At Rise of AI 2025 in Berlin, Gilles Backhus, Head of AI at Recogni, argues that AI adoption is approaching a tipping point defined by cost and speed rather than raw capability. Recogni builds AI chips and data-center systems, which gives the company a bottom-up view of what AI actually costs. Backhus explains that global AI compute is shifting from training to inference, projected to reach 95-99% inference by 2030, and that inference costs for advanced multi-model agent stacks have risen roughly 100x per million tokens because accuracy compounds negatively across chained models. He walks through enterprise dilemmas around where to integrate AI, what it costs, and how to manage internal and external change, citing Shopify and Klarna's hiring reversals as examples of unstable dynamics. He introduces real-time GenAI, especially real-time video generation, as the next unlock, using a hypothetical $18 trillion AI-education market to illustrate current costs of roughly $100 per hour of AI video. He closes by describing Recogni's approach to cutting cost 10-100x through logarithmic math, faster chip-to-chip networking, and purpose-built systems.
Key takeaways
- Global AI compute is projected to shift to 95-99% inference over training by 2030 as models grow too large for smaller companies to train.
- Chaining multiple 98%-accurate models into an agent stack can drop combined accuracy to about 90%, driving up model prices roughly 100x.
- Shopify and Klarna both told staff to stop hiring and use AI, but Klarna partly reversed course and rehired human service agents because top AI remains too expensive.
- An hour of real-time AI video currently costs about $100 because a $5 million Nvidia NVL72 rack supports only a few simultaneous video streams.
- Recogni's logarithmic math approach replaces costly multiplications with additions, cutting AI compute cost by roughly 4x at the transistor level.
Detailed notes
Recogni's bottom-up view of AI cost
Backhus introduces himself as co-founder at Recogni, an AI chip and systems company that he describes as, in a broader sense, an AI infrastructure company. Because its data-center rack systems cost millions and take years to build, Recogni has to understand both the true cost of AI and where the technology is heading. He notes the company has innovations such as logarithmic math and a hyperfast chip-to-chip network that make its systems more efficient than Nvidia's, without going into technical detail.
From training to inference, and the accuracy problem of agent stacks
Backhus recalls the early excitement with simple LLM tasks giving way to the realization that much still does not work. He says that in 2023 global AI compute was dominated by training, but this has flipped and will keep flipping: by 2030 he expects 95%, possibly 99%, of global AI compute to be inference, since models are growing too large for smaller companies to train. Everyone wants the best models with mixture of experts, reasoning and multimodality, and compute cost per task is exploding. He gives an example: a text-to-text model that is 98% accurate in isolation drops to about 90% accuracy when five such models act together as an agent stack, which is why some models now cost $50 per million tokens instead of 50 cents, a roughly 100x price increase.
Enterprise adoption dilemmas and hiring dynamics
Backhus describes the questions enterprises face: where to integrate AI, what it will cost, and how to manage change internally and externally. He says internal developers often reject a hardware vendor, including AMD, because they do not want to learn a new programming language, which is part of why Nvidia has such a large market share. He cites Shopify's CEO telling staff to stop hiring and use AI instead, reducing hiring of junior computer science graduates in favor of senior developers using AI, and the Klarna CEO's similar call that was later partly reversed by rehiring human customer service agents because top AI is still too expensive to use everywhere. He warns that companies who delay AI adoption risk falling behind competitors who have already raised their margins through it, and suggests looking for new products rather than only AI-ifying an existing business.
Real-time GenAI and the cost of video
Backhus introduces the term "real-time GenAI," describing how tasks such as five-second video clips or complex reasoning by ChatGPT o3 currently take minutes rather than seconds, and how Recogni is building systems to bring this down to seconds. He argues that once video becomes real time, interactions such as an AI twin joining a video call become possible, and describes video-to-video as the last modality needed for AI to understand a highly visual world. As an illustration, he estimates a market of $18 trillion if every one of 1.8 billion students received two hours per week of AI video-based teaching at $100 per hour of AI video, calling the number completely unrealistic but useful to show how much cost still needs to fall. He also cites AI companionship and therapy as already the largest GenAI use case.
Why an hour of AI video costs $100, and how to cut the cost
Backhus breaks down the economics: an Nvidia NVL72 rack system costs roughly $5 million, runs at about 100 kilowatts, and can support only a handful of state-of-the-art video streams at once. Over a five-year depreciation period this comes to about $1.1 million per year, or roughly $125 per hour at full utilization, close to the $100 per hour price companies charge with margin. He identifies three levers to reduce cost 10-100x: keeping high numerical accuracy (16-bit rather than 8- or 4-bit) rather than shrinking models, increasing speed since usability often follows a step function once an application becomes fast enough, and cutting raw hardware cost. Recogni's approach includes logarithmic math, which replaces expensive multiplications with additions and cuts cost roughly 4x, a partnership with Juniper Networks on ultra-fast chip-to-chip networking to parallelize across up to 144 chips, and purpose-built systems stripped of unneeded generality.
Q&A highlights
Asked which industry will transform most in the next two to four years, Backhus says therapy and companionship, partly because errors matter less there and it is already, per some studies, the largest GenAI use case. Asked about the state of chip development in Germany, he says Germany's great silicon days are largely over but notes some startups working on analog AI chips, and that Recogni itself is split 50/50 between Silicon Valley and Munich; developing its first chip cost $25 million, which he calls cheap by industry standards. Asked for advice for young people, he says being a specialist in one thing will matter less than becoming an excellent user of AI tools. On education, prompted by an audience member from Exponential University, he agrees it is a major opportunity, particularly for bringing translated, localized instruction to rural areas, though he expects parental concern to fade over roughly a decade. Asked where Recogni will improve most in the next three to five years, he says software, despite being a hardware company, because usability is why Nvidia dominates over technically comparable but less usable AMD chips.
Quotes
It's a giant mathematical formula with hundreds of billions of variables. That's really what an AI model is.
Gilles Backhus
You'll be looking at this paradox that if things become cheaper actually it'll be used more and not less.
Gilles Backhus
The great silicon days of Germany are a little bit over.
Gilles Backhus
Questions answered in this talk
Why is AI compute shifting from training to inference?
According to Gilles Backhus at Rise of AI 2025, models are growing too large for most companies to train themselves, so by 2030 he expects 95%, possibly 99%, of global AI compute to be inference rather than training.
Why do multi-model AI agent stacks become less accurate?
Backhus explains that if five text-to-text models, each 98% accurate alone, act together in an agent stack, the combined chance all of them behave correctly drops to about 90%, which is why capable models now cost far more per token.
What is real-time GenAI and why does it matter?
Backhus defines real-time GenAI as generating outputs like video or complex reasoning in seconds instead of minutes, which he says is a step-function unlock: once an application becomes fast enough, such as real-time video-to-video interaction, it suddenly becomes usable in ways it was not before.