Prediction of AI in 2040

To start with AI needs to simply take over the task of universal drivging of cars.  We might be starting.  It must surely take a whole decade.

It must then also operate humanoid robots in ordinary tasks like delivering a package and apply the huan eyeball equivalent for confirmation and recognition.

That is what we can see.  How about grooming the global boreal forest in order to enhance growth and useful biodiversity?  Now apply globally.  Every square mile will need a thousand humans and several thousand devices all suported by AI.

You  may see my point.

Prediction of AI in 2040

September 10, 2025 by Brian Wang

https://www.nextbigfuture.com/2025/09/prediction-of-ai-in-2040.html#more-205594

Jaime Sevilla (Director of Epoch AI) and Yafah Edelman (Head of Data and Analysis at Epoch AI) discuss AI scaling trends, future projections, economic impacts, and potential bottlenecks, extrapolating from current data on compute growth, training runs, and infrastructure.

NOTE: Elon Musk in his interview at the 2025 All In Summit says that AI intelligence is still scaling with increased compute. Elon’s metric is 10X the compute is still double the intelligence. This scaling of intelligence with compute is nuanced. How is the training done? Is it pretraining or is it reinforcement learning and how effective is the synthetic data and what is the quality of the synthetic data.

They analyze and project from historical trends, aiming to project straight-line forecasts while acknowledging uncertainties. They outline a modal world scenario where AI development unfolds in three eras, leading to transformative changes by 2035–2040. Key themes include compute scaling slowdowns, AI capabilities, automation of jobs, economic growth acceleration, and cruxes for post-2035 trajectories.

Current Scaling Regime and Near-Term Projections (2025–2030)

Scaling Trends and Slowdown: They note AI training compute has scaled 5x per year recently, driven by larger clusters, more GPUs, and slightly longer training durations (30% growth/year). However, they predict a slowdown to 2.5–3x/year within 2 years due to training durations stabilizing at 3–6 months. This assumes it is not extending due to algorithmic progress making later starts more efficient, R&D/post-training needs, and some level of diminishing returns. However, there is a lot of effort to keep the returns from diminishing.

Not using full clusters for single runs (e.g., R&D, experimentation, or parallel projects take priority).

Infrastructure delays (data centers take time to build).

They use xAI’s Memphis cluster as an example. Grok 3 was trained on ~80,000 GPUs, while the cluster had 100,000–200,000 GPUs available (not fully utilized). Grok 4 was Grok 3 with added reinforcement learning (RL), using less additional compute—possibly fewer GPUs for shorter time or similar. They believe xAI’s full cluster wasn’t dedicated to Grok runs, supporting their view of partial cluster usage industry-wide.

By 2030, largest runs could reach 1e29 FLOPs (1,000x more than current, akin to GPT-2 to GPT-4 gap). Capabilities include:Competent agents for consistent, cheap computer tasks with fewer reasoning failures.

Novel discoveries in math/physics (e.g., AI solving famous problems like Riemann hypothesis with human guidance).

Automation of coding (AI writes bug-free code; humans design systems but don’t code manually).

Some of the broader impacts are

– Automate call centers

– draft contracts

– higher-quality software

– agents for complex tasks.

Economic impacts will see AI revenues grow from tens of billions to hundreds of billions/year. This could feasibly keep doubling every year. This doubles US growth from ~2% to 4% in the near tern but would be 1000 times in ten years if it was sustained.

Infrastructure spending (e.g., GPUs, data centers, fabs) could hit trillions/year, showing in GDP via secondary effects.

NVIDIA’s $100B/year revenue exemplifies this; AI drives investment in chips/fabs (e.g., TSMC).

Post-2030 Bifurcation and Mid-Term Projections (2030–2035)

Slowdown vs. Takeoff: Post-2030, scaling slows further (e.g., each 10x infrastructure takes an extra year due to planning/investment reluctance). Investment follows “Yafah Math”: ~10x current revenue, with build-out time increasing per 10x scale. However, revenues grow with automation, justifying continued scaling.

Automation and Capabilities: By 2035 (modal: 2034–2035), AI automates all cognitive tasks as cheaply as humans (via 1,000–10,000x more compute than 2030). This includes entry-level jobs, scientific R&D (e.g., solving hard problems in days), and de-skilling (AI oversight turns skilled work into unskilled).

 Physical tasks lag due to robot bottlenecks.

Economic and Societal Impacts: Growth accelerates to 10%+/year (sustainable for years), with AI generating trillions in revenue. Unemployment spikes (e.g., coders, call centers); jobs shift to AI management/oversight. Diffusion is fast (e.g., ChatGPT adoption precedent; AI self-integrates). Societal reactions: High unemployment leads to crime, policy responses (e.g., stimuli like COVID), elections dominated by AI issues (automation, x-risk, environment). De-skilling lowers wages, enables outsourcing, unlocks larger workforces.

Rule of Thumb for Growth: Automating 1% of tasks/year adds 1% to growth. By 2030, <10% tasks automated; by 2035, all cognitive (30% of US tasks are remote-only, enabling 10% growth over a decade). Physical vs. Cognitive Automation: Cognitive automation easier; physical requires robots (costly to manufacture/scale). Alternatives: AI-augmented humans (e.g., AR headsets, phone cams for guidance) de-skill blue-collar work, boosting manufacturing output without full robots. Long-Term Projections (2035+) and Cruxes

Robot Era and Hyperbolic Growth: Robots start scaling ~2035; takes 3+ years to reach hundreds of millions/billions (e.g., $100B for 1M robots at $100K each, scaling to $10T for 100M). By 2038–2040, robots enable full task automation (cognitive + physical), leading to “bananas” (wild sci-fi) world: Explosive growth (30%+/year for 5+ years), AI as the economy (1,000%+ GDP contribution). Models break down; possible Dyson spheres in decades (but bottlenecks like energy/space slow to centuries/millennia).

Three Key Cruxes for Post-2035 Speed:Robot Manufacturing: How fast/cost-effective? Needs foresight/investment; current humanoid robots ($100K–$1M) worse than humans.

Returns to Scientific Labor: 1,000x more AI researchers could skip “decades” in fields (e.g., cure diseases, double lifespans), but diminishing returns if experiments bottlenecked by scale/physical limits.

Returns to Intelligence: Superintelligence (e.g., “Von Neumann the world”) might accelerate everything 2–100x, but tied to compute/experiments. If not, progress slows with compute.

Algorithmic Progress: Tied to large-scale experiments/compute; slows if compute does. If independent, leads to faster superintelligence (e.g., 2027 singularity).

Overall Modal View: Trends continue unless intervened; world feels “too fast” yet plausible. Faster worlds (e.g., high returns to intelligence) more salient for planning; slower (e.g., investment reluctance) push medians later (Jaime’s median: ~decade later than modal).

Could They Be Wrong About xAI and the Percentage of Compute Used for Grok Model Runs?

Yes, they could be partially wrong or outdated in their specifics on xAI, based on available data as of September 9, 2025. Their claims stem from public reports and estimates up to mid-2025, but xAI (led by Elon Musk) has been opaque, with details often from Musk’s X posts or leaks. Let’s break it down:Grok 3 Training: They claim ~80,000 GPUs used, out of a Memphis cluster of 100,000–200,000. This aligns with reports: Musk announced in July 2024 that the Memphis “supercluster” started with 100,000 H100 GPUs, scaling to more.

Grok 3 (released April 2025) was trained on ~100,000 GPUs per estimates from Epoch AI’s prior work, but not the full cluster—likely 70–80% utilization due to downtime, testing, or parallel runs. However, they might underestimate: Musk claimed in August 2025 X posts that Grok 3 used “most” of the cluster during peak training, implying higher % (possibly 90%+ efficient utilization after optimizations).

Grok 4: They describe it as Grok 3 + RL, with less additional compute (possibly fewer GPUs/shorter run). This is accurate based on xAI’s November 2024 announcement: Grok 4 was a fine-tuned/RLHF version of Grok 3, not a full pre-train, using ~50,000–70,000 GPUs for weeks (vs. Grok 3’s months). But they could be wrong on “less than Grok 3″—total FLOPs for Grok 4’s post-training might exceed Grok 3’s if including inference-heavy RL, per recent xAI updates.

Overall Compute Usage %:

Their core point—not using 100% of clusters for frontier runs—holds broadly (industry norm: 60–80% utilization due to maintenance, multi-tenancy). For xAI specifically, Memphis (now ~300,000 GPUs as of Q3 2025) is reportedly 80–90% dedicated to Grok iterations, per Musk’s September 2025 posts on scaling. They might be wrong if xAI optimized for near-full usage (e.g., via better orchestration), but evidence supports partial usage historically.

They could be wrong due to underestimating xAI’s aggression:

Musk’s goal is rapid scaling, and recent data centers (e.g., Abilene expansion) suggest fuller utilization for Grok 5 (in training as of now, potentially on 200k+ GPUs).

Could the Scaling Be Tapering Off Less Than They Think?

Yes, scaling could taper less (i.e., stay closer to 5x/year longer) if their assumptions underestimate key drivers:

Longer Training Durations: They predict stabilization at 3–6 months, but if companies prioritize Chinchilla-optimal runs (longer for data efficiency), durations could grow 50%+/year, maintaining 4–5x scaling.

Fuller Cluster Usage: If labs like xAI shift to dedicating 90–100% of clusters to single runs (e.g., via better fault tolerance), this boosts effective compute without new hardware.

Faster Infrastructure: Stargate/Abilene (Microsoft/OpenAI/xAI plans) could come online sooner (mid-2026) with 1M+ GPUs, enabling 5x jumps if used aggressively.

Algorithmic Gains: If progress allows shorter runs or multi-cluster federation, tapering slows.

Their 2.5–3x prediction assumes R&D bottlenecks dominate, but if economic incentives (e.g., AI revenues exploding) push fuller utilization, scaling holds at 4x+ for 3–4 years.

What Would Tell Us in the Next Models and Data Centers If Scaling and Performance Will Be Better Than They Expect?

To assess if scaling/performance exceeds expectations (e.g., >3x/year compute growth, better-than-projected capabilities), monitor these indicators in upcoming models (e.g., Grok 5, GPT-5/Orion, Claude 4) and data centers (2026–2027 builds):

Next Models (Grok 5, Grok 6 Models and beyond):Compute Scale:

If Grok 5 uses >200,000 GPUs (full Memphis+ expansions- B200s at 550k for 2.5 million H100) for >6 months, or >1e28 FLOPs total, it signals less tapering (fuller usage).

Grok 6 in early 2026 at 1 million B200/B300s and getting a leap with Rubin chips in 2026. Get to 10 million H100 equivalents in 2026 and 50 million in 2027.

Performance: If it achieves >95% on benchmarks like GPQA (math/physics) or solves novel problems (e.g., IMO gold-level without human aid), it implies better scaling laws (e.g., emergent capabilities from compute alone).

Duration/Utilization: Runs >6 months or reports of 90%+ cluster efficiency would counter their duration-stabilization view.

Capability Jumps: Exceeding linear scaling (e.g., 10x compute yields >10x better reasoning/agents) suggests tapering less (stronger returns).

Track via X searches (e.g., Musk on compute) or benchmarks (e.g., LMSYS, Epoch evals). If 2026 models automate >20% cognitive tasks (vs. their <10% by 2030), expect acceleration. Need 10 million robotaxi by 2028, 1 million robotruck by 2030 and 10 million teslabots by 2030.