometimes you don't have to have the perfect, super accurate black box that will fix everything.
Thomas Schmidt
Questions answered in this talk
What return on investment does generative AI deliver?
According to Thomas Schmidt at Rise of AI 2025, generative AI delivers an average return of 3.5 US dollars for every dollar spent, with leading companies achieving up to 8 times return.
Why is data storage and networking often a bottleneck in AI projects?
According to Thomas Schmidt at Rise of AI 2025, customers frequently budget heavily for compute but overlook storage and networking, comparing it to buying a Ferrari and fitting it with tiny wheels, then having to request additional budget for networking and storage later.
What is the biggest pitfall companies face when trying to make AI useful?
According to Thomas Schmidt at Rise of AI 2025, the biggest pitfalls are over-engineering solutions instead of solving one real problem simply, prioritizing data quantity over quality, and running repeated small pilots without management commitment to scale them.