PARIS — Artificial intelligence is still in the early stages of development, with major advances ahead in models, chips, energy efficiency and quantum computing, according to Amazon senior executive Peter DeSantis.
DeSantis, a senior vice president who leads AI models, custom silicon and quantum computing at Amazon, said the next phase of AI will depend not only on more powerful models but also on cheaper, faster and more energy-efficient infrastructure to run them.
“I think we are just at the beginning of the innovation at all layers of the stack,” DeSantis said during a fireside chat at the VivaTech conference in Paris, which sponsored this journalist’s trip.
Need more clues? Ask the Sherlock chatbot in the lower right corner to summarize this story, explain technical concepts or answer other questions.
The larger question, he said, is how quickly AI improves from here. He said future gains will come from several directions at once: more capable models, more useful agents and better infrastructure.
He said the industry has already seen “an order of magnitude improvement” in the efficiency of tasks models can perform, but more progress is needed before AI can become broadly useful.
“We need a couple more orders of magnitude before this gets interesting in any way,” DeSantis said.
At the model level, DeSantis said transformers remain a durable architecture, but they are not the whole story. He said new architectures will be needed as AI moves beyond text-based chat into more natural forms of interaction involving speech, vision, gesture and ambient context.
“We’re all very comfortable chatting with models with text,” he said. “But as humans we have many ways of communicating.”
That shift will require systems that can respond far faster than today’s models, he said. Human conversation involves interruptions, gestures, nods and visual cues that current AI systems only partially understand.
As AI becomes more multimodal, users may eventually interact with devices that continuously observe their surroundings, listen for context and understand human behavior in a more natural way, he said.
Custom chips become increasingly important
As AI becomes more multimodal, DeSantis said, the hardware underneath it will also have to evolve. Chip design and model development are deeply connected because both require long planning cycles and substantial capital investments.
“If the chips are not telling the model designers what capabilities are coming and where they can optimize the models to take advantage of those capabilities, then we’re not doing the science that’s necessary,” he said.
The same is true in reverse, he added. Model developers need to tell chip designers what capabilities future AI systems will require.
DeSantis said Amazon has worked closely with partners such as Anthropic on that kind of coordination. Amazon has invested heavily in Anthropic, while AWS has developed its own Trainium family of AI chips.
He said forecasting demand for AI infrastructure is unusually difficult because adoption is growing so rapidly.
“There’s no magic bullet for predicting what demand is going to look like six months from now or two years from now,” DeSantis said.
Specialized hardware could cut costs and power use
One major question for chip designers is whether to build general-purpose AI platforms or specialized chips for specific workloads. General-purpose systems are flexible but can waste memory, bandwidth or compute depending on the task. Specialized chips can deliver major efficiency gains, but they also increase forecasting and deployment complexity.
“If you do it, it’s very tempting, because maybe you can save 40% of the power and 40% of the cost on the platform,” DeSantis said.
He said the industry is moving toward more specialization as AI workloads become better understood.
“I think you’re going to certainly see more specialization from us, and I think across the industry, you’re going to see more specialization.”
Much of that innovation will focus on memory, he said. AI chips spend enormous amounts of time moving data between memory and compute resources. New packaging approaches that bring memory physically closer to processors could significantly improve performance and efficiency.
Power consumption is another major challenge. DeSantis argued that the industry should focus less on tokens generated and more on the amount of useful intelligence produced per watt of energy consumed.
“We’re so early in this optimization cycle,” he said.
He predicted several additional orders of magnitude of efficiency improvements over the next few years through advances in hardware, models and software.
AI helps accelerate chip development
AI itself could become an important tool for designing better chips.
Chip engineers already rely heavily on simulations to test ideas before manufacturing expensive hardware. By making those simulations faster and cheaper, AI could allow developers to explore a much larger number of design options.
“If your models make that process of exploring the solution space and building those simulators an order of magnitude cheaper, you can explore a lot more solution space, and you’re going to find better chips,” DeSantis said.
That creates what he described as a flywheel effect. Better chips improve AI models, and better AI models help engineers create even better chips.
“Just like we’re seeing software development speed up because of AI, we’re going to see model development and chip development speed up as well,” he said.
Despite the growing role of AI, DeSantis emphasized that Amazon still relies heavily on human expertise.
“There’s nothing we’re doing that doesn’t have extremely talented humans in the loop,” he said.
Quantum computing’s promise lies in chemistry and materials
As for quantum computing, DeSantis said it is unlikely to make everyday software applications run dramatically faster. Instead, he believes quantum computers will be most valuable for simulating the physical world.
One promising area is chemistry and materials science, where quantum computers could model molecular interactions that remain too complex for classical computers.
DeSantis pointed to nitrogen fixation, the process used to create fertilizer, as one example. Current methods require large amounts of energy and generate significant carbon emissions.
“If we can find better chemical processes for fixing nitrogen, we can produce more fertilizer with less CO2,” he said.
The biggest challenge remains keeping quantum bits, or qubits, stable long enough to perform useful calculations. Quantum systems are highly sensitive to noise and must often operate at temperatures close to absolute zero.
Amazon’s strategy has focused first on quantum error correction before attempting to scale up the number of qubits in its systems.
For now, DeSantis said the industry remains at the beginning of several overlapping revolutions in AI, custom silicon and quantum computing. The technologies are advancing together, and success in one area increasingly depends on breakthroughs in the others.
The common thread is infrastructure. While AI may be defined publicly by chatbots and agents, DeSantis argued that the next wave of progress will come from the less visible systems powering them behind the scenes.
