Agent-Based AI Sparks 24-Fold Token Demand; Profit Turnaround Looms

Overview

Goldman Sachs’ latest report focuses on the economic inflection point brought by 'agent-based AI': as AI agents transition from concept to large-scale deployment, token consumption will grow exponentially, while sustained declines in token costs and stabilizing prices will improve profit margins across hyperscale vendors, model providers, and the entire AI value chain. Using bottom-up simulation experiments, the report quantifies a 24-fold increase in global token demand by 2030 compared to 2026 and introduces the 'profitability turning point' as a key investment theme for the first time.

Core views

Demand Side: Agent-based AI will trigger an explosion in token demand - By 2030, global token demand could reach 120 trillion/month—24 times the 2026 baseline—with corporate agents accounting for 56 trillion and consumer agents for 60 trillion. - By 2040, corporate agent token demand alone may surge to 278 trillion/month, growing at a compound annual rate of 55%. - On the consumer side, AI use cases will shift from 'question-and-answer' to 'always-on' personal assistants, with token intensity rising from 1,715 tokens per session to over 100,000 tokens daily. Supply Side: Costs are falling faster than prices, signaling a profitability turning point - Improvements in chip efficiency are driving a 60–70% annual decline in per-token costs, while leading LLM token prices have stabilized or even slightly rebounded after a 40% annual drop. - The report estimates that starting in the first half of 2026, a 'scissors gap' between token selling prices and costs will emerge, marking a gross margin inflection point. - This turning point will ease capital expenditure pressures on hyperscale vendors, making infrastructure investments more sustainable. Enterprise ROI: More workflows enter the 'profitable' zone - Taking coding, customer service, and data entry as examples, when token costs continue to fall, even with millions of tokens consumed daily, AI agent costs remain lower than corresponding human labor costs. - Coding agents cost about $13/day, far below the $300 human alternative; data entry agents cost $59/day versus $80 human labor; voice-based customer service remains relatively expensive due to real-time requirements, but the gap is narrowing. - Falling costs will continually expand the scope of automatable workflows, boosting software TAM and IT service demand. Valuation and Recommendations: Nine Core Beneficiary Stocks - Semiconductors: Broadcom (leading custom ASIC provider), Nvidia (GPU performance leader), AMD (dual CPU+GPU driver for data centers). - Internet: Alphabet (full-stack cloud + search), Amazon (AWS + self-developed chips), Meta (advertising + AI tools). - Software: Microsoft (enterprise workflow Copilot), Cloudflare (edge AI inference), Accenture (enterprise AI integration services).

Analysis framework

The report employs a 'bottom-up' methodology to build demand and cost models: 1. Using pseudocode, it constructs realistic scenarios such as 'travel booking agents' and 'coding agents,' systematically tracking token consumption to validate daily and monthly token intensities. 2. Combining chip performance, benchmark data, and pricing information, it calculates per-token cost curves for mainstream solutions like Nvidia, AMD, Google TPU, and Trainium. 3. Comparing token costs with human labor expenses, it identifies enterprise ROI inflection points and extrapolates penetration rates and market sizes. 4. Simulating consumer and corporate adoption rhythms via S-curves, it generates high-, medium-, and low-scenario projections for token demand between 2026 and 2040. This approach integrates technical feasibility, economic viability, and commercial implementation timelines, avoiding simplistic linear extrapolation.

Methodology notes

• Industry/sector analysis frameworkPenetration Rate S-Curve

  S-curve

  The report uses the S-curve to depict the progression of new technologies from early adoption to mass-market penetration, modeling consumer and corporate agent token demand along S-curves to avoid linear extrapolation.

• Industry/sector analysis frameworkCost curve analysis

  Token Cost Curve

  By analyzing chip performance, benchmarks, and pricing data, the report computes per-token cost curves for various accelerator solutions and compares them with selling price trends to identify profitability turning points.

• Company fundamentals and financial frameworkFree cash flow analysis

  ROI vs. Labor Costs Comparison

  By comparing AI agent costs with comparable human labor costs on a daily and monthly basis, the report determines which workflows have entered the 'positive ROI' zone, thereby estimating potential expansion in software TAM.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Alphabet(GOOGL.US)

  Full-stack cloud and search capabilities, benefiting from dual increases in token demand and profitability

  Strengths

  Owns TPU self-developed chips, Gemini models, and robust search distribution channels, enjoying full-stack cost advantages

  Weaknesses

  Regulatory and antitrust risks

  Comparison

  Ranked alongside Amazon and Meta as preferred picks in the internet sector

  Risks

  Regulation and competition could push token prices down again

• Amazon(AMZN.US)

  AWS AI workloads plus self-developed Trainium/Graviton chips

  Strengths

  With $364 billion in revenue and strong order backlog, driven by both AI and custom chips

• Meta(META.US)

  Advertising plus AI tools, leveraging AI computing power to enhance ad monetization efficiency

  Strengths

  Significantly outpacing industry growth in advertising, with AI models and hardware investments generating incremental monetization

• Broadcom(AVGO.US)

  Leading custom ASIC provider, with hyperscale clients increasingly adopting its cost optimization solutions

  Strengths

  Highest market share, with ongoing additional orders from Google and other major customers

• Nvidia(NVDA.US)

  GPU performance and ecosystem moat

  Strengths

  Leading training and inference performance, with high software ecosystem stickiness

• AMD(AMD.US)

  Dual CPU+GPU drive for data centers, with increasing corporate agent market share

  Strengths

  New MI450/MI5XX GPU product cycles, growing X86 server CPU market share

• Microsoft(MSFT.US)

  Copilot enterprise workflow ecosystem

  Strengths

  Microsoft 365 upgrade cycles synergize with Copilot, creating a closed-loop ecosystem

• Cloudflare(NET.US)

  Edge AI inference performance and cost advantages

  Strengths

  Network architecture and isolates software deliver latency and cost benefits

• Accenture(ACN.US)

  Enterprise AI integration and governance services

  Strengths

  From AI pilots to large-scale deployment, consulting and integration needs are surging

Key data

• Global Token Demand by 2030120 trillion/month24 times higher than in 2026
• Corporate Agent Token Demand by 2040278 trillion/month55% CAGR
• Annual Decline in Token Unit Costs60–70%Driven by improved chip efficiency
• Daily Token Consumption for Coding Agents7 millionCorresponding cost about $13, far below $300 human labor
• Daily Token Consumption for Customer Service Agents2 millionCorresponding cost about $93, still above $90 human labor

Impact & implications

The report argues that agent-based AI represents not only a demand-side 'token explosion' but also a supply-side 'profitability turning point.' Hyperscale vendors and model providers will shift from 'burning cash for expansion' to 'expanding profits,' making capital expenditures more sustainable; semiconductor and software companies will benefit from larger addressable markets and higher gross margins; meanwhile, IT service providers will gain new growth drivers thanks to increased complexity in enterprise deployments. Overall, the AI value chain stands poised to enter a virtuous cycle of 'demand–profitability–reinvestment.'

Risks

• Competition may force token prices to decline faster than costs, eroding the profitability turning point
• Some text-based chatbots have already begun commoditizing, adding further price pressure
• Voice-based agents require high real-time performance, keeping costs significantly above human labor, slowing commercialization pace

What to watch

• Whether the token price and cost curves will indeed show the expected 'scissors gap' in the first half of 2026
• How closely capital expenditure guidance aligns with revenue realization among hyperscale vendors
• The pace at which corporate clients transition from AI pilots to large-scale agent deployments