The AI Futures Q2.5 2026 timelines update attempts to reconcile the headline numbers from earnings season with the harder question of when AI systems will actually automate economically meaningful work at scale. The picture that emerges is one of massive infrastructure commitment, steady model progress, and persistent gaps between promised capability and delivered economic uplift.
Infrastructure: The Capex Bet Crystallizes
The most concrete validation of AI infrastructure forecasts came from Q2 2026 earnings disclosures. Microsoft, Meta, Amazon, and Alphabet collectively guided to approximately **200 billion in 2024. TSMC raised its 2026 capex guidance to 64 billion, driven partly by agentic AI demand, and now expects revenue growth above 40 percent in USD terms.
TSMC's management was explicit: the AI accelerator business will have an even stronger CAGR for 2024-2029 than previously communicated, and the next three years of capex will be "even more significantly higher" than the last three. This supports the view that infrastructure investment is accelerating, not plateauing.
Model Progress: Time Horizons Expand, Benchmarks Lag
OpenAI's GPT-5 launched with a pre-deployment METR evaluation showing a 50 percent time horizon of approximately 2 hours and 17 minutes on software engineering tasks. This is one of the clearest data points yet for the prediction that autonomous agents will soon handle meaningful slices of professional work. At roughly two hours and seventeen minutes, GPT-5's METR time horizon sits at the threshold of single-task autonomy, yet remains far short of the full workday spans that would justify wholesale replacement of engineering roles.
Revenue and Uplift: The Translation Problem The harder question is when capability converts to revenue. Current AI revenue remains concentrated in infrastructure providers (NVIDIA, TSMC, cloud hyperscalers) and a narrow set of high-value applications such as coding assistants, customer support automation, and creative tooling. The promised broad-based productivity uplift across white-collar work has been slower to materialize.
Several factors explain the gap. Enterprise integration timelines are longer than consumer adoption curves. Regulatory review, security certification, and change management add quarters to deployment schedules. And many tasks that look automatable in a demo require subtle contextual judgment that current models still struggle to maintain across long horizons.
Forecasts for 2026 and 2027 vary widely. Bull cases assume that agentic systems with two-to-four-hour time horizons will compound rapidly, handling end-to-end workflows in legal review, financial analysis, and software maintenance. Bear cases note that each additional hour of reliable autonomy requires disproportionately more capability and safety investment, and that current trajectories may plateau before reaching the eight-hour threshold that would trigger large-scale labor substitution.
What to Watch in Q3 and Q4
Three signals will clarify which scenario is more likely. First, whether GPT-5 and comparable models show reliable multi-hour autonomy in production settings, not just benchmark conditions. Second, whether enterprise AI spend shifts from pilot budgets to recurring operational expenditure, which would indicate sustained integration rather than experimentation. Third, whether open-weight models with similar time horizons emerge, which would accelerate adoption through lower pricing and on-premise deployment options. he infrastructure is being built on a scale that assumes massive uptake. Whether that uptake arrives on the timelines implied by current capex plans is the central uncertainty shaping the second half of 2026.