Future of AI is being shaped by a decade-long shift toward larger, more powerful models driven primarily by scale. For the past fifteen years, artificial intelligence progress has relied heavily on increasing computational power, with the compute used to train major AI models growing around 4–5× annually since 2010. For frontier language models, training requirements have roughly doubled every six months since 2020. This rapid expansion of scale has been a major factor behind the breakthroughs of deep learning, with around two-thirds of the improvement in language model capabilities over the last decade linked to increases in model size and computing resources.

The next decade will look different. Scaling continues, but three things are changing at once: the frontier is hitting physical limits (power, capital, data); a second scaling axis has opened up in the form of test-time or “reasoning” compute; and AI is shifting from a passive tool that answers prompts to agentic systems that plan, use tools and act. For engineers and researchers, this is the interesting moment: the questions are no longer only “how big?” but “how efficient, how reliable, and how do we build on top of it?”

The stakes are commensurate. Independent forecasts put the global AI market on a path from roughly $240B in 2025 to $0.8–1.8T by 2030 (a ~28–37% CAGR depending on definition), and PwC estimates AI could add as much as $15.7T to world GDP by the end of the decade. Figure 3 later in this briefing shows one mainstream trajectory.

This briefing maps ten technology areas that will define AI to roughly 2035, grouped from the physical substrate up to applications, with a cross-cutting layer for trust and safety. Figure 1 shows how they fit together.

Figure 1 — A layered view of the AI stack. Capability flows upward; physical and economic constraints propagate downward.

The scaling substrate: compute, data, and cost

The economics of the frontier have become staggering. Epoch AI estimates GPT-4 used on the order of 2×10²⁵ FLOP at roughly $80M in hardware-time; frontier 2026 runs sit between 10²⁶ and 10²⁷ FLOP with reported cost ranges of $200–500M, and credible projections put late-2027 runs at $1–3B. Meanwhile the global stock of AI compute is growing about 3.3× per year and the largest single site already rivals a million high-end GPUs.

Figure 2 — Training compute of representative models on a log scale. The slope corresponds to ~5× per year.

Two counter-currents matter for the next decade. First, algorithmic efficiency improves about 3× per year the same capability needs one-third the compute roughly every twelve months which partially offsets rising hardware demand. Second, Epoch and others expect frontier compute growth to decelerate from ~10× toward 3–4× per year through 2028, bounded by power and capital rather than silicon. The research implication is clear: the marginal returns are shifting from raw scale toward data quality, architecture, and efficiency.

From pattern-matching to reasoning: test-time compute

The defining shift since 2024 is reasoning models that spend additional computation at inference time, generating, exploring and verifying intermediate steps rather than answering in a single forward pass. This opens a second scaling axis: for many hard problems, letting a model “think longer” improves accuracy much as training a bigger model would, but at inference. Epoch’s capability index has advanced ~14 points per year since reasoning models appeared.

For engineers, the practical consequences are concrete: latency and cost now trade against accuracy on a per-query basis; verification and search become first-class parts of the system; and evaluation must probe multi-step reliability, not just single-shot correctness.

Open questions

• How do we verify long reasoning chains cheaply and reliably, ideally with formal or learned verifiers?
• Where is the compute-optimal boundary between training-time and test-time investment for a given task?
• Can reasoning learned in one domain (maths, code) transfer to open-ended scientific or engineering problems?

Figure 3 — One mainstream projection of the global AI market. Analyst estimates vary widely, but all agree on the direction and magnitude of growth.

Agentic AI: from tools to autonomous systems

An agent wraps a model in a loop: perceive, plan, call tools or APIs, observe results, and iterate toward a goal with memory across steps. This is the fastest-moving application layer in industry. The engineering challenges are those of distributed systems as much as of ML: orchestration, state and memory, error recovery, and safe action boundaries.

A useful mental model is a ladder of autonomy. Most deployed systems today sit at levels 1–2; the research and reliability frontier is the climb to levels 3–4.

Multimodal systems, world models, and embodiment

Frontier models increasingly ingest and produce text, images, audio and video jointly. The deeper research direction is world models, systems that learn a predictive, physical model of an environment and can simulate consequences of actions. World models are the connective tissue between today’s language-centric AI and embodied / physical AI: robots and autonomous systems that need to perceive, predict and act in the messy physical world.

For mechanical, electrical, control and robotics engineers this is where AI becomes tangible: learned controllers, sim-to-real transfer, differentiable simulation, and perception stacks that fuse cameras, LiDAR and event sensors. The decade’s prize is data efficiency, learning robust physical behaviour without millions of costly real-world trials.

AI for science: the research accelerator

For graduate researchers, this is the most consequential area. AI has moved from a data-analysis convenience to a genuine engine of discovery. AlphaFold predicted the structures of nearly all known proteins and has been used by over three million researchers; its creators shared the 2024 Nobel Prize in Chemistry, the same year the Physics prize recognised the foundations of neural networks. In materials science, DeepMind’s GNoME proposed ~2.2 million candidate crystal structures, of which tens of thousands are predicted stable and hundreds have already been synthesised experimentally.

The frontier now is agentic, autonomous science: systems that read the literature, propose hypotheses, design experiments, and coupled to robotic “self-driving laboratories” run and interpret them in a closed loop. Early platforms integrate dozens of scientific databases and tools and are being piloted with national labs and pharmaceutical partners. The likely near-term reality for most researchers is a collaborator, not a replacement: AI that compresses months of literature review, code and analysis into days.

Efficient AI: small models, edge, and algorithmic gains

Not every advance is bigger. A parallel track is making models smaller, cheaper and local. Techniques mixture-of-experts (activating only part of a network), quantisation (lower-precision weights), distillation (training compact students from large teachers), pruning and speculative decoding combine with the ~3×/year algorithmic-efficiency trend to push capable models onto phones, laptops, vehicles and sensors.

Edge and on-device AI matters for latency, privacy, cost and resilience, and it is where much engineering employment will sit. The research questions are the compression frontier (how small can a model be for a given capability?), hardware-aware training, and continual/on-device learning without catastrophic forgetting.

Beyond the GPU: next-generation AI hardware

As Moore’s law slows and Dennard scaling has ended, the energy cost of moving data through the conventional von-Neumann architecture is the dominant bottleneck. Three post-GPU paradigms are maturing in the lab and worth tracking:

Neuromorphic computing — brain-inspired, event-driven “spiking” chips (e.g. the TrueNorth and Tianjic lineages) that process information sparsely and only when signals change, cutting energy for the right workloads.

Photonic computing — performing matrix multiplication with light for ultra-wide bandwidth and near-zero-latency, low-energy linear algebra. Integrated photonic accelerators have demonstrated on the order of 160 TOPS/W and hundreds of TOPS on-chip in recent research.

Analog in-memory & quantum ML — computing inside the memory array to avoid data movement, and — further out — quantum machine learning for specific structured problems.

Takeaway for hardware and EE students: the next decade rewards co-design, algorithms shaped to the physics of the accelerator, and accelerators shaped to the workload. The interface between ML and device physics is unusually open right now.

The energy and infrastructure wall

Capability is increasingly bounded by electricity. The IEA projects global data-centre electricity demand roughly doubling from about 415 TWh in 2024 to ~945 TWh by 2030, close to Japan’s entire consumption today and rising toward ~1,200 TWh by 2035, with AI the primary driver. In the United States, data centres are projected to consume more electricity by 2030 than the production of aluminium, steel, cement and all other energy-intensive goods combined.

Figure 4 — IEA Base Case for global data-centre electricity demand. AI-optimised demand more than quadruples to 2030.

This makes energy efficiency an AI research problem, not just an operations one. Expect intense work on efficient architectures and inference, waste-heat reuse, carbon- and latency-aware scheduling across data centres, and the co-location of compute with new generation (renewables, nuclear, geothermal). Gigawatt-scale facilities that once took a decade are now built in about two years, so the grid not the GPU may set the pace of progress.

Trustworthy AI: interpretability, alignment, safety, governance

As systems gain autonomy and reach, the cross-cutting layer of Figure 1 becomes decisive. Four strands will shape the decade:

Mechanistic interpretability — reverse-engineering the internal computations of networks so behaviour can be understood, not just observed. A deep, tractable open problem well suited to graduate research.

Alignment & control — ensuring systems pursue intended goals and remain correctable as they grow more capable and autonomous.

Evaluation & robustness — rigorous benchmarks for reasoning, agents and safety, plus resistance to distribution shift and adversarial input.

Governance & policy — compute-based regulation, disclosure, and standards; the technical and policy communities increasingly overlap.

Ten open problems worth your PhD

If you are choosing a direction, these are durable, high-impact problems where progress is measurable and the field is not yet saturated:

1. Sample- and data-efficiency: learning far more from far less data.

2. Reliable long-horizon reasoning with cheap, trustworthy verification.

3. Continual and on-device learning without catastrophic forgetting.

4. Compression frontiers: the smallest model for a target capability.

5. Agent reliability and safe action boundaries in open environments.

6. World models and data-efficient sim-to-real for robotics and control.

7. Energy-efficient architectures and hardware–algorithm co-design.

8. Mechanistic interpretability at the scale of frontier models.

9. Closed-loop AI-for-science: autonomous hypothesis → experiment loops.

10. A predictive theory of deep learning: why scaling laws hold, and where they break.

Outlook

The last decade of AI was a story of scale; the next will be a story of scale meeting limits and what we build around them. Compute will keep growing but more slowly and at greater cost; reasoning and agents will extend capability along new axes; efficiency, new hardware and energy will decide who can actually deploy at scale; AI-for-science will compress discovery; and trustworthiness will determine how far autonomy is allowed to go.

For students and professionals planning their next career move, building practical knowledge in areas such as machine learning, automation, data science, and AI development can create a strong foundation for adapting to the changing technology landscape. Resources covering the top AI skills everyone should learn for future career growth can help identify the skills that are becoming increasingly valuable across industries.

For engineering and graduate students, the message is optimistic and concrete: the interesting problems have moved from “train a bigger model” to efficiency, reliability, physical grounding, and scientific application, exactly the places where rigorous engineering and research talent compounds. The field is large enough that a well-chosen thesis today can still shape where it goes.