AI in 10 Years: Three Plausible Scenarios With Named Drivers

Predicting AI in 2036 is a fool's game. Three scenarios are not.

In May 2016, exactly ten years ago, the consensus view held that human-level AI was at least fifty years away. AlphaGo had just beaten Lee Sedol, and most economists treated the result as a chess-club curiosity rather than a leading indicator.

That track record should make anyone humble about predicting 2036. Humility doesn't mean silence — it means trading point forecasts for bounded scenarios. This guide gives you three: each internally consistent, each with named drivers, each falsifiable. The goal is not to be right. The goal is to give you a frame so that when 2027 surprises you, you already know which world you've ended up in.

Scenario 1 — The Plateau

Scaling laws — the empirical finding (Kaplan et al., 2020) that the performance of an LLM (Large Language Model, the engine behind GPT, Claude, and Gemini) improves predictably with more compute, more data, and more parameters — quietly stop delivering. GPT-6 lands in 2027 and is, somehow, just GPT-5 with better manners. The benchmark curves bend. The next trillion dollars of capex earns single-digit returns.

The drivers: data exhaustion (Villalobos and co-authors estimated the high-quality public text corpus runs out somewhere between 2026 and 2032), diminishing returns from synthetic data, and the simple fact that most economically valuable cognition isn't bottlenecked on raw model intelligence. It's bottlenecked on data plumbing, integration, and trust — none of which scale with parameter count.

In this world, AI becomes commodity infrastructure. Like cloud computing after 2014, the layer below stops being interesting and the layer above — vertical SaaS, agents wired into specific workflows, domain models — eats the value. Margins on raw inference (the act of running a model to produce an answer) compress to single digits. NVIDIA still sells chips, but the multiple compresses toward Cisco-after-2002 territory.

Who wins: the hyperscalers (AWS, Azure, Google Cloud) selling commoditized inference; vertical application companies that own distribution and proprietary data; the open-weights ecosystem (Llama, Mistral, DeepSeek lineage) that absorbs frontier capability within twelve months of any closed release. Who loses: pure-play frontier labs whose entire valuation rests on a capability gap that no longer exists.

What most people miss about the plateau is that it would still be the largest technology rollout of the decade. Boring is not small.

Scenario 2 — Continued Takeoff

The other end. Capabilities keep doubling on the metrics that matter — task length, reasoning depth, tool-use reliability — every six to twelve months. METR's 2025 study, which found the length of tasks frontier agents (autonomous AI systems that chain multiple steps to complete a job) can complete reliably was doubling roughly every seven months, turns out to have been the leading indicator. By 2030, agents handle work that takes a human a week. By 2032, a month.

The drivers stack. Post-training — the reasoning, alignment and tool-use work applied on top of a model's initial pretraining — keeps extracting more from existing base models (the o1, o3 and Claude reasoning families in 2024–2025 were the proof of concept). Inference-time compute, the trick of giving a model more "thinking time" before it answers, buys measurable IQ. And the capex cycle continues: Microsoft, Google, Meta and Amazon collectively committed something north of $200 billion to AI infrastructure in 2025, and announced 2026 numbers are higher. OpenAI's Stargate project, unveiled in January 2025, alone targets $500 billion over four years.

If even half of that compute lands and translates into capability gains at the same rate as the past three years, AGI-ish capability — Artificial General Intelligence, the threshold where AI matches or exceeds human cognitive performance across most domains — is plausible by 2032. Not necessarily superintelligence. But enough to break the labor-cost assumptions of every white-collar industry.

Who wins: the three or four labs at the frontier (OpenAI, Anthropic, Google DeepMind, plus probably one Chinese lab), the chip stack (NVIDIA, TSMC, ASML), and whoever controls power generation near the data centers. Who loses: any business whose moat is white-collar labor arbitrage, and most tier-2 AI companies that priced themselves as if they would be at the frontier.

What most people miss about this scenario is that the social and political shock matters more than the technical milestones. A world where AI can do a junior lawyer's job in 2030 is not a world where the legal industry adapts smoothly. It is a world where something breaks.

Scenario 3 — Regulatory Bifurcation

The most likely scenario, because it is already happening. The October 2022 US export controls on H100-class chips were the starting gun. Expanded in October 2023 and again in December 2024, they have split the global AI supply chain in two. China responded with DeepSeek-V3 (December 2024) and DeepSeek-R1 (January 2025) — frontier-competitive models trained on domestically available compute, proving the constraint is harder to enforce than Washington hoped, but not by enough to undo the split.

Project that forward ten years. Capabilities converge — both stacks ship roughly equivalent models, give or take six months — but the ecosystems diverge. Different chips (NVIDIA Blackwell versus Huawei Ascend), different cloud platforms, different app stores, different safety regimes, different training-data languages, different alignment philosophies. The EU sits in between, mostly consuming the Western stack but constrained by the AI Act (in force since August 2024) in ways neither US nor Chinese providers are.

This is not a Cold War analogy in the sense of two airtight blocs. It is more like the mobile OS world: iOS and Android both work, both have apps, both move forward, but no developer of consequence ignores either. AI vendors learn to ship two SKUs. Multinationals keep two model providers under contract. Governments learn that "AI sovereignty" means owning a stack, not building one from scratch.

Who wins: dual-stack vendors (consulting firms, cloud aggregators, NVIDIA for as long as it can sell into both sides), defense and intelligence contractors on both sides, and nation-states that pick the right side early. Who loses: the open-internet model of one global tech stack, and any company that built its product on the assumption that compliance is a checkbox rather than a structural constraint.

The thing to notice is that bifurcation is compatible with either of the first two scenarios. You can have plateau-and-bifurcation, or takeoff-and-bifurcation. The split is structural, independent of capability trajectory.

What to watch in 2026 and 2027

Five signals tell you which scenario is winning. None of them are model-launch press releases.

Frontier benchmark deltas. When GPT-6, Claude 5 and Gemini 3 ship — likely 2026 — do they move the needle on long-horizon reasoning and agentic task completion, or are the gains concentrated in style, multimodality and cost-per-token? Plateau looks like the latter. Takeoff looks like the former.

Capex curves. Combined AI capex of the four US hyperscalers landed around $200 billion in 2025. If 2026 prints at $300 billion and 2027 at $400 billion, the takeoff thesis is being voted on with real money. If capex flattens, the market is calling plateau before the benchmarks do.

Power and grid. US data-center power demand is on track to roughly double by 2028, by some estimates (Lawrence Berkeley Lab, late 2024 report). If grid-interconnect queues, gas-turbine lead times and local opposition cap that — and they might — the plateau is partly forced by physics, not algorithms.

Chinese frontier parity. Track the gap between the best Chinese open-weights model and the best Western closed model. In early 2026 it sits at roughly six months. If it closes to under three, bifurcation accelerates and Western export controls are formally a failure. If the West reopens a meaningful gap, bifurcation softens.

Regulation that bites. The EU AI Act's main provisions come online in August 2026. Watch the first enforcement actions, and watch whether the US passes federal AI legislation or leaves it to a California-led state patchwork. Watch whether China's interim generative-AI measures (in force since August 2023) start being applied to foreign models accessed via VPN.

If only one of those signals matters most, it is capex. Money is the most honest forecast anyone makes.

What this means for you

If you are a builder: pick an architecture that survives all three scenarios. Build on commodity APIs, hide the model provider behind a thin abstraction, and assume both that capability will keep improving (so don't over-engineer fallbacks for current limitations) and that regulatory constraints will tighten (so don't ship anything you can't audit). The companies that died in the dot-com bust mostly bet on a single trajectory.

If you are an investor: NVIDIA wins in scenarios 1 and 2 but on very different multiples — Cisco-style derating is the plateau risk. The frontier labs are a takeoff bet, priced as such. Vertical AI applications are the plateau bet. Defense and dual-stack consulting are the bifurcation bet. The portfolio that survives is the one that doesn't require any single scenario to be right.

If you are a citizen, employee, or human: the bifurcation scenario probably affects your daily life more than the capability ceiling does. Your data is governed by which stack your employer or country uses. The interesting question is not whether AI gets smarter — it almost certainly does — but who you trust to govern it, and whether you have any say in the answer. Most people will not. That is the politics of the next decade.

Ask Gradient

Three scenarios is a starting point, not a forecast. To pressure-test them against current data, try Ask Gradient with something like "What did METR's 2025 task-length study actually measure, and what's the latest update?" or "How much did the four largest US hyperscalers spend on AI capex in 2025, and what are the announced 2026 numbers?" The next surprise will come from the data, not the takes. For the groundwork, see our companion guides on what an LLM actually is and why scaling laws matter.