The trillion-dollar question for AI infrastructure

2025 has been a dizzying year for artificial intelligence. Around the world, technology companies, banks, private equity, sovereign wealth funds and governments are making unprecedented AI infrastructure investments — on the order of trillions of dollars in data centers, land, fiber networks and power expansions. OpenAI alone has outlined plans for well over a trillion dollars in long-term compute and energy capacity.

Beneath all this activity lies a single assumption: that the current AI technical architecture, centered on transformer-based models, such as large language models (LLMs), will keep scaling indefinitely.

But can it? Or are we nearing the edge of what this design can deliver — technically, physically and economically?

At Adeia, we have seen moments like this before. In early 2022, our analysis of conversational AI, published before ChatGPT’s launch, helped analysts and investors anticipate a major inflection point. Today, we are once again asking the hard question that many in the industry are circling around: what comes after the transformer architecture?

The transformer: the architecture that built modern AI

The transformer architecture, introduced in 2017, became the foundation of nearly every large language model in use today. Its encoder-decoder structure converts input sequences into numerical representations and reconstructs outputs through a mechanism called self‑attention.

Self‑attention allows the model to consider relationships among all tokens in an input simultaneously, capturing long-range dependencies that earlier recurrent or convolutional networks could not.

A further innovation, multi-headed attention, lets transformers focus on different parts of a sequence at once, improving learning and generalization. Just as important, the design aligned almost perfectly with GPU architectures, which excel at matrix operations and parallel computation.

Together, these advances made it possible to train enormous models on vast datasets and to discover an empirical rule of thumb that has defined the last decade of AI.

Riding — and bending — the AI scale curve

Over the past half-decade, AI progress has largely followed this scaling law. Models have been trained on nearly all public internet text, fine-tuned on curated data, reinforced through human feedback, and expanded with synthetic datasets.

To extend the life of this paradigm, researchers have added:

• Diffusion models for image and media generation
• Mixture-of-experts (MoE) architectures to route tokens through specialized subnetworks
• State-space models to better handle long sequences
• Chain-of-thought prompting and reasoning techniques to extract more from the same base models

These innovations have squeezed more performance from the same foundation. But the approach is hitting natural limits.

• Hardware performance — measured in throughput per watt and per die area — is no longer keeping pace with the computational demands of frontier models.
• Training state-of-the-art systems now consumes hundreds of megawatts of power and requires vast physical footprints.
• Even with optimized chips, custom interconnects, and aggressive parallelization, cost curves are bending upward faster than efficiency gains, for training the latest and greatest next model.

The transformer, while extraordinarily capable, is nearly a decade old. What propelled us here may not be enough to take us further.

Beyond the text era: why transformers may not be enough

Despite the rapid progress in generative AI, today’s most widely deployed systems remain largely text-bound. In many ways, this resembles the command-line era of computing — powerful, but abstracted from the real world.

The next generation of AI will need to integrate multiple modalities:

• Visual and spatial data
• Physical and sensor data
• Biological and environmental signals

This is the foundation for world models — architectures capable of understanding context, causality, and sensory input together, not just predicting the next token in a sequence.

Here is the core challenge:

• The transformer’s roots in sequence prediction make it inherently optimized for text and token streams.
• Multimodal, causal, and interactive tasks — such as robotics, real-time decision-making, and digital twins — demand architectures that can process time, space, and interaction as fluently as current models process language.

To move beyond the text era, we will need AI architectures that are natively multimodal and causal, not just extended transformers with additional adapters.

The S-curve of AI innovation

Every technology evolves along an S-curve: slow beginnings, exponential acceleration, then a plateau before the next paradigm shift.

AI’s transformer era is approaching that inflection point. Much of today’s investment focuses on scaling and optimizing rather than reinventing. Efficiency tweaks, improved training pipelines, and larger GPU clusters extend performance, but they do not change the underlying physics or economics.

At Adeia, we believe the next phase of AI growth will require a fundamental architectural rethink — a new foundation that:

• Sustains exponential improvement
• Remains compatible with existing software stacks and developer workflows
• Breaks the linear relationship between performance and cost

This shift must go beyond brute-force transistor scaling. It will depend on:

• Hardware-algorithm co-design, not just faster chips
• Innovation at the intersection of compute, memory, and model efficiency
• Hybrid learning methods that change how models represent knowledge and reason, not just how many parameters they have

In search of a new AI foundation

So what could that new foundation look like? Across the research community, several promising directions are emerging for post-transformer AI architectures:

• Neuromorphic and event-driven computing
  
  • Mimic aspects of biological systems
  • Offer orders-of-magnitude energy savings for certain workloads
• Hybrid symbolic–neural architectures
  
  • Integrate logical reasoning with deep learning’s pattern recognition
  • Combine the strengths of symbolic AI and connectionist approaches
• World and physics-based foundation models
  
  • Grounded in sensor data, simulation, and digital twin environments

These are not incremental changes, and this list is not comprehensive. Each represents a potential leap — and the race is on to discover which of these approaches (or which combination) or a totally asymmetric and fundamentally different novel architecture will define the next era of AI.

‍Adeia’s perspective: where scaling laws stall, invention begins

At Adeia, we focus on moments when industries meet architectural walls. Whether in semiconductors, media, connectivity, or AI infrastructure, our work begins where traditional scaling laws stall. We look for the invention that unlocks the next exponential curve.

We believe the future of AI will depend not only on capital and compute, but on creativity in architecture:

• New ways to represent knowledge
• New models that can reason about the physical world, not just text
• New bridges between hardware and software innovation

That is why Adeia continues to engage directly with the global research ecosystem. We recently attended NeurIPS 2025 in early December to exchange ideas, meet fellow inventors and collaborators, and explore what comes after the transformer — in AI models, in compute architectures, and in the broader infrastructure stack that will support them.

The next decade of AI starts with invention

The world has effectively staked trillions of dollars on the assumption that today’s AI architecture will keep scaling. History suggests otherwise. Each generation of technology eventually encounters limits — physical, economic, or conceptual — that demand a new approach.

As AI approaches that boundary, the challenge is clear:

• We must invent beyond the transformer, beyond brute-force scaling, and toward architectures that combine efficiency, multimodality, and real-world reasoning.
• We must align AI models, semiconductor design, memory architectures, and networking in a coherent, co-designed hardware and software stack.

Adeia has always thrived at these turning points. As CTO, I see this moment as an opportunity — not a ceiling. We are ready once again to help reimagine the structures that will power the next decade of intelligent systems and to partner with other industry innovators.

Adeia attended NeurIPS 2025 last week, and we enjoyed seeing old colleagues and meeting new folks in the space.