Medium-Term Trajectories: 2028–2035

Learning Objectives

• Describe how AI is expected to compress scientific research timelines in the 2028–2035 window
• Explain what "personal AI with full life context" means and why it raises both opportunity and serious privacy concerns
• Identify the energy and infrastructure challenges that large-scale AI creates and the solutions being developed

The Difference Between Near-Term and Medium-Term

The 2026–2028 near-term window is largely an extrapolation of observable trends. The 2028–2035 medium-term window is different in character — more uncertain, more dependent on research breakthroughs that are plausible but not guaranteed, and more contingent on policy and regulatory decisions still being made.

What makes these medium-term trajectories worth examining is that they are grounded in active, funded research programs — not pure speculation. The investments, the institutional commitments, and the early results that point toward each development described here exist today. The uncertainty is not about whether these directions are being pursued, but about whether they will reach the scale and form described on the timescales discussed.

Scientific Acceleration: Compressing Discovery Timelines

The most consequential medium-term impact of AI may not be in consumer products or enterprise software — it may be in science itself.

Drug discovery is the most advanced example. AlphaFold 2 (2020) solved protein structure prediction — a problem that had occupied structural biologists for fifty years. AlphaFold 3 (2024) extended this to protein-ligand interactions, directly relevant to drug design. In February 2026, Isomorphic Labs (DeepMind's drug discovery spin-off) announced its Drug Design Engine (IsoDDE) — described by scientists as "on the scale of an AlphaFold 4" — which doubles the accuracy of AlphaFold 3 on protein-ligand structure prediction and identifies novel binding pockets from amino acid sequences alone. Isomorphic Labs has partnerships with Eli Lilly, Novartis, and Johnson & Johnson. By 2028–2035, AI systems are expected to:

• Generate and screen novel drug candidate molecules at speeds that were previously impossible
• Predict biological activity, toxicity, and drug interactions before physical synthesis
• Personalize treatment protocols based on individual patient biology
• Dramatically reduce the average 12–15 year timeline from compound identification to approved drug

Materials science is following a similar trajectory. DeepMind's GNoME discovered 2.2 million new crystal structures, expanding the known set of stable inorganic compounds by approximately ten times. AI is accelerating discovery of new battery materials, superconductors, industrial catalysts, and climate-relevant materials.

Climate modeling benefits from AI's ability to run ensemble simulations at a fraction of the cost of traditional numerical weather prediction. AI-enhanced models can downscale global climate projections to the local scales needed for infrastructure planning, agriculture, and disaster preparedness.

Personal AI with Full Life Context

Today's AI assistants are largely stateless: each conversation starts fresh, with generic knowledge and no understanding of who you are or what your actual situation is. The medium-term trajectory — driven by longer context windows, persistent memory systems, and integration with personal data — points toward AI that knows your life.

The capability being built toward: an AI with access to your calendar, email history (with explicit permission), health data, financial situation, ongoing projects, relationships, and long-term goals. Not a generic chatbot — a context-aware advisor that synthesizes your specific circumstances.

Practical examples of what this enables:

• "You have a high-stress week, your sleep data shows four nights under 6 hours, and you have a major presentation Friday — want me to reschedule your optional Tuesday meetings?"
• "This contract has terms that differ from the last three you signed with this vendor — here are the specific changes to review."
• "You mentioned wanting to save for a house by 2028 — you're currently 18% behind your stated target, and here are three adjustments worth considering."

Several companies are building toward this: Apple's on-device AI strategy is explicitly designed for personal context with privacy-preserving local processing. Limitless (formerly Rewind AI) built a wearable AI pendant for continuous conversation capture and recall before being acquired by Meta in December 2025 — likely to power personal AI features in Meta's smart glasses. OpenAI's memory features in ChatGPT are early steps in this direction.

AI-Designed AI: Closing the Feedback Loop

One of the more significant medium-term developments is the emergence of AI systems that accelerate the development of AI itself.

The current pattern: AI tools help researchers write and debug training code, generate synthetic training data, and analyze experimental results faster. Neural Architecture Search (NAS) uses AI to explore the space of possible model architectures, sometimes producing designs human researchers would not have considered.

The medium-term trajectory: AI systems that identify research directions, design and run experiments, interpret results, and generate the next round of experiments — with human researchers directing at higher levels of abstraction rather than executing at the implementation level.

This is not a fully closed loop yet. Most researchers believe genuine autonomous AI research at the frontier requires capabilities that current systems lack. But the trend is clear, and each generation of AI models has been trained with more AI assistance than the previous one.

Why this matters: If AI can meaningfully accelerate its own development, the pace of capability gains could be faster than historical trends would suggest. This is one of the reasons some researchers are concerned about the speed of the transition — not just the destination.

Brain-Computer Interfaces: Early Medical Deployment

Neuralink implanted its first human patient with a BCI (brain-computer interface) in January 2024. By early 2026, approximately 20 patients have received N1 implants across the US, UK, and Canada. The first patient (Noland Arbaugh) achieved over 9 bits per second of cursor control — doubling the previous BCI record — enabling him to browse the internet, play video games, and communicate using only neural signals. An upgraded N1 architecture with 128 thinner threads reduces tissue damage. Neuralink's Blindsight device (visual cortex implant to restore rudimentary vision) received FDA Breakthrough Device Designation in September 2024 and is recruiting patients. The company is scaling toward mass manufacturing with its R1 surgical robot.

The near-to-medium-term BCI roadmap:

• 2024–2027: Rapid clinical expansion — Neuralink targeting hundreds of implants; improving signal stability and surgical automation; multiple companies approaching pivotal FDA trials
• 2028–2035: Broader medical approval; beginning to explore non-medical applications with healthy subjects; potential for richer bidirectional communication

Other companies are advancing rapidly. Synchron uses an endovascular approach (threaded through blood vessels, no open brain surgery) and has implanted approximately 10 patients globally. Its COMMAND trial met its primary safety endpoint with zero device-related serious adverse events, and Synchron is working with the FDA toward a pivotal trial — potentially the first BCI to reach this regulatory stage. One Synchron patient became the first person to control an Apple Vision Pro using a BCI. Precision Neuroscience takes a third approach: a thin-film Layer 7 Cortical Interface with 1,024 electrodes placed on the brain surface without penetrating tissue. It received FDA 510(k) clearance in April 2025 and partnered with Medtronic in January 2026 to co-develop an integrated surgical navigation system.

The Energy Infrastructure Problem

AI data centers are extraordinary energy consumers, and this is becoming a genuine constraint on AI development — not just a sustainability concern.

The scale of the problem:

• Training a large frontier model consumes as much energy as thousands of homes use over a year
• Inference at scale — serving hundreds of millions of queries daily — compounds this enormously
• Global AI data center power demand has reached an estimated 96 GW and is growing rapidly; by 2030, AI is projected to account for a significant percentage of total US electricity consumption
• The Stargate Project alone plans $400 billion+ in AI infrastructure across five US sites, with the Abilene, Texas facility already operational
• Many proposed data center locations cannot connect to existing power grid infrastructure on the timescales required

Solutions being actively developed:

Small Modular Reactors (SMRs) are nuclear power plants scaled from gigawatt-scale traditional plants to 50–300 MW factory-manufactured units deployable at specific sites:

• Microsoft signed an agreement to restart Unit 1 at Three Mile Island specifically for data center power (2024)
• Google signed agreements with SMR developers for future clean power supply
• Amazon acquired a data center campus adjacent to a nuclear plant
• NuScale Power signed a 6 GW deal with TVA, and Oklo signed a 12 GW deal with Switch — representing massive commitments to nuclear-powered AI infrastructure
• The first commercial SMR deployments specifically for AI campuses are expected in the late 2020s to early 2030s

Space-Based Solar Power (SBSP) is further along as a concept than as deployed infrastructure:

• Satellites collect solar energy in orbit (no night, no weather, no atmospheric loss) and beam it to Earth as microwaves
• The European Space Agency and several private ventures have demonstration programs
• Commercial-scale SBSP for data centers is realistically a 2035+ development, but early demonstrations may occur earlier

Orbital AI Compute inverts the space-solar idea: rather than collecting power in orbit and beaming it down, put the data centers themselves in space, where sunlight is near-continuous and waste heat radiates into the cold of space. SpaceX's plan to fly up to one million AI-compute satellites — and Musk's framing of orbital data centers as "the primary means by which AI can be expanded" — is the most aggressive version of this bet (explored in the near-term predictions lesson). Like SBSP, it is unproven at scale, but it reframes AI's energy constraint as a location problem rather than a generation problem.

Direct Air Cooling and Immersion Cooling are already being deployed to manage the thermal output of dense GPU clusters, which generate heat at densities that traditional air conditioning cannot handle.

International Governance: A Contested but Necessary Project

By 2028–2035, the international governance of AI will look substantially different from 2026's patchwork of national frameworks — though the direction is uncertain.

The possible governance futures:

• A UN AI safety framework with meaningful commitments from major AI-producing nations — ambitious but historically precedented for technologies of global concern (nuclear non-proliferation, chemical weapons)
• A fragmented multipolar regime with competing standards from US, EU, China, and others — probably the base case
• A governance failure in which competitive pressure prevents any effective coordination — possible if the US-China competition makes coordination politically impossible

Institutional foundations are being built: the EU AI Act is now in phased enforcement (penalties up to EUR 35 million or 7% of global turnover); the UK renamed its AI Safety Institute to the AI Security Institute in February 2025, narrowing its focus toward national security and misuse risks; and the US, through initiatives like the Stargate Project and NIST AI Agent Standards, is taking a more investment-led approach. The questions are genuinely hard — what counts as dangerous capability, who enforces limits, how to handle cross-border deployment of AI systems — and the institutions capable of answering them are still being built.

Labor Market Structural Shift

The medium-term labor market impact of AI is one of the most contested questions in economics. Credible forecasts range from "AI creates as many jobs as it displaces, just different ones" to "the structural shift is large enough that existing retraining mechanisms are inadequate."

The economic consensus points are:

• Impact will be uneven across occupations and sectors
• Some tasks within most jobs will be automated while others remain human
• New jobs will be created, but the skills required differ from those being displaced
• The transition speed may exceed the adaptation speed in some labor markets

The policy question — whether governments will implement retraining programs, adjusted safety nets, or income support at adequate scale — is genuinely open as of 2026. The UBI (universal basic income) debate is gaining intensity but remains unresolved politically anywhere at scale.

Key Takeaways

• AI is compressing scientific discovery timelines — most tangibly in drug discovery and materials science — with real but uncertain implications for when benefits reach patients and markets
• Personal AI with full life context offers significant value but requires solving privacy architecture problems that do not yet have satisfactory answers
• The energy infrastructure challenge for AI data centers is a genuine physical constraint requiring nuclear SMRs and potentially novel generation sources — not just a sustainability consideration
• Brain-computer interfaces are advancing through the medical pathway toward broader deployment; non-medical mass adoption remains highly speculative for this timeframe
• Medium-term labor market impacts will be significant and structurally uneven; the adequacy of policy responses is a major open question